10.1007@978 981 15 6403 1

Lecture Notes in Electrical Engineering 682
Sofiane Bououden
Mohammed Chadli
Salim Ziani
Ivan Zelinka Editors
Proceedings of the
4th International
Conference
on Electrical
Engineering and
Control Applications
ICEECA 2019, 17–19 December 2019,
Constantine, Algeria
Lecture Notes in Electrical Engineering
Volume 682
Series Editors
Leopoldo Angrisani, Department of Electrical and Information Technologies Engineering, University of Napoli
Federico II, Naples, Italy
Marco Arteaga, Departament de Control y Robótica, Universidad Nacional Autónoma de México, Coyoacán,
Mexico
Bijaya Ketan Panigrahi, Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, Delhi, India
Samarjit Chakraborty, Fakultät für Elektrotechnik und Informationstechnik, TU München, Munich, Germany
Jiming Chen, Zhejiang University, Hangzhou, Zhejiang, China
Shanben Chen, Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai, China
Tan Kay Chen, Department of Electrical and Computer Engineering, National University of Singapore,
Singapore, Singapore
Rüdiger Dillmann, Humanoids and Intelligent Systems Laboratory, Karlsruhe Institute for Technology,
Karlsruhe, Germany
Haibin Duan, Beijing University of Aeronautics and Astronautics, Beijing, China
Gianluigi Ferrari, Università di Parma, Parma, Italy
Manuel Ferre, Centre for Automation and Robotics CAR (UPM-CSIC), Universidad Politécnica de Madrid,
Madrid, Spain
Sandra Hirche, Department of Electrical Engineering and Information Science, Technische Universität
München, Munich, Germany
Faryar Jabbari, Department of Mechanical and Aerospace Engineering, University of California, Irvine, CA,
USA
Limin Jia, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland
Alaa Khamis, German University in Egypt El Tagamoa El Khames, New Cairo City, Egypt
Torsten Kroeger, Stanford University, Stanford, CA, USA
Qilian Liang, Department of Electrical Engineering, University of Texas at Arlington, Arlington, TX, USA
Ferran Martín, Departament d’Enginyeria Electrònica, Universitat Autònoma de Barcelona, Bellaterra,
Barcelona, Spain
Tan Cher Ming, College of Engineering, Nanyang Technological University, Singapore, Singapore
Wolfgang Minker, Institute of Information Technology, University of Ulm, Ulm, Germany
Pradeep Misra, Department of Electrical Engineering, Wright State University, Dayton, OH, USA
Sebastian Möller, Quality and Usability Laboratory, TU Berlin, Berlin, Germany
Subhas Mukhopadhyay, School of Engineering & Advanced Technology, Massey University,
Palmerston North, Manawatu-Wanganui, New Zealand
Cun-Zheng Ning, Electrical Engineering, Arizona State University, Tempe, AZ, USA
Toyoaki Nishida, Graduate School of Informatics, Kyoto University, Kyoto, Japan
Federica Pascucci, Dipartimento di Ingegneria, Università degli Studi “Roma Tre”, Rome, Italy
Yong Qin, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China
Gan Woon Seng, School of Electrical & Electronic Engineering, Nanyang Technological University,
Singapore, Singapore
Joachim Speidel, Institute of Telecommunications, Universität Stuttgart, Stuttgart, Germany
Germano Veiga, Campus da FEUP, INESC Porto, Porto, Portugal
Haitao Wu, Academy of Opto-electronics, Chinese Academy of Sciences, Beijing, China
Junjie James Zhang, Charlotte, NC, USA
The book series Lecture Notes in Electrical Engineering (LNEE) publishes the
latest developments in Electrical Engineering - quickly, informally and in high
quality. While original research reported in proceedings and monographs has
traditionally formed the core of LNEE, we also encourage authors to submit books
devoted to supporting student education and professional training in the various
fields and applications areas of electrical engineering. The series cover classical and
emerging topics concerning:
• Communication Engineering, Information Theory and Networks
• Electronics Engineering and Microelectronics
• Signal, Image and Speech Processing
• Wireless and Mobile Communication
• Circuits and Systems
• Energy Systems, Power Electronics and Electrical Machines
• Electro-optical Engineering
• Instrumentation Engineering
• Avionics Engineering
• Control Systems
• Internet-of-Things and Cybersecurity
• Biomedical Devices, MEMS and NEMS
For general information about this book series, comments or suggestions, please
contact [email protected].
To submit a proposal or request further information, please contact the
Publishing Editor in your country:
China
Jasmine Dou, Associate Editor ([email protected])
India, Japan, Rest of Asia
Swati Meherishi, Executive Editor ([email protected])
Southeast Asia, Australia, New Zealand
Ramesh Nath Premnath, Editor ([email protected])
USA, Canada:
Michael Luby, Senior Editor ([email protected])
All other Countries:
Leontina Di Cecco, Senior Editor ([email protected])
** Indexing: The books of this series are submitted to ISI Proceedings,
EI-Compendex, SCOPUS, MetaPress, Web of Science and Springerlink **
More information about this series at http://www.springer.com/series/7818

Sofiane Bououden Mohammed Chadli
• •
Salim Ziani Ivan Zelinka

•
Editors
Proceedings of the 4th

International Conference
on Electrical Engineering
and Control Applications
ICEECA 2019, 17–19 December 2019,
123
Editors
Sofiane Bououden Mohammed Chadli
Faculty of Sciences and Technology University of Evry, IBISC
University Abbes Laghrour Khenchela University of Paris-Saclay
Khenchela, Algeria Evry, France
Salim Ziani Ivan Zelinka

Department of Electronics Faculty of Electrical Engineering
University of Frères Mentouri and Computer Science
Constantine 1 VŠB TU Ostrava
Constantine, Algeria Ostrava, Czech Republic
ISSN 1876-1100 ISSN 1876-1119 (electronic)

Lecture Notes in Electrical Engineering
ISBN 978-981-15-6402-4 ISBN 978-981-15-6403-1 (eBook)
https://doi.org/10.1007/978-981-15-6403-1
© Springer Nature Singapore Pte Ltd. 2021
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Organization
Conference
The International Conference on Electrical Engineering and Control Applications

(ICEECA) provides a forum for specialists and practitioners to present and discuss
their research results in several areas of the conference and also state-of-the-art
findings in using the applied electrical engineering and automatic control to solve
national problems that face developing countries. The conference ICEECA pub-
lishes papers on theoretical analysis, experimental studies, and applications in the
domain of automatic control and computer engineering. The objective of the con-
ference is not only the exchange of knowledge and experience, since the conference
is an open door to students, but also provides opportunities for researchers to target
future collaboration on current issues.
General Chairs
Salim Ziani, University of Frères Mentouri Constantine 1, Algeria

Mohammed Chadli, University Paris-Saclay, IBISC Laboratory, University of
Evry, France
Program Chairs
Sofiane Bououden, University of Khenchela, Algeria

Hamid Reza Karimi, Polytechnic of Milan, Italy
Ivan Zelinka, Technical University of Ostrava, Czech Republic
v
vi Organization
Steering Committee
Sofiane Bououden, University of Khenchela, Algeria

Mohammed Chadli, University Paris-Saclay, UEVE-IBISC Laboratory, France
Peng Shi, Victoria University, Australia
Salim Ziani, University of Frères Mentouri Constantine 1, Algeria
Local Committee
ICEECA established a local committee from the department of electronics as an

organizer, faculty of sciences of the technologies, University Frères Mentouri
Constantine 1, Algeria. The list below comprises the local committee members.
Aris Skander (Chair), University of Frères Mentouri Constantine 1, Algeria
Ameur Ikhlef, University of Frères Mentouri Constantine 1, Algeria
Kheireddine Cheikh, University of Frères Mentouri Constantine 1, Algeria
Kheireddine Lamamra, University of Oum El Bouaghi, Algeria
Abdelmalek Zahaf, University of Frères Mentouri Constantine 1, Algeria
Merzoug Ammari, University of Frères Mentouri Constantine 1, Algeria
Mohamed Chemachema, University of Frères Mentouri Constantine 1, Algeria
Mohammed Raslain, University of Frères Mentouri Constantine 1, Algeria
Ouassila Bourbia, University of Frères Mentouri Constantine 1, Algeria
International Advisory Committee
Abedelfetah Charef, University of Frères Mentouri Constantine 1, Algeria

Abedlhamid Tayebi, Lakehead University, Canada
Hamid Reza Karimi, Polytechnic of Milan, Italy
L. X. Zhang, Harbin Institute of Technology, China
N. B. Braiek, Ecole Polytechnique de Tunis, Tunisia
Said Mammar, University Paris-Saclay, UEVE-IBISC Laboratory, France
Scientific Program Committee
ICEECA established an international committee of selected well-known experts in

electrical engineering who are willing to be mentioned in the program and to review
a set of papers each year. The list below comprises the scientific program committee
members.
Organization vii
Abdelaziz Hamzaoui, University of Reims, France

Abdeldjalil Ouahabi, Polytech Tours, France
Abdelhafid Zeroual, University of Skikda, Algeria
Abdelhak Bennia, University of frères Mentouri Constantine 1, Algeria
Abdellah Kouzou, University of Djelfa, Algeria
Abedlhamid Tayebi, Lakehead University, Canada
Ahmed Chemori, University Montpellier 2, France
Ahmed Hafaifa, University of Djelfa, Algeria
Ahmed Rachid, University of Picardie Amiens, France
Ali Zemouche, Université de Lorraine, Nancy, France
Amar Djouak, Université Catholique de Lille, France
Ameur Ikhlef, University of frères Mentouri Constantine 1, Algeria
Aouaouda Sabrina, Université Souk Ahras, Algeria
Aziz Naamane, Université de Marseille Nord, France
Barkat Ouarda, University of frères Mentouri Constantine 1, Algeria
Bhekisipho Twala, University of Johannesburg, South Africa
Bin Haji Hassan Masjuki, University of Malaya, Malaysia
Carlos Astorga Zaragoza, CENIDET, Mexico
Chérif Chibane, Massachusetts Institute of Technology, Boston, USA
Chériti Ahmed, Université Trois-Rivières, Canada
Chokri Ben Salah, Université de Sousse, Tunisie
De Lima Neto Fernando, University of Pernambuco, Recife, Brazil
Djillali Bouagada, University of Mostaganem, Algeria
Faouzi Bouani, Ecole Nationale d’Ingénieurs de Tunis, Tunisia
Farid Marir, University of frères Mentouri Constantine 1, Algeria
Faycal Megri, Université of Oum El Bouaghi
Fella Hachouf, University of frères Mentouri Constantine 1, Algeria
Fernando Tadeo, University of Valladolid, Spain
Fethi Demim, Ecole Militaire Polytechnique, Algiers, Algeria
Fouad Allouani, University of Khenchela, Algeria
Fouad Kerrour, University of frères Mentouri Constantine 1, Algeria
Fouzi Soltani, University of frères Mentouri Constantine 1, Algeria
Gloria Osorio Gordillo, CENIDET, Mexico
Hamid Reza Karimi, Ecole Polytechnique de Milan, Italy
Hervé Coppier, ESIEE, France
Ilyes Boulkaibet, University of Johannesburg, South Africa
Jean-Jacques Loiseau, IRCyN, France
Kamel Guesmi, University of Djelfa, Algeria
Kheireddine Lamamra, University of Oum El Bouaghi, Algeria
Lamir Saidi, University of Batna, Algeria
Maamer Bettayeb, University of Sharjah, UAE
Malika Kandouci, University of Sidi Bel-Abbes, Algeria
Manuel Adam Medina, CENIDET, Mexico
Mehadji Abri, University of Tlemcen, Algeria
viii Organization
Mohamed Boumehraz, University of Biskra, Algeria

Mohamed Boutayeb, Centre de Recherche en Automatique de Nancy, France
Mohamed Chemachema, University of frères Mentouri Constantine 1, Algeria
Mohamed Darouach, University of Lorraine, France
Mohamed Elleuch, University of Tunis, Tunisia
Mohamed Lahdi Riabi, University of frères Mentouri Constantine 1, Algeria
Mohamed Lashab, University of Oum El Bouaghi, Algeria
Mohamed Seghir Boucherit, Ecole Nationale Polytechnique, Algeria
Mohamed Toufik Benhabiles, University of frères Mentouri Constantine 1, Algeria
Mohammed Msaad, University of Caen Normandie, France
Mostapha Ziad, University of Suffolk, Boston, USA
Nadjet Zioui, Université du Québec à Trois-Rivières, QC, Canada
Najib Essounbouli, University of Reims Champagne Ardenne, France
Noureddine Slimane, University of Batna 2, Algeria
Ouassila Bourbia, University of frères Mentouri Constantine 1, Algeria
Peng Shi, Victoria University, Australia
Pierre Borne, Ecole Centrale de Lille, France
Rachid Illoul, Ecole Nationale Polytechnique de Alger, Algeria
Roman Senkerik, Tomas Bata University in Zlin, Czech Republic
Salim Ziani, University of frères Mentouri Constantine 1, Algeria
Said Benierbeh, University of frères Mentouri Constantine 1, Algeria
Said Hassaine, University of Tiaret, Algeria
Zidani Mosbah, University of Biskra, Algeria
Zoheir Hammoudi, University of Constantine 1, Algeria
Preface
This proceedings book about the advanced control engineering methods in elec-
trical engineering systems conference contains accepted papers presenting the most
interesting state of the art on this field of research. Presented topics are focused on
classical as well as modern methods for modeling, control, identification, and
simulation of complex systems with applications in science and engineering. Topics
are (but not limited to) control and systems engineering, renewable energy, faults
diagnosis–fault-tolerant control, large-scale systems, fractional-order systems,
unconventional algorithms in control engineering, signal and communications, and
much more. The control of complex systems dynamics, analysis, and modeling of
its behavior and structure is a vitally important problem in engineering, economy,
and generally in science today. Examples of such systems can be seen in the world
around us and are a part of our everyday life. Application of modern methods for
control, electronics, signal processing, and more can be found in our mobile
phones, car engines, home devices as, for example, washing machine is as well as
in such advanced devices as space probes and communication with them. The main
aim of the conference is to create a periodical possibility for students, academics,
and researchers to exchange their ideas and novel methods. This conference will
establish a forum for the presentation and discussion of recent trends in the area of
applications of various modern as well as classical methods for researchers, stu-
dents, and academics. The accepted selection of papers was extremely rigorously
reviewed in order to maintain the high quality of the conference that is supported by
organizing universities and related research grants. Regular and student’s papers
have been submitted to the conference, and in accordance with review process, have
been accepted after a positive review. We would like to thank the members of the
program committees and reviewers for their hard work. We believe that this con-
ference represents a high-standard conference in the domain of control, modeling,
and analysis of dynamical and electronic systems. We would like to thank all the
contributing authors, as well as the members of the program committees and the
local organizing committee for their hard and highly valuable work. Their work has
definitely contributed to the success of the conference.
ix
x Preface
This event is supported by the University of frères Mentouri Constantine 1,

Algeria, and the University Paris-Saclay, IBISC Laboratory, University of Evry,
France, and mainly financed by DGRSDT, SONATRACH society, and
SONELGAZ society.
Khenchela, Algeria Sofiane Bououden

Evry, France Mohammed Chadli
Constantine, Algeria Salim Ziani
Ostrava, Czech Republic Ivan Zelinka
Contents
Control and Systems Engineering (CSE)

Sensorless Speed Control of Induction Motor Used Differential
Flatness Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Yassine Beddiaf, Fatiha Zidani, and Larbi. Chrifi-Alaoui
Multi-UAVs Coverage Path Planning . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Abdelwahhab Bouras, Yasser Bouzid, and Mohamed Guiatni
Reducing the Collision Checking Time in Cluttered Environment
for Sampling-Based Motion Planning . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Amine Belaid and Boubekeur Mendil
Random Scaling-Based Bat Algorithm for Greenhouse Thermal
Model Identification and Experimental Validation . . . . . . . . . . . . . . . . 47
Mounir Guesbaya, Hassina Megherbi, and Ahmed Chaouki Megherbi
Control and Observation of Induction Motor Using First-Order
Sliding Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Salah Eddine Farhi, Djamel Sakri, and Noureddine Goléa
Cuckoo Search Algorithm for Solving the Problem
of Unit-Commitment with Vehicle-to-Grid . . . . . . . . . . . . . . . . . . . . . . 77
Amel Terki and Hamid Boubertakh
A Comparative Study of MPPT Techniques for Standalone Hybrid
PV-Wind with Power Management . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
Tarek Boutabba, Hamza Sahraoui, Mouhamed Lamine Bechka, Said Drid,
and Larbi Chrifi-Alaoui
Adapted Search Equations of Artificial Bee Colony Applied
to Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Hayet Djellali, Souad Guessoum, and Nacira Ghoualmi-Zine
xi
xii Contents
Networked Cooperation-Based Distributed Model Predictive Control

Using Laguerre Functions for Large-Scale Systems . . . . . . . . . . . . . . . 123
Kamel Menighed, Ahmed Chemori, Boumedyen Boussaid,
and Joseph Julien Yamé
Direct Torque Control Using Fuzzy Second Order Sliding Mode
Speed Regulator of Double Star Permanent Magnet Synchronous
Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Louanasse Laggoun, Brahim Kiyyour, Ghoulemallah Boukhalfa,
Sebti Belkacem, and Said Benaggoune
Fractional Order Integral Controller Design Based on a Bode’s Ideal
Transfer Function: Application to the Control of a Single Tank
Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Chahira Boussalem, Rachid Mansouri, Maamar Bettayeb,
and Mustapha Hamerlain
Modeling of Torsional Vibrations Dynamic in Drill-String
by Using PI-Observer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
R. Riane, M. Kidouche, M. Z. Doghmane, and R. Illoul
Power System Generator Coherency Identification for Large
Disturbances by Koopman Modes Analysis . . . . . . . . . . . . . . . . . . . . . 187
Zahra Jlassi, Khadija Ben Kilani, Mohamed Elleuch, and Lamine Mili
Analysis and Experimental Validation of Single Phase Series
Resonance Inverter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Alla Eddine Toubal Maamar, M’hamed Helaimi, and Rachid Taleb
Optimal Decentralized State Control of Multi-machine Power System
Based on Loop Multi-overlapping Decomposition Strategy . . . . . . . . . . 217
M. Z. Doghmane and M. Kidouche
Large Synchronverter Integration in Power Electrical System:
Impacts on SCR and CCT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Raouia Aouini, Khadija Ben Kilani, Mohamed Elleuch,
and Quoc Tuan TRAN
Immersion and Invariance Based Adaptive Dynamic Surface Control
for Parametric Strict-Feedback Nonlinear Systems . . . . . . . . . . . . . . . . 247
Y. Soukkou, S. Labiod, M. Tadjine, Q. M. Zhu, and M. Nibouche
Design of an Adaptive Fuzzy Backstepping Synergetic Control
Scheme for a Class of Strict-Feedback Nonlinear Systems . . . . . . . . . . 263
Aissa Rebai, Kamel Guesmi, and Mohamed Bougrine
Improvement of the Stability Performance of a Quad-Copter
Helicopter by a Neuro-Fuzzy Controller . . . . . . . . . . . . . . . . . . . . . . . . 279
Djalal Baladji, Kheireddine Lamamra, and Farida Batat
Contents xiii
The Importance of Applying Artificial Intelligence on Unmanned

Aerial Vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
Amine Mohammed Taberkit, Ahmed Kechida,
and Abdelmalek Bouguettaya
Fuzzy H∞ Delay-Independent Stabilization of Depth Control
for Underwater Vehicle with Input Constraint . . . . . . . . . . . . . . . . . . . 305
Mohamed Nasri, Dounia Saifia, Salim Labiod, and Mohammed Chadli
Petri Type 2 Fuzzy Neural Networks (PT2FNN) for Identification
and Control of Dynamic Systems—A New Structure
and a Comparative Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
Youssouf Bibi, Mohamed Seghiri, Omar Bouhali, Abdelkarim Nemra,
and Tarek Bouktir
Design Fractional Order PI Controller with Decoupler
for MIMO Process Using Diffusive Representation . . . . . . . . . . . . . . . . 331
Sami Laifa, Badreddine Boudjehem, and Hamza Gasmi
Direct Sliding Mode Control of Transient Power in Microgrid During
Grid Failure ‘‘Unintentional Islanding’’ . . . . . . . . . . . . . . . . . . . . . . . . 347
I. Ameur, N. Gazzam, and A. Benalia
New Virtual Synchronous Generator Control Technique
of Distributed Generator Unit to Improve Transient Response
of the Microgrid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
Yacine Daili and Abdelghani Harrag
Design and Analysis of Robust Nonlinear Synergetic Controller
for a PMDC Motor Driven Wire-Feeder System (WFS) . . . . . . . . . . . . 373
Noureddine Hamouda, Badreddine Babes, and Amar Boutaghane
Simulation Study of the Dual Star Permanent Magnet Synchronous
Machine Using Different Modeling Approaches . . . . . . . . . . . . . . . . . . 389
Elyazid Amirouche, Khaldi Lyes, Ghedamsi Kaci,
and Aouzellag Djamal
Modeling and Experimental Identification of Salient-Pole
Synchronous Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
Khalil-Errahmane Sari and Bilal Sari
Power Quality Improvement of PWM Rectifier-Inverter
System Using Model Predictive Control for an AC Electric
Drive Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
Abdelkarim Ammar
Novel Smart Air Quality Monitoring System Based
on UAV Quadrotor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
Mehdi Zareb, Benaoumer Bakhti, Yasser Bouzid, Hamza Kadourbenkada,
Kamel Bouzgou, and Wahid Nouibat
xiv Contents
GA Based Adaptive Fuzzy Sliding Mode for 3 DOF Spatial Arm

Manipulator Trajectory Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
Rabie Belloumi and Noureddine Slimane
Robust Adaptive Fuzzy Approach with Unknown Control Gain
Direction and External Disturbance . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
Ouassila Bourebia and Nassira Zerari
Adaptive Neural-Network Control Design for Uncertain CSTR
System with Unknown Input Dead-Zone and Output Constraint . . . . . 477
Zerari Nassira and Chemachema Mohamed
Adaptive Fuzzy Fault-Tolerant Control Using Nussbaum Gain
for a Class of SISO Nonlinear Systems with Unknown Directions . . . . 493
Abdelhamid Bounemeur, Mohamed chemachema, Abdelmalek Zahaf,
and Sofiane Bououden
Renewable Energy (RE)

Optimal Integration of Renewable Distributed Generation
Using the Whale Optimization Algorithm for Techno-Economic
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
Samir Settoul, Rachid Chenni, Mohamed Zellagui, and Hassan Nouri
Enhancement of Power System Transient Stability with a Large
Penetration of Solar Photovoltaic Using Facts . . . . . . . . . . . . . . . . . . . 533
Ikram Boucetta, Naimi Djemai, Salhi Ahmed, and Zellouma Laid
Minimization of the Energy Consumption of an Aircraft . . . . . . . . . . . 549
Abbes Lounis, Kahina Louadj, and Mohamed Aidene
Study to Improve the Technical Parameters for the Optimising
the Injection of the Photovoltaic Energy into an Electrical Network
via a Line 30 KV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
Bahriya Badache, Hocine Labar, and Mounia Samira Kelaiaia
Second-Order Super-Twisting Control of an Autonomous Wind
Energy Conversion System Based on PMSG for Robustness
and Chattering Elimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 579
Abderrahmane Abdellah, Djilali Toumi, and M’hamed Larbi
Open-Switch Faults Based on Six/Five-Leg Reconfigurable
AC-DC-AC Converter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595
Sahraoui Khaled and Gaoui Bachir
A Robust Model Predictive Control for a DC-DC Boost Converter
Subject to Input Saturation: An LMI Approach . . . . . . . . . . . . . . . . . 615
O. Hazil, S. Bououden, and M. Chadli
Contents xv
Strategy for Optimization of Energy Management Based on Fuzzy

Logic in an FCEV with a Contribution of a Photovoltaic Source . . . . . 631
Said Belhadj, Kaci Ghedamsi, and Zina Larabi
Modeling and Analysis of an Electromagnetic Vibration Energy
Harvester for Automotive Suspension . . . . . . . . . . . . . . . . . . . . . . . . . . 645
Mustapha Zaouia, Bachir Ouartal, Nacereddine Benamrouche,
and Arezki Fekik
Experimental and Numerical Study of Hybrid PV/Thermal Solar
Collector Provided with Self Ventilation and Tracking Structure . . . . . 659
Mohamed El-Amine Slimani, Rabah Sellami, Achour Mahrane,
and Madjid Amirat
Maximum Power Point Tracking (MPPT) for a PV System Based
on Artificial Neural Network ANN and Comparison with P&O
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 671
Aicha Djalab, Nassim Sabri, and Ali Teta
Development of a Stand-Alone Connected PV System Based on Packe
U Cell Inverter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683
Khaled Rayane, Mohamed Bougrine, Atallah Benalia, and Kamel Guesmi
A Novel Hybrid Photovoltaic/Thermal Bi-Fluid (Air/Water) Solar
Collector: An Experimental Investigation . . . . . . . . . . . . . . . . . . . . . . . 697
Mohamed El-Amine Slimani, Rabah Sellami, Mohammed Said,
and Amina Bouderbal
A DC/DC Buck Converter Voltage Regulation Using an Adaptive
Fuzzy Fast Terminal Synergetic Control . . . . . . . . . . . . . . . . . . . . . . . 711
Noureddine Hamouda and Badreddine Babes
Study of Electrical Field Distribution in the Insulation
of High-Voltage Cables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 723
Rouini Abdelghani, Kouzou Abdellah, and Larbi Messaouda
Self Tuning Fuzzy Maximum Power Tracking Control of PMSG
Wind Energy Conversion System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735
Chaib Housseyn, Mihoub Youcef, and Hassaine Said
A Backstepping Controller for Interleaved Boost DC–DC Converter
Improving Fuel Cell Voltage Regulation . . . . . . . . . . . . . . . . . . . . . . . . 751
Ali Dali, Samir Abdelmalek, and Maamar Bettayeb
The Impact of the Solar Radiation Profile on Sizing and Performance
of Photovoltaic Systems, Case Study Tamanrasset, Algeria . . . . . . . . . 763
Madjid Chikh, Aicha Degla, and Achour Mahrane
xvi Contents
Investigation and Prototyping Implementation of a Novel Solar

Water Collector Based on Used Engine Oil as HTF . . . . . . . . . . . . . . . 779
Oussama Touaba, Salah Mohamed AitCheikh,
Mohamed El-Amine Slimani, Ahmed Bouraiou, and Abderrezzaq Ziane
Piezoelectric Energy Harvesting Based Autonomous
Vehicles’ Vibrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 791
Nadjet Zioui, Sousso Kelouwani, and Gilbert Lebrun
Faults Diagnosis-Faults Tolerant Control (FTC)

Fault Detection of Uncertain Systems Based on Interval Data
Driven Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803
Chouaib Chakour
Discrimination of Unbalanced Supply and Stator Interturn Faults
in Induction Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815
D. Kouchih, R. Hachelaf, and N. Boumalha
Impact of the Stator Winding Topology on the Fault Harmonic
Components in Induction Motors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 831
Seddik Tabet, Adel ghoggal, and Salah Eddine Zouzou
Sensor Fault Detection for Uncertain T-S DC Model with Descriptor
Observer Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 843
Moussaoui Lotfi, Aouaouda Sabrina, and Righi Ines
Detection of Stator Interturn Fault in Double Star
Induction Motors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 857
R. Hachelaf, N. Boumalha, D. Kouchih, M. Tadjine, and M. S. Boucherit
Fuzzy-Expert System Fault Tolerant Control of IGBT Open Switch
Fault for RSC in Wind Turbine System . . . . . . . . . . . . . . . . . . . . . . . . 871
Amina Bouzekri and Tayab Allaoui
Multiclass Support Vector Machine Based Bearing Fault Detection
Using Vibration Signal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 885
Issam Attoui, Nadir Fergani, Nadir Boutasseta, Brahim Oudjani,
Mohammed Salah Bouakkaz, and Ahmed Bouraiou
Processing Signal Parameters Based Fuzzy Inference System
Classifier for Analog Circuit Single Parametric Faults . . . . . . . . . . . . . 897
Imad Laidani and Naceredine Bourouba
Line Current Spectrum Analysis as a Technique to Diagnose
the Demagnetization Fault in Permanent Magnet
Synchronous Motor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 915
Zakaria Gherabi, Djilali Toumi, Noureddine Benouzza,
and Mostefa Becheikh
Contents xvii
Fault-Tolerant Path-Tracking Control with PID Controller

for 4ws4wd Electric Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 931
Youcef Zennir, Sami Allou, and J. J. Fernandez-Lozano
Discrete Faults Diagnosis for the Multicellular Converter:
Algebraic Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 947
K. Haddadi, N. Gazzam, I. Ameur, and A. Benalia
Fault Diagnosis of Uncertain Hybrid Actuators Based Model
Predictive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 961
Zahaf Abdelmalek, Bounemeur Abdelhamid, Bououden Sofiane,
and Ilyes Boulkaibet
Fault Tolerant Predictive Control for Constrained Hybrid Systems
with Sensors Failures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 973
Zahaf Abdelmalek, Mohamed Chemachema, Bououden Sofiane,
Active Adaptive Fuzzy Fault-Tolerant Control for a Class
of Nonlinear Systems with Actuator Faults . . . . . . . . . . . . . . . . . . . . . . 985
Abdelhamid Bounemeur and Mohamed Chemachema
Signal and Communications (SC)

Simultaneous Localization and Mapping Algorithm based
on 3D Laser for Unmanned Aerial Vehicle . . . . . . . . . . . . . . . . . . . . . . 1003
Fethi Demim, Abdelkrim Nemra, Oumima Mouali, Mehdia Hedir,
Abdenebi Rouigueb, Mustapha Hamerlain, Mohamed Amine Bendoumi,
and Abdelouahab Bazoula
Comparison Between Gabor Filters and Wavelets Transform
for Classification of Textured Images . . . . . . . . . . . . . . . . . . . . . . . . . . 1021
Abdelkader Zitouni, Fatiha Benkouider, Fatima Chouireb,
and Mohammed Belkheiri
High Level Modeling and Hardware Implementation of Image
Processing Algorithms Using XSG . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1033
Kamel Messaoudi, Hakim Doghmane, Hocine Bourouba,
E. B. Bourennane, and S. Toumi
Proposal for a Radar Detection Architecture Based
on the Knowledge Based Systems Exploitation . . . . . . . . . . . . . . . . . . . 1047
Abdellatif Rouabah, Hamza Zeraoula, M’hamed Hamadouche,
and Kamal Tourche
Comparative Analysis and Simulation of Propulsion and Energy
Systems for Satellite Communication . . . . . . . . . . . . . . . . . . . . . . . . . . 1061
Jalal Eddine Benmansour, Boulanouar Khouane, and Tarek Badis
xviii Contents
Narrowband Sensor-Antenna Based on an Interdigital Capacitor . . . . 1071

Rafik Khelladi, Mustapha Djeddou, and Farid Ghanem
Secure Color Image Transmission Based on the Impulsive
Synchronization of Fractional-Order Chaotic Maps Over
a Single Channel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1081
Ouerdia Megherbi, Hamid Hamiche, Saïd Djennoune,
and Maamar Bettayeb
Performance of FBMC for Backhauling 5G . . . . . . . . . . . . . . . . . . . . . 1097
Abdelhamid Bounegab and Mohammed Boulesbaa
Flexible FPGA-Based RFID Emulation Platform
for Experimental Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1111
Ibrahim Mezzah, Omar Kermia, and Hamimi Chemali
Human Activity Recognition Based on Ensemble Classifier Model . . . . 1121
Souhila Kahlouche and Mahmoud Belhocine
Security in the Internet of Things: Recent Challenges
and Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1133
Hamza Belkhiri, Abderraouf Messai, Mohamed Belaoued,
and Farhi Haider
Characterization of Melanoma Using Convolutional Neural
Networks and Dermoscopic Images . . . . . . . . . . . . . . . . . . . . . . . . . . . 1147
Abdelghani Tafsast, Mohamed Laid Hadjili, Ayache Bouakaz,
and Nabil Benoudjit
Advanced Feedforward-and-Feedback Decorrelation Algorithms
for Speech Quality Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1157
Rédha Bendoumia
New Log CFAR Radar for Compound Sea Clutter . . . . . . . . . . . . . . . 1171
Nouh Guidoum, Faouzi Soltani, and Khadidja Belhi
Stochastic Analysis of the Crosstalk Between Transmission Lines . . . . 1181
Hemza Gheddar, M. Melit, and B. Nekhoul
Moving Objects Detection and Tracking with Camera
Motion Compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1193
Bousta Mohamed Akli, Nemra Abdelkrim, Hamami Fatima,
and Demim Fethi
Potential of Using Adaptive Complex Flower Pollination Algorithm
to Interferometric Coherence Optimisation . . . . . . . . . . . . . . . . . . . . . . 1211
Sofiane Tahraoui and Mounira Ouarzeddine
Intelligent Cooperative Method for Medical Image Segmentation . . . . 1221
Karima Benhamza and Hamid Seridi
Contents xix
An Image Segmentation Model Using a Level Set Method Based

on Improved Signed Pressure Force Function SPF . . . . . . . . . . . . . . . . 1233
Larbi Messaouda, Zoubeida Messali, Rouini Abdelghani,
and Samira LARBI
New Convergence Architecture Between 5G Mobile
Telecommunication Networks and Satellite Networks to Enhance
Their Capacities and Improve Their Performance . . . . . . . . . . . . . . . . 1247
Salheddine Sadouni, Malek Benslama, André-luc Beylot,
Mohammed Nassim Lacheheub, and Ouissal Sadouni
Comparative Study of Noise Robustness in Visual
Attention Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1259
Sarra Babahenini, Foudil Cherif, and Fella Charif
The Fast-NLMS Algorithm for Acoustic Echo Cancellation
with Double-Talk Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1271
Islam Hassani, Madjid Arezki, and Ahmed Benallal
Design of Sensor Based on CRLH for Liquid
Mixture Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1281
Sekkache Hocine, Mohamed Lashab, and Salim Ouchtati
Control and Systems Engineering (CSE)
Sensorless Speed Control of Induction
Motor Used Differential Flatness Theory
Yassine Beddiaf, Fatiha Zidani, and Larbi. Chrifi-Alaoui
Abstract In this paper, differential flatness theory and reduced order observer are
used to develop sensorless control for induction motor. The concept of flatness is
based on the description of flat outlets. The first step is showing that the induc-
tion motor is a differentially flat system, since all state variables of the circuits
describing the motor’s dynamics can be expressed as functions. In the second step,
when the chosen flat outputs involve non available state variable measurements
the state observer is used to estimate them. Simulation and experimental results
are presented to illustrate the effectiveness of the proposed approach for sensorless
control of induction motor.
Keywords Induction motor · Differential flatness · Sensorless vector control
1 Introduction
The induction machine, is the most robust and low cost machine. The progress
made in control and the considerable technological advances, both in the field of
power electronics and microelectronics, have made possible the implementation of
the powerful controls of this machine.
However, many problems remain. As the parameters variation, the presence of the
mechanical sensor and the degraded functioning are all difficulties that have resulted
from the eagerness of the researchers, as evidenced by the ever-increasing number
of publications that study the subject.
Y. Beddiaf (B)
Department of Electrical Engineering, University of Khenchela, Khenchela, Algeria
e-mail: [email protected]; [email protected]
F. Zidani
Laboratory LSPIE, Department of Electrical Engineering, University of Batna-2, Batna, Algeria
e-mail: [email protected]
Larbi. Chrifi-Alaoui
Laboratoire des technologies innovantes (LTI EA 3988), Amiens, France
© Springer Nature Singapore Pte Ltd. 2021 3

S. Bououden et al. (eds.), Proceedings of the 4th International Conference on Electrical
Engineering and Control Applications, Lecture Notes in Electrical Engineering 682,
https://doi.org/10.1007/978-981-15-6403-1_1
4 Y. Beddiaf et al.
Several strategies have been proposed in the literature to realize the sensorless
control of induction motor. A large part of the proposed methods is based on observers
depending on the model of the asynchronous machine [1–5, 12]. Other research is on
the contribution of artificial intelligence to improve sensorless control of the machine
[6–8]. These techniques do not exactly replace the speed sensor, mainly during in
the load application.
In this context, we propose the sensorless control based on flatness theory.
The flatness theory is a more or less recent notion in automatic that was proposed
from 1992 by Lévine et al. [9], Dannehl and Fuchs [10], Fliess et al. [11]. This
concept makes it possible to control of dynamic system. The concept of flatness
theory is based on the highlighting of flat outlets. The first step of control by the
flatness method is to generate a desired trajectory that takes into account the model
of system. In the second step, this control requires the design of a loopback control
allowing the continuation of this trajectory.
The validity of the proposed method is verified using simulation and experimen-
tation.
2 Differential Flatness Theory
Consider a dynamic system defined by the following state equation:
dX
= f (X, U ) (1)
dt
with state vector X (t) ∈ Rn input vector U (t) ∈ Rm where f is a smooth vector
field, is differentially flat if there exists a vector y(t) ∈ Rm in the form

y(t) = A X (t), U (t), U̇ (t), . . . . . . , U q (t) (2)
Such that

X (t) = B y(t), ẏ(t), . . . . . . , y r (t) (3)

U (t) = C y(t) ẏ(t), . . . . . . , y r +1 (t) (4)
where A, B and C are smooth functions. Therefore that the new systems description
is given by the m algebraic variables y(t).
Since the components of y(t) are differentially independent, the output plate
groups all the free variables of the system. But one can also say, through the Eq. (2),
which does not depend on the state and the control, what makes an endogenous
Sensorless Speed Control of Induction Motor … 5
variable system, for example in the state of an observer who is an exogenous variable
of the system.
Moreover, the notion of differential equivalence in the Lie-Bäcklund [11] sense
shows it well, the number of components of y(t) is given by that of the command that
is to say dim y(t) = dim U(t). This property allows knowing a priori the number of
free variables that must be found on a model to highlight its flatness. To shed more
light on the notion of flatness, one considers the following example:

Ẋ 1 (t) = X 2 (t)
(5)
Ẋ 2 (t) = U (t)
we define the following variables
y(t) = X 2 (t) and U (t) = ẏ(t)
So, we deduce
t
X 1 (t) = X 1 (t0 ) + y(τ )dτ (6)
t0
So y(t) cannot be considered a flat output because the relation (2) is not verified.
If now, we defines, y(t) = X 1 (t) then:
X 2 (t) = ẏ(t) and U (t) = ÿ(t)
In this case, we can say that X 1 (t) is a flat output, therefore this system is flat and
flat output y(t) = X 1 (t).
2.1 Control Law Synthesis Methodology for Flat Systems
The synthesis of a control law for a flat system requires a strategy based on the
following approaches:
• Generation of the trajectories of the reference flat output yr e f .
• Generation of the corresponding reference input u r e f trajectory (controls).
• Synthesis of a strategy of stabilization of the control around the planned reference
trajectories.
6 Y. Beddiaf et al.
2.2 Trajectory Planning
From the relation (4), if one wishes to obtain for the flat system (1), the trajectory:
z d (t) for a time t from t0 to t f , it suffices to impose, on the same time segment, the
following open-loop control:

(β)
Ud (t) = B z d (t), ż d (t), . . . . . . . . . . . . z d (t) (7)
In the hypothesis of a perfect model, we will then have, for t from t0 to t f ,

y(t) = yd (t), therefore:

X (t) = X d (t) = A z d (t), ż d (t), . . . . . . . . . . . . z d(α) (t) (8)

(γ )
y(t) = yd (t) = C z d (t), ż d (t), . . . . . . . . . . . . z d (t) (9)
2.3 Stabilization Around Reference Trajectories
A control developed from the flatness concept is established and defined for an
open-loop trajectory tracking. The perfect chase is ensured when the system is not
disturbed. However, non-linear physical systems are subjected on the one hand to
disturbances inherent to their working context, and on the other hand to uncertainties
about the parameters. It is therefore necessary to provide a solution to stabilize the
system around the trajectories if they are disturbed. Figure 1 illustrates the principle
of stabilization of the control by insertion of feedback loop.
3 Flatness Control of Im
In this section, we present the analyze of the flatness of IM model then the design its
control.
Fig. 1 Bloc diagram of

stabilization
3.1 Analysis of the Flatness of IM Model
The IM modeling stationary reference frame is given by the following equations:
d
V̄s = Rs ī s + ϕ̄s (10)
dt
d
0 = Rr īr + ϕ̄r (11)
dt
ϕ̄s = L s ī s + M īr e j pθ1 (12)
ϕ̄r = L r īr + M ī s e j pθ1 (13)
pM
Ce = Jm ī s · ϕ̄r∗ (14)
Lr
where * represents conjugated complex variables.

From Eq. (14), one can write:
d 2 θ1 pM
J = Jm ī s · ϕ̄r∗ − Cr − f θ̇ (15)
dt 2 Lr
We set (ρ) the modulus of the rotor flux and (δ) its position in the stationary
reference frame (α, β), the rotor flux is defined as:
ϕ̄r = ρe jδ (16)
From Fig. 2 we can write
ϕr = ϕ̄r e j pθ1 = ρe jα1 (17)
Fig. 2 Angle spotting

8 Y. Beddiaf et al.
We obtain:

1 L r d ϕ̄r
ī s = ϕ̄r + (18)
M Rr dt
p 2
ce = ρ α̇1 (19)
Rr
From Eq. (15) we define the following equation:

J · R f θ̇1 + θ̈1 + Cr
ρ= (20)
p α̇1
Therefore

ϕ̃r = ρe jα1 = B θ̈1 , θ̇1 , α1, α̇1 , Cr (21)
If we choose y = (α1 , θ1 ) as a flat output, we obtain:
ϕ̃r = B(y, ẏ, ÿ) (22)
1 d
ĩr = − ϕ̄r
R dt
r
= C θ̇1 , θ̈1 , θ1(3) , α̇1, α̈1 , Cr , Ċr (23)
e j pθ1
ī s = ϕ̄r − L r īr
M

= C θ̇1 , θ̈1 , θ1(3) , α̇1, α̈1 , Cr , Ċr (24)
ϕ̄s = L s ī s + Me j pθ1 īr

= E θ̇1 , θ̈1 , θ1(3) , α̇1, α̈1 , Cr , Ċr (25)
d
V̄s = Rs ī s ϕ̄s
dt
= F θ̇1 , θ̈1 , θ1(3) , θ1(4) , α̇1, α̈1 , α1(3) , Cr , Ċr , C̈r (26)
We note that Eqs. (17)–(26) satisfy the flatness conditions (2)–(4), so we can say
that the induction motor is a differentially flat system.
3.2 Design of Flatness Control Based of IM
The mechanical model of the induction motor in the (d–q) reference frame is given
by the following equations system:
⎧
⎪
⎨ dt =
dω p
ρ M I − fJ ω − CJr
J Rr Tr sq
dρ
= −1 ρ + TMr Isd (27)
⎪
⎩
dt Tr
dθ1
dt
=ω
where ρ = ϕr = ϕr d .
Replace (Isd , Isq , ρ, Cr ) by (Isd−r e f , Isq−r e f , ρr e f , Cr )
We obtain:
1
Isd−r e f = Cr ρ̇r e f + ρr e f (28)
M
L r J θ̈1−r e f + f θ̇1−r e f + Ĉr

Isq−r e f = (29)
pM ρr e f
where Ĉr represents the estimated charge torque.

To estimate the load torque, we use the following observer [3]:

= −l1 ω̂ − ω
d Ĉr
d ω̂
dt (30)
dt
= JpM ρ I − Jf ω̂ − ĈJr + l2 ω̂ − ω
L r r e f sq
where: l1 and l2 are the observer gains.

The flux (ρ) and the estimated speed (ω̂) are obtained by the reduced order state
observer [12].
The dynamic error of the observer (30) is given by:

d 1 0 −l1 1
= (31)
dt 2 − 1J − Jf + l2 2
where: 1 = Ĉr − Cr and 2 = ω̂ − ω.

The observer gain (30) is chosen in such a way that the dynamic error converges to
zero. More details about the reduced order state observer applicable to the induction
motor are given in [12].
10 Y. Beddiaf et al.
Finally, the control of stator currents of the reduced model is to ensured by the PI
regulators.
t
∗

Isd = k pρ ρ̂ − ρr e f + kiρ (ρ̂ − ρr e f )dτ (32)
0
t
∗

Isq = k pω ω̂ − ωr e f + kiω (ω̂ − ωr e f )dτ (33)
0
3.3 Planning Reference Trajectories
Applying the flatness approach, reference trajectories are defined by (ρr e f , θ1−r e f ).
The objective of this planning is:
– Respect the electromechanical constraints of the IM.
– Guarantee the existence of bounded derivatives up to order two.
Therefore, the methodology chosen is to apply to the speed and flux set points a
second-order filter making it possible to obtain the final reference trajectories that
are derivable twice.
3.4 Complete Model Control
The proposed flatness-based control scheme with the use of reduced observer for
estimation of the no measurable parameters of the motor’s state vector is shown in
Fig. 3. Equations (10)–(14) can be written as follows:
d I¯s M
Lsσ = V̄s − Rs I¯s − ρ̇ + j δ̇ρ e jδ (34)
dt Lr
Tr (ρ̇ + j α̇1 ρ) = −ρ + Me− jδ I¯s (35)

Fig. 3 Bloc diagduram of reduced-order state observer
The nominal control planned voltages are defined as follows:

∗∗ dδ
∗
d I ∗∗ M
Vsd∗ = σ L s γ Isd
∗∗
− Isq + sd − ρr e f (36)
dt dt σ L s L r Tr
∗ ∗∗
∗ ∗∗ ∗∗ dδ
d Isq M ∗
Vsq = σ L s γ Isq + Isd + + pω ρr e f (37)
dt dt σ Ls Lr
t
where: δr e f = pθ1−r e f + α̇1−r e f dτ .
0
∗∗ ∗∗
Currents Isd and Isq are given by:
t
∗∗ ∗

Isd = Isd + k pρ ρ̂ − ρ ∗ + kiρ ρ̂ − ρ ∗ dτ (38)
0
t
∗∗ ∗
∗

Isq = Isq + k pω ρ̂ − ρ + kiω ω̂ − ω∗ dτ (39)
0
From Eqs. (38), (39) and (27), we can write:

kiρ ρ̂ − ρr e f
∗∗
d Isd d Isd−r e f Tr
= + (40)
dt dt M +k pρ −Trρ̂ + TMr Isd − ρ̇r e f

∗∗
d Isq ∗
d Isq
Tr kiω ω̂ − ω∗ +
= + ∗∗ (41)
dt dt M k pω pM
J Lr
ρ̂ Isq − Jf ω − Ĉr
J
− ω̇∗
Finally the control voltages are given by:

⎛ ⎞
t

Vsd = Vsd∗ + σ L s ⎝k p I sd Isd − Isd + ki I sd ∗∗ ∗∗
Isd − Isd dτ ⎠
0
⎛ ⎞
t

Vsq = Vsq∗ + σ L s ⎝k p I sq Isq − Isq + ki I sq ∗∗ ∗∗
Isq − Isq dτ ⎠
0
4 Reduced State Observer
In this paper, a reduced state observer is used to estimate the rotor speed and rotor
flux [12]. Model (42) is used for realize the observer, that’s magnetizing currents
i mr α and i mrβ are used as state variables. stator currents i sα and i sβ being the input
variables. The correction of the observer is done from the variables Y and Ŷ which
depend on the input vectors and its derivative.
⎧ di
⎪
⎪
sα
= − R
i + M 1
ϕ α + ωϕ + 1
Vsα
⎪ dt
⎪
σ Ls sα σ L s L r Tr r rβ
σ Ls
⎨ disβ
= − σ RL s i sβ + σ LMs L r T1r ϕrβ − ωϕr α + 1
σ Ls
Vsβ
dt (42)
⎪
⎪ dϕr α
= TMr i sα − T1r ϕr α − ωϕrβ
⎪
⎪
⎩ dt
dϕrβ
dt
= TMr i sβ − T1r ϕrβ + ωϕr α
The observer is synthesized by following equation:

Ẋ = A X̂ + BU + G Y − Ŷ
(43)
Ŷ = DU + Ji U̇ + C X̂

−1 −1
−ω 0 k1 k2 R 0
where: A = Tr
; B = Tr
;G = ; D = ; Ji =
ω −1 0 −1 −k2 k1 0 R

Tr
Tr

−1
σ Ls 0 M2 −ω î mr α i sα Vsα
; = Lr Tr
; X̂ = ;U = ;Y = ; Ŷ =
0 σ Ls ω −1 Tr î mrβ i sβ Vsβ

V̂sα
.
V̂sβ
In case the speed is not measurable. It is replaced by its estimated value ω̂, as
in our case, it comes from a speed estimator that will be developed later. The state
equation of system (42) is given by:
Ẋ = AX + BU (44)
Y = DU + Ji U̇ + C X (45)
Equation (43) becomes:

Ẋ = (A − GC)X + BU + G Z (46)
where: Z = C X .
Let’s put A − GC = Aobs
It comes that:

Ẋ = Aobs X̂ + BU + G Z (47)
Aobs Define the observer’s dynamics which is chosen as a function of the

convergence speed of the observation error towards the zero value.
The gain matrix G is determined using the pole imposition method. the structure
of the observer is shown in Fig. 3.
For the estimation of speed, several methods have been proposed [12]. We can
use the mechanical equation to estimate the rotor speed as shown in Figs. 4 and 5.
Fig. 4 Bloc diagram of speed estimator

Fig. 5 Schematic diagram the proposed flatness-based control scheme with the use of reduced
observer
5 Simulation and Experimental Results
The sensorless control used flatness theory is simulated under the environment
(Matlab/Simulink).The parameters of the machine are: Rr = 4.2 ; Rs = 5.72 ;
Ls = 0.462 H; Lr = 0.462 H; M = 0.44 H; J = 0.0049 kg m2 ; f = 0.003 Nm s/rad;
P = 2; Power of asynchronous motor is 1.5 kW.
The results of simulation are shown in Fig. 6. These results are obtained by
simulation with a load (1, 5 Nm).The speed is shown in Fig. 6a. The estimated speed
always followed the reference. The load applied does not affect the performance of
the system. Good speed estimation was obtained. We can seen in Fig. 6d that the flux
(ϕ̂r αβ ) installs correctly and well sinusoidal, Fig. 6c shows that the estimated torque
follows the reference value (Figs. 7 and 8).
Fig. 6 a Simulation result of rotor speed. b Simulation result of rotor speed (zoom). c Simulation
result of estimated torque. d Simulation result of estimated rotor flux. e Simulation result of estimated
rotor flux (zoom)
Fig. 6 (continued)
Figure 9 shows the DSP based implementation of the proposed speed sensorless
control. The sampling frequency is fixed at 10 kHz and the controller receives the
stator currents measurements through two 8-bit A/D converters. Then, using the
PWM technique, the reference voltages are sent to the machine via the voltage-source
inverter whose switching frequency is fixed at 10 kHz.
Fig. 7 a Experimental result of rotor speed. b Experimental result of rotor speed (zoom). c Exper-
imental result of estimated torque. d Experimental result of estimated rotor flux. e Experimental
result of estimated rotor flux (zoom)
Fig. 7 (continued)
Figure 7 shows the experimental results obtained for the proposed sensorless
control by differential flatness method. A rotor speed reference is imposed with a
load torque equal to 1.5 Nm applied at time t = 4.2 s.
By comparing the results, we can say that both results are very similar. It will be
noticed that the estimated and the real rotor speed signals are very close. The results
confirm the validity of the proposed method. According to the results we can say
that the speed response obtained by the flatness technique is significantly improved
compared to that obtained by the classical sensorless vector control, Fig. 8.
Fig. 8 a Experimental result of rotor speed (classical FOC). b Experimental result of rotor speed
(classical FOC) (zoom). c Experimental result of Estimated Rotor Flux. d Experimental result of
Estimated Rotor Flux (Zoom)
Fig. 8 (continued)
Fig. 9 The photograph of the experimental test system

6 Conclusion
The paper has studied sensorless control, for induction motors using differential
flatness theory and reduced state observer.
The experimental results have shown that an adequate sensorless speed control
by differential flatness method of IM drive can be achieved at rated, low, and zero
reference speed control. Obtained results show that this control strategy assures a
perfect linearization regardless trajectory profiles physically imposed on the induc-
tion machine. It can be noticed that the reduced state observer is an efficient approach
for the implementation of state estimation-based control of induction motor. Finally
the proposed control has given very satisfactory results in terms of load disturbance
rejection and tracking rotor speed.
References
1. Morand F (2005) Techniques d’observation sans capteur de vitesse en vue de la Commande

des machines asynchrones, Thèse de doctorat, soutenue à l’université de Lyon, France, le 7
janvier 2005
2. Al-Rouh I (2005) Contribution à la commande sans capteur de la machine asynchrone, Thèse
de doctorat, Faculté des Sciences & Techniques—54506 Vandœuvre-lès-Nancy, Juillet 2005
3. De Wit PAS, Ortega R, Mareels I (1996) Indirect field-oriented control of induction motors is
robustly globally stable. Automatica 32(10):1393–1402
4. Zbede YB, Gadoue SM, Atkinson DJ (2016) Model predictive MRAS estimator for sensorless
induction motor drives. IEEE Trans Indus Electron 63(6):3511–3521
5. Comanescu M (2016) Design and implementation of a highly robust sensorless sliding mode
observer for the flux magnitude of the induction motor. IEEE Trans Energy Conv 31(2):649–657
6. Chang G-W, Espinosa-Pérez G, Mendes E, Ortega R (2000) Tuning rules for the PI gains of
field-oriented controllers of induction motors. IEEE Trans Ind Electr 47(3):592–602
7. Lorenz RD, Lipo TA, Novotny DW (1994) Motion control with induction motors. In:
Proceeding of IEEE, power electronics and motion control, vol 82, no 8, pp 1215–1240, August
1994
8. de Doncker RW, Novotny DW (1994) The universal field oriented controller. In: IEEE
transactions on industry applications, vol 30, no 1, pp 92–100, January–February 1994
9. Fliess M, Lévine J, Martin P, Rouchon P (2007) Flatness and defect of non-linear systems,
introductory theory and examples. Int J Control 61(6):1327–1361
10. Dannehl J, Fuchs FW (2006) Flatness-based control of an induction machine fed via voltage
source inverter—concept, control design and performance analysis. In: Proceedings of the 2nd
annual. IEEE industrial electronics conference, pp 5125, 5130, 6–10 November 2006
11. Fliess M, Lévine J, Martin P, Rouchon P (1999) A Lie-Bäcklund approach to equivalence and
flatness of nonlinear systems. IEEE Trans Autom Control 44:922–937
12. Beddiaf Y, Zidani F, Alaoui LC, Drid S (2016) Modified speed sensorless indirect field-oriented
control of induction motor drive. Int J Identif Control 25(4):273– 286
Multi-UAVs Coverage Path Planning
Abdelwahhab Bouras, Yasser Bouzid, and Mohamed Guiatni
Abstract In this literature review, we attempt to summarize and analyze some recent
works on Coverage Path Planning (CPP) strategies. We are interested by those that
consider UAV fleets. Covering a region of interest is like sweeping a space in either
2D or 3D, and visiting specific points just once each, to cover as much space as
possible with obstacle prevention. These strategies are classified according to their
evolution trends, such as grid-based methods, evolutionary algorithms, sampling
based methods, etc.
Index Terms Path planning · Multi-robots · Multi-UAVs · Coverage path planning
1 Introduction
Autonomous aerial vehicles have become a very attractive field for researchers
over the last two decades. Their implication in wide range of applications (e.g.
aerial photography, agricultural imaging, search and rescue operations, etc.), brings
several benefits. Moreover, the multi-UAVs configurations compensate the limitation
of flight autonomy and minimize mission time.
This survey deals with the Coverage Path Planning (CPP) strategies using a fleet
of UAVs. Usually, each UAV should cover its assigned sub-region by visiting a finite
number of points. The locations of these points must be well distributed to ensure the
full coverage and satisfy certain constraints such as: flight time, energy consumption,
obstacles avoidance, feasibility, etc. Most of published CPP algorithms proceed by
four steps that are:
A. Bouras (B) · Y. Bouzid · M. Guiatni

CSCS Laboratory, Ecole militaire polytechnique, Bordj El Bahri, Algeria
Y. Bouzid
M. Guiatni

https://doi.org/10.1007/978-981-15-6403-1_2
24 A. Bouras et al.
• (S1) The decomposition of the Region of Interest (RoI) into sub-regions with
obstacle isolation using several strategies (e.g. Trapezoidal decomposition, Morse
decomposition, Voronoi diagram, etc.).
• (S2) Squares, triangles, hexagons, etc. are the usual discretization forms (depends
on the footprint of the used sensor) of the sub-regions (obtained in S1) for a good
localization and an adaptation to the algorithms of the next step.
• (S3) A local path planning algorithm that requires a starting cell and a goal cell is
needed to deal with the discrete map provided by S2. Many strategies have been
proposed in the literature, which are discussed here after.
• (S4) A step of smoothing is used to make the planned trajectory feasible by the
drone. Generally, this step does not pose a real problem for UAVs with Vertical
Take-Off and Landing (VTOL) ability.
The Multi-UAVs CPP can be done in several ways. It depends mainly, on the algo-
rithms and the onboard sensors. A family of these algorithms requires a discretiza-
tion of the space such as A*, D*, etc. However, this category cannot be apply in the
case of multi-UAVs, unless they can be applied for each UAV independently. Active
Rapidly-exploring Random Trees (Active RRT), RRT*, Dynamic Domain RRT, Arti-
ficial Potential Field (APF), etc. do not require a discretization of the space. A load
distribution between UAVs is necessary when an iterative path planning algorithm
is adopted.
In fact, the updates on the multi case of coverage are rare. Thus, this paper gives a
summary of recent works, during the last three years, dealing with the CPP of multi-
UAVs. It will be a good introduction for the interested researchers on the multi-UAVs
CPP field.
The remaining of the paper is organized as follows: Sect. 2 provides a review of
the literature, presenting the current state of the art on. Sections 3, 4 and 5 describes
Grid based methods, Evolutionary Algorithms and Sampling based methods of CPP,
whereas Sect. 6 closes the paper with a conclusion and future works.
2 Related Works
In the literature, many path planning strategies for one or multi-UAVs have been
developed. Herein, we are interested by the second case (multi-UAVs) knowing that
the generalization of many algorithms of the first class for a fleet of UAVs is a
subject of research, or, simplifying and rendering a problem of multi agent to simple
problems already known and optimized for the singular case, for this reason, the
works dealing with this multi agent coverage problem must take into account the
division of the RoI.
Choset [1] has classified the CPP algorithms according to the decomposition class:
non-decomposition, exact and approximate decomposition. The author presents for
each category the progress of the corresponding methods intended to improve the
performance of the coverage. This classification is widely used later in similar works.
Multi-UAVs Coverage Path Planning 25
Galceran and Carreras [2] illustrate the most redundant classical offline methods
of exact decomposition. They focus on Boustrophdon strategy, which aims to fuse the
small adjacent cells resulting from the Trapezoid decomposition method (see Fig. 1a),
and reduce the global number of cells (see Fig. 1b). Then, they show some variants of
Morse-based cellular decomposition method and other popular approaches of decom-
position such as the Grid-based method. After discretization, several algorithms are
used for the coverage of the workspace, such as Wave Front algorithm, Spanning
Tree Coverage (STC), Backtracking Spiral Algorithm and Neural Networks, etc.
For the case of multi-UAVs, the authors discuss the use of Boustrophdon decom-
position with a division of roles of UAVs, some of them work as scouts of the
borders, while the others make the round trip in the inner space in order to complete
an Online coverage mission. The generalization of coverage algorithms from single
case to multi such as Spanning Tree Coverage (STC) which has become Multi STC
(MSTC), this category fits into the class of approximation decomposition, see Fig. 2.
The next point on techniques using the neural networks each robot sees as dynamic
Fig. 1 Variants of Morse decomposition method: a Trapezoid decomposition. b Boustrophdon

decomposition
Fig. 2 Approximate cellular decomposition with square cells of different sizes

26 A. Bouras et al.
obstacles. Then; Graph-based, bioinspired algorithms are limited in practical applica-

tions. Finally, multi homogeneous or heterogeneous aerial robots have had a variety
of applications and constraints for this task.
After three definitions (clear, short and accurate) on the planning and the gener-
ation of trajectories, Yang et al. [3] arrange the algorithms of path planning (3D
environment) in five categories. These algorithms are well analyzed and discussed.
A summary is given in the Table 1. For more details, the reader may refer to [4, 5].
In paper [6], RRT variants have been combined with other strategies in order to
improve the computation time, the required memory, etc. Moreover, future challenges
have been exposed.
In our work, a new criterion for the classification of CPP techniques has been
proposed, this choice being in relation to current work trends, such as: grid-based
methods, evolutionary algorithms, sampling based methods. These techniques are
more adapted to the constraints imposed by this theme. Furthermore, since 2016,
this type of update does not exist.
3 Grid Based Methods (Approximate Cellular

Decomposition)
3.1 Grid of Elementary Cells of Square Shapes
The decomposition of the RoI into elementary cells of the same or different sizes, this
is relative to the importance of each sub-region. In general, this is reflected by different
flight altitudes for the UAVs, which gives different dimensions of the footprint of the
sensors by choosing the right compromise with the coverage resolution. In Kapoutsis
et al. [7] propose an offline algorithm named Divide Areas Algorithm for multi-Robot
coverage Path planning (DARP), for the balanced sharing of the RoI according to
the number of robots. For that, five rules have been established to better fit the
multi Coverage Path Planning (mCPP) problem. Indeed, the rules consist of ensuring
non-backtracking, full coverage and balancing efforts between the robots, then, the
Spanning Tree Coverage (STC) algorithm is used to cover and improve the speed
of calculation of robot paths. However, choice of coverage direction for back-and-
forth (BF), memory capacity and the calculation time are still points to improve.
Figure 3 shows a simulation variant of this technique with seven UAVs, and set up
several obstacles of our choice. In the same trend [8] have introduced improvements
to this algorithm (DARP), with an optimization of Spanning Tree’s generation by Ant
Colony Optimization (ACO) algorithm of each sub-region, and the introduction of
nodes exchanges between the sub-regions in a second step (ST nodes), which gave
a remarkable reduction in the number of turns, and reduce the energy consumed.
However, the same problem of computation time persists. In [9], the nodes (vertexes)
of STC are distributed to the STCs of the robots by criteria of priorities adopted
with the auction (Auction-STC) (auctioneer & bidder), the path traveled by each
Table 1 Categories of 3D path planning algorithms
Category Types of algorithms Advantages Disadvantages Notes
Sampling based methods Passive: PRM, Voronoi, – RRT*, DDRRT, RRG, – RRT does not propose the • Requires a pre-knowledge
RRG, K-PRM proposed the issue of issue of re-planning of the environment, then;
Active: RRT, RRT *, re-planning – PRM and Voronoi can not mapping or rand to find the
DDRRT, Artificial potential – Simple and easy structure optimize alone (requires right planning
field (APF) to implement optimization with A* for • Calculation time:
– Suitable for static and example) O(n log n) ≤ T ≤ O n2
dynamic environments, and – Problems of local minima • Several RRT and PRM
online planning conditions for Artificial potential series have been developed
field (APF) for the improvement of
previous versions
Node based methods Dijkstra, A *, Tetha *, Lazy – Easy to be combined with – Do not reproduce the exact • An objective function (cost
Tetha*, D *, Harmonic other methods to improve form of the environment function) is necessary for
Search, etc. optimization – Are not suitable for multi the optimization
– Suitable to work online agents • Calculation time:
O(n log n) ≤ T ≤ O n2
Mathematical methods Linear programming and – Takes into account all – Calculation time is • Calculation time: depends
optimal control dynamic and kinematics important on the complexity of the
Linear binary programming constraints by – No analytical resolution problem
Non linear programming differential equations methods • Maybe useful more in
Mixed integer linear – Suitable for working offline advancing in computer
programming (MILP) technologies
(continued)
27
Table 1 (continued)
28
Category Types of algorithms Advantages Disadvantages Notes

Bio-inspired methods EA: Evolutionary algorithms – Able to solve NP-hard – Are not practical in real • Calculation
time:
(ACO), Particle swarm problems, complex situations O n2 ≤ T
optimization (PSO), Shuffled environment – Are not suitable for online • Generally simulations work
frog leaping algorithm – Suitable for work on multi work
(SFLA), Memetic algorithm agents
(MA)
NN: Neural networks
Multi-fusion-based methods Embedded (mixed): Reactive – Enjoy the advantages of the – Keeps the disadvantages of • Calculation time:
deformation road maps previous methods the combined algorithms O(n log n) ≤ T
(RDR), Visibility voronoi – Possibility to solve a large
potential (VVP), etc. part of the problems
Ranked (succession – Choice of the combination
algorithms): Probabilistic depending on the problem
Road Maps (PRM)
optimized, Visibility Graphs
optimized, etc.
A. Bouras et al.
Fig. 3 DARP simulation of seven drones with multiple obstacles
robot is estimated online (in the same time) of this process by quantifying vertexes,
including those of big cells, partially filled with obstacles. This approach shows
a robust resolution of the problem of mCPP, either in complex environments or
with the increase of the number of robots. However, small redundancies of coverage
appear, the problem of the initial conditions and the number of turns and/or the power
consumption are not discussed.
3.2 Triangles Grid
Other work used a considerable tools for a full simulation of the space coverage
problem with a fleet of UAVs [10], Balampanis et al. solved the problem in three
steps:
• The triangulation for the decomposition taking into account the size of the triangles
so that it is compatible with the capture (footprint) of the sensor used, and to make
the triangles equilateral as possible (Fig. 4);
• In a second step, the authors proposed two algorithms called Antagonizing Wave
front Propagation and Reverse Watershed Schema (AWP & RWS) for the distri-
bution of sub-regions to be covered by UAVs, taking into account the capabilities
of UAVs and getting rid of dead zones by Dead Lock Handling (DLH) algorithm;
• Finally; “Waypoint list computation for coverage” is used for coverage.
30 A. Bouras et al.
Fig. 4 Approximate cellular

decomposition of a square
area with a circular obstacle
using triangular cells of the
same size
RWS has shown its advantage in a comparative study. This work was accompanied
by an assessment of the complexity of the algorithms, the type of sensor and the flight
capabilities of the UAVs. However, the resulting trajectories for the coverage are not
discussed in terms of the number of turns and/or energy consumed.
3.3 Hexagons Grid
In [11], Azpúrua et al. take advantage of the decomposition of the region of interest
in hexagons. Where this type of decomposition offers a minimization of the inter-
cellular distance compared to that of small squares, and also a good orientation of
the cells compared to the case of the triangles, as well as this technique takes into
consideration the range of communications between the robots. Indeed, the division
of the space to cover for each robot is ensured with the K-means algorithm. After
that, ensuring coverage using TSP. Finally, a Finite State Machine algorithm (FSM)
has been proposed to ensure synchronization between the robots.
After the introduction of some dynamic parameters (α, h: Orientation of the
hexagons and distance between parallel lines of coverage respectively), and the
respect of the feasibility of the mission (size of the hexagons (radius r), autonomy
of the robots, obstacles), their simulation work shows that:
• The coverage time decreases as the number of robots or the inter-coverage lines
(h) increases;
• The optimal orientation of the hexagons reduces the intercellular distance. Hence,
the reduction of the trajectory length, therefore decreased coverage time.
However, the comparison of the proposed methodology with cooperative

approaches (proposed in [12]) shows that the coverage time in the latter is less
than in the proposed technique, either for the case of individual flight or in formation
(Multi-robot). This difference is due to the extra-terrain coverage caused by the prin-
ciple of the space decomposition (exact and non-exact decomposition). On the other
hand, this disadvantage becomes an advantage in the case of the evaluation of the
quality of detection (magnetic map), this is translated by a collection of more data
compared to [12]. An experimental study of the proposed technique was performed.
4 Evolutionary Algorithms (Swarm Intelligence)
Among the most used algorithms is ACO, given its efficiency in complex environ-
ments, and the possibility of using several ants in parallel for different constraints.
Chaotic ACO of Coverage (CACOC) [13], is a version of ACO, the authors use
a chaotic process based on Rösler system instead of Random process, this change is
introduced when there is an absence of pheromone help with the decision (guidance)
of UAVs. The same authors extend this algorithm in [14], by adding a mechanism of
Collision Avoidance (CA), which gives CA-CACOC. In this case, except that other
drones for each of them that are considered obstacles. This algorithm is simulated in
a V-REP (Drone Model Simulation) environment using the Model Predictive Control
(MPC), which gives results in terms of coverage-which we are interested in-better
than previous versions, but, this type of algorithm is always influenced by the choice
of the initial parameters.
Other works use a Multi ACO for the CPP problem, in [15], the update of the
pheromone of each colony is relative to the fitness of the ten best ants of the previous
colony, in another word, each colony benefit from the results of the other colonies,
which improves the convergence rate of the algorithm on the one hand, and out of the
local minima on the other hand. Nevertheless, computation time and implementation
in the real world are the major disadvantage of this algorithm.
5 Sampling Based Methods
These methods are based on the use of the algorithms of the first category of Table 1,
which need to have a preliminary knowledge on the RoI (in the form of points
or cells etc.), both the starting points and the arrival points to conduct an iterative
process (sampling) creating a tree of nodes connected to each other, then obtaining
the optimal trajectory towards the arrival points. The way in which these nodes are
selected or connected at each iteration is the key to the success of the resulting
trajectories. Therefore, the calculation time, the narrow passages are the essential
points to improve.
32 A. Bouras et al.
Fig. 5 The expanded trees and a solution path found by the RRT Connect algorithm (a) (9.75 s,
7347 nodes) and LM-RRT approach (b) (1.35 s, 3286 nodes)
Wang et al. [16] propose the Learning-based Multi-RRT algorithm (LM-RRT) as

a solution for narrow passages, this approach is structured in four steps, the first one
for sampling by the improve Bridge Test algorithm. Then, the hierarchical group
obtains the multiple trees in a second step. The third step consists in choosing the
tree by an adaptive process. The last step is the extension and the connection of the
chosen tree. If the final trajectory is not yet obtained, the algorithm returns to the
beginning of the third step. The performances of this algorithm are clear in Fig. 5 in
terms of computing time, number of nodes, trajectory quality.
In [17], Bouzid et al. use multi-directional algorithm Rapidly-exploring Random
Trees Star Fixed Nodes (RRT * FN) to optimize inter-point-of-interest distances, take
the optimal path and avoid obstacles. The coverage is solved by the combination of
TSP with genetic algorithms for the passage of all selected points. The numerical
simulation of a single UAV shows the effectiveness of this solution.
6 Other Techniques
The coverage problem is started in [18] in three steps, the approximate reproduction
of the terrain (NURBS: Non Uniform Rational B-Spline), this provides the necessary
data for the determination of control points, taking into account the coordinates of cell
centers and altitude, in order to cover the terrain by Multi-UAVs Vehicle Routing.
To minimize the mission time of the UAVs, the authors adopt a criterion of the
minimization of the distance traveled by each UAV at the same time, instead of
minimizing the sum of their traveled distances. This trick makes it possible to reduce
the overall time of the mission with similar and/or equivalent traveled distances
by UAVs, these distances are adapted to the constraints of the UAVs, that instead
of having a minimum sum of the distances traveled with loads unbalanced on the
UAVs.
Modares et al. [19] introduce a second term of optimization of energy compared
to the TSP, in addition to optimizing the energy compared to the distance traveled by
the UAV, the authors also take into account the angle described by the path between
adjacent cells (0°, 45°, 90°, 135°, and 180°), i.e.; take into account the image of the
number of turns (the energy consumed during turns) in the optimization criterion,
which gives more meaning to the energy consumed. This term has been introduced in
Lin-Kernighan Heuristic (LKH) used in TSP, which has named LKH-D for drones.
The comparison with LKH shows that the chosen path is more economical in energy
despite small losses in computing time, this time decreases more by increasing the
number of agents (UAVs).
7 Conclusion
This works presents a recent research dealing the problem of coverage path plan-
ning for aerial detection applications using UAV fleets. This works tends towards
the combination of several algorithms to benefit from their advantages in order to
remedy the constraints posed by the coverage problem. These cell decomposition and
connection techniques between cells (Table 2) must take into account the require-
ments of the application being processed and the available working means such as
the flight capabilities of UAVs and the types of on-board sensors. On another side;
computation time, required memory space and the diversity of applications are still
research topics. Improvements in Sampling based methods have opened up another
broad line of research, given the challenges and constraints resulting each time, and
the wide scope of their uses. Despite that; the use of these algorithms for coverage
is not much used yet.
In future works, the research base will be broadened by the introduction of several
practical works dealing with this theme, as well as the proposal of the techniques
in the same subject by the treatment of the different terrains with several kinds of
obstacles, focusing on sampling based methods.
Table 2 Categories of path planning approaches with used algorithm
34
References Multi/Single Decomposition 2D/3D Decomposition Coverage Offline/Online Notes

approaches algorithm algorithm
Kapoutsis et al. [7] Multi Approximate cellular 2D Grid based DARP STC Offline • Critics of STC persist
decomposition • Requires large
(Squares) memory capacity
Balampanis et al. Multi Approximate cellular 2D Grid based AWP & RWS + Offline • The number of turns
[10] decomposition Waypoint list and/or energy
(Triangles) computation consumed is not
discussed
Gao et al. [8] Multi Approximate cellular 2D Grid based DARP STC + ACO + Offline • Reduction of the
decomposition node exchanges number of turns
(Squares) • Requires large
memory capacity
Azpúrua et al. [11] Multi Approximate cellular 2D K-means TSP Offline • FSM: pour
decomposition synchronisation
(Hexagons) • Extra field problem
Dentler et al. [14], Multi Approximate cellular 2D Grid based CA-CACOC + Offline • Influenced by the
Cekmez et al. [15] decomposition Multi ACO choice of initial
(Squares) parameters
• Problems with
computing time and
implementation in
the real world
(continued)
A. Bouras et al.
Table 2 (continued)
References Multi/Single Decomposition 2D/3D Decomposition Coverage Offline/Online Notes
approaches algorithm algorithm
Choi et al. [18] Multi NURBS: Non Uniform 3D NURBS: Non Multi-UAVs Offline • Compromise: total
Rational B-Spline Uniform Rational Vehicle Routing distance and time of
B-Spline the mission
Gao and Xin [9] Multi Exact cellular 2D Grid based Auction-STC Offline • Increase in the
decomposition number of turns
Modares et al. [19] Multi Approximate cellular 2D Grid based STC Offline • Take into
decomposition consideration the
(Squares) number of turns in
the optimization of
the TSP
35
36 A. Bouras et al.
References
1. Choset H (2001) Coverage for robotics—a survey of recent results. Ann Math Artif Intell
31(1):113–126
2. Galceran E, Carreras M (2013) A survey on coverage path planning for robotics. Robot Auton
Syst 61(12):1258–1276
3. Yang L, Qi J, Song D, Xiao J, Han J, Xia Y (2016) Survey of Robot 3D path planning algorithms.
J Control Sci Eng 2016:1–22
4. Goerzen C, Kong Z, Mettler B (2010) A survey of motion planning algorithms from the
perspective of autonomous UAV guidance. J Intell Robot Syst 57(1–4):65–100
5. LaValle SM (2006) Planning algorithms. Cambridge University Press, Cambridge
6. Noreen I, Khan A, Habib Z (2016) Optimal path planning using RRT* based approaches: a
survey and future directions. Int J Adv Comput Sci Appl 7(11)
7. Kapoutsis ACH, Chatzichristofis SA, Kosmatopoulos EB (2017) DARP: divide areas algorithm
for optimal multi-robot coverage path planning. J Intell Robot Syst 86(3–4):663–680
8. Gao C, Kou Y, Li Z, Xu A, Li Y, Chang Y (2018) Optimal multirobot coverage path planning:
ideal-shaped spanning tree. Math Prob. Eng 2018:1–10
9. Gao G-Q, Xin B (2019) A-STC: auction-based spanning tree coverage algorithm formotion
planning of cooperative robots. Front Inf Technol Electron Eng 20(1):18–31
10. Balampanis F, Maza I, Ollero A (2017) Coastal areas division and coverage with multiple
UAVs for remote sensing. Sensors 17(4):808
11. Azpúrua H, Freitas GM, Macharet DG, Campos MF (2018) Multi-robot coverage path planning
using hexagonal segmentation for geophysical surveys. Robotica 36(8):1144–1166
12. Maza I, Ollero A (2007) Multiple UAV cooperative searching operation using polygon area
decomposition and efficient coverage algorithms. In: Alami R, Chatila R, Asama H (eds)
Distributed autonomous robotic systems 6. Springer, Tokyo, Japan, pp 221–230
13. Rosalie M, Danoy G, Chaumette S, Bouvry P (2016) From random process to chaotic behavior
in swarms of UAVs. In: 6th ACM symposium on development and analysis of intelligent
vehicular networks and applications, Malta, pp 9–15
14. Dentler J et al (2019) Collision avoidance effects on the mobility of a UAV swarm using chaotic
ant colony with model predictive control. J Intell Robot Syst 93(1–2):227–243
15. Cekmez U, Ozsiginan M, Sahingoz OK (2018) Multi-UAV path planning with multi colony
ant optimization. In: Abraham A, Muhuri PK, Muda AK, Gandhi N (eds) Intelligent systems
design and applications, vol 736. Cham: Springer International Publishing, pp 407–417
16. Wang W, Zuo L, Xu X (2018) A learning-based multi-RRT approach for robot path planning
in narrow passages. J Intell Robot Syst 90(1–2):81–100
17. Bouzid Y, Bestaoui Y, Siguerdidjane H (2017) Two-scale geometric path planning of quadrotor
with obstacle avoidance: First step toward coverage algorithm. In: 2017 11th international
workshop on robot motion and control (RoMoCo), Wasowo Palace, Poland, pp 166–171
18. Choi Y, Chen M, Choi Y, Briceno S, Mavris D (2019) Multi-UAV trajectory optimization
utilizing a NURBS-based terrain model for an aerial imaging mission. J Intell Robot Syst.
https://doi.org/10.1007/s10846-019-01027-9
19. Modares J, Ghanei F, Mastronarde N, Dantu K (2017) UB-ANC planner: energy efficient
coverage path planning with multiple drones. In: 2017 IEEE international conference on
robotics and automation (ICRA), Singapore, pp 6182–6189
Reducing the Collision Checking Time
in Cluttered Environment
for Sampling-Based Motion Planning
Amine Belaid and Boubekeur Mendil
Abstract In the Sampling-based motion planning, the collision checking procedure

called many times and this operation takes the most planning time. The time spend
in the collision checking increases with the number of obstacles. This paper aims
to reduce the number of calls of the collision checking procedure in environment
containing high number of obstacles. Our approach shows that the complexity of
collision checking does not depend on the number of obstacles. The idea is to perform
a pre-procedure that localizes the obstacles by segmenting the workspace into a grid
and save it in a matrix form. This information is used in the collision checking
procedure. We have used the Probabilistic Roadmaps (PRM) motion planner with a
punctual robot and polygonal obstacles in 2D environment. Simulation results show
the efficiency of our algorithm in reducing the complexity of collision checking and
minimizing the planning time.
Keywords Path planning · Probabilistic roadmap · Collision checking · Motion

planner
1 Introduction
Motion planning is widely used in robotics, games, manufacturing and many related
areas, to find free collision path of objects, characters and robots. Sampling-
based motion planning methods, such as Probabilistic Roadmaps [1], and Rapidly
Exploring Random Tree (RRT) [2], are developed especially for problems with high
degree of freedom. These algorithms avoid explicit computation of obstacle bound-
aries in the configuration space, C-space. Instead it uses sampling techniques to
compute paths in the free space Cfree .
A. Belaid (B) · B. Mendil

Industrial Technologies and Information Laboratory, Faculty of Technology,
Bejaia University, 06000 Bejaia, Algeria
B. Mendil
https://doi.org/10.1007/978-981-15-6403-1_3
38 A. Belaid and B. Mendil
Collision checking procedure in sampling-based path planners is considered like a

black box called each time needed [3]. The first presented sampling motion planning
method uses the uniform distribution to generate samples in the configuration space.
However, this strategy has bad performance in complex spaces, like the presence
of narrow passages. To deal with this problem, many sampling strategies have been
proposed like Gaussian sampling [4], Bridge sampling, etc. These strategies use
an intensive number of calls to collision checking procedure, and as a result, the
sampling phase takes a long time.
Many algorithms have been proposed to reduce the number of calls to collision
checking procedure and, thus, to minimize the planning time in workspaces with
high number of obstacles. The most of them use prior information about the sampled
configurations [6, 7] or collision queries [5, 8, 9]. In other words, they modify the
sampling strategy to reduce the complexity of the approach.
A wide variety of approaches and techniques for geometric modeling exist. The
choice usually depends on the application and the difficulty of the problem. Two
models, with different collision checking procedures are often used in path planning:
Bitmap and Polygonal or Polyhedral representation for 2D or 3D environments [3].
For the polygonal representation, the obstacles and the robot are represented by
polygons, each polygon is described by its vertices. in the other hand, for the Bitmap
representation, the environment is discretized into rectangular cells that may or may
not be occupied. This information is saved in two or three-dimensional matrix. Each
geometric modeling technique presents advantages and drawbacks. But when the
workspace is large the bitmap needs a huge storage space [3].
In this paper we present a new technique to address the problem of collision
checking in 2D cluttered environment for sampling-based methods. Unlike the
previous presented approaches, we don’t use prior information about the sampled
configurations, and we don’t modify the sampling strategy. The main contribution of
our method is to disassociate the dependence between the number of obstacles and
the complexity of the algorithm. The idea is to localize the obstacles by discretizing
the environment into cells in the beginning of the motion-planning algorithm and to
save them as a matrix, which is used in the collision checking procedure. This pre-
procedure allows the avoidance of the collision checking with all obstacles existing
in the workspace.
In order to test the performance of our method, we integrated it with the PRM
path planner. The 2D environment is considered. The obstacles are modeled by poly-
gons and the robot is punctual. The results show the efficiency of our approach
in environment with high number of obstacles, comparing with the traditional
techniques.
The rest of this paper is organized as follows. The Sect. 2 presents the Probabilistic
Roadmap path planner. Section 3 gives the details of our approach. The simulation
results and discussion are presented in Sect. 4. Section 5 concludes this paper.
Reducing the Collision Checking Time in Cluttered … 39
2 Probabilistic Roadmap
The planners based on an explicit representation of the configuration space cannot

solve problems with high degree of freedom. Because, as the dimension of the config-
uration space grows, these planners become impractical, due to the required time.
The sampling-based planners are developed to solve this kind of problems [10].
The probabilistic roadmap [1] was the first proposed sampling based-algorithm.
It proceeds in two phases: learning phase and query phase. In the learning phase, the
sampled configurations are stored in a graph data structure called a roadmap. The
latter is constructed in a probabilistic way. The roadmap is an unidirectional graph
R = (V, E), where V is a set of free configurations sampled randomly and E is a set
of edges corresponding to local paths. For every configuration q ∈ V, a set N q of k
closest neighbors to the configuration q according to some metric distance, dist, is
chosen from V. The local planner is called to connect q to each configuration q0 ∈
N q . If the local planner succeeds in computing a feasible path between q and q0 , the
edge (q, q0 ) is added to the roadmap.
Algorithm 1 Learning phase of the Probabilistic Roadmap

Input:
n : number of nodes to put in the roadmap
k : number of closest neighbors to examine for each
configuration
Output:
A roadmap G(V,E)
E ←{}
V ← set of free configurations generated by a sampling
method
for all q V do
Nq ← the k closest neighbors of q chosen from V
according to dist
for all q0 Nq do
if (q,q0) E and ∆(q,q0) ≠ NIL then
E ← E {(q,q0)}
end if
end for
end for
In the query phase, the initial and the end configurations qinit , qgoal are connected
to the roadmap using the local planner. Then, an algorithm of finding the shortest
path in the graph, like A∗ or Dijkstra, is used to find the path between the start and the
end configuration. A post-processing step can be performed to improve the quality
of resulting path by checking whether nonadjacent configurations q and q0 along the
path can be connected with the local planner.
3 Collision Checking in Cluttered Environment
Collision testing takes the most of time in path planning, especially with high number
of obstacles, since the temporal complexity increases with the number of obstacles.
Our main contribution in this paper is to disassociate the dependence of collision
testing complexity with the number of obstacles in the workspace. To avoid the
collision test with all the obstacles, we add a pre-processing to localize the obstacles
in the workspace by decomposing the workspace into rectangular cells. We give for
each obstacle a rectangular region by taking the maximum and the minimum of X
and Y abscise for the obstacle vertices (Fig. 1). Then, we construct the matrix M
representing the workspace. Each matrix element contains the number of obstacle
regions overlapping the corresponding cell. Once the construction of the matrix M
is done, it can be used now in the collision checking procedure. After localizing the
sampled configuration in the workspace, we do the collision checking with just the
obstacles found in the cell of the matrix M that corresponding to the location of the
sampled configuration.
Fig. 1 A non-convex obstacle A with his rectangular region found by taking the maximum abscises
according to X and Y
Algorithm 2 Query phase of the Probabilistic Roadmap

Input:
qinit : the initial configuration
qgoal : the goal configuration
k : number of closest neighbors to examine for each
configuration
G(V,E) : the roadmap constructed by the Algorithm 1
Output:
A path from qinit to qgoal or failure
V ← V {qinit} {qgoal}
for q ←{qinit,qgoal} do
Nq ← the k closest neighbors of q chosen from V
according to dist
q0 ← the closest neighbor of q from Nq
while Nq is not empty do
if ∆(q,q0) ≠ NIL then
E ← E (q,q0)
N ← N −{q0}
else
N ← N −{q0}
end if
q0 ← the closest neighbor of q from
Nq
end while
end for
3.1 Pre-processing (Obstacles Localization)
The idea of a pre-processing algorithm is to perform the most of collision checking

tests in the beginning. In this paper, we consider the environment is in 2D, and the
obstacles are non-convex polygons, so the matrix M that represent the environment
is of dimension (n × m). First of all, the algorithm divides the environment into a
grid as is shown in Fig. 2. The number of the grid cells along the X axis is n and the
number of the grid cells along the Y axis is m. Then, the algorithm creates the (n ×
m) empty matrix M which have the same dimension of the grid. The obstacles are
Fig. 2 2D environment divided into grid with its corresponding matrix M. the bold rectangles are
the obstacles regions. One cell can be covered by more than one obstacle region. NIL means the
cell is not cover by any obstacle’s region
numbered from 1 to k. The algorithm gives for each obstacle Oi a rectangular region
(Fig. 1), where the vertices coordinates of this region in clockwise direction are
X Oir egion = [min(X Oi) min(X Oi) max(X Oi) max(X Oi)],
Y Oir egion = [min(Y Oi) max(Y Oi) max(Y Oi) min(Y Oi)]
In the workspace, each obstacle region covers some cells of the grid, so for each
obstacle region, the algorithm will write its number in all the elements of the matrix
M that correspond to the cells of the grid covered by this obstacle region. Note that
a cell of the grid can be covered by more than one obstacle region. An example of
creating the matrix M for a 2-dimension environment is shown in Fig. 2.
3.2 Collision Checking Procedure
By considering the robot is punctual, in this procedure, a configuration is sampled,

then, the algorithm finds the cell of the grid that contain this configuration. Next, the
collision checking will perform with just the obstacles whose number are found in
the element of the matrix M that correspond to this cell. By applying this strategy,
we avoid the collision checking with all obstacles in the workspace. This strategy
can be used with polygonal robot, where we find all the cells that contain the robot
then we perform the collision checking with the union of obstacles numbers found
in the elements of the matrix M corresponding to those cells. And this algorithm can
be generalized to three dimensions environment, which is our perspective.
Algorithm 3 pre-processing (obstacles localization)

Input:
Obs : obstacles vertices coordinate
Output:
The matrix M
for A in Obs do
Imin,Imax ← indices of cells that contain min(XA), max(XA)
respectively according to X axis
Jmin,Jmax ← indices of cells that contain min(YA), max(YA)
respectively according to Y axis
for i ← Imin to Imax do
for j ← Jmin to Jmax do
if N(A) in M(i,j) then
Continue
else
M(i,j) ← M(i,j)&N(A)
end if
end for
end for
end for
Algorithm 4 Collision checking procedure

Input:
Obs : obstacles vertices coordinate
conf : the configuration to test for collision
M : the matrix that produced by algorithm 3
Output:
true: if the configuration is in collision ; false: otherwise
i,j ← indices of cells that contain conf according to X axis
and Y axis
Obst ← M(i,j)
for A in Obst do
if conf in collision with A then
return true
end if
end for
return false
4 Simulation and Results
To state the performance of our approach, we integrated it with the probabilistic

roadmap planner. Then, we compared its performance with PRM using a traditional
collision checking. A 2D environment (Figs. 3, 4, and 5) and a punctual robot model
are considered where the scene is taken to be [−10, 10] along both, X axis and Y axis.
The obstacles are convex or non-convex polygons generated randomly. We compared
the simulation results of just the learning phase of the PRM. The Gaussian sampling
Fig. 3 Execution time in function of number of obstacles for learning phase of PRM path planning
with traditional collision checking, using the uniform sampling and the Gaussian sampling strategies
Fig. 4 Execution time in function of number of obstacles for learning phase of PRM path planning
with our collision checking approach, using the uniform sampling and the Gaussian sampling
strategies
Fig. 5 The 2D environment considered with 300 obstacles generated randomly

strategy with σ = 0.2, and the uniform sampling strategy are chosen. The number of
configurations to generate in each simulation is set to n = 500. In each simulation,
we take the average time of 3 attempts. All the simulations have done in MATLAB
2014a with 2.3 GHz DualCore Intel PC.
Figure 3 shows the variation of execution time of learning phase in function of
number of obstacles for PRM with a traditional collision checking procedure. For the
Gaussian sampling the time execution increase according to the number of obstacles
until t = 112 s with around 150 obstacles, then it decreases to stay roughly constant t
= 96 s, this decreasing is because the scene is cluttered of obstacles, so the free space
will narrow, and this gives high chances in finding the second configuration for the
Gaussian sampling. On the other hand, in Fig. 4 the execution time of the learning
phase with Gaussian sampling and using our algorithm decrease according to the
number of obstacles from t = 6.6 to t = 2.3 s with 300 obstacles, and here is the
power of our algorithm. In other comparison with uniform sampling, the execution
time with the traditional collision checking increase linearly according to the number
of obstacles until t = 7.96 s with 300 obstacles. In other hand, when we used our
algorithm with the uniform sampling, the execution time stay roughly constant,
around t = 0.28 s.
5 Conclusion
This paper presents a new collision-checking algorithm for sampling-based motion

planning in environment with high number of obstacles. We have integrated our
algorithm with the PRM path planner and compared its execution time of the learning
phase with that of PRM using a traditional collision checking. Gaussian and uniform
sampling strategies are used to test the performance of our algorithm. Simulation
results show the efficiency of our approach in collision checking especially with
high number of obstacles. In future works, we will investigate our approach in the
case where the robot is a polygon in 2D environment then in 3D where the robot and
the obstacles are modeled by polyhedrons.
References
1. Kavraki LE et al (1996) Probabilistic roadmaps for path planning in high-dimensional

configuration spaces. IEEE Trans Robot Autom 12(4):566–580
2. Kuffner JJ, LaValle SM (2000) RRT-connect: an efficient approach to single-query path plan-
ning. In: Proceedings ICRA 2000, IEEE international conference on robotics and automation,
vol 2. IEEE
3. LaValle SM (2006) Planning algorithms. Cambridge University Press, Cambridge
4. Boor V, Overmars MH, van der Stappen AF (1999) The Gaussian sampling strategy for prob-
abilistic roadmap planners. In: Proceedings of the 1999 IEEE international conference on
robotics and automation, vol 2. IEEE
5. Kim D, Kwon Y, Yoon S (2018) Adaptive lazy collision checking for optimal sampling-based
motion planning. In: 2018 15th international conference on ubiquitous robots (UR). IEEE
6. Hauser K (2015) Lazy collision checking in asymptotically-optimal motion planning. In: 2015
IEEE international conference on robotics and automation (ICRA). IEEE
7. Bialkowski J et al (2013) Efficient collision checking in sampling-based motion planning. In:
Algorithmic foundations of robotics X. Springer, Berlin, Heidelberg, pp 365–380
8. Pan J, Manocha D (2016) Fast probabilistic collision checking for sampling-based motion
planning using locality-sensitive hashing. Int J Robot Res 35(12):1477–1496
9. Daee P, Taheri K, Moradi H (2014) A sampling algorithm for reducing the number of collision
checking in probabilistic roadmaps. In: 2014 22nd Iranian conference on electrical engineering
(ICEE). IEEE
10. Choset HM (2005) Principles of robot motion: theory, algorithms, and implementation. MIT
Press, Cambridge
Random Scaling-Based Bat Algorithm
for Greenhouse Thermal Model
Identification and Experimental
Validation
Mounir Guesbaya, Hassina Megherbi, and Ahmed Chaouki Megherbi
Abstract In this paper, we propose a Random Scaling-based Bat Algorithm (RSBA)

for parametric identification of a greenhouse thermal model. The proposition includes
modifying the exploitation of the standard BA by making the scaling parameter
changes randomly over the iterations. The proposed thermal model identification
method has been assessed first on a simulated greenhouse thermal model with known
parameters. The simulation results have shown the superiority of the proposed RSBA
compared to the standard BA in term of convergence and performance accuracy. To
experimentally investigate the proposed identification method, we used a greenhouse
prototype under arid climate conditions located in M’ziraa, Biskra, Algeria. The
obtained prediction results are found to be in a good agreement with the measured
ones which show the effectiveness of the proposed RSBA in identifying the real
greenhouse thermal model.
Keywords Bat Algorithm · Exploitation and exploration · Metaheuristic

optimization · Greenhouse thermal modelling · Parameter identification
1 Introduction
Greenhouse systems have become a prominent mean in the agricultural field. It is

constantly developed in order to reach the quantity and quality of crops that meet
the world growing population and the strict standards of the competitive markets
today. The inside air temperature of greenhouse plays an important role in affecting
M. Guesbaya · A. C. Megherbi
Laboratory of Identification, Command, Control and Communication, University of Mohamed
Kheider, Biskra, Algeria
A. C. Megherbi
H. Megherbi (B)
Department of Electrical Engineering, University of Mohamed Kheider, Biskra, Algeria

https://doi.org/10.1007/978-981-15-6403-1_4
48 M. Guesbaya et al.
the yield qualitatively and quantitatively. Analysing and controlling this physical
process needs an optimized model that describes it with sufficient accuracy and low
complexities and costs [1]. Provision of such a model will increase the implementa-
tion of the new intelligent control approaches to improve agriculture production in
greenhouses.
The Bat Algorithm (BA) is a popular population-based and stochastic algorithm
developed by Yang [2]. It shares the feature of swarm intelligence with the most
popular nature-inspired algorithms. BA is characterized by quick convergence at an
earlier stage of the search process, frequency-tuning which increases the diversity
of solutions during the search and the parameter control feature of BA that leads to
achieve smart moves from explorative to local intensive exploitative search steps.
However, the standard BA includes problems with local and global dimensional
search space. Among these problems; the low accuracy and the slow convergence
rate when the search is near to optimal solution, which are indicated in many research
studies [3–5].
In this paper, we propose a random scaling-based bat algorithm (RSBA) to iden-
tify the static parameters of a greenhouse thermal model under arid desert climate
conditions. In the RSBA, the scaling parameter that determines the step size changes
randomly over iterations. The idea behind this proposition is to update the scaling
parameter in a way that makes the step size hit relatively large and small values
to maintain an effective step size control in relative to the closeness of the optimal
solution. This mechanism aims to affect and enhance the exploitation mechanism
especially in the final stage of the search. The greenhouse thermal model used in this
work is proposed in [6]. This model is devoted to predict the inside air temperature.
It describes this phenomenon as a set of heat exchanges generated by the differences
in energy content between the inside and outside air.
The rest of this paper is organized as follows. In Sect. 2, the used grey-box thermal
model is described. Section 3 includes the details of the RSBA-based identification
method of the static parameters of the used thermal model. In Sect. 4, simulation
and experimental results of the thermal model identification process are discussed.
Section 5 concludes this paper.
2 Greenhouse Thermal Model
The grey-box thermal model used in this work is presented in [6]. Greenhouse models
represent the heat exchange processes between all the different subsystems of the
greenhouse. However, in the proposed model only two components are studied; the
cover and the internal air. It is considered as a mean that quantitatively describes the
energy exchanges. The temperature dynamic behaviour is a combination of physical
interactions including the conduction, the convection, the solar and thermal radiation
and the infiltration as depicted in Fig. 1. These processes are basically affected by
the external environmental conditions and the structure of the greenhouse [7]. The
constructed greenhouse is used as a nursery. Hence, the canopy which is constituted
Random Scaling-Based Bat Algorithm for Greenhouse Thermal … 49
Fig. 1 Heat exchanges affecting the inside air temperature
of seedlings has a small effect on inside air temperature due to its small reserved
area. The nursery includes trays full of treated soil without planted seedlings. The
following assumptions have been taken into account:
• The temperature of the canopy is neglected.
• The temperature of the treated soil in seedling trays is neglected because it has a
small effect compared with the other elements due to the reserved small surface.
• The heat exchange between external soil and the inside air is neglected, due to
the role of the wooden floor as a separator.
• Transfer coefficient of the convective heat and the absorbed solar radiation are
uniform throughout the cover.
• There is no stratification in greenhouse air temperature.
• The heat storage of greenhouse elements can be neglected.
• The greenhouse is East-West oriented.
The inside heat balance equation [6] is given by:
V dTin
ρa = Q solar − Q cnv,cnd − Q loss − Q ther mal (1)
S f dt
where t is the time in seconds, Tin is the inside air temperature (K), V is the volume
of the greenhouse (m3 ), S f is the floor surface (m2 ) and ρa is the inside air density
(kg m−3 ) which is given by:
ρa = γ0 + Ha (2)
where, γ0 is the inside dry air density (kg m−3 ) and Ha is the absolute humidity (kg
m−3 ), which has been obtained by converting the relative humidity as follows:
(0.62198 Pw )
Ha = (3)
Patm − Pw
where, Patm (kPa) is the atmospheric pressure, Pw is the instant vapour pressure,
calculated as follows:
Hr
Pw = Pws · (4)
100
where, Hr is the relative humidity of greenhouse (%) and Pws is the saturation vapour
pressure (kPa °C−1 ); it is defined by:
(17.2694(Tin −273.15))
Pws = 0.61078 e Tin
(5)
Q solar is the shortwave radiation absorbed by the greenhouse given by:
Q solar = Csolar I (6)
where, Csolar is the solar energy coefficient and I is the external global solar radiation
(W m−2 ).
Q cnv,cnd , is the convection and conduction heat exchange rate given by:
Q cnv,cnd = Ccnd,cnv U (Tin − Ta ) (7)
where, Ccnd,cnv is the convection and conduction energy coefficient, Ta is the ambient
temperature (K) and U is the overall heat transfer coefficient through the greenhouse
cover calculated as follows:
−1
1 1
U= + (8)
ho hi
where, h o and h i are respectively, the convective heat transfer coefficients of the
outside and inside greenhouse cover (W m−2 K−1 ). They are calculated by:
h o = 2.8 + 1.2 Wv (9)
hi = 5.2 · |Tin − Ta |0.33 (10)
where, Wv presents external wind velocity (m s−1 ).

Q loss is the leakage rate of air through the greenhouse given by:
Q loss = Closs ρa (Tin − Ta ) (11)

Table 1 Input parameters

Symbol Value Unit Description
used for computation
Sf 0.68 m2 Greenhouse floor surface
V 0.76 m3 Greenhouse volume
γ0 1.205 Kg m−3 The inside dry air density
Patm 101 kPa The atmospheric pressure
where, Closs is the heat loss coefficient.

Q ther mal is the longwave radiation absorbed by the greenhouse given by:

Q ther mal = Cther mal h o Tin − Tsky (12)
where, Cther mal is the thermal energy coefficient and Tsky is the sky temperature (K).
The linearly dependent static parameters: Csolar , Ccnv,cnd , Closs and Cther mal have
to be accurately known, which highlights the need for their identification. The other
static parameters are fixed and given in Table 1.
3 Thermal Model Identification Based on Bat Algorithms
This work concerns the application of the proposed RSBA in an off-line parametric
identification process. The problem of identification in this paper consists of finding
the optimal static parameters (Csolar , Ccnv,cnd , Closs and Cther mal ) of the thermal
model described in Eqs. (1)–(12), that fit the data samples of the inside air temperature
of the greenhouse.
BA is a bio-inspired population-based metaheuristic algorithm. It has been devel-
oped based on bats behaviour of how they search on their targets using the echoloca-
tion capability [2]. Virtual bats are used in simulation and a set of bases are defined
for clarification as follows:
• All bats sense distance and differentiate between prey and background barriers
by using echolocation.
• Bats fly randomly (Random walks technique) with velocity vi at position xi . The
frequency (or wavelength) of their emitted pulses is automatically adjusted.
• The rate of pulse emission r can also be tuned automatically according to the
closeness of target, or considered as a constant.
• The loudness can be considered variant, starting from a large positive value A0 to
a minimum value Amin or it can also be a constant.
Bats position (solution) xit and velocity vit in a predefined d-dimensional search
space are updated according to the following equations:
f i = f min + ( f max − f min )β (13)


vit+1 = vit + xit − x∗ f i (14)
xit+1 = xit + vit+1 (15)
where, f i ∈ [ f min , f max ] is the randomly assigned frequency to each bat. β ∈ [0, 1]
is a random variable drawn from a uniform distribution, x∗ is the current global best
solution which is selected after comparing all the solutions of all the n bats.
When a solution is chosen among the current best solutions, a new solution
for some bats (according to pulse emission rate) is generated locally during the
exploitation stage using random walk based on the following equation:
xnew = xold + σ t At (16)
where, xnew and xold are respectively the new and the old best local solutions, t ∈
[−1, 1] is a random number and σ ψ is a scaling parameter to control the step size.
The scaling parameter is declared constant for the standard BA. It should be
linked to the scalings of the design variables of the problem under consideration
[2]. However, this is not enough to reach the optimality, because when the optimal
solution is near to be reached, the constant step size of search (local random walk)
remains relatively large even with the effect of loudness on local search. It results
in a reduction in convergence speed and low-accurate solution as a sub-stagnation
state at the end. Based on this fact, we proposed to make the scaling parameter σ:
1. Fully responsible about step size control by eliminating the effect of loudness on
local search equation given by:
xnew = xold + σ t t (17)
2. Dynamically updated over iterations based on random scaling parameter mech-

anism, given by:
σ t+1 = σmin + (σmax − σmin )β (18)
where the scaling parameter σ ∈ [σmin , σmax ].

As in standard BA, the loudness Ai and the pulse emission rate ri have to be
updated as the iterations proceed. The loudness generally decreases once a bat finds
the target, whereas the rate of pulse emission increases. They are updated based on
the following equations:
Ait+1 = α Ait (19)


rit+1 = ri0 1 − ex p(−γ t) (20)
where α and γ are constants. For any 0 < α < 1 and γ > 0, the change in loudness
and pulse rate is directed as follows:
Ait → 0, rit → ri0 , as t → ∞
In order to replace the parameter values xit j out of the search range

max x j , min x j with random ones, an algorithm has been implemented, it is given
as follows:

I f xit j > max x j or xit j < min x j then

xit j = min x j + max x j − min x j · rand(0, 1)
The objective function to be minimized by BAs is the commonly used least squares
criterion, defined as:

N
2
J= yri eal − y ipr e (21)
i=1
where, yri eal and y ipr e are respectively the real and the predicted data samples and N
is the number of data samples.
Based on the aforementioned assumptions and rules the essential stages of the
proposed RSBA are summarized as the schematic pseudo code presented in Fig. 2.
4 Simulation and Experimental Validation
The aim of this section is two folds. Firstly, it is to investigate the efficiency of the
RSBA against the standard BA in identifying simulated greenhouse thermal model
with known parameters. Secondly, it is to use the RSBA to identify the thermal model
of an experimental greenhouse prototype. The obtained model will be then validated
experimentally for predicting the inside air temperature of this prototype.
A. Identification setup and specifications
The searching ranges of the model parameters are listed in Table 2. The common
control parameters between standard BA and RSBA are: the number of population
is n = 100 bats, the minimum and maximum frequency respectively are f min = 0
and f max = 1.5, the loudness of the initial bats is A0i = 1, the rate of pulse emission
Random Scaling-based Bat Algorithm

Initialize the bat population and
Initialize frequencies , pulse rates and the loudness
while (Current iteration < Maximum number of iterations)
Generate new solutions (positions) by frequency tuning
Update velocities and solutions Eqs. (14-15)
if (rand > )
Select a solution among the best solutions
Update scaling parameter value Eq. (18)
Generate a local solution around the selected best
solution Eq. (17)
end if
Generate a new solution by flying randomly
if (rand < and < )
Accept the new solutions
Increase and reduce Eq. (19-20)
end if
Find the current best
end while
Fig. 2 Pseudo code of RSBA
Table 2 The search range of the parameters to be identified

Parameter Csolar Ccnv,cnd Closs Cther mal

Range 0, 10−4 0, 10−3 0, 10−3 0, 10−4
of the initial bats is generated randomly r i ∈ [0, 1] and ri0 = 0.1, and the constants
ψ ← ψ ← 0.9. The scaling parameter of the standard BA is constant
σ = 10−2 .
−7 −2
Whereas the range of its variations regarding RSBA is 10 , 10 .
B. Analysis of search dynamics
The analysis is done based on ten runs for each BA. The results of identification are
analysed according to a limited number of iterations (500 iterations) and achieved
using different random initial populations. The evolution of the best fitness function
values for all the runs of BAs along the iterations are illustrated in Fig. 3. It can be
noticed that standard BA seems to have a weak performance for all the runs in term
of solutions precision. In contrast, RSBA runs have achieved better performance in
term of convergence speed at late stage and solution accuracy. Table 3 includes all
the fitness function values and illustrates that RSBA scored the best fitness value, as
a result of the influence of random scaling parameter on the local search step size.
Moreover, it shows a good balance between convergence speed and result quality.
As a result, RSBA proves to be more efficient than the standard BA in finding the
parameters of the thermal model.
Fig. 3 Evolution of the best fitness function values
Table 3 Best fitness function

Identification process Bat Algorithms
values of identification
processes Standard BA RSBA
1st run 3.4003e−04 9.7949e−08
2nd run 1.4187e−03 2.6286e−05
3rd run 3.6587e−04 1.0008e−06
4th run 2.1736e−04 1.4771e−04
5th run 2.0633e−04 1.3245e−05
6th run 2.2142e−04 2.9300e−05
7th run 2.8384e−05 2.2254e−04
8th run 3.7457e−05 7.1692e−07
9th run 1.7585e−04 3.5900e−04
10th run 1.0056e−04 1.5055e−06
C. The evaluation of parameter values found by RSBA
The relative error criterion is used to evaluate the best identified parameter values by
RSBA against the optimal ones. It is given by:
| p R S B A − pr eal |
RE = · 100 (22)
pr eal
where, pr eal is the assumed parameter value and p R S B A is the parameter value
identified by the proposed RSBA
The evolution of the relative error of every identified parameter value from the
best RSBA execution is shown in Fig. 4. It is obvious that the RSBA succeeded at
the end to find the parameter values. The best parameter values are listed in Table 4,
next to their corresponding relative errors. It shows that the value of Closs is the
less accurate identified parameter due to its small effect on the output, in contrast
to Cther mal and Csolar which are the most accurate identified parameters because of
their strong effect on the inside air temperature. It is recognized that the acceptable
identification result corresponds to the relative errors of less than 0.5%. Hence, the
results show that the parameters have been identified with satisfactory accuracy.
D. Experimental setup
(a) Description of the greenhouse prototype
A small-size single span wooden-structured greenhouse has been
constructed as a nursery prototype covered by polyethene with 0.2 mm
of thickness as it appears in Fig. 5. It is located in the municipality of
M’ziraa, affiliated to Biskra province in Algeria (34°43 19.7 N 6°17 39.2
Fig. 4 Evolution of the relative error of every parameter from the best identification run of the
RSBA
Table 4 Comparison
Parameter Assumed value Identified value Relative error
between the assumed and the
(%)
identified model parameter
values using rsba Csolar 5.8e−05 5.800004e−05 6.007544e−05
Ccnv,cnd 5.3e−04 5.300057e−04 1.067788e−03
Closs 1.7e−04 1.699894e−04 6.251072e−03
Cther mal 2.2e−05 2.199999e−05 5.369802e−05
Fig. 5 External and internal view of greenhouse prototype
E); it is an arid region with a moderate desert climate. The experimental

nursery has a wooden soilless floor that embraces three seedling bunches
full of treated soil.
(b) Data acquisition system
A low-cost data acquisition system has been designed based on two
programmable boards; Arduino Mega 2560 and NodeMCU v0.1. It gathers
measurements from all the sensors then send them to the PC via Wi-Fi.
The data is treated by a MATLAB code to create the database used in this
study. The internal and external climate variables are measured using the
following sensors:
– Two DHT22 equipped with a low-cost design of a plate-shaped radia-
tion shield [8]. They are installed indoor the greenhouse at high of 0.3 m
and outdoor the greenhouse at high of 1.25 m to measure air tempera-
ture and relative humidity. The accuracy of humidity and temperature
respectively are ± 5%RH and < ±0.5 °C.
– A low-cost pyranometer based on BPW34 silicon photodiode with an
accuracy of 2.3% error [9]. It is installed outdoor the greenhouse at a
height of 1.4 m, to measure the global solar irradiation.
– A low-cost anemometer based on DC motor is installed outside at a
height of 1.55 m [10].
The sensitiveness of each of the sensors is listed in Table 5.
(c) Greenhouse database
The database consists of measurements of the inside and the outside air
temperature, the inside relative humidity, the outside global solar radiation
and the wind velocity. These measurements are recorded in five consecutive
days in the winter season (26th to 30th of January 2019) as it appears in
Table 5 Sensors sensitiveness

Sensor Temperature Humidity Pyranometer Anemometer
Sensitiveness 0.1 °C 0.1% 9.6 w/m2 0.47 m/s
Figs. 6 and 7. The sampling time was 1 min. The data has been divided
into two parts; the data of 2nd and 3rd days are devoted to the identification
process. The diversification of their climate (calm and turbulent) will ensure
Fig. 6 Inside and outside air temperature and inside relative humidity
Fig. 7 External global solar radiation and wind velocity

Table 6 The identified

Parameter Experimentally identified value
parameters of the real
greenhouse model Csolar 6,83455e−05
Ccnv,cnd 4,14191e−04
Closs 2,55726e−04
Cther mal 2,55627e−05
an effective selection of model parameter values and prediction flexibility

for various climate states. whereas, the data of the other three days are kept
for the experimental validation.
E. Experimental results
We aim in this section to validate the simulation outcomes of the proposed thermal
model identification strategy, using an experimental database obtained from the
greenhouse prototype.
The RSBA-based identification of the experimental thermal model has been
achieved using the same control parameters as for the identification of the simu-
lated thermal model. Whereas, in this section, the target is the real measured inside
air temperature. The identified parameter values using the data of the 2nd and 3rd
days of the experimental database are listed in Table 6. The results of thermal predic-
tion using these identified parameter values with the measured inside temperature of
the 2nd and 3rd days are shown in Fig. 8.
The validation of the identified model is achieved by predicting the measured
inside air temperature of the 1st, 4th and 5th days. The variation of predicted and
measured inside air temperature are illustrated in Fig. 9. The performance of the
identified model is evaluated based on five criteria: Mean Absolute Error (MAE), Max
Fig. 8 Measured and predicted inside temperature of the 2nd and 3rd days
Fig. 9 Measured and predicted inside temperature of the 1st, 4th and 5th days
Table 7 Statistical evaluation of predicted inside air temperature

Criterion MAE R2 EF MaxAE
Value 0.793692 0.995624 0.995127 3.391057
Absolute Error (MaxAE), Coefficient of determination (R2 ) and Model Efficiency

(EF). They evaluate the difference between predicted and measured air temperature
in different techniques based on Eqs. (23)–(26).
1
n
M AE = Tm (i) − T p (i) (23)
n i=1

MaxAE = max Tm (i) − T p (i) (24)

n 2
Tm (i) − Tm T p (i) − T p
i=1
R =
2
n n (25)
i=1 Tm (i) − Tm · i=1 T p (i) − T p
n 2 n 2
i=1 Tm (i) − Tm − i=1 T p (i) − Tm (i)
EF = n 2 (26)
i=1 Tm (i) − Tm
where Tm is the measured temperature, Tm is the mean value of the measured tempera-
ture, T p is the predicted temperature, T p is the mean value of the predicted temperature
and n is the number of data samples.
Table 7 gives the results of this statistical analysis as an evaluation of the iden-
tified model. The MAE shows a satisfactory evaluation represented by a very small
difference between the measured and predicted inside air temperature. We can also
see that R 2 gives a very high value for regression analyses meaning that 99.5% of
the variance in the measured inside air temperature of the greenhouse prototype has
been predicted by the real thermal model. Moreover, EF shows a very good evalua-
tion where the closer the value is to “1” the more accurate the model is. The MaxAE
shows that the biggest error value among all the data samples equals to 3.39 °C which
is logically small. The aforementioned evaluation criteria illustrate that the model
is a promising tool to study this phenomenon in term of usefulness, efficiency and
applicability.
5 Conclusion
In this paper, we proposed a variant of BA called RSBA to identify the thermal

model of a greenhouse. The proposed RSBA differ from the standard BA in using
a random scaling parameter mechanism. A comparative study has been conducted
between the proposed RSBA and the standard BA on a simulated thermal model with
known parameters. By examining the evolution of the search and the results of the
optimization, we found that the RSBA outperforms the standard BA and its capability
of the exploitation has been enhanced. Specifically, the accuracy of the solution and
the speed of convergence when the optimal solution is near to be found have been
effectively increased. The RSBA was then applied to identify a real greenhouse
thermal model. The results of the identified model validation using the experimental
database exhibited a very good fit between the measured and predicted inside air
temperature.
This study paves the way for future investigation on developing a novel mechanism
that will make the scaling parameter changes adaptatively with the closeness of the
optimal solution.
References
1. Ma D, Carpenter N, Maki H, Rehman TU, Tuinstra MR, Jin J (2019) Greenhouse environment
modeling and simulation for microclimate control. Comput Electron Agric 162:134–142
2. Yang XS (2014) Nature-inspired optimization algorithms, 1st edn. Elsevier
3. Lyu S, Li Z, Huang Y, Wang J, Hu J (2019) Improved self-adaptive Bat Algorithm with step-
control and mutation mechanisms. J Comput Sci 30:65–78
4. Ramli MR, Abas ZA, Desa MI, Abidin ZZ, Alazzam MB (2018) Enhanced convergence of Bat
Algorithm based on dimensional and inertia weight factor. J King Saud Univ Comput Inform
Sci (In press)
5. Yilmaz S, Kucuksille EU (2013) Improved Bat Algorithm (IBA) on continuous optimization
problems. Lect Notes Softw Eng 1(3):279–283
6. Guesbaya M, Megherbi H (2019) Thermal modeling and prediction of soilless greenhouse in
arid region based on particle swarm optimization: experimentally validated. In: International
conference on advanced electrical engineering, Algiers, Algeria
7. Rodríguez F, Berenguel M, Luis Guzmán J, Ramírez-Arias A (2015) Modeling and control of

greenhouse crop growth. Springer International Publishing, Switzerland
8. Holden ZA, Klene AE, Keefe RF, Moisen GG (2013) Design and evaluation of an inexpensive
radiation shield for monitoring surface air temperatures. Agric Meteorol 180:281–286
9. Čekon M, Slávik R, Juráš P (2016) Obtainable method of measuring the solar radiant flux based
on silicone photodiode element. Appl Mech Mater 824:477–484
10. Horsey J (2016) DIY Arduino wind speed meter anemometer project. Geeky Gadgets. https://
www.geeky-gadgets.com/arduino-wind-speed-meter-anemometer-project-30032016/
Control and Observation of Induction
Motor Using First-Order Sliding Mode
Salah Eddine Farhi, Djamel Sakri, and Noureddine Goléa
Abstract A sliding mode controller is one of the non-linear techniques that have
advantages of being robust and simple to implement. In this paper, a First Order
Sliding Mode (FOSM) is chosen to provide decoupled control of the induction
machine. In addition and in order to eliminate the use of sensors, flux and speed
observers based on First Order Sliding Mode are introduced. The proof of stability of
the algorithm is also presented using Lyapunov theory. The obtained results through
simulations verifies that the designed control scheme have strong robustness toward
the variation of the speed, load torque and parameter uncertainty.
Keywords Induction Motor (IM) · Lyapunov · Observer · Sensors · Sliding mode

controller · First order sliding mode (FOSM)
1 Introduction
The asynchronous or induction motor is the most widely used electrical propulsion
motor, his invention at the end of the last century gave a strong impetus to the transi-
tion from DC to AC motor [1]. It is the most popular machine in many fields such as
chemistry, medicine, aerospace and various industries, the induction motor has many
advantages more than the other types of electrical machines, among which: robust-
ness, low cost, power maximum rating, maximum top speed, ease of maintenance.
Another aspect is more important in the realization of electric motors is the notion
of robustness. The models used are approximate and the variation of electrical and
S. E. Farhi (B) · D. Sakri · N. Goléa

Laboratory of Electrical Engineering and Automatic , Department of Electrical Engineering,
LGEA, Université Larbi Ben M’hidi, Oum El Bouaghi, Algeria
D. Sakri
N. Goléa

https://doi.org/10.1007/978-981-15-6403-1_5
64 S. E. Farhi et al.
mechanical parameters degrades the performance of the controls and can lead, in
some cases, to unstable operations. Powerful nonlinear control techniques have been
presented widely in electric drives. Among them, we mention sliding mode control
[2].
Sliding mode control theory, given its ability to circumventing the effects of
parameter variation along with low implementation complexity [3] and disturbance
rejection, strong robustness and simplicity of execution, is one of the future control
methodologies for electrical machines [4, 5]. Sliding mode setting which by its nature
is a nonlinear control, is a powerful tuning technique is basically a method that forces
the response to slide along a predefined path.
For getting high performance with an asynchronous machine, requiring reliable
information from process control [6], Most of the research in induction motors control
during the last years has been on sensorless solutions [7]. Observer design is an
important topic in control system [8]. Sensorless control means that an encoder is
replaced by estimated speed and flux from an observer or an estimator in order to
reduce cost and increases the reliability of the system [9]. Sliding mode observers
have unique characteristics, that a produces a set of state estimates that are strictly
proportional with the actual output of the plant [10] and makes the flux and speed
estimation robust to all variations, such as sensors and measurement noises, off set
in the measurement system, phase shift in the measured value [11].
High-performance controllers are achieved when speed and flux are readily avail-
able, it seems better to estimate the speed and flux [12]. To replace the sensor, the
information is extracted on the rotor speed and flux from measured stator voltages
and currents at the ends of the motor [13].
In this paper, the sliding mode approach is introduced both for control and for
observers design basing only on the measurements of stator voltages and currents.
2 Model of Induction Motor
The mathematical model of IM in the stationary reference (α, β) can be written as

follows [14]:
1 M
ϕ̇r α = − ϕr α − ωr ϕrβ + i sα
Tr Tr
1 M
ϕ̇rβ = ωr ϕr α − ϕrβ + i sβ
Tr Tr
1
i̇ sα = a1 ϕr α + a1 Tr ωr ϕrβ − a2 i sα + Vsα
σ Ls
1
i̇ sβ = −a1 Tr ωr ϕr α + a1 ϕrβ − a2 i sβ + Vsβ (1)
σ Ls
Control and Observation of Induction Motor … 65
The torque equation that describes the relationship between the motor torque and
the rotor speed is depicted as:
˙ = Tm − TL − F
J (2)
where the electromagnetic torque of motor is given as follows:
3 M
Tm = p i sβ ϕr α − i sα ϕrβ (3)
2 Lr
where

Lr Ls M2 M 1 1 1−σ
Tr = , Ts = ,σ = 1 − , a1 = , a2 = +
Rr Rs Ls Lr σ Tr L s L r σ Ts Tr
3 Speed and Flux Control
The method of speed and flux control uses the sliding mode technique to design
discontinuous controls without explicitly involving current control. The advantage
of this approach is its robustness with respect to the mechanical parameters [15].
The method of selecting the sliding surfaces in the majority of controls is the errors
of controlled states and according to the relative degree required. The expression of
the speed and flux control surface is defined by:

s1 = c1 ϕr∗ − ϕr + (ϕ̇r∗ − ϕ̇r )

s2 = c2 ωr∗ − ωr + (ω̇r∗ − ω̇r ) (4)
where

ϕr = ϕr2α + ϕrβ
2
(5)
Equation (5) is the magnitude of the rotor flux and the sign (*) in (4) represents the
references inputs. c1 , c2 are positive constants determining the motion performance
in sliding mode [15]. The temporary derivative of (4) is:

ṡ1 = c1 ϕ̇r∗ − ϕ̇r + ϕ̈r∗ − ϕ̈r

ṡ2 = c2 ω̇r∗ − ω̇r + ω̈r∗ − ω̈r (6)
After calculates and reductions, ṡ1,2 written on the matrix form simplified as:

ṡ1 F1 Vsα
= − A1,2 (7)
ṡ2 F2 Vsβ
where

a cos θ a sin θ
s1,2 = [s1 s2 ] , F1,2 = [F1 F2 ] ,
T T
A1,2 =
−b sin θ b cos θ
where F1,2 are function of the state variables, the reference inputs and their
time
ϕ
derivatives. But they do not depend on the controls Vsα , Vsβ ; θ = tan−1 ϕrβra and
a, b are defined by:
M 3 ϕr M
a = Rr b= P
Ls Lr − M 2 2 J Ls Lr − M 2
With L s L r − M 2 > 0
We extract relation between ṡ1,2 and the switching controls Vabc :
ṡ1,2 = F1,2 − D1,2 Vabc (8)
With D1,2 = A1,2 Aabc abc

αβ , matrix Aαβ is defined as the Clarke transformation.
Equation (8) is the starting point of the sliding mode control design. Since matrix
D1,2 is time-varying. The controls defined in Vabc cannot perform their switching
depending on s1,2 . In this case a transformed vector of new switching function is
required to achieve a time-invariant control matrix. Design the transformation as
[15]:
+
s∗ = D1,2 s1,2 (9)
T
where s∗ = s∗1 s∗2 s∗3 is the vector representing the transformed switching
+
function, D1,2 is the pseudo inverse of matrix D1,2 given by:
+
T −1
3
D1,2 = D1,2
T
D1,2 D1,2 = (Aabc )T (A1,2 )T Q (10)
2 αβ
+
With Q is a positive definite time-varying matrix, Q and D1,2 being defined by
⎡1 ⎤
1
cos θa − b1 sin θa
a
0
Q= a2
1
+
D1,2 = ⎣ cos θb − b1 sin θb ⎦
1
a
(11)
0 b2 cos θc − b1 sin θc
1
a
where θa = θ , θb = θ − 23 π , θc = θ + 23 π , parameter b is proportional to ϕr . The

control Vabc law is given by:

Vabc = V0 sign s∗ (12)
T
With sign(s∗ ) = sign(s∗1 ) sign s∗2 sign s∗3 and V0 is the DC link voltage.
To ensure the operation of control well, we need to prove the stability, by the
Lyapunov function we provide the format:
1 T
V = s1,2 Qs1,2 (13)
2
T 1 T
V̇ = s1,2 Qṡ1,2 + s1,2 Q̇s1,2
2
1 T
= (s∗ )T (D1,2 )T Q F1,2 − (s∗ )T (D1,2 )T Q D1,2 Va,b,c + s1,2 Q̇s1,2 (14)
2
Can be simplified:
T abc
V̇ = (s∗ )T F ∗ − (s∗ )T Aabc
αβ Aαβ V0 sign s∗

− f ϕra , ϕrβ , ϕ̇ra , ϕ̇rβ s22 (15)
where
T
F ∗ = F1∗ F2∗ F3∗ = (D1,2 )T Q F1,2
1 T
s1,2 Q̇s1,2 = − f ϕra , ϕrβ , ϕ̇ra , ϕ̇rβ s22 (16)
2
T abc
(D1,2 )T Q D1,2 = Aabc
αβ Aαβ
2
3 J Ls Lr − M 2 ϕra ϕ̇ra + ϕrβ ϕ̇rβ
f ϕra , ϕrβ , ϕ̇ra , ϕ̇rβ =
2
2p M
ϕr2α + ϕrβ
2
According to (15) there exists a DC link voltage V0 high enough such that V̇ < 0.
This means that system (6) converges to the origin s1,2 = 0 in finite time under
control (12) and transformation (9) [16]. This means is system asymptotic stable.
4 Sliding Mode Observer
At present, we will proceed to design an observer to facilitate the maximum control

and neglected the sensors and to establish a good compromise between the stability
and the simplicity of control, it is necessary to take a reference axis linked to the
stator (α, β), we suggest observer submitted in [15]. The observer model is described
by the following state equation:
1 M
ϕ̂˙r α = − ϕ̂ra −ω̂r ϕ̂rβ + i sα
Tr Tr
1 M
ϕ̂˙rβ = ω̂r ϕ̂r α − ϕ̂rβ + i sβ
Tr Tr
˙î = a ϕ̂ + a T ω̂ ϕ̂ − a î +
1
Vsα
sα 1 ra 1 r r rβ 2 sα
σ Ls
î˙sβ = −a1 Tr ω̂r ϕ̂ra + a1 ϕ̂rβ − a2 î sβ
1
+ Vsβ (17)
σ Ls
ω̂r is discontinuous and given by:
ω̂r = ω0 sign(s3 ) (18)
where ω0 is a positive control gain, the sliding surface s3 is defined as:
s3 = i sβ ϕ̂ra − i sα ϕ̂rβ (19)
where i sα = î sα − i sα ; i sβ = î sβ − i sβ
Defines the errors:
ϕ̄r α = ϕ̂r α − ϕr α ; ϕ̄rβ = ϕ̂rβ − ϕrβ ; ω̄r = ω̂r − ωr
The temporary derivative of s3 is:

ṡ3 = a1 Tr ωr ϕ̂r α ϕr α + ϕ̂rβ ϕrβ − a1 Tr ω̂r ϕ̂r2α + ϕ̂rβ 2

− ω̂r ϕ̂rβ ī sβ + ϕ̂r α ī sα + a1 ϕ̂r α ϕ̄rβ − ϕ̂rβ ϕ̄r α
1
− s3 + M i sβ ī sα − i sα ī sβ − a2 s3 (20)
Tr
So ṡ3 can be simplified:
ṡ3 = −a1 Tr ωr e1 − a1 Tr ω̄r ϕr2 − ω̂r e3 + a1 e2

1
− s3 + M i sβ ī sα − i sα ī sβ − a2 s3 (21)
Tr
where e1 , e2 and e3 is defined as:

e1 = ϕ̄r α ϕ̂r α + ϕ̄rβ ϕ̂rβ = ϕ̂r2α + ϕ̂rβ
2
− ϕ̂r α ϕr α + ϕ̂rβ ϕrβ
e2 = ϕ̄rβ ϕ̂r α − ϕ̄r α ϕ̂rβ = −(ϕ̂r α ϕrβ − ϕ̂rβ ϕr α )

e3 = ϕ̂rβ ī sβ + ϕ̂r α ī sα (22)
Replaces ω̂r in (21) with (18) where the condition of ω0 is

−a T ω e + a T ω ϕ 2 + a e − a s
1 r r 1 1 r r r 1 2 2 3
ω0 > e3 + a1 Tr ϕr2 (23)
− Tr s3 + M i sβ ī sα − i sα ī sβ
1

where

e3 + a1 Tr ϕr2 > 0 (24)
If condition (23) check such that s3 ṡ3 < 0, this means that s3 enforced equal a
zero after a finite time interval and this means the estimated current converges to
reference current.
The condition (24) for sliding mode to occur on the surface is s3 = 0 not very
restrictive because the stator currents i sα and i sβ are measurable. We can always
choose the initial conditions î sα (0) and î sβ (0) close enough to the true stator currents
i sα (0) and i sβ (0) such that the initial errors ī sα = 0 and ī sβ (0) and hence e3 are small
enough to guarantee that this condition holds [16].
Sliding mode trajectories lie in the manifold s3 = 0 and the equivalent control
ω̂r _eq is a solution to the equation ṡ3 = 0 it is necessary to replace discontinuous
control with such continuous control that the state velocity vector lies in the tangential
manifold [17].
From (19) see that if ī sα = 0 and ī sβ = 0 is a solution of s3 = 0 replaces in (21),
the equivalent control ω̂r _eq is given by:

ω̂r _eq = ωϕ 2r ϕ̂r α ϕr α + ϕ̂rβ ϕrβ + 1
e
Tr ϕr2 2
r (25)
ϕr2 = ϕ̂r2α + ϕ̂rβ 2
= 0∀t
The Eq. (25) implies that if the estimated rotor flux converges to the true flux then
the equivalent value of ω̂r _eq is equal to the true rotor speed, lim ω̂r _eq = ωr .
ϕ̂→ϕ
ω̂r _eq can be obtained through a low pass filter with the discontinuous value [15]:
1
ω̂r _eq = ω̂r (26)
1 + tc P
where tc = f1r is the time constant of the filter, and P is the Laplace variable. Where
the time constant of the low-pass filter tc should be small compared with the slow
component of ω0 sign(s3 ) but large enough to filter out the high rate component [16].
The control structure of in IM is (Fig. 1):
Fig. 1 Induction motor controls with sliding mode
5 Simulation Results
To assess the performance of control studied, the simulation is applied for the IM
drives. The software used is Simulink Matlab 2016b. The following figures shows
the behavior of the induction motor at different values of speeds and load torque
variation.
Figure 2 show the speed response, at startup the speed is changed from 0 rad/s
to their reference value 148.7 rad/s at t = 0.3 s change to −148.7 rad/s at second
t = 0.6 Back to their reference value. Note that the real speed and estimated have
the same behavior, the tracking performance is good that means the observer can
follow the speed command strictly. When the load is suddenly increased at t = 0.8 s
valued at Cr = 10 N m there is no change on the speed and the estimated value and
which is illustrated by speed zoom.
Fig. 2 Speed response

Figure 3 show the time response of the actual stator current components, we note
the classic call of the current when powering up the induction motor. The current
reaches a maximum value of about 8–10 times the nominal current of the machine as
well as when the speed change then stabilizes at the value of off-load current about
4 A once the load is applied an increase is noted the current reaches about 5.35 A
and which is illustrated by current zoom, due to the increase of the motor torque.
Figure 4 is the results of the rotor flux, we note the components are perfectly in
quadrature and their amplitudes are identical. At seconds to change values of speeds
Fig. 3 Stator current response
Fig. 4 Rotor flux response

there is only a slight change on the flux and it goes back to steady state quickly. See
that the convergence between the rotor flux and the estimated of this value it occurs
at the startup and good tracking performance the motor torque with reference load
torque and good robustness to external load torque variation, which is illustrated by
Fig. 5. Which is proves ability and strength to this control and observer.
Figures 6, 7, and 8 give the result when stator resistance is increase 50% in motor
model. The sliding mode control and observer was tested with difference speed and
the load torque:
Fig. 5 Zoom flux and torque response
Fig. 6 Speed response in case of increased 50% stator resistances

Fig. 7 Rotor flux response in case of increased 50% stator resistances
Fig. 8 Flux zoom and torque response in case of increased 50% stator resistances
In this test speed response for sliding control and observer there is a slight differ-
ence through a slow response during speed changes and then a good tracking follow-
up of its reference we note their robustness with respect to the variation of the load
torque this is what appears in speed zoom. The estimate flux is bigger than the actual
value at seconds to change values of speeds and the real value are lowers than their
actuals values and their return to steady state, note that the load torque does not
affect them this is what appears in flux zoom. As a result, the control and observer
small error in transient speeds only when stator resistance change, and good tracking
performance the motor torque with reference load torque.
6 Nomenclature
Vsα , Vsβ Stator voltages

V0 DC link voltage
i sα , i sβ Stator currents
ϕra , ϕrβ Rotor fluxes
ωr Electrical rotor speed
Ls , Lr , M Stator, rotor and mutual inductance
Rs , Rr Stator, rotor resistance
Tr Rotor time constant
p, F Number of pole pairs, coefficient of friction
J, σ Moment of inertia, coefficient of dispersion
Tm , TL Motor torque and load torque.
ϕ̂ra , ϕ̂rβ Estimated for rotor flux
î sα , î sβ Estimated for stator currents
r e f , mes , est Reference, measured and estimate speed.
7 Conclusion
In this paper, First order sliding mode strategy was applied to ensure a decoupled
control for induction motor.
To avoid the need of sensors, the speed and the rotor flux were estimated using
non-linear observers based on sliding mode.
Based on the Lyapunov theory, the convergence of control and observer was
proved.
The simulation results show the robustness of the proposed control scheme and
confirm that the FOSM is more efficient and present a low sensitivity to disturbances
and motor parameters variation.
Appendix
Table 1presents the parameters utilized in the control:

Table 1 Parameters of
Parameters of induction motor Parameters of sliding mode
induction motor and control
[18] control
Pn = 1500 W n = 1420 rpm c1 = 200 c2 = 100
Rs = 4.85 Rr = 3.805 ω0 = 1200 fr = 45 Hz
L s = 0.274 H L r = 0.274 H
M = 0.258 H p = 2
V0 = 537 V F =
0.001136 N ms
References
1. Leonhard W (2001) Control of electrical drives, 3rd edn. Springer, Berli

2. Ammar A, Bourek A, Benakcha A (2017) Nonlinear SVM-DTC for induction motor drive using
input-output feedback linearization and high order sliding mode control. ISA Trans 67:428–442
3. Alsmadi YM, Utkin V, Haj-Ahmed M, Xu L, Abdelaziz AY (2018) Sliding-mode control of
power converters: AC/DC converters & DC/AC inverters. Int J Control 91(11):2573–2587
4. Derdiyok A, Guven M, Utkin V (2001) A sliding mode speed and rotor time constant
observer for induction machines. In: IECON’01, 27th annual conference of the IEEE industrial
electronics society (Cat. No. 37243), vol 2, pp 1400–1405
5. Oliveira CMR, Aguiar ML, Monteiro JBA, Pereira WC, Paula GT, Santana MP (2015) Vector
control of induction motor using a sliding mode controller with chattering reduction. In:
2015 IEEE 13th Brazilian power electronics conference and 1st southern power electronics
conference (COBEP/SPEC), pp 1–6
6. Dris A, Bendjebbar M (2016) Different technics of observation of the flux and the speed of an
induction motor (KUBOTA and MRAS observers). In: 2016 8th international conference on
modelling, identification and control (ICMIC), pp 61–66
7. Bullo D, Ferrara A, Rubagotti M (2011) Sliding mode observers for sensorless control of
current-fed induction motors. In: Proceedings of the 2011 American control conference, pp
763–768
8. Bernat J (2018) Multi observer structure for rapid state estimation in linear time varying
systems. Int J Control Autom Syst 16(4):1746–1755
9. Dong CST, Tran CD, Ho SD, Brandstetter P, Kuchar M (2008) Robust sliding mode observer
application in vector control of induction motor. In: 2018 ELEKTRO, pp 1–5
10. Fridman L, Li S, Man Z, Wang X, Yu XH (2017) Advances in variable structure systems and
sliding mode control—theory and applications. Springer
11. Nadh G, Syamkumar U, Jayanand B (2016) Sliding mode observer for vector control of induc-
tion motor. In: 2016 international conference on next generation intelligent systems (ICNGIS),
pp 1–6
12. Ghanes M, Zheng G (2009) On sensorless induction motor drives: sliding-mode observer and
output feedback controller. IEEE Trans Indus Electron 56(9):3404–3413
13. Holtz J (2002) Sensorless control of induction motor drives. Proc IEEE 90(8):1359–1394
14. Rao S, Buss M, Utkin V (2009) Simultaneous state and parameter estimation in induction
motors using first- and second-order sliding modes. IEEE Trans Indus Electron 56(9):3369–
3376
15. Utkin V, Guldner J, Shijun M (1999) Sliding mode control in electro-mechanical systems. CRC
Press, Philadelphia
16. Utkin V, Guldner J, Shi J (2009) Sliding mode control in electro-mechanical systems. CRC
Press, Boca Raton
17. Utkin VI (1993) Sliding mode control design principles and applications to electric drives.
IEEE Trans Indus Electron 40(1):23–36
18. El hani S, Essaaidi E (2019) Recent advances in electrical and information technologies for
sustainable development. In: Proceedings of the 3rd international conference on electrical and
information technologies — ICEIT 2017 Morocco. Springer International Publishing
Cuckoo Search Algorithm for Solving
the Problem of Unit-Commitment
with Vehicle-to-Grid
Amel Terki and Hamid Boubertakh
Abstract This paper addresses the optimization of unit commitment (UC) with
Vehicle to Grid (V2G) problem using a Binary-real coded Cuckoo Search (BCS)
algorithm. The problem lies in the efficient planning of gridable vehicles in restricted
constrained parking lots. Cuckoo Search (CS) algorithm is an intelligent meta-
heuristic algorithm that implement local and global search mechanisms via Lévy
flights approach. It has been effectively used to solve several optimization problems.
In order to verify the efficiency of the described algorithms; numerical simulation are
implemented for 10-unit system with V2G in constraints parking lots to minimize
the operating costs and has compared to the other existing algorithms.
Keywords Unit commitment problem · Power system · Renewable energy ·

Vehicle to grid (V2G) · Cuckoo Search algorithm
1 Introduction
The Unit Commitment (UC) problem is one of difficult and complex optimization
problem in power system. The main role of the UC is to determine when to start-up
or shut-down of each generating unit for a scheduling horizon in order to minimize
the operating costs and satisfy the prevailing constraints, including minimum up and
down time limits, ramp rate limits, generator power output limits and proper spinning
reserves [1, 2].
Vehicle to grid (V2G) technology is collecting a large research attention in the
generation of energy storage units; it’s one of the many energy storage technologies
A. Terki (B) · H. Boubertakh

Laboratoire d’Automatique de Jijel-LAJ, University MSB of Jijel, BP 98 Ouled Aissa 1800, Jijel,
Algeria
H. Boubertakh

https://doi.org/10.1007/978-981-15-6403-1_6
78 A. Terki and H. Boubertakh
[3]. The V2G have the ability of minimizing the cost of the power generation in small
expensive units, and they could provide a distributed storage to power [4].
The Unit Commitment with Vehicle to Grid (UC-V2G) is the perspectives that
involve the efficient planning of gridable vehicles in restricted constrained parking
lots to meet out the demand. In this context, the UC-V2G problem is among of the
most complex optimization problem in power systems than classic UC problem [5],
because they should meet the forecasted load demand and spinning reserve require-
ments at each time interval such that the total cost is minimum [6]. Various optimiza-
tion methods have been employed to solve the UC problem, such as Priority List
(PL) [7] Dynamic Programming (DP) [8], and Lagrangian Relaxation (LR) [9]. These
methods have a simple representation and fast convergence but the solution quality is
not guaranteed. Recently, intelligent meta-heuristic optimization approaches attract
much attention to solve the UC problem. This attraction is due to the parallel search
mechanism of such approaches which offers a best way to search the global optimum.
Among them, one find, Genetic Algorithm (GA) [10], Differential Evolution (DE)
[11], Evolutionary Programming (EP) [12], Gravitational Search Algorithm (GSA)
[13], Ant Colony Optimization (ACO) [14], Particle Swarm Optimization (PSO)
[15, 16]. Moreover, since each method has its own advantages over the others, so, in
order to combine the advantages of these approaches, some hybridized approaches
have been proposed; Hybrid GA–PSO [17], Genetic Algorithm Based on Priority
List (GABPL) [18] and Hybrid LRGA [19].
The Cuckoo Search (CS) approach, has been inspired by the natural behaviour
of cuckoos, is one of the most powerful emerging approaches, initially proposed by
Yang and Deb [20]. The CS has been applied successfully to different optimization
problems [21–28], particularly to the UC problem [29–32].
Binary Cuckoo Search (BCS) algorithm it’s a binary version of the CS algo-
rithm. BCS has never been used to solve the UC-V2G problem until now. In [29,
31] the On/Off status of a unit is based on a signed integer representation. Chan-
drasekaran and Sishaj [30] developed a fuzzy integrated binary real coded cuckoo
search Lagrangian (BCSL) algorithm, which is used for multi-objective UC problem.
In [32] weight pattern based cuckoo search algorithm is used for 3 and 6 units system.
In this paper, we propose a new CS algorithm for solving the UC-V2G problem
by using a binary code to deal with the units scheduling and a real code to deal with
the produced power levels of the committed units. The CS algorithm is the only
approach which implements separately a local and global search mechanism.
The rest of paper is organized as follows. Section 2 provides the description of
the UC-V2G problem, the CS and the BCS algorithm will be successively detailed
in Sect. 3, the technique of handling UC-V2G constraints is described in Sect. 4.
Results are presented in Sect. 5 and the conclusions in Sect. 6.
Cuckoo Search Algorithm for Solving the Problem … 79
2 Unit Commitment with V2G Problem
To improve the quality level of power system by integrating new generation and
storage opportunities, the electric vehicles will be seen as a charge in the unit commit-
ment. The objective in the UC problem with the V2G is to find the optimal schedule
of committed/uncommitted units and generated power per each unit over a study
period of time while satisfying a variety of constraints. The objective function F to
minimize comprises the fuel costs Cit ( pit ) of the generating units, the start-up costs
STit of the committed units and the V2G cost during the scheduling hours.

T
N

T

F= Cit ( pi ) + STit 1 − u i,t−1 u it + Pv N V 2G,t (1)
t=1 i=1 t=1
where u it is the On/off status of unit i at hour t, pit is the power output of unit i at
hour t in (MW), Pv is the Capacity of each vehicle in (MW), N V 2G,t is the number
of vehicles connected to the grid at time t, N is the number of units and T the time
horizon.
Cit ( pit ) = ai + bi pit + ci pit2 (2)
where ai , bi and ci the fuel cost coefficients of unit.

of f
di , i f idown ≤ τit ≤ idown + f i
STit = of f (3)
ei , i f τit > idown + f
where di and ei are the hot and cold startup cost of unit i, in ($), f i is the cold start
up
hours of unit i, in h (hour), idown and i are the Minimum down and up time of
of f
unit i, in h and τit is the continuously off time of unit i up to time t.
A. Constraints
The system and unit constraints plus the V2G constraints must be satisfied during
the optimization process are as follows.
(1) Power balance constraint: The total generated power from the gridable vehicles
and all committed units at each hour must be equal to the load demand plus the
system losses at hour t.

N
pit u it + Pv N V 2G,t = Dt + Losses (4)
i=1
where Dt is the load demand at hour t, in (MW)
(2) Spinning reserve constraint: To assure reliable operation, the total generating
unit and the gridable vehicles capacity must be greater than or equal to the load
demand and the specified spinning reserve.

N
pimax u it + Pvmax N V 2G,t ≥ Dt + Rt (5)
i=1
where Rt is the spinning reserve at hour t, in (MW)
(3) Minimum up/down time: A unit must remain on or off for a certain number of
hours before it can be, respectively, shut off or put on. The constraints are given
by
on up
τi,t−1 − i u i,t−1 − u it ≥ 0

of f
τi,t−1 − idown u it − u i,t−1 ≥ 0 (6)
where τiton is the continuously on time of unit i up to time t
(4) Power generating limit: The active powers that can be generated by the
generating units have minimum and maximum values.
pimin ≤ pit ≤ pimax (7)
where pimin and pimax Minimum and Maximum capacity of unit i, in (MW).
(5) Gridable vehicle balance in V2G: The total numbers of vehicles are registered
to charge from renewable sources and discharge to the grid during 24 h is
predefined for V2G technology.

T
N V 2G,t = N Vmax
2G (8)
t=1
where N Vmax
2G is the maximum number of vehicle in the system.
(6) Vehicle parking limits: The number of vehicles will discharge at a given time is
limited. The constraints are given by
N V 2G,t ≤ N Vmax
2G,t (9)
(7) Charging-discharging frequency: The daily charging/discharging frequency is

assumed as 1 per day. Each vehicle should have a suitable departure State of
charge Soc level. Charging and inverter Efficiencies Effi should be considered.
3 Overview of the Cuckoo Search Metha-Heuristic
A. Standard Cuckoo Search
The CS is a swarm-intelligence optimisation method which has been initially devel-

oped by Yang and Deb [20]. It is inspired by the unavoidable brood parasite of cuckoo
species. As it is well-known, the cuckoo bird never builds its personal nest and lays
their eggs in the nest of another host’s. The implementation of CS algorithm follows
the rules given bellow [21].
– Each cuckoo lays one egg at a time, and puts it in a random nest;
– The best nests with high quality of eggs are carried to the next generations;
– A host bird can discover a foreign egg with probability pa ∈ [0, 1]. In this case,
they discard it away or abandon the nest to build a new nest.
The main steps of the CS algorithm are recapitulated in the Algorithm 1. The first
nest population is placed randomly and the generation of the new solution for the ith
cuckoo is updated using random walk via Lévy flights [28] such as
xi(t+1) = xi(t) + α ⊕ L évy(λ) (10)
where, xi for i = 1, 2, . . . , N is the position of a nest i, N the population size, α > 0

the scaling factor of the step size t and ⊕ the entry-wise product. The Lévy flights
provide a random walk which is drawn from a Lévy distribution given by
L évy ∼ u = t−λ , (1 < λ ≤ 3) (11)

The new solution xi(t+1) can be generated by local and global search. The CS is
used to represent N × T real variables of the pit quantity of power to
be generated by

committed units, where pit assigned a random real value in the range pimin , pimax .
B. Binary Cuckoo Search
Binary Cuckoo Search BCS algorithm is developed by [33]. It is a binary version of

the CS algorithm, in which the search space is defined as a binary n-bit string. The
BCS algorithm represents each nest as a binary vector. Equations (12) and (13) are
used in order to build binary vectors. In this study, we adopt the BCS algorithm to
determine the On/Off status of the generating units u it over each interval of the time
horizon with 0/1 decision variables with equal probability [33].
1
S xi(t) = (t)
(12)
1 + e−xi

1 i f s xi(t)
xi(t+1) = (13)
0 other wise
where, σ ~ U (0, 1) and xi(t+1) stands for the new eggs.
4 Procedure for Handling the UC-V2G Problem

Constraints
In this paper, a new binary- real coded CS algorithm is proposed to solve the UC-
V2G problem. In fact, the problem is divided in two sub-problems. A procedure to
handle constraints and an optimization method to solve the UC-V2G problem.
A. Handling Constraints
Several methods have been proposed in the literature for handling infeasible solu-
tions; rejecting methods, repairing methods and penalty methods [10]. Unfortunately
these algorithms cannot satisfy all the problem constraints. In this paper, the repairing
and penalty methods are applied to solve UC-V2G constraint problems. In fact,
when the constraints are not satisfied after the repairing methods, the penalty method
is added to discourage the invalid solutions.
(10) (11)
(1) Gritable vehicle balance in V2G and Vehicle parking limits: A randomly
distribution is applied to satisfy (10) and (11) during 24 h.
(2) Spinning reserve constraint repairing: To satisfy spinning reserve constraint, a
heuristic-based repair method is applied, through which de-committed units are
committed in an ascending order of their average full load costs αi , i = 1, 2, …,
N; until the spinning reserve constraint is satisfied [34].
Cit ( pit ) ai
αi = max = max + bi + ci pimax (14)
pi pi
(3) Minimum up/down time constraints repairing: The minimum uptime/downtime

constraints impose a minimum number of time periods that must elapse before
the unit can change its status. The continuous On/Off times of units at hour t
can be decided as follows.
on
τ + 1 i f u i,t = 1
τi,t
on
= i,t−1
0 other wise
of f
of f τ + 1 i f u i,t = 0
τi,t = i,t−1 (15)
0 other wise
To satisfy these time periods, a heuristic technique procedure is summarized in

[35]. Its purpose is to handle the minimum up-time and down-time constraints.
(4) De-commitment of excess units: The high operation cost leads to undesirable
excessive spinning reserve. In this case, a heuristic algorithm based on priority
list is applied to de-commit some units one by one, in a descending order of their
average full load costs αi , without violating the minimum uptime/downtime and
spinning reserve constraints, which were checked after de-committing a unit
[35, 36].
(5) Power balance constraint repairing: Each unit must generate power within
the interval given by Eq. (7), but should lead to a higher production cost and
satisfy the load demand constraint of Eq. (4). The quantity of divergence in the
generated power with respect to the power demand at time t is given by

N
et = u it pit − Dt (16)
i=1
To satisfy the power balance constraint the algorithm is demonstrated in [34].
B. The Proposed Approach for Solving UC-V2G Problem
In the proposed method the decision variables include binary variables for generating
unit, real variables for power output and integer variables for gridable vehicle. The
proposed algorithm is shown in Algorithm 2.
To demonstrate the efficiency and feasibility of the proposed algorithm, we consider

a basic system of 10 units with 50,000 gritable vehicles which are charged from
renewable source for 24 h scheduling horizon. The basic unit data and the load
demand D t of system are taken from [15]. In all cases, the spinning reserve require-
ment is set to 10% of the total load demand. The parameters’ settings of proposed
algorithm are a population size of 100, the Capacity of each vehicle Pv is 15 kWh,
the maximum and the minimum battery capacity are 25 kWh and 10 kWh respec-
tively, the maximum parking lot capacity at hour t N Vmax 2G is 10% of total vehicles, Effi
is 85% and SOC is 50%. Load demand is dispersed among generating units and grid-
able vehicles. At each hour t, the power from vehicles is E f f i × Soc × N V 2G,t × Pv
and capacity of each vehicle Pv is constant. All simulations have been run on
Matlab2016a environment. Tables 1 and 2, show the obtained results for all the test
system without and with gritable vehicles respectively. The total cost (fuel cost plus
start-up cost) is $554,01660 when gritable vehicles are considered in the UC system
and is $561,28456 when gritable vehicles are not considered in the UC system. There-
fore, it is observed that the total cost is reduced by $7.2680 for only gritable vehicles,
this reduction due to, addition of vehicle power to the grid On the other hand the
reserve of unit is increased when using V2G technology. It is obvious from Tables 2
and 3 that there is a saving of $7,2680 of the total cost comparison of the generating
units with and without the V2G parking lot during 24, which shows the effectiveness
of the proposed method for UC-V2G problem. It can be concluded that theV2G tech-
nology maintaining reliability of the power system. The best results obtained with
other methods are compared to confirm the performance of the proposed algorithm.
Table 3 shows a numerical comparison of proposed algorithm and other existing
methods for the same number of unit system. It is clear that the total cost obtained
by the proposed algorithm is less than the other approaches.
6 Conclusion
In this paper, a new binary-real coded Cuckoo Search BCS algorithm is proposed
for solving UC–V2G problem. The CS algorithm has been proposed to determine
the real variables representing the power generated by committed units and a binary
version of the CS algorithm is used to determine the binary variables representing the
status of the generating units during each interval of the scheduling period. Results
of the proposed algorithm have been compared with other methods reported in the
literature. The production costs of the proposed algorithm are less than the existing
algorithms. In conclusion, we have shown through simulations that the proposed
algorithm outperforms other methods in the minimization of the total generation
cost of the UC and the effective V2G scheduling can reduces this generation cost,
increase reserve and improve the power system reliability.
Table 1 Dispatch schedule and reserve power of generating units without V2G
Time (h) Power generated by units (MW) Max. capacity (MW) Demand (MW) Reserve (MW)
1 2 3 4 5 6 7 8 9 10
1 455 245 0 0 0 0 0 0 0 0 910 700.0 210
2 455 295 0 0 0 0 0 0 0 0 910 750.0 160
3 455 370 0 0 25 0 0 0 0 0 1072 850.0 222
4 455 455 0 0 40 0 0 0 0 0 1072 950.0 122
5 455 448.16 0 71.83 25 0 0 0 0 0 1202 1000.0 202
6 455 360.57 129.42 130 25 0 0 0 0 0 1332 1100.0 232
7 455 455 85 130 25 0 0 0 0 0 1332 1150.0 182
8 455 455 130 130 30 0 0 0 0 0 1332 1200.0 132
9 455 455 130 130 85 20 25 0 0 0 1497 1300.0 197
10 455 455 130 130 162 33 25 10 0 0 1552 1400.0 152
Cuckoo Search Algorithm for Solving the Problem …
11 455 455 130 130 162 73 25 10 10 0 1607 1450.0 157

12 455 455 130 130 162 80 58 10 10 10 1662 1500.0 162
13 455 455 130 130 162 43 25 0 0 0 1497 1400.0 97
14 455 455 130 130 85 20 25 0 0 0 1497 1300.0 197
15 455 455 130 130 30 0 0 0 0 0 1332 1200.0 132
16 455 356.16 0 115 25 0 0 0 0 0 1202 1050.0 152
17 455 404.37 0 115.62 25 0 0 0 0 0 1202 1000.0 202
18 455 455 0 130 60 0 0 0 0 0 1202 1100.0 102
19 455 455 0 130 160 0 0 0 0 0 1202 1200.0 2
20 455 455 0 130 162 80 85 33 0 0 1422 1400.0 22
(continued)
87
Table 1 (continued)
88
Time (h) Power generated by units (MW) Max. capacity (MW) Demand (MW) Reserve (MW)
1 2 3 4 5 6 7 8 9 10
21 455 455 0 130 162 73 25 0 0 0 1367 1300.0 67
22 455 455 0 0 145 20 25 0 0 0 1237 1100.0 137
23 455 455 0 0 0 0 0 0 0 0 910 900.0 10
24 455 345 0 0 0 0 0 0 0 0 910 800.0 110
Total Cost ($) = 5.6128e + 05
A. Terki and H. Boubertakh
Table 2 Dispatch schedule and reserve power of generating units with V2G
Time (h) Power generated by units (MW) No.of Capacity of Max. Demand Reserve
1 2 3 4 5 6 7 8 9 10 vehicle vehicles capacity (MW) (MW)
(MW) (MW)
1 455 237.74 0 0 0 0 0 0 0 0 1138 7.25 917.3 700.0 217.25
2 455 282.60 0 0 0 0 0 0 0 0 1945 12.39 922.4 750.0 172.39
3 454 374.70 0 0 0 0 0 0 0 0 3183 20.29 930.3 850.0 80.29
4 455 454.70 0 32.26 0 0 0 0 0 0 1259 8.02 1048 950.0 98.02
5 455 454.98 0 58.81 25 0 0 0 0 0 973 6.20 1208.2 1000.0 208.20
6 455 455 24.27 130 25 0 0 0 0 0 1683 10.72 1.42.7 1100.0 242.72
7 455 455 77.35 130 25 0 0 0 0 0 1200 7.65 1339.7 1150.0 189.65
8 455 455 114.77 130 25 0 0 0 0 0 3173 20.22 1352.2 1200.0 152.22
9 455 455 130 130 97.49 20 0 0 0 0 1961 12.50 1424.5 1300.0 124.50
10 455 455 130 130 162 25.70 25 10 0 0 1145 7.29 1559.3 1400.0 159.29
Cuckoo Search Algorithm for Solving the Problem …
11 455 455 130 130 162 53.49 25 10 10 0 3059 19.50 1626.5 1450.0 176.50
12 455 455 130 130 162 80 33.08 10 10 10 3908 24.91 1686.9 1500.0 186.91
13 455 455 130 130 157.26 20 25 0 0 0 4351 27.73 1524.7 1400.0 124.73
14 455 455 130 130 97.49 20 0 0 0 0 1967 12.53 1424.5 1300.0 124.53
15 455 455 124.29 130 25 0 0 0 0 0 1680 10.71 1342.7 1200.0 142.71
16 455 455 20 113.91 0 0 0 0 0 0 955 6.08 1176.1 1050.0 126.08
17 455 455 20 58.18 0 0 0 0 0 0 1853 11.81 1181.8 1000.0 181.81
18 455 455 51.72 130 0 0 0 0 0 0 1298 8.27 1178.3 1100.0 78.27
19 455 455 129.90 130 0 20.06 0 0 0 0 1574 10.03 1260 1200.0 60.03
20 455 455 130 130 0 80 85 43.12 0 0 3431 21.87 1411.9 1400.0 11.87
(continued)
89
Table 2 (continued)
90
Time (h) Power generated by units (MW) No.of Capacity of Max. Demand Reserve
1 2 3 4 5 6 7 8 9 10 vehicle vehicles capacity (MW) (MW)
(MW) (MW)
21 455 455 130 130 0 80 28.17 0 0 0 3424 21.82 1356.8 1300.0 56.82
22 455 455 129.85 0 0 26.03 25 0 0 0 1430 9.11 1214.1 1100.0 114.11
23 455 435.51 0 0 0 0 0 0 0 0 1488 9.48 919.5 900.0 19.48
24 455 332.74 0 0 0 0 0 0 0 0 1922 12.25 922.3 800.0 122.25
Total Cost ($) = 5.5402e + 05
A. Terki and H. Boubertakh
Table 3 Comparison
Method Total cost ($)
between proposed algorithm
and other algorithm for 10 Without V2G With V2G
unit systems HSA, [37] 573,87934 554,13459
CRO, [38] – 564,72787
PM, [39] 564,714 557,5549
PSO, [6] 563,74183 559,08136
Proposed 561,28456 554,01660
References
1. Abookazemi K, Mustaf MW, Ahmad H (2009) Structured genetic algorithm technique for unit
commitment problem. Int J Recent Trends Eng Technol 1(3), November 2009
2. Park HG, Lyu JK, Kang YC et al (2014) Unit commitment considering interruptible load for
power system operation with wind power. Energies 7:4281–4299
3. Zhile Y, Kang L, Xiandong X (2016) A hybrid meta-heuristic method for unit commitment
considering flexible charging and discharging of plug-in electric vehicles. In: 2016 IEEE
Congress on Evolutionary Computation (CEC), Vancouver, BC, 2016, pp 2014–2020. https://
doi.org/10.1109/CEC.2016.7744035
4. Mohammad EK, Lei W, Shahidehpour M (2012) Hourly coordination of electric vehicle
operation and volatile wind power generation in SCUC. IEEE Trans Smart Grid 3(3):1271–1279
5. Ghanbarzadeh T, Goleijani S, Parsa M (2011) Reliability constrained unit commitment with
electric vehicle to grid using hybrid particle swarm optimization and ant colony optimization.
In: 2011 IEEE Power and Energy Society General Meeting, Detroit, MI, USA, 2011, pp 1–7.
https://doi.org/10.1109/PES.2011.6039696
6. Saber AY, Venayagamoorthy GK (2009) Unit commitment with vehicle-to-grid using particle
swarm optimization. In: IEEE Bucharest power tech conference, Romania
7. Manisha G (2016) Solution to unit commitment with priority list approach. Int J Adv Res Electr
Electron Instr Eng (IJAREEIE) 3(6):5114–5121
8. Snyder WL, Powell HD, Rayburn JC (1987) Dynamic programming approach to unit
commitment. IEEE Trans. Power Syst. 2(2):339–347
9. Chuang CS, Chang GW (2013) Lagrangian relaxation-based unit commitment considering fast
response reserve constraints. Energy Power Eng 5:970–974
10. Farag MA, El-Shorbagy MA, El-Desoky IM et al (2015) Binary-real coded genetic algorithm
based k-means clustering for unit commitment problem. Appl Math 6:1873–1890
11. Jeong YW, Lee WN, Kim HH et al (2009) Thermal unit commitment using binary differential
evolution. J Electr Eng Technol 4(3):323–329
12. Juste KA, Kita H, Tanaka E et al (1999) An evolutionary programming solution to the unit
commitment problem. IEEE Trans Power Syst 14:1452–1459
13. Yuan X, Ji B, Zhang S et al (2014) Anew approach for unit commitment problem via binary
gravitational search algorithm. Appl Soft Comput 22:249–260
14. Zand A, Bigdeli M, Azizian D (2016) A modified ant colony algorithm for solving the unit
commitment problem. Adv Energy: Int J (AEIJ) 3, July 2016
15. Guo YR, Wu YH, Zhang JR et al (2015) Improved PSO approach for the solution of unit commit-
ment problem. In: International conference on artificial intelligence and industrial engineering
(AIIE), pp 510–513
16. Kumar S, Singh HD (2017) Optitimization of unit commitment problem using classical soft
computing technique (PSO). Int Res J Eng Technol (IRJET) 4(11), November 2017
17. Archana N (2015) Unit commitment and economic load dispatch using hybrid genetic—particle
swarm optimization algorithm. Int J Appl Res 1(6):109–113
18. Sarjiya B, Arief M, Andi S (2015) Unit commitment solution using genetic algorithm based
on priority list approach. J Theor App Inform Technol (JATIT) 72(3), February 2015
19. Chuan-Ping C, Chih-Wen L, Chun-Chan L (2000) Unit commitment by lagrangian relaxation
and genetic algorithms. IEEE Trans Power Syst 15(2):107–114
20. Yang XS, Deb S (2009) Cuckoo search via levy flights. In: World congress on nature and
biologically inspired computer (NaBIC), pp 210–214
21. Jiang M, Song H et al (2016) A cuckoo search-support vector machine model for predicting
dynamic measurement errors of sensors. IEEE Access 4:5030–5037
22. Svetlana A, Jussi S, Olli P et al (2015) Cuckoo search for wind farm optimization with auxiliary
infrastructure. Wiley, pp 1–21
23. Feng Y, Zhou J, Mo L et al (2018) A gradient-based cuckoo search algorithm for a reservoir-
generation scheduling problem. MDPI 11(36):1–21
24. Normansyah A, Erick R et al (2017) Cuckoo search algorithm for environmental/economic
dispatch problem. IOSR-JEEE 12:59– 63
25. Intissar K, Taoufik L, Faouzi M et al (2018) Cuckoo search approach for parameter identification
of an activated sludge process. Comput Intell Neurosci 1–9, January 2018
26. Andrés I, Akemi G, Patricia S (2018) Cuckoo search algorithm with Lévy flights for global-
support parametric surface approximation in reverse engineering. MDPI 10(8):1–25
27. Liu L, Liu X, Wang N et al (2018) Cuckoo search algorithm with variational parameters and
logistic map. MDPI 11(30):1–11
28. Zhao J, Liu S, Wang Y (2017) Cuckoo search algorithm with interactive learning for economic
dispatch. In: 2017 36th Chinese control conference (CCC), pp 2904–2909, July 2017
29. Chitra J, Ravichandran CS (2014) Cuckoo and Levy flights algorithm applied to unit-
commitment problem. Int J Adv Res Electr Electron Instr Eng (IJAREEIE) 3(12):13670–13677
30. Chandrasekaran K, Sishaj PS (2012) Multi-objective unit commitment problem using Cuckoo
search Lagrangian method. Int J Eng Sci Technol (IJEST) 4(2):89–105
31. Chitra J, Ravichandran CS (2014) Performance comparison of integer coded cuckoo and Levy
flights algorithm applied to unit-commitment problem. J Theor Appl Inform Technol(JATIT)
66(3):876–883
32. Sharma S, Mehta S, Verma T (2015) Weight pattern based cuckoo search for unit commitment
problem. Int J Res Adv Technol 3(5):102–110
33. Rodrigues D, Pereira LA, Almeida T et al (2013) BCS: a binary cuckoo search algo-
rithm for feature selection. In: 2nd International Conference on Knowledge Engineering and
Applications (ICKEA), London, pp 6–12. https://doi.org/10.1109/ICKEA.2017.8169893
34. Datta D (2013) Unit commitment problem with ramp rate constraint using a binary-real-coded
genetic algorithm. Appl Soft Comput, Elsevier 13:3873–3883
35. Yuan X, Su A, Nie H (2009) Application of enhanced discrete differentia evolution approach
to unit commitment problem. Energy Convers Manag 50:2449–2456
36. Yuan X, Su A, Nie H (2011) Unit commitment problem using enhanced particle swarm
optimization algorithm. Soft Comput 15:139–148
37. Pavithra Priya R, Sivaraj N, Muruganandam M (2015) A solution to unit commitment problem
with V2G using Harmony Search algorithm. Int J Adv Res Electr Electron Instr Eng 4(3):1208–
1214
38. Yu JJQ, Li VOK, Lam AYS (2013) Optimal V2G scheduling of electric vehicles and
unit commitment using chemical reaction optimization. In: IEEE congress on evolutionary
computation (CEC), Cancun, Mexico, pp 392–399, June 2013
39. Hossein B, Zareian M, Jahromi M, Rashidinejad M (2012) A combinatorial artificial intel-
ligence real-time solution to the unit commitment problem incorporating V2G. Electr Eng,
Springer, 341–355, December 2012
A Comparative Study of MPPT
Techniques for Standalone Hybrid
PV-Wind with Power Management
Tarek Boutabba, Hamza Sahraoui, Mouhamed Lamine Bechka, Said Drid,

and Larbi Chrifi-Alaoui
Abstract In this work a comparative study of maximum power point tracking tech-
niques for a standalone hybrid generation power system with power management is
presented, this system comprising a double fed induction generator branch (DFIG)
based wind turbine and photovoltaic generator branch compared to the conventional
perturbation and observation (P&O) and the incremental conductance algorithms
respectively. The fuzzy logic controller is designed to vary the duty-cycle of the DC–
DC converter automatically such that to maintain the load voltage constant. MPPT
provides a high precision in current transition and keeps the voltage without any
changes. In addition, to improve the performance of the hybrid system, a strategy
of energy management has been used, as well as the control of the charging and
discharging of the battery and the setting of the priority of the load and the control of
the secondary charge. The proposed scheme ensures optimal use of the photovoltaic
(PV) array and DFIG wind proves its efficacy in variable load conditions, unity and
lagging power factor at the inverter output (load) side. A dynamic and steady-state
mathematical model and simulations for the entire scheme is presented. The model
is implemented in the MATLAB/Simulink platform.
T. Boutabba (B) · H. Sahraoui

LSPIE Batna Laboratory, University of Abbès Laghrour Khenchela, Khenchela, Algeria
H. Sahraoui
Faculty of Technology, University of Hassiba Benbouali University of Chlef, UHBC, Ouled Fares,
Algeria
M. L. Bechka · S. Drid
LSPIE Laboratory, University of Batna 2 (Mostefa Ben Boulaïd), Fesdis, Algeria
L. Chrifi-Alaoui
University of Picardie Jules Verne, LTI (EA 3899), 13 av F. Mitterrand, 02880 Cuffies, France

https://doi.org/10.1007/978-981-15-6403-1_7
94 T. Boutabba et al.
Keywords MPPT · Hybrid energy · Stand alone system · Photovoltaic · Wind

energy · Power management · Fuzzy logic control
1 Introduction
The constantly increasing consumption of energy, the exhaustible nature of fossil fuel
and its ever increasing cost and global environment worsening have created interest
in green power generation systems.
However, due to the random nature of the wind speed and solar radiation, the
reliable power supply cannot be provided with only the use either a wind turbine
generator (WTG) or a photovoltaic generator (PV) [2]. However, it can be improved
by using a hybrid association of photovoltaic and wind systems, where the storage
is the solution for standalone applications [3–10].
Moreover, In addition, numerous studies and economic analyzes have indicated
that hybrid PV-wind and PV-wind-diesel systems with or without battery backup for
power generation are now considered cost-effective technologies [7, 9, 11.
The major limitation for these hybrid systems is the control requirement for
optimal efficiency. For that, the control systems used a maximum power point
tracking (MPPT) control algorithm to track the maximum available power from the
high penetrating renewable sources (PV-wind). In Solar conversion system, MPPT
is achieved with the DC to DC converter which operates PV module at its maximum
power point by using the conventional algorithms such us P&O (Perturb and Observe)
and incremental conductance algorithm. This algorithm may fail to track the point of
maximum power since performance depends largely on the choice of the perturbation
size value used. The value of this disturbance is usually calculated by testing and/or
simulation tests [5]. In recent years, some control techniques not conventional and
artificial such as the fuzzy logic algorithm and the neural networks (ANN) have been
developed [5, 6]. The application of these algorithms varies with their complexity
and convergence speed of the maximum power point.
In order to improve energy conversion efficiency of the photovoltaic generator,
this paper presents a comparative study between an classical method of maximum
power point tracking (MPPT) and intelligent controller as fuzzy logic controller
(FLC), under various climatic conditions. Also, the power management strategy
that enhances the hybrid system performance by introducing battery charging
and discharging limit control, load priority setting and secondary load control is
discussed. This proposed strategy has been implemented on a PV-wind hybrid energy
system model with battery backup.
A Comparative Study of MPPT Techniques for Standalone Hybrid … 95
2 Modeling of Hybrid System
The hybrid scheme with wind turbine and photovoltaic system, doubly fed induction
generator models proposed is presented in this section.
2.1 Wind Turbine Modeling
Wind turbines convert the kinetic energy present in the wind into mechanical energy
by means of producing torque. Since the energy contained by the wind is in the form
of kinetic energy, its magnitude depends on the air density and the wind velocity.
The wind power developed by the turbine is given by the Eq. (1)
1
Pw = C P Av3 (1)
2
where C P is the Power Co-efficient, ρ is the air density in kg/m3 , A is the area of
the turbine blades in m2 and v is the wind velocity in m/s. The power coefficient C P
gives the fraction of the kinetic energy that is converted into mechanical energy by
the wind turbine. It is a function of the tip speed ratio λ and depends on the blade
pitch angle for pitch-controlled turbines. The tip speed ratio may be defined as the
ratio of turbine blade linear speed and the wind speed
Rω
λ= (2)
V
Substituting (2) in (1), we have:
3
1 R
Pw = C P (λ)ρ A ω3 (3)
2 λ
The output torque of the wind turbine is calculated by the following Eq. (4).
1
Ttur bine = C P ρ AV /λ (4)
2
where R is the radius of the wind turbine rotor (m) There is a value of the tip speed
ratio at which the power coefficient is maximum [1]. Variable speed turbines can be
made to capture this maximum energy in the wind by operating them at a blade speed
that gives the optimum tip speed ratio. This may be done by changing the speed of
the turbine in proportion to the change in wind speed.
2.2 Photovoltaic Array Model
Figure 1 shows the equivalent circuit of photovoltaic (PV) cell. A PV cell can be
represented by an equivalent circuit [9] as shown in Fig. 1. The characteristics of this
PV cell can be obtained using standard Eq. (5).

V +Rs I V + Rs I
I = I pv − I0 e Vt a
−1 − (5)
Rp
I PV = photovoltaic current
I O = saturation current
V t = N S k T/q, thermal voltage of array
Ns = cell connected in series
T = is the temperature of the p–n junction
k = Boltzmann constant
q = electron charge
RS , RP = equivalent series/parallel resistance of the array.
The general equation of a PV cell describes the relationship between current and
voltage of the cell.
Since the value of shunt resistance RP is high compared to value of series resistance
RS the current through the parallel resistance can be neglected. The light generated
current of the photovoltaic cell depends linearly on the solar irradiation and is also
influenced by the temperature [10] given by the Eq. (6).
G
I pv = I pv,n + K I T (6)
Gn
I PV = is the light generated current at nominal condition (25 °C and 1000 W/m2 )
Fig. 1 Equivalent circuit of

PV cell
Table 1 Parameter of
I mp 4.40 A V oc 21.20
KPC-12075 solar array at
25 °C,1000 W/m2 V mp 17.00 V a 1.3
Pmax 74.8 W Rsc 0.511
I sc 5.02 A Rsh 44.25
Ns 36 Kv −74.7 mv/°C
I 0, n 9.83 × 10−8 KI 2.80 mA/°C
ΔT = T −Tn (7)
T, T n = actual and nominal temperature [K]

K I = current coefficients
G = irradiation on the device surface [W/m2 ]
Gn = nominal irradiation.
The current and voltage coefficients KV and KI are included as shown in Eq. (8) in
order to take the saturation current IO which is strongly dependent on the temperature
(Table 1).
Isc,n + K I T
Io =
Voc,n +K V T
(8)
e aVt
−1
K V = voltage coefficients
K I = current coefficients.
2.3 Doubly-Fed Induction Generator Dynamic Model
The application of Concordia and Park’s transformation to the three-phase model of

the DFIG permits to write the dynamic voltages and fluxes equations in an arbitrary
d–q reference frame:
⎧
⎪
⎪ Vds = Rs Ids + dψds
− ωψqs
⎪
⎨ dt
dψqs
Vqs = Rs Iqs + dt
+ ωψds
(9)
⎪
⎪ V = Rr Idr + dψdr
− ωψqr
⎪
⎩V
dr dt
dψqr
qr = Rr Iqr + dt
+ ωψdr
⎧
⎪
⎪ ψds = L s Ids + M Idr
⎨
ψqs = L s Iqs + M Iqr
(10)
⎪
⎪ ψ = L r Idr + M Ids
⎩ dr
ψqr = L s Iqr + M Iqs
The stator and rotor angular velocities are linked by the following relation:
ωs = ω + ωr (11)
where the electromagnetic torque c e can be written as a function of stator fluxes and
rotor currents:
M
Γe = p ψqs Idr + ψds Iqr (12)
Ls
3 DC–DC Boost Converter
A dual stage power electronic system comprising a boost type DC–DC converter
and an inverter is used to feed the power generated by the PV array to the load. To
maintain the load voltage constant a DC–DC step up converter is introduced between
the PV array and the inverter. The block schematic of the proposed scheme is shown
in Fig. 2.
In this scheme a PV array feeds DC–DC converter used in step-up configuration.
For a DC–DC boost converter, by using the averaging concept, the input–output
voltage relationship for continuous conduction mode is given by
Vo 1
= (13)
Vin (1 − D)
Fig. 2 A MPPT controller

in a PV system
where,
D = duty cycle.
4 The Control of Hybrid System
In this section the strategy of controls of the photovoltaic system and the DFIG are
discussed.
4.1 Control Strategy of DFIG
By choosing a reference frame linked to the stator flux, rotor currents will be related
directly to the stator active and reactive power. An adapted control of these currents
will thus permit to control the power exchanged between the stator and the grid. If
the stator flux is linked to the d-axis of the frame we have
ψds = ψs and ψqs = 0 (14)
and the electromagnetic torque can then be expressed as follows.

The active power and consequently the torque depend on the rotor current compo-
nent of the q-axis. If the statorique resistance per phase is neglected, which is a real-
istic approximation used in wind energy conversion, the statorique vector voltage is
in quadrature.
The stator voltages are
Vds = 0 and Vqs = Vs = ωs ψs (15)
Using Eqs. (10), (14) and (15) the stator active and reactive power can then be
expressed only versus these rotor currents as

P = Vs Iqs = −Vs LMs Iqr
(16)
Q = Vs Ids = VLs ψs s − VLs M
s
Iqr
In steady state, the second derivative terms of the two equations in (16) are nil.
The third term, which constitutes cross-coupling terms, can be neglected because of
their small influence. Knowing relations (9) and (16), it is possible to synthesize the
regulators and establish the global block-diagram of the controlled system (Fig. 3).
Fig. 3 Block-diagram of the controlled system
4.2 Control Strategy of Photovoltaic System
Several MPPT control algorithms in which the classical algorithms are the technique
of disturbance and observation (P & O) and the technique of incremental conductance
where they have been developed in many research works. This work involves applying
the Fuzzy Logic Controller (FLC) as an intelligent MPPT method for PV system
control.
4.2.1 MPPT Using Fuzzy Logic Controller
Fuzzy logic has been used for tracking the MPP of PV systems [7] because it has the
advantages of being robust, relatively simple to design and do not require the knowl-
edge of an exact model. The proposed FLC for MPPT in PV systems is based on the
model developed by [11] but with a few modifications made in the fuzzification and
rule base. The main components in a FLC are fuzzification, rule-base, and inference
and defuzzification as shown in Fig. 4. The input variables to the FLC are the change
in PV array voltage (ΔV pv ) and change in current (ΔI pv ) whereas the output of FLC
is the duty cycle (ΔD).
Using seven fuzzy subsets in terms of their linguistics variables for the inputs
(ΔV pv ) and (ΔI pv ) to define Their universe of discourse. Figures 5, 6, and 7 shown
inputs and output variable membership functions.
The fuzzy system rule base is created as shown in Table 2 with (ΔV pv ) and
(ΔI pv ) as inputs while D is the output. The fuzzy inference of the FLC is based
Fig. 4 Components of a fuzzy logic controller
VSS VS S M B VB VVB
Fig. 5 Membership functions of the 1st input variable 1st (V pv )
VSS VS S M B VB VVB
Fig. 6 Membership functions of the 2nd input variable 2nd (I pv )
on the Mamdani’s method which is associated with the max–min composition. The
defuzzification technique is based on the centroid method which is used to compute
the crisp output, D.
VSS VS S M B VB VVB
Fig. 7 Membership functions of the output variable (D)
Table 2 Fuzzy rule

V P V I P V
VVS VS S M B VB VVB
VVS VVS VVS VVS VS VVS VVS VS
VS VVS VVS VVS VS S M B
S VVS VVS VS S M B VB
M VVS VS S M B VB VVB
B VS VVB VVB B VB VVB VVB
VB S M B VB VVB VVB VVB
VVB M B VB VVB VVB VVB VVB
5 Hybrid Power Management Strategy
To operate the stand alone PV-wind hybrid energy system a customized control
strategy is implemented ensuring higher reliability of this system under different
atmospheric conditions. The control strategy proposed is developed based on the
state of charge (SOC) of the batteries which includes AC and DC load control,
secondary AC and DC load control during higher generation and most importantly
the battery charge and discharge limit control. The proposed control, subdivided
into five different stages, considers all possible operating conditions to run a hybrid
system efficiently and reliably under different weather conditions. The block diagram
representation of the suggested control strategy is shown in Fig. 8. The major five
control stages are:
1. Initialization
2. Normal operating condition
3. Lower limit of battery SOC control (High load or low generation)
Fig. 8 Block diagram representation of the diagram of the management strategy
4. Upper limit of battery SOC control (Low load or high generation)

5. Wind energy extreme case control.
The dynamic simulation model for the system is described. The system consists of
various units, PV power and wind power units as primary sources of energy, battery
bank unit as auxiliary source of energy, DC–DC and DC–AC converters, load unit
and control unit. The function of controller unit is to ensure the management of the
power, which is delivered by the hybrid system to satisfy the load and to charge the
battery. The inverter unit is used to convert the DC generated power from renewable
energy sources to feed the load with the required AC power. The excessive charge
from the battery will be dumped to the dump load unit. The dump load in this case
is the battery storage Figs. 8 and 9.
In Fig. 10 we can see the two characteristics (I-V) and (P-V) with the application of
three different techniques to maximize the power of panels through a boost converter.
Figures 11 and 12 shows respectively that the output voltage of the photovoltaic
panel is well controlled to follow the optimal value according to the variation of
the sunlight and the load, and while the panel always produce a maximum power
under all conditions. Our controller has optimized the production of the photovoltaic
system.
The report named “Gate” is controlled to ensure proper operation of the converter
“bidirectional chopper” voltage applied to the battery (Fig. 13).
Until the moment of 1.5 s, the climatic conditions deteriorate “a kiss of the
sunshine with a weak wind”, but the requested power remains in surplus the battery
Convertisseur BOOST ONDULEUR
DC/DC Gat
DC/AC
Panneaux solaire MPPT Control

Charges
Charge dérivative direct du
MPPT puissance
AC/DC DC/DC
Redresseur Convertisseur BOOST

Eolienne
Superviseur S1
S2
Batterie
EDC« SOC » DC/DC

ΔP
+ +-
Convertisseur Buck/BOOST
bidirectionnel
Fig. 9 Block diagram representation of PV-wind hybrid power system with supervisor
5
x 10
2.5
400
Inc P&O
P&O Fuzzy Logic
300 Logique floue 2 Inc
Current Ipv (A)
200
Power (W)
1.5
100
0 1
-100
0.5
-200
-300 0
-400 -200 0 200 400 600 800 -400 -200 0 200 400 600 800
Voltage (V) Voltage (V)
Fig. 10 (I-V) and (P-V) MPPT characteristics
is discharged (S1 = 1) because the load request exceeds the renewable power, and
when this power is sufficient, the battery will be deactivated (S2 = 1) in order to use
all the renewable power produced (Fig. 14).
In the case where the load power exceeds the power supplied, the battery must
compensate for this difference.
In the 1 s to the 1.5 s, the wind speed increases, at the same time the load demand
decreases, the surplus becomes too big (Ppv + Peel-Pdem = 3.2 kW), then to avoid
the gaseous release of the battery, the load shed resistance is switched on (Gate = 1)
to dissipate the excess energy Fig. 12.
1800
1600
DC bus Voltage Vdc(V)
1400
1200
Fuzzy loagic
1000 P&O
Inc
800
600
400
200
0
0 0.2 0.4 0.6 0.8 1 1.2
Time (s)
Fig. 11 DC bus voltage
5
x 10
2.5
P&O
Inc
Fuzzy logic
2
Power (W)
1.5
0.5
0
0 0.2 0.4 0.6 0.8 1 1.2
Time (s)
Fig. 12 Power of the PV panel
The voltage quality is always respected (f = 50 Hz, V = 380 V) in all climatic

and operating conditions.
These simulations show that our controller has good results. It has the load
demand, despite the variations in weather conditions, with good power while
respecting the charging process of the battery. Other simulations with other more
or less important variations made it possible to verify these good results (Fig. 13).
1000 12
SOLAR IRRADIANCE (W/m≤)

900 11
800 10
Wind speed (m/s)

700 9
600 8
500 7
1.5 1.5
Time (s) Time (s)
400 6
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Fig. 13 Wind speed and solar irradiance variation and load
x 10 5
2.5
Power hyb=Ppv+Pwind
Power Battery
Power Load
2
1.5
Power (W)
0.5
-0.5
0 0.5 1 1.5 2 2.5 3
Time (s)
Fig. 14 Renewable power, battery power and load power
7 Conclusion
This method is capable of extracting maximum power from each of the independent
PV arrays connected to DC link voltage The hybrid power management strategy is
developed to control the power flow of the system and maintain battery charging and
discharging limits under any operating conditions. Fuzzy Logic Maximum Power
Point Tracking (MPPT) algorithms, that provide maximal voltage extracted from the
PV system and fed the inverter, have been included.
We have introduced a power generation system for sites that cannot be connected
to the grid. Indeed, we used a continuous bus which receives the energy produced by
the photovoltaic and wind power sources and then delivered to the consumer using an
inverter. One of the advantages of this structure is the use of the necessary batteries
in case of need to compensate for a possible lack of power. We have also used a load
shedding resistor which allows to dissipate the surplus energy in the event of a drop
in demand and full batteries.
To have optimal behavior of the installation from a power flow point of view, we
have developed a supervisor or. It enables efficient and rational energy management
to meet the energy needs of the consumer. Several simulation results have been
presented to illustrate the performance of our facility in the presence of climate
change and changes in energy consumption.
Based on the above analysis, it can be concluded that the proposed hybrid scheme
offers a reliable and alternative solution of renewable energy to meet the increasing
power demand, overcoming any intermitting challenges and allowing widespread
development of such systems especially in locations where no grid is available.
For Reviewer
• Lack of explanation on the control of the DFIG and the rectifier, and Why the
boost chopper is used to control the power produced by the DFIG.
Static converters are indispensable in the structure of such a system. They make it
possible not only to operate at variable speed but also to extract the maximum of the
power produced. A structure often used is that which uses a diode rectifier bridge
associated with a controllable converter (of the Boost type, buck or buck_boost).
• Lack of explanation on the operation and control of the DC/AC converter
supplying the load.
For simplification reason, the operation and control of the DC–AC converter feeding
the three-phase load will be applied to prove that our controller is working well.
References
1. Lei Y, Mullane A, Lightbody G, Yacamini R (2006) Modeling of the wind turbine with a doubly
fed induction generator for grid integration studies. IEEE Energy Conv 21:257–264
2. Messaïf I, Berkouk E-M, Saadia N (2010) A study of DTC-power electronic cascade fed by
photovoltaic cell-three-level NPC inverter. Smart Grid Renew Energy 1:109–118
3. Vijayalakshmi R, Nazar Ali A (2012) Hybrid power generations (wind/solar by PV)—an effi-
cient output with reduced total harmonics distortions using multi level inverter. Int J Commun
Eng 4(4), 4 March 2012
4. Balasubramanian G, Singaravelu S (2012) Fuzzy logic based controller for a stand-alone hybrid
generation system using wind and photovoltaic energy. Int J Adv Eng Technol, May 2012. ISSN
2231-1963
5. Jain SK, Agrawal P, Gupta HO (2002) Fuzzy logic controlled shunt active power filter for
power quality improvement. IEE Proc Electr Power Appl 149(5):317–328
6. Esram T, Chapman PL. Comparison of photovoltaic array maximum power point tracking
techniques. IEEE Trans Energy Conver 22(2):439–449 July 2007
7. Boutabba T, Drid S, Chrifi-Alaoui L, Mehdi D, Benbouzid MH, Alibi A, Ouriagli M (2018)
dSPACE—Real-time implementation of sliding mode maximum power point tracker for photo-
voltaic system. In: 7th international conference on systems and control, ICSC 2018, València,
Spain, 24–26 October 2018
8. Tarek B, Drid S, Benbouzid MEH (2013) Maximum power point tracking control for photo-
voltaic system using adaptive neuro-fuzzy “ANFIS”. In: 2013 ecological vehicles and renew-
able energies (EVER), 8th international conference and exhibition on ecological vehicles and
renewable energies (EVER). IEEE, Monaco
9. Tarek B, Drid S, Benbouzid MEH (2014) A multi-output boost converter (MOB) controlled by
fuzzy logic technique supplied by a photovoltaic system with grid-connected and fed by three
level inverter. In: International conference on electro-energy, ICEE’14, Skikda, Algeria, 10–11
November 2014
10. Boutabba T, Drid S, Benbouzid MH (2014) A hybrid power generations system (wind
turbine/photovoltaic) to driving a DFIG fed by a three inverter. In: STA’2014, the 15th inter-
national conference on sciences and techniques of automatic control & computer engineering,
Hammamat, Tunisia
11. Boutabba T, Drid S, Chrifi-Alaoui L, Ouriagli M, Benbouzid MEH (2016) dSPACEs—Real-
time implementation of maximum power point tracking based on ripple correlation control
(RCC) structure for photovoltaic system. In: 2016 IEEE ICSC, Marrakesh, Morocco, pp 371–
376
Adapted Search Equations of Artificial
Bee Colony Applied to Feature Selection
Hayet Djellali, Souad Guessoum, and Nacira Ghoualmi-Zine
Abstract This article proposes a Novel Search Equation of Artificial Bee Colony
called NSABC. This study introduces new search equations for employed and
onlooker bees involving more parameters for each food source (more features) and
avoid local optima. NSABC outperforms various ABC variants and particle swarm
optimization PSO in term of reduced size of feature and accuracy. Experimental
results validate the efficiency of NSABC method on machine learning UCI data.
Keywords ABC · Feature selection · Search equation · Exploitation · Exploration
1 Introduction
Feature selection (FS) aim to select the best collection of features that ensures the
highest accuracy and minimal number of features. Four categories of FS approaches
are described in literature: filters, wrappers, embedded and hybrid approaches [1–3].
To solve feature selection task, optimization algorithms such as Artificial Bee
Colony [4] (ABC), Genetic algorithm [5] (GA); Ant Colony Optimization (ACO)
[6]; Particle swarm optimization [7] (PSO), and differential evolution [8] (DE) are
efficient methods. However, they often stuck in local optima.
ABC is considered as wrapper method because of using machine learning
algorithms (classifier) to assess the proposed solutions quality defined as fitness.
Employed, Onlooker and Scoot bees are different types of Artificial Bee colony
ABC ensuring different functions. Employed bees convey via waggle dance the
position and quality of food sources to onlooker. Onlooker bees select food source
to be exploited using received information. If the source is exhausted, the employed
become scout and start to explore a new food source. ABC is simple to implement
with few parameters, ABC benefits from high exploration compared to GA, and ACO
[6]; PSO [7] However, it converges slowly because of the poor local exploitation.
H. Djellali (B) · S. Guessoum · N. Ghoualmi-Zine

Department of Computer Science, LRS Laboratory, Badji Mokhtar University, Annaba, Algeria

https://doi.org/10.1007/978-981-15-6403-1_8
110 H. Djellali et al.
The search equation of ABC focuses on exploration, but it neglects exploitation.

ABC does not profit from the prior information of the best solution and the promising
evolution direction. To enhance ABC, new solutions search equations were proposed
in ABC variants to realize a tradeoff between exploration and exploitation [9, 10].
Various previous search equations exploited the information of the Current best
solution. For instance, Kiran and Findik [11] design five search strategies and
different solutions update rules in ABC. Influenced by PSO and searching to improve
the exploitation ability of ABC, Zhu and Kwong [10] developed a new search equa-
tion called Gbest-guided ABC (GABC) which benefit from prior information of
the actual good solution. The comparative study proved that GABC is better than
standard ABC.
The best so far ABC [12] algorithm conveyed the best solution to the whole colony
and improved the solutions of onlooker bees. Evolutionary based similarity search
process is introduced in binary ABC for feature selection investigated by Hancer
et al. [1]. It is compared to PSO, GA, discrete ABC and the experiments carried out
concluded that the accuracy is better on 10 datasets.
A modified onlooker bee is described by karaboga [13] where a new search equa-
tion is developed and benefit from the information of the best solution. The algorithm
is called quick ABC (qABC). Gao and Liu [14] developed a novel search equation
with crossover operation of GA (called CABC) improving the local search of ABC.
The experiments validated that CABC is effective.
ABC algorithm generates a new food source affecting only one parameter (feature)
of the corresponding previous solution. This modification makes the new solution
disturbed weakly and founded in the neighborhood of the previous solution.
We propose a novel ABC method called Novel search Algorithm ABC (NSABC)
to enhance the performance of ABC and realize a tradeoff between the exploration
and the exploitation. Designed search equations create better candidate solutions
involving more features for each food source of employed and onlooker bees.
Our contributions focus on:
– Reducing the features space to the efficient features realizing high accuracy;
– Adapting the search equations to ameliorate the exploitation process;
– Enhancing the ABC convergence and comparing the proposed NSABC with PSO
and different variants of ABC.
The manuscript is organized as follows. Section 2 describes original ABC and
related work. Section 3 explains the proposed FS methods: Novel search ABC. Exper-
imental results and discussions are carried out in Sects. 4 and 5. finally conclusion
in Sect. 6.
Adapted Search Equations of Artificial Bee Colony … 111
2 Methods and Related Works
2.1 Artificial Bee Colony
In Artificial bee colony (ABC) [4], foraging behaviors of honey bee colony is simu-
lated. The candidate solutions represent food sources and their fitness is the nectar
amounts.
Algorithm 1. ABC algorithm[4].

1: Initialize the candidate solutions of the population Xi,
i =1, ..., NBsource.
2: Evaluate the population Xi using fitness function,
i, i = 1, .., NBsource.
For it = 1 to Max iterations do
Employed Bees(); Onlooker_bees();
Scout _Bees(): Memorize the best solution realized.
iter =iter + 1
End for.
Employed Bees() {
For each Employed Bees i
Compute and evaluate new solutions Vi using Eq. (1).
V(i,j) = X(i,j) + phi(i,j) * (X(i,j) – X( p, j))
(1)
Where p=1…NBsource and p≠i;
parameter j ϵ [1..NBfeatures];
5: Apply the greedy selection. Update trial counter.
Endfor }
Onlooker Bees(){
For each Onlooker Bee i
Compute the probability Pi for the solution Xi by Eq. (2).
P(i) = fitness (i) / ∑ fitness(t) , t=1..N. (2)

Calculate and evaluate new solutions Vi for the solution Xi
depending on Pi.
Apply the greedy selection. Update trial counter}
Scout Bees () {
The abandoned solutions are replaced with a new Xi
described by Eq. (3).
X(I,j) =X(min, j) + rand(0,1) * ( X(max,j) – X(min, j)) (3)
Reset trial to zero. }
2.2 Related Work
Gao et al. [15] developed a new algorithm combining DE with Gbest guided ABC
(GABC) proposing a new strategy. The information of the previous search explo-
ration is applied to enhance the exploitation. The experimental results validated that
DGABC outperforms the other algorithms.
Authors [16] developed a hybrid artificial bee colony algorithm. During the search
process, each bee can reproduce or die leading to dynamic size of bee population. The
search equation is redesigned with insertion of global best. Experiments consolidate
the effectiveness of this algorithm tested on the CEC 2014 benchmarks compared to
other mimetic algorithms.
This study [17] introduces swap and crossover operators in a new binary version
of ABC algorithm. The improvement of local and global search of ABC is realized.
The obtained results confirm that GBABC is suitable algorithm in binary optimiza-
tion. Gao et al. [18] developed a new ABC algorithm named Enhanced EABC. The
search equations of employed and onlooker bees are modified. Experiments validate
that EABC significantly produces faster global convergence and better performance
compared to traditional FS methods.
Mao et al. [19] implemented a modified artificial bee colony algorithm (ABCEM)
with extended memory in employed and onlooker bees. From Experiments, ABCEM
algorithm converges faster and realizes a tradeoff between exploration and exploita-
tion compared with other ABC versions. Authors [20] developed an approach named
depth first search (DFS). The elite information is added to two novel search equations
in employed bee phase. The experiments conducted consolidate that DFS enhance
both the convergence and accuracy of ABC.
Authors [21] propose two hybrid FS approaches based on particles swarm and
artificial bee colony ABC-PSO and ABC with Genetic algorithm ABC-GA. Muta-
tion operators and PSO are implemented in ABC phases. This architecture attempts
to improve the search ability of Bees. To increase exploitation and found better solu-
tions, mutation is introduced in both Onlookers and Scout Bees and PSO particles
are inserted in Employed Bees. It has been found that the proposed method ABC-
GA outperforms the proposed ABC-PSO method and other approaches on WDBC,
hepatitis, and DLBCL.
Akay and Karaboga [22] enhanced the search ability of onlookers by incorporating
the parameters: modification rate (MR) and scaling factor (SF) to adjust the frequency
and the perturbation.
3 Proposed Novel Search Equations of ABC Applied

to Feature Selection
To overcome the insufficiency of poor solution search equation at exploitation; we

propose two new search equations. The solution search equation of employed and
onlooker bees of (ABC) algorithm are adapted to change more features.
In this paper, we propose a novel ABC method called Novel search ABC (NSABC)
to improve the performance of ABC in Fig. 1. New search equations are designed
for the employed and onlooker bees involving more features for each source.
In fact, ABC generates the new food source modifying one parameter of the
current solution. This change conducts to disturb weakly the new solution and it will
be located in the neighborhood of the associated solution. We increase the number
of parameters chosen randomly set to parameter called parlimit to achieve better
exploitation.
Firstly, the solution search equation generated concerns more than one parameter j
(1…,NBfeatures) to increase the search space in the neighborhood for the employed
bees.
- Pseudo code A
For c=1: parlimit
Generate j randomly between [1… NBFeatures];
V(i,j) = X(I, j) + φ(i,j) (X(r1,j) – X(r2,j) ) (4)

End
Fig. 1 Novel search

Initialize_ABC() ; cycle=1 ;
equations NSABC algorithm Begin Computeparlimit() ;
Cycle =Max
Iter Best
More
Enhanced Employed Bee
feature
Enhanced Onlooker Bee

Best
solution
Scout Bees:
Save best solution ;

cycle=cycle+1;
Where parlimit is the maximum number of features allowed to change. r1 and r2

are two different neighbors for the same parameters j.
Example: Source food vector V(3,:) is modified on 3 values of j. index 1; 3 and
6 after Pseudocode A executed.
J = 1..NBfeatures. where NBfeatures = 7;
V(3,j); j = 1..7 before Pseudocode A

1 1 0 1 1 0 0
V(3,j); j = 1..7 after Pseudocode A
0 1 1 1 1 1 0
Parlimit is a maximum number of parameters (features) to change depending on

the feature size of dataset. We choose various dataset with different size of features
which means that the value of parlimit must defined according to the size of feature
datasets.
Parlimit parameter must be subject to deep investigation according to the size of
features in each data. We adopt the strategy to vary the index c of parlimit (pseudocode
A) between 1 to parlimit randomly. High value of parlimit means that exploration is
prioritized than exploitation. Nbfeatures is the size of feature
Pseudo code Computeparlimit(Nbfeatures){

If (Nbfeatures <100) Parlimit=3;
Else if (Nbfeatures <=1000) Parlimit=7;
Else Parlimit=15;
End;
}
NBFeatures is the number of parameters.
Secondly, the prior information about the best solution is
added to solution search equations of the onlookers to
improve the candidate solution.
NBFeatures is the number of parameters.

Secondly, the prior information about the best solution is added to solution search
equations of the onlookers to improve the candidate solution.
Pseudo code B
For c=1: parlimit
Generate j randomly between [1.. NBFeatures];
V(i,j) = X(best, j) + φ(i,j) (X(r1,j) – X(r2,j) ) (5)

End
The main ABC functions are described below:
Initialize_ABC() {
1.1: The population of food sources is created using Eq. (3)
1.2: Evaluate the quality of population using Eq. (6)
Fit(i)= accuracy (6)
Enhanced Employed Bees

The search equation Eq. (5) proposed is generated using food source X(i) added
to the weighted factor with the difference of two random neighbors r1 and r2.
Instead of using one single value of parameter j m [1..NBFeatures], in traditional
ABC, we modify multiples values of j, making the variability of solution faster and
consequently accelerate the exploitation ability of solutions.
To solve the convergence problems, Akay and Karaboga [22] introduced a modi-
fication rate (MR) in Eq. (1). Then, for each parameter (feature) of candidate solution
a generated random number (0, 1) is compared to MR. if the random number is less
than MR, the Xij is modified.
Enhanced Employed_Bees() {
2. For i=1 to SourceNb
2.1 Produce a new food source Vi from Xi (based on Xi,
i≠r1≠ r2) using Eq. (4):
For c=1: parlimit
V(i,j) = X(I, j) + φ(i,j) (X(r1,j) – X(r2,j) ) (4)

Where r1 and r2 are randomly chosen with (r1≠ r2).
End }
2.2 Compute the fitness of the food source Vi using Eq. (6)
2.3 Apply the greedy selection between the new food source
and the previous one.
Set trial counter to zero if improvement
Otherwise increment by 1.
If (trial counter > limit) then new random food sources
are created and trial counter =0.
end for ; }
Enhanced Onlooker Bees

The information of the best previous solutions guides the search of candidate solu-
tions. The search equation of ABC is adapted by firstly, changing a few parameters
(j = 1…parlimit) and secondly, changing the source I with the index of best previous
solution [23]; we present a modified search Eq. (7) as follow:
V(i, j) = X(best, j) + ϕ(i, j)(X(r1, j) − X(r2, j)); r1 = r2 (7)
best is the index of previous good solutions at iteration −1; . We assume that the
solution V is calculated and affect multiples features.
Modified Onlooker_Bees() {
3.1 Compute the probability values for food sources using
Eq. (2)
3.2 for i=1 to SourceNB
3.3 if (random< p(Xi ) )
3.4 Produce a new food source Vi from Xi ,
Where best≠r1≠r2) using Eq. (7)
For c=1: parlimit
V (i, j) = X (best , j) +φ (i, j).(X (r1, j) - X (r2, j) ) ; (7)

r1,r2= 1..SourceNB ; r1 ≠r2;
End
3.5 Calculate the fitness of the food source Vi using Eq. (6);
3.6 Apply the greedy selection.
Compared trial counter to limit and update its value.
end if ;
endfor }
Scout Bees
Scout bees are a few employed bees exploiting their food sources without improve-
ment for multiples iterations. So, to avoid wasting energy; the corresponding bee
becomes the scout and the food sources exhausted are reinitialized in the search
space randomly.
Scout Bees() {
4.1 Search the exhausted food sources and replace it with
random sources.
For each exhausted Bee
X(I,j) =X(min, j) + rand(0,1) * ( X(max,j) – X(min, j)) (3)
Evaluate new sources food
End
Table 1 Datasets tested

Dataset Number of features Class Number of
examples
Hepatitis 19 2 80
WDBC 30 2 569
Ionosphere 34 2 351
Colon Tumor 2000 2 62
DLBCL 4026 2 45
4 Experimental Results
In this section, we describe the classifier used, the parameters of different algo-
rithms, we test the proposed feature selection algorithm; Novel search equation ABC.
Comparative study is carried out with FS methods: standard ABC variants of ABC:
Wang and al [24], Zhu and Kwong [10] and PSO [2].
4.1 Classifier
The classifier used is Support Vector machine [25] and generalizes well in various
applications. SVM accuracy is calculated as fitness function to evaluate accuracy of
ABC and its variants including our proposed method and PSO. Classification task is
accomplished with SVM.
4.2 Databases
We assess the proposed FS algorithm NSABC on Wisconsin Breast cancer Dataset

WBCD, hepatitis, ionosphere, colon cancer, DLBCL of machine learning repository
UCI [26]. The databases are presented in Table 1.
4.3 Comparative Feature Selection Algorithms
We conducted a learning process as follows: 70% for training and 30% for test.
We present the parameters of standard ABC, Zhu ABC [10], Wang ABC [24] and
Particle swarm PSO. All tested algorithms act as wrapper and apply accuracy as
fitness function. These Feature Selection (FS) methods are implemented by using
MATLAB 7.0. The parameters of FS algorithms are detailed in Table 2. Twenty
independent runs are performed and the average accuracy of NSABC is computed.
Table 2 Setting parameters

FS algorithms Parameters
ABC Max generations = 100; Colony size = 100
NSABC MaxIteration = 50; limit = 100; food NBSource = 50 Parlimit = 3 and
random value tested
NSABCprand with parlimit equal to random value between [1…parlimit]
NSABCp3: parlimit = 3
ABCZhu2010 MaxIteration = 50; limit = 100 NbOnlooker = NbEmployed = 50
Wang et al [24] MaxIteration = 50; limit = 100 NbOnlooker = NbEmployed = 50
PSO Population size = 100; inertia weight = 0.7
c1 = 2.01; c2 = 2.01; max iteration = 100
Ten fold cross-validation methods on training data are executed to compute accuracy.
To calculate the test classification accuracy, the chosen features are evaluated on test
set [27].
5 Discussions
The experiments in Table 3 on Hepatitis demonstrate that NSABCp3 achieves the

mean accuracy (train 93.75%) and (test 80.95%) accuracy with low number of
features involved (8). NSABCp3 and NSABCprand outperform all methods for train
accuracy and also for test accuracy. For the number of features, proposed algorithm
is better except for standard ABC and Zhu ABC (6 features).
It can be learn that WDBC from Table 4 that ABC algorithm realizes the best
SVM train accuracy (97.25%) but NSABCp3 and NSABCprand achieve highest
test accuracy respectively (95.69% and 96.47%) where NSABCp3 outperform all
methods for minimal number of features(12) except for ABC (11 features).
It is observed for Ionosphere in Table 5, NSABCprand accomplishes best tradeoff
between classification accuracy (train 83.38% and test 98.49%) and number of
Table 3 HEPATITIS
FS methods Train SVM Test SVM # features
performance of proposed FS
accuracy (%) accuracy
methods
NSABCp3 93.75 ± 0.00 80.95 ± 0.05 8
NSABCprand 93.75 ± 0.00 76.79 ± 0.08 8
ABC 71.43 ± 0.02 71.40 ± 0.06 6
ABCzhu2010 91.47 ± 0.014 70.23 ± 0.08 6
Wang et al. [24] 93.75 ± 0.02 76.19 ± 0.02 9
PSO 84.53 ± 0.01 80.86 ± 0.02 9
Table 4 Performance of Proposed FS methods on WDBC

FS methods Train SVM accuracy (%) Test SVM accuracy # features
NSABCp3 94.73 ± 0.017 95.69 ± 0.007 12
NSABCprand 96.92 ± 0.040 96.47 ± 0.017 14
Wang et al. [24] 96.76 ± 0.01 95.18 ± 0.01 14
ABC 97.25 ± 0.017 95.16 ± 0.0119 11
ABCzhu2010 95.69 ± 0.01 95.42 ± 0.02 13
PSO 96.80 ± 0.01 94.60 ± 0.02 17
Table 5 Performance of
Proposed and Existing FS
methods on Ionosphere
NSABCp3 83.20 ± 0.023 97.63 ± 0.014 13
NSABCprand 83.38 ± 0.03 98.49 ± 0.0144 13
Wang et al. [24] 82.74 ± 0.0153 98.28 ± 0.0096 16
ABC 96.77 ± 0.00 97.20 ± 0.059 12
ABCzhu2010 84.93 ± 0.05 96.77 ± 0.10 12
PSO 90.40 ± 0.02 96.52 ± 0.012 8
features (13). Standard ABC outperforms all methods for train accuracy (96.77%).
Minimal number of features (8) is realized with particle swarm PSO.
Colon cancer shows in Table 6 that Wang and al method produces the best SVM
accuracy (97.22%). Zhu ABC achieves the best tradeoff between classification and
number of features (91.67%; 70.00%.and 761 features).
The results on DLBCL data in Table 7 demonstrate that NSABCp3 achieve the
best tradeoff between the highest accuracy (train 94.00%; test 100.00%) and number
of features (1634).
Experiments show that NSABCp3 and NSABCprand improve the accuracy and
achieve minimal number of features applied to hepatitis WDBC, ionosphere and
DLBCL. Zhu ABC is better than proposed algorithm for colon data.
Table 6 Performance of
proposed methods on colon
NSABCp3 90.00 ± 0.046 60.00 ± 0.20 814
NSABCprand 90.56 ± 0.040 60.00 ± 0.20 765
Wang et al. [24] 97.22 ± 0.0032 48.00 ± 0.0989 792
ABC 70.00 ± 0.002 60.10 ± 0.01 980
ABCzhu 91.67 ± 0.03 70.00 ± 0.12 761
PSO 82.4 ± 0.013 81.78 ± 0.04 1023
Table 7 DLBCL performance of proposed FS methods

FS methods Train SVM accuracy (%) Test SVM accuracy # features
NSABCp3 94.00 ± 0.0203 100.00 ± 0.0 1634
NS ABCprand 94.00 ± 0.031 100.00 ± 0.00 1662
Wang et al. [24] 90.00 ± 0.02 100.00 ± 0.00 1676
ABCzhu 94.00 ± 0.01 100.00 ± 0.00 1641
PSO 90.43 ± 0.02 89.50 ± 0.03 2068
ABC 90.00 ± 0.01 88.30 ± 0.02 1541
It has been found that the novel search ABC NSABCprand and NSABCp3
(parlimit=3) outperform the other ABC variant and exiting methods (ABC, PSO)
for most cases. We adopt the strategy to vary the index of parlimit randomly.
6 Conclusions
In this paper two search equations are adapted in employed and onlooker bees based
on information of the previous successful solution and modifying more parameters
(features) for each food source. The results indicate that increasing the number of
features to change is appropriate under condition that the number of features involved
still low.
NSABC presents promising results in term of accuracy and minimal number of
features for the UCI datasets compared to the state of the art algorithms. Extending
the number of parameters must be tested on others established ABC variants.
References
1. Hancer E, Xue B, Karaboga D, Zhang M (2015) A binary ABC algorithm based on advanced
similarity scheme for feature selection. J Appl Soft Comp 36:334–348
2. Djellali H, Ghoualmi-Zine N, Azizi N (2016) Two stages feature selection based on filter
ranking methods and SVMRFE on medical applications. In: Modelling and implementation of
complex systems. Lecture notes in networks and systems, pp 281–293. vol 1. Springer, Cham
3. Djellali H, Guessoum S, Layachi S, Ghoualmi-Zine N (2017) Fast correlation based filter
combined with genetic algorithm and particle swarm on feature selection. The 5th IEEE
international conference on electrical engineering. Boumerdes (ICEE-B)
4. Karaboga D (2005) An idea based on honey bee swarm for numerical optimization. Technical
report TR06, Erciyes University, Engineering Faculty
5. Holland JH (1975) Adaptation in natural and artificial systems. Ann. University of Michigan
Press, Arbor, MI
6. Dorigo M, Maniezzo V, Colorni A (1996) Ant system: optimization by a colony of cooperating
agents. IEEE Trans on Systems, Man Cybern 26(1): 29–41
7. Kennedy J, Eberhart RC (1995) Particle swarm optimization. Proceedings IEEE Int Conf Neural
Netw 4:1942–1948
8. Storn R, Price K (1997) Differential evolution: A simple and efficient heuristic for global
optimization over continuous spaces. J Glob Optim 11(4):341–359
9. Gao WF, Chan FTS, Huang LL, Liu SY (2015) Bare bones artificial bee colony algorithm with
parameter adaptation and fitness-based neighborhood. Inf Sci 316:180–200
10. Zhu G, Kwong S (2010) Gbest-guided artificial bee colony algorithm for numerical function
optimization. Appl Math Comput 217:3166–3173
11. Kiran MS, Findik O (2015) A directed artificial bee colony algorithm. Appl Soft Comput
26:454–462
12. Banharnsakun A, Tiranee A, Booncharoen S (2011) The best-so-far selection in artificial bee
colony algorithm. Appl Soft Comput J 11(2): 2888–2901
13. Karaboga D, Gorkemli B (2014) A quick artificial bee colony (qABC) algorithm and its
performance on optimization problems. Appl Soft Comput 23:227–238
14. Gao WF, Liu SY (2011) Improved artificial bee colony algorithm for global optimization. Inf
Process Lett 111(17):871–882
15. Gao WF, Huang LL, Liu SY, Dai C (2015) Artificial bee colony algorithm based on information
Learning. IEEE Trans Cybern 45(12):2827–2839
16. Ma L, Zhu Y, Zhang D, Niu B (2016) A hybrid approach to artificial bee colony algorithm. J
Neural Comput Appl 27:387–409
17. Ozturk C, Hancer E, Karaboga D (2015) A novel binary artificial bee colony algorithm based
on genetic operators. Inf Sci 297:154–170
18. Gao WF, Liu SY, Huang LL (2014) Enhancing artificial bee colony algorithm using more
information-based search equations. Inf Sci 270:112–133
19. M. Mao, Q. Duan, and L. Zhang. “Artificial Bee Colony Algorithm Based on Adaptive Search
Equation and Extended Memory”. (2017). Journal Cybernetics and Systems
20. Cui LZ, Li GH, Lin Q, Du Z, Gao WF, Chen JY, Nan Lu (2016) A novel artificial bee colony
algorithm with depth-first search framework and elite-guided search equation. J Inf Sci (367–
368): 1012–1044
21. Djellali H, Djebbar A, Zine NG, Azizi N (2018) Hybrid artificial bees colony and particle
swarm on feature selection, pp 93–105. CIIA Algeria
22. Akay B, Karaboga D (2012) A modified artificial bee colony algorithm for real–parameter
optimization. Inf Sci 192:120–142
23. Gao WF, Liu SY (2012) A modified artificial bee colony algorithm. Comput Oper Res
39:687–697
24. H. Wang, Z. Wu, S, Rahnamayan, H. Sun, Y, Liu and J. Pan (2014) Multi strategy ensemble
artificial bee colony algorithm. Inform. Science 279: 587–603
25. Vapnick V (1998) Statistical learning theory. Wiley
26. Machine Learning Repository UCI. https://archive.ics.uci.edu/ml/datasets.html
27. Kohavi R, John JH (1997) Wrappers for feature selection. J Artif Intel 97(1/2):273–324
Networked Cooperation-Based
Distributed Model Predictive Control
Using Laguerre Functions
for Large-Scale Systems
Kamel Menighed, Ahmed Chemori, Boumedyen Boussaid,

and Joseph Julien Yamé
Abstract This paper proposes a novel cooperative distributed control system archi-
tecture based on unsupervised and independent Model Predictive Control (MPC)
using discrete- time Laguerre functions to improve the performance of the whole
system. In this distributed framework, local MPCs algorithms might exchange and
require information from other sub-controllers via the communication network
to achieve their task in a cooperative way. In order to reduce the computational
burden in the local rolling optimization with a sufficiently large prediction horizon,
the orthonormal Laguerre functions are used to approximate the predicted control
trajectory. Simulation results show that the proposed architecture could guar-
antee satisfactory global performance even under strong interactions among the
subsystems.
Keywords Model predictive control · Larg-scale systems · Distributed control
K. Menighed (B)
Department of Petrochemicals and Process Engineering, University of 20 août 1955, Skikda,
Algeria
A. Chemori
LIRMM, University of Montpellier, CNRS, Montpellier, France
B. Boussaid
MACS-LR16ER22, University of Gabes, Gabès, Tunisia
J. J. Yamé
Université de Lorraine, CNRS, CRAN, F-54000, Nancy, France

https://doi.org/10.1007/978-981-15-6403-1_9
124 K. Menighed et al.
1 Introduction
During the past two decades a lot of interest has been given to the control of a class
of large-scale systems which are composed by many physically or geographically
divided subsystems. However, these subsystems are interconnected and character-
ized by significant interactions. Each subsystem interacts with the other subsystems
through their states and their inputs. At the same time, due to the high performance
requirements, modern control systems are becoming more and more complex. For
these processes, different control solutions have be developed to achieve some global
performance of the whole system.
In centralized MPC architecture, a control agent is able to acquire the informa-
tion of the global system and all the input sequences are optimized with respect to
one given performance index in a single optimization problem. However, when the
number of the state variables and inputs of the system becomes larger, the compu-
tation burden of the centralized optimization problem may increase significantly.
Therefore, it is often impractical to apply the centralized MPC to large-scale systems
due to their size and their higher risk of failure of the MPC controller due to its central-
ized nature. However, the centralized control scheme in these industrial process
has gradually given way to decentralized and distributed control strategies for their
remarkable decrease in system dimensionality using the local MPC control [1, 7].
The main idea in both decentralized and distributed control in large intercon-
nected systems is to approximate the overall system with several separated subsystem
models, where each subsystem is controlled by a local controller. Indeed, each
local controller uses the so-called local performance index. On the one hand, these
control architectures have the advantages of being flexible to system structure, error-
tolerance, less computational efforts and no global information requirements [11, 18,
19]. The only difference between them is that distributed control schemes give due
consideration to the external state and/or control signal information communicated
between sub-controllers from other subsystems while decentralized ones follow a
completely decoupled approach [2, 4, 17].
On the other hand, the development of DCS (Distributed Control Systems), field-
bus, communication network technologies in process industries allows the control
methodologies to utilize their potentials for improving control performance. In this
case, the improvement of the global control performance of the entire system using
the network information exchange for large-scale systems is a valuable open problem.
In order to fulfill the global objective for the global system, cooperation between
the controllers through a digital communication network might be necessary. Thanks
to the digital network, the required cooperation can be achieved by means of a
proper information exchange between the sub-controllers. From the control point of
view, it is well known that MPC allows to deal with linear, nonlinear, multivariable
and constrained systems [3]. Moreover, MPC technique benefits greatly from both
advances in communication technology and advances in computational resources.
Recalling that the MPC strategy is based on an on-line optimization problem and
uses a process model to predict the effect of the control sequence on the behavior of
Networked Cooperation-Based Distributed Model … 125
the plant. Typically, MPC is implemented in a centralized way. The complete system
is modeled and all the control inputs are computed in one optimization problem.
However, for large interconnected systems, it may be necessary to have a distributed
control scheme as mentioned above, where local control inputs are computed using
local measurements and small order models of the local dynamics. But with infor-
mation exchange between the controllers, the objective is to achieve some degree
of cooperation between sub-controllers that are solving MPC problem with locally
relevant variables, costs and constraints.
Previous works on distributed MPC are reported in [6, 13–15]. A preliminary
analysis of the control performance of distributed MPC has been addressed in [13].
Menighed et al. [6] and Razavinasab et al. [10] proposed a distributed state estima-
tion strategy, developed for supporting distributed state feedback MPC for large-
scale interconnected systems. In [15], two approaches for coordination between
sub-controllers are proposed leading to the so-called communication and cooper-
ation based MPC. In the cooperation- based MPC, each sub-controller knows the
global objective in order to improve optimality and stability and makes the decen-
tralized strategy very close to the centralized one. When only the local objectives are
known, a hierarchical decentralized control architecture uses a supervisor to compute
the global optimum and to coordinate the sub-controllers, in communication-based
MPC [8].
In this paper, we deal with unconstrained distributed model predictive control
of large-scale interconnected systems to achieve global performance based on the
use of a cooperative strategy between the sub-controllers. Thanks to the flexibility
and the on-line optimization process inherent to MPC algorithms, we propose a
distributed framework based on MPC with Laguerre functions. In order to reduce
the computational burden in the local rolling optimization with a sufficiently large
prediction horizon, the orthonormal Laguerre functions are utilized to approximate
the predicted control trajectory. It is worth to emphasize that the MPC/Laguerre
approach is becoming an important topic in modern control theory and practice [16].
2 Complex Process Interaction Modeling
We consider the overall system represented as a discrete, linear time-invariant (LTI)

model of the form

x(k + 1) = Ax(k) + Bu(k)
S (1)
y(k) = C x(k)
where x(k) ∈ n x , u(k) ∈ n u and y(k) ∈ n y are the state, the control input and
the output, respectively with their corresponding dimensions, n x , n u , n y .
In the decentralized modeling framework, it is assumed that subsystem-subsystem
interactions have a negligible effect on system variables, i.e., we suppose that the
previous overall model S is composed of N subsystems Sii , each subsystem is
represented by the following state-space equations

xi (k + 1) = Aii xi (k) + Bii u i (k)
Sii (2)
yi (k) = Cii xi (k) i = 1, 2, . . . , N
where xi (k) ∈ n xi , u i (k) ∈ n ui and y

i (k) ∈ i are
ny
the local state,
control
input and output, respectively and n x = n
i xi , n u = n
i ui and n y = i n yi .
Frequently, the components of the interconnected system are tightly coupled due the
material/energy and/or information flow between them. In such cases, the decentral-
ized assumption leads to a loss in achievable control performance. It is natural to
view the previous overall model S composed of N subsystems Si interacting with
each other through linear interconnections. Each subsystem is represented by the
following state-space equations

xi (k + 1) = Aii xi (k) + Bii u i (k) + wi (k)
Si (3)
yi (k) = Cii xi (k) + vi (k) i = 1, 2, . . . , N
where the state and output interaction vectors wi and vi are given by

N
N
wi (k) Ai j x j (k) + Bi j u j (k)
j=1; j=i j=1; j=i
(4)
N
vi (k) Ci j x j (k)
j=1; j=i
These vectors represent the interaction of subsystem j = i on subsystem i.

The proposed distributed control architecture fosters implementation of cooperation-
based strategy for several interacting processes (3)–(4) in order to emulate the benefits
achievable with centralized control.
3 Problem Statement
By means of (3) and (4), the future state and output prediction over a horizon p are
given by

l
x̂ i (k + l|k) = Alii x̂ i (k|k) + Aiis−1 Bii u i (k + l − s|k)
s=1

l
+ Aiis−1 ŵi (k + l − s|k − 1) (5a)
s=1
ŷ i (k + l|k) = Cii x̂ i (k + l|k) + v̂i (k + l|k − 1) (5b)

Given the overall system S composed by N interactive subsystems Si , i =

1, 2, . . . , N , the unconstrained Distributed Model Predictive Control (DMPC)
problem with prediction horizon p > 0 and control horizon m > 0 two integer
values, with m ≤ p consists of finding, at time k, a set of independent sub-controllers
Ci such that each Ci minimizes the local cost function Ji .
p
2
m
d
Ji = yi (k + l|k) − ŷi (k + l|k) + u i (k + l − 1|k)2Ri for i = 1, . . . , N
Qi
l=1 l=1
(6)

model constraints given by (5)
subject, for l = 1, . . . , p, to
initial condition : x̂i (k|k) = x̂i (k)
where
α2Λ α T α is the norm of vector α induced by matrix .
yid (k + l|k) the desired output.
u i (k + l − 1|k) future control increment at time k, with u i (k) = u i (k − 1) +
u i (k|k).
Qi a symmetric and positive semi-definite matrix of appropriate
dimension (Q i ≥ 0).
Ri a symmetric and positive definite matrix of appropriate dimension
(Ri > 0).
It should be noted that u j (k + l − 1|k) is excluded in the performance index,
since it is independent of the future control sequence of S j .
4 Cooperation Based Distributed MPC Problem
In order to find an explicit solution to the first version of DMPC problem, each sub-
controller Ci is decomposed into three connected function blocks: an optimizer, a
state predictor and an interaction predictor.
Assumptions 1
• the prediction and control horizons are the same for all sub-controllers, i.e., m i =
m j =, pi = p j = p, ∀i, j = 1, 2, . . . , N , j = i;
• all the sub-controllers are synchronous;
• all the sub-controllers communicate only once within a sampling interval;
• all the communication channel introduces a delay of one sampling period.
To simplify the mathematical expressions, the following notations are adopted
• 0a×b is the a × b null matrix;
• Ia (0a ) is the a × a identity (null) matrix;
• diaga {A} is a diagonal block matrix made by a blocks equal to A.
A. Interaction Prediction
Under Assumptions 1, at step k, the predictions of the interaction vectors are given
by
Ŵ i (k, p|k − 1) = Ãi X̂ (k, p|k − 1) + B̃ i Γ˜ i U (k − 1, m|k − 1)

(7)
V̂ i (k, p|k − 1) = C̃ i X̂ (k, p|k − 1)
H̃ i [diag p {Hi,1 } . . . diag p {Hi,i−1 } 0 diag p {Hi,i+1 } . . . diag p {Hi,N }]
where H̃ i ∈ { Ãi , B̃ i , C̃ i }.
B. State Predictor
Under Assumptions 1 at step k, the local state prediction for the sub-controller Ci is
expressed by

l
l
s−1 s−1
x̂ i (k + l|k) = Alii x̂ i (k|k) + Aii Bii u i (k + l − s|k) + Aii ŵi (k + l − s|k − 1)
s=1 s=1
(8)
C. Optimal Control Sequence
Under Assumptions 1, at step k, based on the exchanged information, the interaction

prediction together with the local measurement is used by the optimizer to solve the
MPC optimization problem (without constraints). Once computed the optimal control
sequence [u i (k|k), . . . , u i (k + m − 1|k)], which minimize the local cost function
(6), only the first element of the optimal sequence u i (k|k) is selected and the control
action u i (k) = u i (k − 1) + u i (k|k) is computed and applied as control action to the
subsystem Si , following the receding horizon strategy. Generally, the parameters of
the above optimization problem are p and m; these parameters are directly affecting
computational load in MPC. One of the MPC formulation is the classic approach
presented in [5]. In this approach, for the case of rapid sampling, complicated process
dynamics and/or high demands on closed-loop performance, satisfactory approxi-
mation of the control signal u(k) may require a very large number of parameters
(large m), leading to poorly numerically conditioned solutions and heavy computa-
tional load when implemented on-line. Instead, a more appropriate technique would
be to use Laguerre network in the design of MPC presented in [9, 16].
5 Description of the Control Signal Trajectory
The Z-transforms of the discrete-time Laguerre networks are written as follows
z −1 − a
n (z) = n−1 (z) n = 2, 3, . . . , M (9)
1 − az −1
√
1−a 2
with 1 (z) = 1−az −1
.
where M is the number of Laguerre functions in the network and a is the pole of the
Laguerre network. The scaling factor a is required to be selected by the user, where
0 ≤ a < 1 for the stability of the network. Note that the Laguerre networks are well
known for their orthonormality.
With the relation (9), the Laguerre network is illustrated in Fig. 1.
The discrete-time Laguerre functions are obtained through the inverse Z-transform
of the Laguerre network. The set of Laguerre functions can be expressed as
L(k) = [l1 (k)l2 (k) . . . l M (k)]T , where li (k) denotes the inverse Z-transform of
i (z, a). Taking advantage of the network realization (1), the set of discrete-time
Laguerre functions satisfies the following difference equation
L(k + 1) = Al L(k) (10)
where Al ∈ M×M and is a function of the parameters a and β = (1 − a 2 ), and the

initial condition is given by

L(0)T = β[1 − a a 2 − a 3 . . . (−1) M−1 a M−1 ]
The orthonormality can be expressed by

∞

li (k)l j (k) = 0 for i = j (11a)
k=0
∞

li (k)l j (k) = 1 for i = j (11b)
k=0
The orthonormal property of the Laguerre functions will be used in the design of
MPC. The key idea in MPC based on Laguerre functions lies in the approximating
member of control sequence by a set of Laguerre functions as
Fig. 1 Illustration of a discrete Laguerre network

⎧
⎪
M
⎪
⎨ u(k + l|k) = ln (l)cn (k) = L(l)T η(k)
n=1
(12)
⎪
⎪ for l = 0, 1, . . . , m − 1
⎩
where u(k|k) = u(k|k) − u(k − 1|k)
with k being the initial time of the moving horizon window and l being the
future sampling. The parameter vector η comprises M Laguerre coefficients: η =
[c1 c2 . . . c M ]T and L(l)T is the transposed Laguerre function vector as defined in the
difference Eq. (10).
By using this approximation, the optimization problem (6) can be expressed in
terms of coefficient vector η, instead of u(k) as in the classic approach. Thus, the
coefficient vector η will be optimized and computed in the design. With this design
framework, the control horizon m from the classical MPC approach has vanished.
Instead, the number of terms M(M < m) is used to describe the complexity of
the trajectory in conjunction with the free parameter a. Furthermore, a long control
horizon m can be achieved without using a large number of parameters, leading to
low computational burden and memory storage. In this paper, the MPC based on
Laguerre functions is used in the proposed DMPC scheme.
6 Use of Laguerre Functions in DMPC Design
6.1 Modified-State Predictor
(1) For SISO subsystem Si , i = 1, 2, . . . , N
Under Assumptions 1, at step k, the future state prediction over the horizon p are
given by

l
x̂ i (k + l|k) = Alii x̂ i (k|k) + Aiis−1 Bii u i (k + l − s|k)
s=1

l
+ Aiis−1 ŵi (k + l − s|k − 1) (13)
s=1

let us denote u i (k) = u i (k−1)+u i (k) and u i (k+l−s) = u i (k−1) + l−sr =0 u i (k+
r ) for l = 1, 2, . . . , p. Then, the prediction of the future state variables at time l
becomes

l
x̂ i (k + l|k) = Alii x̂ i (k|k) + Aiis−1 Bii u i (k − 1)
s=1
l

l
l
+ Al−h
ii Bii L i (s − 1)T ηi + Aiis−1 ŵi (k + l − s|k − 1)
s=1 h=s s=1
f or l = 1, 2, . . . , p (14)
where the function u i (k + s − 1) is replaced by L i (s − 1)T ηi for s = 1, 2, . . . , l.

The prediction of the plant output will be
l

l
l
ŷ i (k + l|k) = Cii Alii x̂ i (k|k) + Cii Aiis−1 Bii u i (k − 1) + Cii Al−h
ii Bii
s=1 s=1 h=s

l
L i (s − 1)T ηi + Cii Aiis−1 ŵi (k + l − s|k − 1) + v̂i (k + l|k − 1)
s=1
(15)
With this formulation, both predictions of state and output variables are expressed
in terms of the coefficient vector ηi of the Laguerre network, instead of u i as in the
classical approach. Thus, the coefficient vector ηi will be optimized and computed
in the control design. To compute the prediction, the convolution sum
l

l
Sci (l) = ii Bii L i (s − 1)
Al−h T
(16)
s=1 h=s
needs to be computed. To this end, note that
Sci (1) = Bii L i (0)T

Sci (2) = [Aii Bii + Bii ]L i (0)T + Bii L i (1)T
= [Aii + I ]Bii L i (0)T + Bii L i (0)T AliT
= Sci (1) + Aii Sci (1) + Sci (1)AliT (17)
Sci (3) = Sci (2) + Aii2 Sci (1) + Aii Sci (1)AliT + Sci (1)(Ali2 )T
Sci (4) = Sci (3) + Aii3 Sci (1) + Aii2 Sci (1)AliT + Aii Sci (1)(Ali2 )T
+Sci (1)(Ali3 )T
Continuing the recursion in (10) reveals that for l = 2, 3, . . . , p

⎧
⎨
l
Sci (l) = Sci (l − 1) + Aiih−1 Sci (1)(Ali(l−h) )T
(18)
⎩ h=1
with Sci (1) = Bii L i (0) T
The difference equation L i (k + 1) = Ali L i (k) is used for generating the set of
Laguerre functions, where Ali ∈ Mi ×Mi and is a function of the parameters ai and
βi = (1 − ai2 ).
(2) For MIMO subsystem Si , i = 1, 2, . . . , N
To extend the description to MIMO subsystem, with full flexibility in the choice of
ai and Mi parameters, let u i (k) be
n T
u i (k) = u i1 (k) u i2 (k) · · · u i ui (k)
and the input matrix be partitioned to

n
Bii = Bii1 Bii2 · · · Biiui
g
where n u i is the number of inputs and Bii for g = 1, 2, . . . , n u i is the g th column
g
of the Bii matrix. We express the g th control signal u i (k) by choosing a scaling
g g g g
factor ai and order Mi , where ai and Mi are selected for this particular input, such
that
g g g
u i (k) = L i (k)T ηi for g = 1, 2, . . . , n u i
g g
where ηi and L i (k) are the Laguerre network description of the g th control, namely
g g g g
L i (k)T = [li(1) (k) li(2) (k) . . . li(M g ) (k)]
i
Based on the partition of the input matrix and given the state variable information
at x̂ i (k), the prediction of the future state at time l can be written as
x̂ i (k + l|k) = Alii x̂ i (k|k) +

i (l)u i (k − 1) + i (l)T ηi + i (l) (19)
with
l

l
l

i (l) = Aiis−1 Bii , i (l)T = Al−h
ii Bii L i (s − 1)
T
s=1 s=1 h=s
l
i (l) = Aiis−1 ŵi (k + l − s|k − 1)
s=1
where the parameter vector ηi and the data matrix i (l)T consist of the following
individual coefficient vectors
nu
ηiT = [ηi1 ηi2 . . . ηi i ]T
l

l nu nu
i (l) =
T
ii [Bii L i (s)
Al−h 1 1 T
Bii2 L i2 (s)T ... Bii i L i i (s)T ]
s=1 h=s
for i = 1, 2, . . . , N and g = 1, 2, . . . , n u i .
Note that the kth block matrix

l
l
ik (l)T = ii Bii L i (s − 1)
Al−h k k T
s=1 h=s
has an identical structure as the single-input case defined by Sci (l), thus it can
be computed recursively using (18). From here on, the convolution sum in each
multi-input subsystem Si is decomposed into computing the subsubsystems, and the
computed results are put together block by block to form the multi-input structure.
It is worth stressing that in the formulation of the multivariable problem, the scaling
g g
factors ai and the number of terms Mi can be chosen independently for each input
signal of subsystem Si .
6.2 Modified-Optimizer
Under Assumptions 1, since the discrete Laguerre functions are orthonormal for
a sufficiently large prediction horizon p, the cost function (6) is equivalent to the
modified local cost function defined as

p
d
Ji = y (k + l|k) − ŷ i (k + l|k)2 + η T R L ηi (20)
i Qi i i
l=1
where R L i ∈ Mi ×Mi is a diagonal matrix with the weighting matrix Ri on its

diagonal. Therefore, the objective is to find the coefficient vector ηi that minimizes
Ji . Without constraints, by substituting (15) into (20), the optimal solution of the
parameter vector ηi is
p −1
opt

p
η̂i = ηi = i (l)Q i i (l) + R L i
T
×( i (l)Q i (yid (k + l|k)
l=1 l=1
− i (l) x̂ i (k|k) −
i (l)u i (k − 1) − i (l) − v̂i (k + l|k − 1))) (21)
with
i (l) = Cii Alii ,

i (l) = Cii
i (l)
i (l)T = Cii i (l)T , i (l) = Cii i (l)
Upon obtaining the optimal parameter vector ηi , the control increment at time k
is as follows
⎡ ⎤
L i1 (0)T 02T . . . 0nTu
⎢ i
⎥
⎢ 01T L i2 (0)T . . . 0nTu ⎥
u i (k) = ⎢
⎢ .. .. .. ..
i ⎥ηi
⎥ (22)
⎣ . . . . ⎦
nu
01T 02T . . . L i i (0)T
where 0rT , for r = 1, 2, . . . , n u i , represents a zero block row vector with identical
dimension of L ri (0)T .
Consequently, the control signal u i (k) can be calculated as
u i (k) = u i (k − 1) + u i (k) (23)
6.3 Modified Interaction Prediction
Under Assumptions 1, at step k, the predictions of the new interaction vectors are
given by
Ŵ i (k, |k − 1) = Ãi X̂ (k|k − 1) + B̃ i Ũ (k − 1|k − 1)

(24)
V̂ i (k, |k − 1) = C̃ i X̂ (k|k − 1)
F̃ i [Fi,1 . . . Fi,i−1 0 Fi,i+1 . . . Fi,N ]
where F̃ i ∈ { Ãi , B̃ i , C̃ i }.
7 NC-MPC Algorithm
For the ith sub-controller Ci , where the desired output yid (k + l|k) is provided by a
proper reference generator, the algorithm for the novel distributed MPC is outlined
in detail:
1. Set k = 1.
2. Acquire by network the predicted future state trajectories X̂ j (k|k −1) and control
inputs U j (k − 1|k − 1) from sub-controllers C j .
3. Build X̂ (k|k − 1) and U (k − 1|k − 1) by combining the local state trajectory
x̂ i (k|k − 1) and control input u i (k − 1|k − 1) with the acquired information, and
compute the corresponding predictions of the interactions according to (24).
4. Acquire the measures x̂ i (k) and the desired trajectory yid (k +l|k) over the horizon
p.
5. Compute the optimal control u i (k) and broadcast it by network to sub-controllers
C j , cf. (21), (22) and (23).
6. Apply the control input u i (k) to subsystem Si .

7. Compute the future state trajectory of subsystem Si over the horizon p and
broadcast it by network to sub-controllers C j , cf. (19).
Increment the sample time index k ← k + 1 and iterates by going to step 2.
In this section, the performance of the proposed NC-DMPC is investigated and

compared to the centralized MPC. Consider the following unstable-non minimum
phase plant S introduced in [12] discretized with a sampling time Ts = 0.2 s. A
state-space realization of S has the form (1), with the following matrices

A11 0 B11 0 C11 C12
A= ,B = ,C =
0 A22 0 B22 C21 C22
The examination of the process model S leads to decomposition into two intercon-
nected subsystems S1 and S2 . The corresponding state-space realizations of S1 and
S2 have the form (2), with matrices {A11 , B11 , C11 } and {A22 , B22 , C22 }, respectively.
The constant parameter α is used to study the effect of the interactions between S1
and S2 .
⎡ ⎤ ⎡ ⎤
2.859 −1.335 0.409 0 0.063
⎢ 2 0 0 0 ⎥ ⎢ ⎥
A11 = ⎢ ⎥, B11 = ⎢ 0 ⎥
⎣ 0 1 0 0 ⎦ ⎣ 0 ⎦
0 0 0 0.819 0.125
⎡ ⎤ ⎡ ⎤
0.819 0 0 0.125
A22 = ⎣ 0 1.637 −0.67 ⎦, B22 = ⎣ 0.250 ⎦
0 1 0 0

C11 = −0.0153 −0.029 −0.007 0 , C12 = α 0.145 0 0

C22 = 0 −0.193 0.292 , C21 = α 0 0 0 0.145
In the proposed test, the stability performances depend on the choice of the tuning
parameters α and p. The control performance of the resulting closed-loop system is
plotted in Figs. 2, and 3, where the black lines correspond to the desired outputs,
the red solid lines correspond to the system outputs using Centralized MPC (C-
MPC) based on Laguerre functions, and the blue solid lines represent the system
outputs using the proposed NC-DMPC. In Fig. 2, the performance of ND-DMPC is
comparable to the one of C-MPC. However, for strong interactions with α = 3.5,
Fig. 2 Control performance with α = 1, p= 10
Fig. 3 Control performance with α = 3.5, p = 10
Fig. 3 shows that the ND-DMPC strategy can achieve a better global performance
of the closed-loop system than C-MPC.
To sum up, for the given example, the NC-DMPC can achieve a satisfactory
global performance even whether the interactions among subsystems are strong or
not. Furthermore, the cost of computation is very small as compared with the classical
centralized controller.
9 Conclusion
In the present study, a formulation of a novel distributed model predictive control

for a class of large-scale systems is proposed, in which the whole system is divided
into many small scale subsystems interacting with each other by both their states
and inputs. The NC-DMPC solution, was proposed for improving the global perfor-
mance of closed-loop system; it is based on Laguerre functions used in MPC formu-
lation. The proposed methodology is demonstrated on an example for a set-point
tracking. Compared to the C-MPC scheme, the proposed NC-DMPC allowed to
achieve an improved performance of the whole system through using a local index
in optimization.
References
1. Christofides PD, Scattolini R, Muñoz de la Peña D, Liu J (2013) Distributed Model Predictive
control: A Tutorial Review and Future Research Directions. Comput Chem Eng 51:21–41
2. Ikeda M, Šiljak DD (1986) Overlapping decentralized control with input state and output
inclusion. Control Theory Adv Technol 2:155–172
3. Jia D, Krogh BH (2001) Distributed model predictive control. In: Proceeding of the American
control conference, arlington, pp 2767–2771
4. Lunze J (1992) Feedback control of large-scale systems. Prentice-Hall, London, UK
5. Maciejowski JM (2002) Predictive control with constraints. Prentice-Hall, London, U.K
6. Menighed K, Aubrun C, Yamé JJ (2009) Distributed state estimation and model predictive
control: application to fault tolerant control.In: th IEEE international conference on control
and automation, pp 936–94
7. Negenborn RR, Maestre JM (2014) Distributed model predictive control: an overview and
roadmap of future research opportunities. IEEE Control Syst 34(4):87–97
8. Patton RJ, Kambhampati C, Uppal FJ (2005) Challenges of networked control systems:
autonomy, reconfiguration and plug and play. In: Proceeding of 1st workshop on network
control systems and fault-tolerant control. Ajaccio, Corsica
9. Qian X, Yin Y, Zhang X, Sun X, Shen H (2016) Model predictive controller using laguerre
functions for dynamic positioning system. In: th Chinese control conference, pp 4436–4441
10. Razavinasab Z, Farsangi MM, Barkhordari M (2017) State estimation-based distributed model
predictive control of large-scale networked systems with communication delays. IET Control
Theory Appl 11(15):2497–2505
11. Vaccarini M, Longhi S, Katebi MR (2009) Ünconstrained networked decentralized model
predictive control. J Process Control 19(2):328–339
12. Vaccarini M, Longhi S, Katebi MR (2006) Stability analysis tool for tuning unconstrained
decentralized model predictive controllers. In: Proceedings of the international conference
control. Glasgow, Scotland, UK. August 2006
13. Vaccarini M, Longhi S, Katebi MR (2006) State space analysis of unconstrained decentral-
ized model predictive control systems. In: Proceeding of the American control conference.
Minneapolis, Minnesota, pp 159–164
14. Ventak AN, Hiskens IA, Rawlings JB, Wright SJ (2006) Distributed output feedback MPC for
power system control. In: Proceedings of the 45th IEEE conference on decision and control,
pp 4038–4045. San Diego, California
15. Ventak AN, Rawlings JB, Wright SJ (2005) Stability and optimality of distributed model
predictive control. In: Proceedings of the 44th IEEE conference on decision and control, and
the European control conference, pp 6680–6685
16. Wang L (2009) Model predictive control system design and implementation using MATLAB.
Advances in industrial control. Springer
17. Zečević AI, Šiljak DD (2005) A new approach to control design with overlapping information
structure constraints. Automatica 41:265–272
18. Zheng Y, Li S, Li N (2011) Distributed model predictive control over network information
exchange for large-scale systems. Control Eng Pract 19(7):757–769
19. Zheng Y, Li S, Qiu H (2013) Networked coordination-based distributed model predictive control
for large-scale system. IEEE Trans Control Syst Technol 21(3):991–998
Direct Torque Control Using Fuzzy
Second Order Sliding Mode Speed
Regulator of Double Star Permanent
Magnet Synchronous Machine
Louanasse Laggoun, Brahim Kiyyour, Ghoulemallah Boukhalfa,

Sebti Belkacem, and Said Benaggoune
Abstract The conventional Direct Torque Control (DTC) strategy using PI regula-
tors has certain disadvantages such as significant flux, torque ripples and sensitivity
to parametric variations. To overcome these drawbacks, we apply a new type with
more robust regulators such as the Second Order Sliding Mode Control (SOSMC)
based on Super Twisting algorithm and fuzzy second order sliding mode control.
This work deals with the modeling and performance improvement study of the DTC
of a Double Star Permanent Magnet Synchronous Machine (DSPMSM) using a
hybrid FUZZY-SOSMC speed regulator. The torque ripple, speed and currents will
be evaluated and compared by the classical PI-DTC and SOSMC-DTC. Simulation
results demonstrate the feasibility and validity of the proposed FUZZY-SOSMC-
DTC system by effectively accelerating system response, reducing torque and a very
satisfactory performance has been achieved.
Keywords Direct torque control (DTC) · Double star permanent magnet

synchronous machine (DSPMSM) · Second order sliding mode control (SOSMC) ·
Fuzzy logic control
1 Introduction
In recent years, variable speed drives consisting of an alternative machine asso-

ciated with a static converter, have attracted much attention from research groups
and industry. They are more and more present in the fields of high-power industrial
L. Laggoun
Department of Industrial Engineering, University Abbas Laghrour, LSTEB Laboratory,
Khenchela, Batna, Algeria
B. Kiyyour (B)
Department of Electrical Engineering, University of Biskra, Biskra, Algeria
G. Boukhalfa · S. Belkacem · S. Benaggoune
Department of Electrical Engineering, University of Batna2, LEB Laboratory, Batna, Algeria

https://doi.org/10.1007/978-981-15-6403-1_10
140 L. Laggoun et al.
applications. However, the constraints on the power components limit the switching
frequency, and consequently reduce its performances. To enable the use of higher
switching frequency components, the power should be divided to realize this, one of
the solutions is to use machines with a large number of phases (multi-phase (n > 3)
or (multi-star) [1–3].
One of the most common examples is the double-star permanent magnetic
synchronous machine (DS-PMSM), which is the most widely used in industry [3–5].
In recent years, several techniques have been developed to improve the performance
of these electrical machines. The Direct Torque Control (DTC) proposed by Depen-
brock and Takahashi in the middle of the eighties is a vector control solution [6]. It
was introduced especially for three-phase machines and several studies allowed to
apply this control technique on multi-phase machines.
Like every command, the DTC has advantages and disadvantages, it has to be less
dependent on the parameters of the machine (the stator resistance is theoretically the
only parameter of the machine, which intervenes in the command) and to provide
a faster response of the torque. Despite these advantages, this control scheme also
has significant disadvantages, the instability problem such as the lack of control of
the switching frequency of the inverter and the use of hysteresis bands generating
electromagnetic torque ripples and noise in the machine [1, 3, 7].
Several research works has discussed this subject, initially based on the control
principles set out in AL. Takahashi and giving rise to various developments of DTC
type strategies.
This structure of control is based on PI type controllers, these regulators suffer
from sensitivity to variations in the motor parameters. This requires a good identi-
fication of the parameters. To ensure the robustness and good performance of the
direct torque control using a PI controller, several approaches have been recently
proposed were adopted to change the PI controllers by other controllers [1, 8].
The sliding mode control (SMC) is a modern control strategy, with higher robust-
ness against load and parameter variations, faster response and higher level of energy
efficiency. It relies on fast switching, which made it difficult to implement it that time
[1, 8, 9].
Nevertheless, the classic sliding mode (standard sliding modes) has a signifi-
cant disadvantage, like the phenomenon of chattering. The main cause of chattering
phenomenon has been identified as the presence of unmodelled parasites in the
switching device [3]. These disadvantages have been minimized by the introduc-
tion of the higher order sliding mode control in the 80 s. The second order sliding
mode control (SOSMC) keeps the main advantages of standard sliding modes and
has the additional advantage that it can be used to remove chattering effect.
In fact, during the last decade, the fuzzy logic control (FLC) has been selected
as suitable control solution in the field of power electronics and drives. Among the
advantages provided by this control approach over the conventional controllers in
other hand it does not require accurate mathematical model.
It can thus work with inaccurate inputs, handle nonlinear model systems and easily
reach performances of PI controllers. Accordingly, this paper aims at combining the
Direct Torque Control Using Fuzzy Second Order Sliding Mode … 141
advantages of Fuzzy and SOSMC. This approach has been successful applied in wide
area [10].
In order to develop a robust DTC, current researchers have proposed the use of
a DTC with a speed controller based on hybrid fuzzy-Second order sliding mode
control to obtain a more flexible control. This last solution allowed the reduction or
even the attenuation of the chattering phenomenon while keeping the properties of
robustness [3, 11].
The strategy proposed in this work is the study of the dynamic behavior of double-
star permanent magnets synchronous machine (DSPMSM) controlled by a DTC
during a speed adjustment by conventional regulators (PI), SOSMC and by regulators
based on fuzzy-second order sliding mode. Simulation results reveal that the FUZZY-
SOSMC-DTC has a very robust behavior against the SOSMC-DTC and PI-DTC.
2 Modeling of the Double Star Permanent Magnet

Synchronous Machine
By applying the Park transformation to the model of the DSPMSM, the equations
are expressed in a reference frame linked to the rotating field as follows [12]:
d
Vds1 = R s1 i ds1 + Ψ ds1 − w Ψ qs1
dt
d
Vqs1 = R s1 i qs1 + Ψ qs1 − w Ψ ds1 (1)
dt
d
Vds2 = R s2 i ds2 + Ψ ds2 − w Ψ qs2
dt
d
Vqs2 = R s2 i qs2 + Ψ qs2 − w Ψ ds2 (2)
dt
And the expressions for stator flux are:
Ψds1 = L ds1 i ds1 + M ds2 i ds2 + Ψ P M

Ψqs1 = L qs1 i qs1 + M qs2 i qs2
Ψds2 = L ds2 i ds2 + M ds1 i ds1 + Ψ P M
Ψds2 = L ds2 i ds2 + M qs1 i qs1 (3)
For studying the dynamic behavior, the following equation of motion was added:
dΩ
J = Te − Tr − f r Ω (4)
dt
Fig. 1 Schematic isb1

representation of DSPMSM Sb1
Vsb1
Vsa2 isa2
windings isb2
Vsb2 ψ PM
Sb2
Sa2
θ Vsa1
isa1
γ=30°
Sa1
Vsc1
isc1
Vsc2
S c1 isc2
where: J is the moment of inertia, fr is the friction coefficient, Te is the electromagnetic

torque, Tr is the load torque and is the mechanical rotation speed of the rotor [3].
The model of the DSPMSM has been completed by the expression of the
electromagnetic torque Te given below [3]:
T e = p(Ψ ds1 i qs1 − Ψ qs1 i ds1 + Ψ ds2 i qs2 − Ψ qs2 i ds2 )

= p(Ψ P M ( i qs1 + i qs2 ) + (L d − L q ) − ( i ds1 i qs1 + i ds2 i qs2 )
+ (M d + M q )( i ds1 i qs2 + i ds2 i qs1 )) (5)
where: p is the number of pole pairs, ψ P M permanent magnet. The structure of the
DSPMSM is represented in the electrical space by Fig. 1.
3 Direct Torque Control (DTC) of the DSPMSM
The principle of the DTC is based on the direct determination of the command
sequences applied to the switches of a voltage inverter to deliver the stator voltage
vectors. These vectors are chosen from a selection table as a function of the flow and
torque errors as well as the position of the stator flow vector.
Two comparators control the status of system control variables, stator flux and
electromagnetic torque. The hysteresis corrector is the simpliest and best suited to
DTC. Its role is to maintain the error between the value to be set and its reference in
a hysteresis band. For a two-state controller, the choice of the voltage vector depends
only on the sign of the error and does not depend on its amplitude. However, a
hysteresis band is added around zero to avoid unnecessary switching when the flow
error is very small [13]. A two-level voltage inverter allows to have 7 distinct positions
in the phase plane, corresponding to the 8 voltage vectors of the inverter. These
positions are illustrated in Fig. 2.
Fig. 2 Voltage vector
Table 1 Vectors voltage localization

Cflx CTe 1 2 3 4 5 6 Comparator
1 1 V2 V3 V4 V5 V6 V1 Two levels Three
1 0 V7 V0 V7 V0 V7 V0 levels
1 −1 V6 V1 V2 V3 V4 V5
0 1 V3 V4 V5 V6 V1 V2 Two levels Three
0 0 V0 V7 V0 V7 V0 V7 levels
0 −1 V5 V6 V1 V2 V3 V4
Moreover, Table 1 presents the sequences corresponding to the position of the

stator flux vector to the different sectors (see Fig. 2). The flux and the torque are
controlled by two hysteresis comparators at 2 and 3 levels, respectively, in the case
of a two-level voltage inverter.
Direct torque control is based on the orientation of the stator flux. The expression
of the stator flux in the park frame of reference is described by [2, 3, 13]:
t
Ψ ds1.2 = (V ds1.2 − R s i ds1.2 )dt
0
t
Ψ qs1.2 = (V qs1.2 − R s i qs1.2 )dt (6)
0
where Vsd1.2 and Vsq1.2 are the estimated components of the vector voltage. They are
expressed from the model of the inverter. Thus, the stator flux module becomes:
Fig. 3 Block diagram of the proposed PI-DTC of DSPMSM

Ψ s1 = Ψ ds1
2
+ Ψ qs1
2

Ψ s2 = Ψ ds2
2
+ Ψ qs2
2
(7)
The electromagnetic torque can be estimated from the estimated magnitudes of

the flux, and the measured magnitudes of the line currents, by the equation below:
T̂ e = p(Ψ ds1 i qs1 − Ψ qs1 i ds1 + Ψ ds2 i qs2 − Ψ qs2 i ds2 ) (8)
he block scheme of the investigated DTC for a voltage source inverter fed DSPMSM
is presented on Fig. 3.
4 Second Order Sliding Mode Control of DSPMSM
The proposed control strategy is based on the Super Twisting Algorithm. This algo-
rithm is an exception that only requires information about the sliding surface [14–18].
The application of this control strategy begins with the determination of the relative
degree of the variable to be regulated. This variable is the speed, so we choose
a surface that is sufficient to make the command appear. We define the following
sliding surface:
S = ref − (9)
d Te − Tr − f r
= (10)
dt J
Te − Tr − f r
S∗ = r∗e f − ∗ = r∗e f − (11)
J
If we define the functions AΩ as follows:
1
A = r∗e f − ( f r − Tr ) (12)
J
∗∗
Thus: S t = A∗ − 1J T e∗
The second-order sliding mode controllers contain two parts:
Ter e f = T eeq + T esut (13)
where: T esut = I 1 + I 2
I 1∗ = −α sin g (S Ω )
1
I 2 = −λ|S Ω | 2 sin g (S Ω ) (14)
The corresponding sufficient conditions for finite time convergence to the sliding
manifold are:

α > KCm0
M (α+C 0 )
(15)
λ2 ≥ 4CK0 3K(α−C 0)
m
λ, α, K m , K M and C 0 are positive gains used to adjust the Super Twisting controller.
[3, 13]
5 Fuzzy Second Order Sliding Mode Direct Torque Control

(FUZZY-SOSMC-DTC)
SOSMC has proven itself in many studies and research applications of its effective-
ness in minimizing chattering effect is mainly caused by the presence of a discontin-
uous control term containing the sign function [19–21]. To improve the SOSMC-DTC
of the DSPMSM and more and more decrease the negative effect caused by the sign
function, Recently, fuzzy logic controllers (FL) have generated a lot of interest in
some applications. In this work we propose to employ a fuzzy-second order sliding
mode. FUZZY-SOSMC is a hybrid development of second order sliding mode control
and fuzzy logic control, where the switching controller term sign, has been replaced
by an inference fuzzy system [3, 22], membership functions are chosen to represent
Fig. 4 Membership
1 NB NM NS ZE PS PM PB
functions for e and ec
0.8
0.6
0.4
0.2
0
-1.5 -1 -0.5 0 0.5 1 1.5
Fig. 5 Membership NB NS ZE
1 NM PS PM PB
functions for Tem
0.8
0.6
0.4
0.2
0 -15 - 10 -5 0 5 10 15
the linguistic variables for the inputs and outputs of the controllers. in the FUZZY-
SOSMC the Eq. (14) becomes:
I 1∗ = −α f uzzy(S Ω )
1
I 2 = −λ |S Ω | 2 f uzzy(S Ω ) (16)
5.1 Fuzzification
The inputs to the Fuzzy-GA have to be fuzzified before being fed into the control
rule and gain rule determinations. The triangular membership functions (MFs) used
for the input (e, ec and, ΔT em ) are shown in Figs. 4 and 5.
5.2 Inference and Defuzzification
The present paper uses MIN operation for the calculation of the degree μ(ΔT em )
associated with every rule, for example, μ(ΔT em ) = Min[μ(e),μ(ec )].
In the defuzzification stage, a crisp value of the electromagnetic torque is obtained
by the normalized output function as:

m
μ T em j T em j
j=1
du = (17)

m
μ T em j
j=1
where: m is the total number of rules (7*7), μ(ΔT em ) is the membership grade
for the n rule, ΔT em is the position of membership functions in rule n in
U(−15,−10,−5,0,10,15).Fig. 4.
5.3 Control Rule Determination
The logic of determining this rule matrix is based on a global knowledge of the
system operation. As an example, we consider the following two rules:
if e is PB and ec is PB then Tem is PB
if e is ZE and ec is ZE then Tem is ZE
They indicate that if the speed is too small compared to its reference (e is PB),
so a big gain (Tem is PB) is required to bring the speed to its reference and if the
speed reaches its reference and is established (e is ZE and ec is ZE) so impose a
small gain Tem is ZE. (Table 2)
Figure 6 illustrates the general structure of the FUZZY-SOSMC-DTC of
DSPMSM.
6 Simulation Results and Discussion
In order to have a better appreciation of the results obtained through the direct
torque control DTC based on the three types of speed regulators PI and SOSMC
Table 2 Inference rules of fuzzy logic controller

Tem ec
NB NM NS ZE PS PM PB
e NB NB NB NB NB NM NS ZE
NM NB NB NB NM NS ZE PS
NS NB NB NM NS ZE PS PM
ZE NB NM NS ZE PS PM PB
PS NM NS ZE PS PM PB PB
PM NS ZE PS PM PB PB PB
PB ZE PS PM PB PB PB PB
Fig. 6 Block diagram of the proposed FUZZY-SOSMC-DTC of DSPMSM
and FUZZY-SOSMC applied to the DSPMSM, it is necessary to compare the static

and dynamic characteristics of the three speed controllers under the same operating
conditions (reference, disturbance loads) and in the same simulation configuration.
A series of numerical simulations under the MATLAB/Simulink environment have
been undertaken for the following two modes of operation:
– Operation with speed reversal.
– Operation with variation of the load torque.
6.1 Robustness Test for Speed Reversal
The purpose of this test is to validate the robustness of the controllers PI-DTC and
SOSMC-DTC and FUZZY-SOSMC for speed reversal. Figure 7 depicts the wave-
forms of the improved performances of speed control in the case of a no-load. It
can be noticed that the use of the FUZZY-SOSMC controller allows the speed to
judiciously follow its reference value of +100 to −100 rad/s at time t = 0.15 s and at
50 rad/s at t = 0.3 s. In fact, this behavior represents a clear improvement in dynamic
response with a FUZZY-SOSMC-DTC, contrary to a drive with a standard PI-DTC
and SOSMC-DTC.
Performance with each controller is also analyzed through these of Integral
Squared Error (ISE), Integral Absolute Error (IAE) and Integral Time Squared
Error (ITSE), and the results described in Appendix Table 3 confirm the improved
performance with the FUZZY-SOSMC-DTC.
120
PI
90 SOSMC
FUZZY-SOSMC
60
Speed (red/s)
30
-30
-60
-90
-120
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t(s)
150
100
100 20 90
80
Torque(N.m)
10
50
0,299 0,314
0
0 -20
-40
-50 -60 PI
-80 SOSMC
0.15 0.175 FUZZY-SOSMC
-100
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t(s)
30
20
Current iq1(A)
10
-10
-20
PI
SOSMC
-30 FUZZY-SOSMC
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

t(s)
Fig. 7 Comparison of the speed regulation of the PI- DTC, SOSMC-DTC and FUZZY-SOSMC-
DTC with speed reversal
Table 3 Comparison of
Controllers IAE ISE ITSE
Performance Index
PI-DTC 0.056 0.1331 3.33 × 10−3
SOSMC-DTC 0.0093 0.0125 5.375 × 10−5
FUZZY-SOSMC-DTC 0.0076 0.001503 4.856 × 10−5
6.2 Robustness Test for a Variation of the Load Torque
Figure 8 shows the simulation result (speed, the torque and the current isq1 ) during
a load torque setpoint variation of 10 N.m at time t = 0.12 s with 20 Nm and at
time t = 0.2 then 10 N.m at time t = 0.28 s and 0 N.m at t = 0.35 s. In this case
of simulation, we realize that the torque is perfectly following the setpoint value,
the speed reaches its reference after a small deformation for the case of PI-DTC
120
100 102.5
80 98
Speed (red/s)
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
60
40 0.2
20 PI
SOSMC
FUZZY-SOSMC
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t(s)
60 55
PI
50 SOSMC
50 FUZZY-SOSMC
45
40 40
Torque(N.m)
35
30
30
0 0.01
20
24
22
10 20
18
0 16
0.12 0.14 0.16
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t(s)
25
PI
SOSMC
20 FUZZY-SOSMC
Current iq1(A)
15
10
-5
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t(s)
Fig. 8 Comparison of the speed regulation with variation of the load torque, PI-DTC, SOSMC-DTC
and FUZZY-SOSMC-DTC
and with a negligible influence which is recovered quickly with its reference for
the SOSMC-DTC, the use of the FUZZY-SOSMC-DTC control presents a high
dynamic performance. Moreover, the tracking performances are improved by the
use of the FUZZY-SOSMC-DTC control, in comparison with others controllers. The
electromagnetic torque produced by the DSPMSM controlled by PI-DTC, SOSMC-
DTC and FUZZY-SOSMC-DTC is presented in Fig. 8, it can be noticed that the
ripple is not the same for the three techniques.
A zoom of the electromagnetic torque for the three strategies is shown. It’s clear
that the classical PI-DTC suffer from two problems: steady state error and high
torque ripple. Howere, the SOSMC-DTC and FUZZY-SOSMC-DTC corrects the
steady state error and reduces the torque ripple.
7 Conclusion
In this paper, a comparative study between the conventional PI-DTC, SOSMC-DTC

and FUZZY-SOSMC-DTC has been carried out for speed controller of DSPSM.
Simulation results reveal an improvement in the control performance of the torque
including ripple, steady state error reduction and a satisfactory performance with the
use of the FUZZY-SOSMC-DTC controller.
Furthermore, the effectiveness of the proposed algorithms is evaluated and justi-
fied from performance indices IAE, ISE and ITSE. So, this algorithm is suitable for
applications requiring a high tracking accuracy when external disturbances occur.
Appendix
(Table 4)
Table 4 DSPMSM
Parameter Symbol
parameters
Nominal voltage 250 V
Stator resistance Rs1 = Rs2 0.12
Stator inductance Ls 0.8 mH
Mutual inductance Lm 0.3 mH
Flux linkage ψPM 0.394 Wb
Pole pairs P 4
Machine inertia J 5 10−5 kg . m2
References
1. Naas B, Nezli L, Naas B, Mahmoudi MO, Elbara M (2012) Direct torque control based three
level inverter-fed double star permanent magnet synchronous machine. Sciverse Sci direct
Energy Procedia 18(1):521–530
2. Moubayed N (2009) Speed control of double stator synchronous machine supplied by two
independent voltage source inverters. Wseas Trans Syst Control 4(6): 253–258
3. Laggoun L, Youb L, Belkacem2 S, Benaggoune S, Craciunescu A (2018) Direct torque control
using second order sliding mode of a double star permanent magnet synchronous machine.
U.P.B Sci Bull Ser C 80(4): 93–106
4. Moubayed N, Bernard B (2009) Comparison between two double stator synchronous machine
supplying strategies. 7th Int Conf Electromech Power Syst, pp 143–147. Iasi, Romania. 8–9
October 2009
5. Dieng A, Benkhoris MF, Mboup AB, Aït-Ahmed M, Le Clairel JC (2016) Analysis of five-
phase permanent magnet synchronous motor. Revue Roumaine des Sciences Techniques - Serie
Électrotechnique et Énergétique 61(2):116–120
6. Boudjema Z, Taleb R, Djeriri Y, Yahdou A (2017) Novel direct torque control using second
order continuous sliding mode of a doubly fed induction generator for a wind energy conversion
system. Turk J Electr Eng Comput Sci 25(2): 965–975
7. Vivek D, Rohtash D (2011) Comparative study of direct torque control of induction motor
using intelligent techniques. Can J Electr Electro Eng 2(11): 550–556
8. Humod AT, Abdullah MN, Faris FH (2016) A comparative study between vector control and
direct torque control of induction motor using optimal controller. Int J Sci Eng Res 7(4):1362–
1371
9. Tria FZ, Srairi K, Benchouia KT, Mahdad B, Benbouzid M (2016) An hybrid control based on
fuzzy logic and a second order sliding mode for MPPT in wind energy conversion systems. Int
J Electri Eng Inf 8(4):711–726
10. Kairous D, Belmadani B (2015) Robust fuzzy-second order sliding mode based direct power
control for voltage source converter. (IJACSA) Int J Adv Comput Sci Appl 8(8): 167–175
11. Elawady WM, Lebda SA, Sarhan AM (2015) Continuous second order sliding mode control
with on line tuned pid. In: Proceedings of the world congress on engineering and computer
science, pp 864–869. San Francisco, USA
12. Benyoussef E, Meroufel A, Barkat S (2015) Three-level DTC based on fuzzy logic and neural
network of sensorless DSSM using extende kalman filte. Int J Power Electron Drive Syst
(IJPEDS) 5(4):453–463
13. Kiyyour B, Naimi D, Salhi A, Laggoune L (2019) Hybrid fuzzy second-order sliding mode
control speed for direct torque control of dual star induction motor. J Fundam Appl Sci
11(3):1440–1454
14. Bounasla N, Hemsas KE (2013) Second order sliding mode control of a permanent magnet
synchronous motor. In: The international conference on sciences and techniques of automatic
control & computer engineering, STA’2013, pp 535–539. Sousse. Tunisia, 20–22, Dec. 2013
15. Meghni B, Dib D, Azar AT (2017) A second-order sliding mode and fuzzy logic control to
optimal energy management in wind turbine with battery storage. Nat Comput Appl 28(6):
1417–1434
16. Guangping Zh, Hostettler JD, Patrick G, Wang X (2016) Robust sliding mode control of perma-
nent magnet synchronous generator-based wind energy conversion systems. Sustainability
8(12):1–20
17. Hostettler J, Wang X (2015) Sliding mode control of a permanent magnet synchronous generator
for variable speed wind energy conversion systems. Syst Sci Control Eng: Open Access J
3(1–3): 453–459
18. Hafiane M, Sabor J, Taleb M (2017) Optimal speed control based on adaptive second order
sliding mode and modified Hsc Mppt algorithm for wind turbine. ARPN J Eng Appl Sci
12(21):5891–5902
19. Boudjema Z, Taleb R, Yahdou A (2016) A new DTC scheme using second order sliding mode
and fuzzy logic of a DFIG for wind turbine system. (IJACSA) Int J Adv Comput Sci Appl 7(8):
49–56
20. Yuan X, Chen Z, Yuan Y, Huang Y (2015) Design of fuzzy sliding mode controller for hydraulic
turbine regulating system via input state feedback linearization method. Energy Elsevier. 93(1):
173–187
21. Ullah N, Shaoping W, Khattak MI, Shafi M (2015) Fractional order adaptive fuzzy sliding mode
controller for a position servo system subjected to aerodynamic loading and nonlinearities.
Aerosp Sci Technol 43(6): 381–387
22. Ardjoun SAEM, Abid M (1686) Fuzzy sliding mode control applied to a doubly fed induction
generator for wind turbine Turk J Electr Eng Comput Sci 23(6): 1673–1686
Fractional Order Integral Controller
Design Based on a Bode’s Ideal Transfer
Function: Application to the Control
of a Single Tank Process
Chahira Boussalem, Rachid Mansouri, Maamar Bettayeb,

and Mustapha Hamerlain
Abstract The paper deals with the water level control of a single tank process
associated feedback linearization (FL) with a novel fractional order integral controller
(FOIC) which is based on the Bode’s ideal transfer function. The first controller is
used to cancel the nonlinearities of the single tank process and the latter used to solve
the tracking problem. The new controller is implemented on a single tank process
and the results are compared with a Linear Quadratic Regulator (LQR).
Keywords Fractional calculus · Nonlinear systems · Single tank · Feedback

linearization · Fractional integral controller · Bode’s ideal transfer function
1 Introduction
The first references to the Fractional calculus refer to Leibniz and L’hopital in 1695,
where a differentiation of order 0.5 was discussed. This theory attracted increasing
interests in the world and research is still underway in all fields, especially in auto-
matic control for modeling of physical systems, stability questions and control.
C. Boussalem (B) · R. Mansouri

Automatic Department, University of Tizi-Ouzou, Tizi-Ouzou, Algeria
R. Mansouri
C. Boussalem · M. Hamerlain
Productive and Robotics Division, Center for Development of Advanced Technologies, Algiers,
Algeria
M. Bettayeb
Center of Excellence in Intelligent Engineering Systems, King Abdulaziz University, Jeddah,
Saudi Arabia
Electrical Engineering Department, University of Sharjah, Sharjah, United Arab Emirates

https://doi.org/10.1007/978-981-15-6403-1_11
156 C. Boussalem et al.
Designing and tuning fractional controllers appear to be very important for their
superior results in control. In fact, many methods have been developed. In [1, 2],
the authors developed the CRONE (Commande Robuste d’Ordre Non Entier) and
the fractional order PID controllers and demonstrated the superior performance and
advantage of both controllers over the classical PID controller. An optimization
method to tune the fractional PI controller based on a nonlinear function minimiza-
tion was used in [3]. The authors also show better performance of this controller
in comparison with a conventional PI controller. In [4], A fractional order fuzzy
control is developed and its parameters are tuned with a particle swarm optimization
algorithm. This controller shows better performance over the classical PID, and the
integer order fuzzy PID controller.
Most physical systems are nonlinear therefore their control attracted attention of
many researchers around the world. Analysis and control of nonlinear systems are in
general extremely difficult problems in both integer and non integer control, where
we find several control methods based on the linearized models: some methods used
the usual Jacobian linearization [5, 6], and some other used feedback linearization
[7–9]. In the fractional control, several works have been addressed for validating the
developed control strategies through the experimental setup. Recentlly, in [10, 11], the
authors proposed a fractional PI-state feedback controller for a linearized system by
using the pole placement technique and it is implemented on an inverted pendulum-
cart system to stabilize the pendulum angle and the cart position to the setpoint zero.
A novel control structure based on the state feedback with fractional integral control
is proposed in [12] if the setpoit is not zero. The controller is also implemented on
an inverted pendulum-cart system. The aim of this paper is to develop a method of
determining the integral gain of the fractional controller in SISO linearized system
(transfer function of a simple integrator) to solve the output tracking problem and
to validate the method by an experimental setup of a single tank system. We use the
feedback linearization (FL) technique to cancel the nonlinearities of the system and
a fractional order integral controller (FOIC) to the tracking. The proposed method is
based on the Bode’s ideal transfer function. The latter has very interesting properties.
The first property refers to the location of the poles for values of γ in the range [1, 2]. In
fact, the poles are complex and the step response have the behavior of an underdamped
second order system with the two important parameters such as overshoot and time
constant which depend directly on the values of the fractional order γ and the gain
crossover frequency ωc , respectively. The second property can be exploited in the
design of the integral gain kI of the fractional integral controller. Another interesting
property is the iso-damping robustness property [13]. The closed-loop systems is
represented as an ideal closed-loop system whose open-loop is given by Bode’s
ideal transfer function. Several control techniques use this type of transfer function
[13–15]. Many works have shown that fractional systems presents best qualities [13]
and very interesting structure for its implementation [12, 15].
Fractional Order Integral Controller Design Based … 157
Single or multiple tanks systems are featured as a bench-mark problem in the

category of nonlinear control systems. Control of liquid level in tanks and fluid flow
between tanks is a fundamental requirement in almost all process industries such as
waste water treatment, chemical, petrochemical, pharmaceutical and beverages.
Single Tank System is used for the experimental verification of designed
controllers. The system is found to be very simple to operate and it is a digital
control system which works on MATLAB/Simulink platform. A simulation study
of fractional order model reference adaptive control applied to level control in a
conic tank was presented in [16]. In [17], modelling and designing of controllers on
real time single tank system and simulation were presented. Three controller tuning
methods such as Relay Auto-Tunning, Ziegler-Nichols, and Tyreus-Luyben are used.
A simulation and experimental results of PI/PID controllers of integer order as well
as of fractional order for level control on a conical tank have been presented in [18].
The contribution of this paper is twofold. We first present a new fractional integral
control strategy based on the Bode’s ideal transfer function for a single tank system
to solve the problem of tracking. Second, we show the effectiveness of the proposed
controller by simulation, experimental results and comparison with an optimal Linear
Quadratic Regulator (LQR).
The rest of the paper is organized as follows: Sect. 2 describes the single tank
system and its model. Section 3 describes the input-output feedback linearization
theory and a fractional integral controller design. Simulation and experimental results
using the proposed controller, and the experimental results using the LQR controller
applied to single tank system are carried out in Sect. 4. We conclude with Sect. 5.
2 Single Tank Modeling
A. Description of Experimental Setup
The experimental setup consists of four cylindrical tanks of uniform cross sections for
water circulation, and a rectangular tank for water storage. It can be configured into
several types of experiments (SISO, MIMO, SIMO, MISO). In this paper only one
tank is considered as shown in Fig. 1. The latter consists of a tank of cross Section A
for water circulation, and a rectangular tank for water storage. The water level in tank
is measured by a pressure sensor. Water is pumped from the storage tank to the tank
by submersed pump. The water flows freely to the storage tank through the orifices.
The system is controlled via the computer and Matlab has been used to simulate the
model. An advantech peripheral communication interface (PCI1711) card is used
for interfacing between the plant and the controller. The water level information is
transferred to the PC via power supply unit and power amplifier (PSUPA) [17, 19].
Fig. 1 Schematic diagram of a single tank experimental setup
B. Single Tank Modeling

As shown in Fig. 2, the water is pumped from the storage tank to tank by pump via
valve MVB providing a flow Q0 . The water is flowing out of the tank through an
orifice of area a with a flow Q1 . The input is the voltage u to the pump, and the output
is the water level H.
Let V be the volume of liquid in the tank, the mass balance is given by:
d V (t) d H (t)
=A = Q0 (t) − Q1 (t) (1)
dt dt
Consider the Bernoulli’s law which defines the outflow rate as the product between
the outlet cross sectional area a and the outflow velocity v. The outflow rate from
tank is given by:

Q1 (t) = av(t) = a 2g H (t) (2)
Fig. 2 Schematic diagram of a single tank system

The flow rate generated by pump is given by:
Q 0 (t) = ηu(t) (3)
where, η is a constant relating the control voltage with the water flow from pump.
Substituting (2) and (3) into (1), we obtain a nonlinear state equation which describe
the system dynamics of the single tank process:
d H (t) a
= ηu(t) − 2g H (t) (4)
dt A
Defining water level x = H(t) of tank as state variable, the voltage u = u(t) to the
pump as the input to the process and y = H(t) as the output, a nonlinear state space
model of the single tank system can be obtained as follows:
d x(t) √
dt
= − aA 2gx(t) + ηu(t)
(5)
y(t) = x(t)
The single tank parameters and variables are: A = 0.01389 is the cross-sectional
area of the tank (m2 ), a = 50.265.10−6 is the outlet area at tank (m2 ), η = 2.4.10−3
is the constant relating the control voltage with the water flow from the pump, Q0
is the input flow rate of the tank (m3 /s), Q1 is the output flow rate of tank (m3 /s), u
is the voltage applied to Pump (V), H is the water level of tank (m), g = 9.81 is the
gravitational acceleration (ms−2 ) [17].
3 Input-Output Feedback Linearization and Fractional

Integral Controller Design
A. Input-Output Feedback Linearization
The theory of the input-output feedback linearization (IOFL) has been presented in
[20, 21]. In this subsection, our goal is the application of the IOFL to the single
tank process (5). For this, the following preliminaries are given. Consider a SISO
nonlinear system given by:

ẋ = f (x) + g(x)u
(6)
y = h(x)
where x ∈ Rn is the state vector, u ∈ R is the control input, y ∈ R is the output, f and
g are smooth vector fields, and h is the smooth scalar function. The basic approach
of input-output Feedback linearization is simply to differentiate the output function
y repeatedly until the input u appears, and then design u to cancel the nonlinearity
[9]. The r time derivatives of y can be expressed as
⎡ ⎤ ⎡ ⎤
ẏ L f h(x)
⎢ ⎥ ⎢
ÿ L 2f h(x) ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥=⎢
.. .. ⎥ (7)
⎣ ⎦ ⎣. . ⎦
r −1
y r
L f h(x) + L g L f h(x)u
r
where Lf h is the Lie derivative of h with respect to f. The number r of differentiations

required for the input u to appear is called the relative degree of the system [21].
The control law:
L rf h(x) 1
u = α(x) + β(x)υ = − + υ (8)
L g L rf−1 h(x) L g L rf−1 h(x)
applied to y r = L rf h(x) + L g L rf−1 h(x)u yields the simple linear relation:
yr = υ (9)
In this form, the system is readily seen to be completely controllable and thus can
be stabilized using a linear feedback.
In the following, the input-output linearization
√ concept is applied to the single
tank nonlinear system (5), where, f (x) = − aA 2gx(t), g(x) = η and h(x) = x.
We consider the voltage u to the pump as the control input and we want the
tracking of the desired output Href . The relative degree of the single tank system is
r = 1. The first time derivatives of y is:
a
ẏ = L f h(x) + L g h(x)u = − 2gx(t) + ηu(t) (10)
A
By using (8) the control law is given by
a√
2gx(t) 1
u(x) = A
+ υ (11)
η η
By replacing u in (10), the system can be transformed into a simple linear system:
ẏ = υ (12)
B. Fractional Integral Controller Design and Stabilization of the Single Tank

Process
Design of fractional-order controllers has been the subject of many researchers for
their superior results in control [1–3]. In this paper, we use the fractional integral
controller to control a system with a pure integrator. The fractional aspect of the
control law give an ideal closed-loop system whose open-loop is given by the Bode’s
ideal transfer function.
• Bode’s ideal transfer function:
Bode [22] has suggested an ideal shape of the open-loop transfer function of the
form:

ω γ
, γ ∈ R+
c
L(s) = (13)
s
where ωc is the gain crossover frequency which satisfies |L(jωc )| = 1. The parameter
γ is the slope of the magnitude curve, on a log-log scale, and may assume integer and
non-integer values. In fact, the transfer function L(s) is a fractional order differentiator
for γ < 0, and a fractional order integrator for γ > 0.
In the Bode diagrams of L(s) (1 < γ < 2), the amplitude curve is a straight line
of constant slope −20γ db/dec, and the phase margin of L(s) is π − γπ/2 rad. That
means the phase margin will not change along with the change of ωc .
The closed-loop system whose open-loop transfer function is L(s) is robust to
gain variations and step responses exhibiting an iso-damping property [13, 23].
The closed-loop transfer function of the Bode’s ideal transfer function is
ωc γ
L(s) 1 (ωc )γ
F(s) = = ωc sγ =
γ = γ (14)
1 + L(s) s
+1 s
+1 s + (ωc )γ
ωc
where γ ∈ R + , and 1/ωc is the time constant of the system.

For 1 < γ < 2, the poles of F(s) in the s-plane are given by
s1,2 = ωc e± jπ/γ (15)
A detailed study of this function is given in [13, 23].

• Fractional Order Controller Design
Consider the linearized model of the single tank system described by (12) and the
fractional integral controller of the form

υ(t) = k I Iα Hr e f (t) − y(t) (16)
where kI ∈ R1×1 , is the designed gain to achieve a desired closed-loop characteristic

polynomial. Iα is the non-integer integration operator of order α. Href is the desired
setpoint.
Fig. 3 Block diagram of input-output feedback linearization with fractional order integral controller
Let e(t) = Href − y(t) be the error between the desired setpoint and the output.
Figure 3 shows the block diagram of the input-output feedback linearization with the
fractional integral controller applied to the single tank process.
The result of the fractional order integral controller design is given by the following
proposition.
Proposition Let’s consider an integrator pur system described by (12) and the control
law given by (16).
The gain kI that can achieve a desired closed-loop characteristic polynomial as
that of the Bode’s ideal transfer function written in the form d (s) = sγ + ωc γ is
given by kI = ωc γ.
Proof Using the control law (16), the resulting closed loop transfer function is
obtained from (12) and (16) by:
y(s) kI
= α+1 (17)
Hr e f (s) s + kI
This closed loop transfer function is similar to the closed loop transfer function of
the Bode’s ideal transfer function (14) (with γ = α + 1), witch is involves two tuning
parameters. The first one is the fractional order γ and the second is the integral gain
kI . The tuning of both parameters consists in the choice of the closed loop reference
specifications such as the gain crossover frequency and the phase margin.
The fractional order γ = α + 1 is deduced by the formula phase margin = π −
γπ/2 rad.
Let ωc and γ be the desired gain crossover frequency and the fractional order of
the closed loop reference respectively. The desired characteristic polynomial of the
closed-loop system is then
d (s) = s γ + ωcγ (18)
The identification term by term between the characteristic polynomial of the

closed loop transfer function (17) and the desired characteristic polynomial d (s)
(18), give:
k I = ωcγ (19)
H ref
0.14
H: γ = 1.1
H: γ = 1.33
0.12 H: γ = 1.5
H: γ = 1.66
Water level (m)

0.1
0.08
0.06
0.04
0.02
0
0 2 4 6 8 10 12 14
Time (s)
60 γ = 1.1
γ = 1.33
50 γ = 1.5
γ = 1.66
40
Control (V)
30
20
10
0
-10
0 2 4 6 8 10 12 14
Time (s)
Fig. 4 Evolution of the output H and control u for a step reference and for several values of order
γ = 1.1, 1.33, 1.5, 1.66, using FL combined with FOIC
Table 1 Fractional order

γ ωc = 1.75 rad/s
integral controller parameters
kI α
1.1 1.8507 0.1
1.33 2.1049 0.33
1.5 2.315 0.5
1.66 2.5319 0.66
ωc (rad/s) M = 60◦ (γ = 1.33)
kI α
0.25 0.1582 0.33
0.75 0.6821 0.33
1.25 1.3455 0.33
1.75 2.1049 0.33
The following algorithm presents the steps for the design of the fractional order
integral controller:
• Apply the Input-Output Feedback Linearization method presented in Sect. 3-A to
the SISO nonlinear system to get a simple linear model with a pure integrator.
• Choose the desired phase margin and the gain crossover frequency of the desired
closed-loop transfer function. Deduce the fractional order γ, where γ = α + 1.
We only take (1 < γ < 2) into consideration which realizes the closed loop transfer
function of the Bode’s ideal transfer function.
• Deduce the fractional order α, (0 < α < 1) of the controller (16), and then the
feedback gain kI by using (19).
4 Simulation and Experimental Results
To evaluate the performance of the proposed control experimentally, we first provid

a series of simulations in the case of step and square references to single tank system
and then an implementation experimentally.
0.12
0.1
H ref
Water level (m)
H: ω c = 0.25
0.08
H: ω c = 0.75
H: ω c = 1.25
0.06
H: ω c = 0.75
0.04
0.02
0
0 5 10 15 20 25 30 35
Time (s)
50
ω c = 0.25
45 ω c = 0.75
40 ω c = 1.25
35 ω c = 1.75
Control (V)
30
25
20
15
10
5
0
0 5 10 15 20 25 30 35
Time (s)
Fig. 5 Evolution of the output H and control u for a step reference and for several values of the
gain crossover frequency ωc = 0.25, 0.75, 1.25, 1.75, using FL combined with FOIC
A. Simulation Results
Consider the linear model of the single tank system given by (5). We use steps 2 and
3 of the algorithm summarized in Sect. 3-A. We use one test for square reference and
two different tests for the step reference. For the latter, in the first test, we choose a
fixed value of the gain crossover frequency ωc = 1.75 and we vary the phase margin
as M = 30°, 45°, 60°, 81° or the fractional order γ as γ = 1.66, 1.5, 1.33, 1.1. In the
Second test, we use a fixed value of the fractional order γ = 1.33 (M = 60°) and we
vary the gain crossover frequency ωc as ωc = 0.25, 0.75, 1.25, 1.75. The controller
parameters are listed in Table 1. Simulation results of those tests are shown in Figs. 4
and 5 respectively. The system responses and control for a square reference signal
for the proposed controller and by using a fixed value of the fractional order γ = 1.1
(M = 81°) and a fixed value of the gain crossover frequency ωc = 1.75 are shown
in Fig. 6.
From Figs. 4, 5 and 6, the output H tracks the setpoint Href with different over-
shoots and almost constant rise time by varying the fractional order γ and for a
fixed coefficient ωc . In the other hand, by varying the coefficient ωc and maintaining
0.14
H ref
0.12 H: γ = 1.1
0.1
Water level (m)
0.08
0.06
0.04
0.02
0
0 100 200 300 400 500 600
Time (s)
45
40
35
30
Control (V)
25
20
15
10
5
0
0 100 200 300 400 500 600
Time (s)
Fig. 6 Evolution of the output H and control u for a square reference signal using FL combined
with FOIC
a fixed order γ, the output H tracks the setpoint Href with different rise times and
a constant overshoot. The dynamic of the system changes according to the choice
of the gain crossover frequency ωc and the fractional order γ. This shows that the
output tracking is achieved and the stability and convergence of the error to zero is
guaranteed.
A. Experimental Results
To show the pratical effectiveness of the proposed controller, performances and
robustness are tested experimentally to single tank process and results are compared
with a robust and optimal Linear Quadratic Regulator (LQR).
For the fractional controller, the fractional order γ or phase margin are chosen as:
γ = 1.1 (M = 81°) and a coefficient ωc is chosen as ωc = 1.75.
For the LQR controller, the weighting matrix Q and the weighting factor R are
chosen as 10 and 0.001 respectively.
Figure 7 shown the experimental results of the system responses and control for
a square reference for both controllers.
From the experimental results shown in Fig. 7, it can be observed that both types
of controllers can guarantee the closed-loop water level stability and realize well
setpoint tracking.
15
H ref
H: FL & FOIC: γ = 1.1

H: FL & LQR
Water level (cm)
10
0
0 50 100 150 200 250 300 350 400 450
Time(s)
6 FL & FOIC: γ = 1.1

FL & LQR
4
Control (V)
0
0 50 100 150 200 250 300 350 400 450
Time (s)
Fig. 7 Experimental results of the water level H and control u in single tank system for a square
reference signal using FL combined with FOIC and LQR controller
15
Water level (cm)

10
H ref
H: FL & FOIC
H: FL & LQR
5
0
0 50 100 150 200 250 300 350 400 450
Time (s)
6
FL & F0IC: γ = 1.1
FL & LQR
4
Contrl (V)
0
0 50 100 150 200 250 300 350 400 450
Time (s)
Fig. 8 Experimental results of the water level H and control u in single tank system for a square
reference signal with disturbance at time t = 150 s and t = 250 using FL combined with FIC and
LQR controller
For the proposed control, the dynamics of the system changes according to the
choice of the gain crossover frequency ωc and the fractional order γ and changes
according to the choice of the weighting matrix Q and the weighting factor R for LQR.
In order to test the robustness of both controllers, we first added one liter of water
in tank 1 at t = 150 s, and then valve MV1 opened at t = 250 s during 20 s. The exper-
imental results for square reference are shown in Fig. 8. This figure shows the robust-
ness of both controllers. Indeed, after adding one liter of water in tank 1 at t = 150 s,
the water level in FOIC and LQR increase instantly to 12.16 cm, and then it is rejected
quickly and reaches the reference value. We also see that after opening valve MV1
at t = 250 s, the water level in FOIC and LQR decrease to almost 9.7 cm and 9.2 cm,
respectively and then it is rejected and reaches the reference value.
5 Conclusion
In this paper, FL with a novel FOIC for setpoint tracking is presented and successfully
implemented to a single tank process. From the simulation and experimental results,
it is shown that the proposed controller is very effective. This shows the feasibility
of the designed controller in practice.
From the comparison of the experimental results, both controllers have a good
performance regarding robustness and tracking effect.
References
1. Oustaloup A (1991) La commande CRONE. Herme‘s, Paris

2. Podlubny I (1999) Fractional-order systems and P I α Dα controllers. IEEE Trans Aut Control
44:208–214
3. Monje CA, Calderoń J, Vinagre BM, Chen YQ (2004) On fractional P I λ controllers: Some
tuning rules for robustness to plant uncertainties. Nonlinear Dyn 38: 369–381
4. Pan I, Das S (2016) Fractional order fuzzy control of hybrid power system with renewable
generation using chaotic PSO. ISA Trans 62:19–29
5. Owczarkowski A, Horla D, Zietkiewicz J (2018) Introduction of feedback linearization to
robust lqr and lqi control analysis of results from an unmanned bicycle robot with reaction
wheel. Asian J Control 21:1–13
6. Grognard F, Sepulchre R, Bastin G (1999) Global stabilization of feedforward systems with
exponentially unstable Jacobian linearization. Syst Control Lett 37:107–115
7. Noah DM, Laheeb M, Roger CF (2017) Using feedback linearization to improve the tracking
performance of a linear hydraulic-actuator. J Dyn Syst Meas Control
8. Moradi H, Hajikolaei KH, Motamedi M, Vossoughi GR (2009) Motion control of a
nonlinear single-rod hydraulic actuator via feedback linearization In: ASME 2009 international
mechanical engineering congress and exposition, American society of mechanical engineers
9. Cesáreo R, Alejandro FV, Antonio B (2014) Adaptive tracking in mobile robots with input-
output linearization. J Dyn Syst Meas Control 136
10. Bettayeb M, Boussalem C, Mansouri R Al Saggaf UM (2014) Stabilization of an inverted
pendulum cart system by fractional PI State feedback. ISA Trans 53: 508–516
11. Mansouri R, Bettayeb M, Boussalem C, Al Saggaf UM (2015) Linear integer order system
control by fractional PI-state feedback. Chapter In: RAZ Daou, X Moreau (eds) Fractional
calculus applications
12. Al-Saggaf UM, Mehedi IM, Mansouri R, Bettayeb M (2015) State feedback with fractional
integral control design based on the Bode’s ideal transfer function. Int J Syst Sci 1–13
13. Barbosa R, Machado JAT, Sneddon IN (2004) Tuning of PID controllers based on Bode’s ideal
transfer function. Nonlinear Dyn 305–321
14. Amoura K, Mansouri R, Bettayeb M, Al-Saggaf UM (2016) Closedloop step response for
tuning PID-fractional-order-filter controllers. ISA Trans 64: 247–257
15. Bettayeb M, Mansouri R (2014) IMC-PID-Fractional order filter controller design for integer
order systems. ISA Trans 53:1620–1628
16. Balaska H, Ladaci S, Zennir Y (2018) Conical tank level supervision using a fractional order
model reference adaptive control strategy. In: 15th International conference on informatics in
control, automation and robotics (ICINCO). Portugal
17. Jagnade SA, Pandit RA, Bagde AR (2015) Modeling, simulation and control of flow tank
system. Int J Sci Res 657–669
18. Jáuregui C, Duarte-Mermoud MA, Oróstica RA, Travieso-Torres JC, Beytía O (2016) Conical
tank level control using fractional order PID controllers: a simulated and experimental study.
J Control Theory Technol 14: 369–384
19. Mahapatro SR (2012) Control algorithms for a two tank liquid level system: an experimental
study. Thesis for the award of master of technology by research in electrical engineering. India.
(2012–2014)
20. Isidori A (1995) Nonlinear control system, 3rd edn. Springer, Berlin
21. ESlotine JJ, Li. Weiping (1991) Applied nonlinear control. Prentice Hall
22. Bode HW (1945) Network analysis and feedback amplifier design. Van Nostrand, New York
23. Liu L, Zhang S (2018) Robust fractional-order PID Controller tuning based on Bode’s optimal
loop shaping. Hindawi 2018:1–14
Modeling of Torsional Vibrations
Dynamic in Drill-String by Using
PI-Observer
R. Riane, M. Kidouche, M. Z. Doghmane, and R. Illoul
Abstract Torsional vibrations, appeared in drill string of oil well, are one of the
major cause of drilling failures, in most of cases the penetration rate reduction and
drilling costs expansion. The driller’s intervention is basically based on parame-
ters manipulation to mitigate torsional vibrations, precisely variation of the angular
velocity of the Top drive. The weight on bit variation is also proven to be practically
effective in the reduction of such vibration but with a considerable time delay. In
this paper, we modeled the nonlinear dynamic of the down hole using a PI Observer,
thus, we estimate the downhole angular speed and perceive the torsional vibrations
as soon as they appear so that the time delay is minimized. The obtained model is
useful to design robust controller to mitigate the torsional vibrations in real-time.
Keywords Torsional vibration · Drill string · Top drive · PI observer · Real-time
1 Introduction
Rotary drilling systems are designed to drill a hole on different layers of the earth to
reach its target called reservoir. They are generally composed of two main important
parts: mechanical part that ensures the propagation of energy delivered by the top
drive to the bit, and hydraulic part to take material from the bottom hole to the surface
and cool the bit. The mechanical system contains a Top drive to rotate the drillstring,
the latter is composed of many pipes and it transfers the energy to the bit which cuts
the rocks and creates the borehole. The hydraulic system ensures the circulation of
drilling fluid, by using pumps and transport channels. During drilling process, the
system can face many types of vibrations, they can be categorized according the
direction of their propagation, we can find: torsional, axial, and lateral [1, 2], they
R. Riane (B) · M. Kidouche · M. Z. Doghmane

Laboratory of Applied Automatic, University M’hamed Bougara, Boumerdes, Algeria
R. Illoul
Department of Automation, National Polytechnic School, Algiers, Algeria

https://doi.org/10.1007/978-981-15-6403-1_12
172 R. Riane et al.
can appear at the same time or separately. Drilling systems, considered in this study,
contain drag bits; it contains a fixed blades and cutter at the biggest surface of bit
body. This type of bit configuration in most of cases leads to torsional vibrations for
which the angular velocity of the bit starts decreasing till it completely stop (stick
phase) the cumulating energy will, after some threshold, free the tool that slips with
angular velocity much higher than the angular velocity of the top drive (slip phase).
This stick-slip phenomenon affects the drilling efficiency, the early damage of drag
bits, and may cause system failure.
In this study, the main objective is to model the nonlinear dynamic of the drillstring
under torsional vibration in order to detect as quickly as possible these vibrations and
mitigate them in real-time. The torsional pendulum equivalent system with different
degree of freedom has been considered to describe mathematically the drilling system
[3, 4]. Like [3, 5, 7, 8], a two degree of freedom mathematical model is consid-
ered. Besides the reliability of the system, a PI observer is proposed to estimate the
nonlinearity with the unmeasured states and inputs.
2 Drilling System Description
Figure 1 shows the main parts of drilling system. One of the most important compo-
nents is the drillstring, it is of drill pipes that are used to guarantee the circulation
of the mud using pumps, and transmission of torque generated by the top drive to
the drill bit. In addition to drill pipes, the drillstring also contains 9 m slender tubes
jointed by threaded connections. The drillstring bottom end is called the borehole
assembly (BHA), it involves a thick walled tubes called drill collars used to avoid
buckling of the drillstring [1]. At the BHA bottom end, there is the drilling tool,
named the bit. Typical bit design is a steel body Traditional bits consist of a steel
body set with three rotating conical cylinder, it contains teeth made of tungsten that
cuts the rocks. The cutters of the bit contained the fixed steel body and enclosed by
diamond tooth that can cut the rock, they have been used in rotary drilling since early
the 19th. The bit diameter can vary from 0.1 to 0.9 m, it depends on the diameter of
the section, and the deeper sections are drilled with smaller bit diameters [1].
The top drive is generally composed of an AC/DC electric motor, a rotary table and
a gearbox. The role of the Kelly is to transmit the torque from the electric motor to the
drillstring through the rotary table. The newest rigs are supported by high technology
top drive that contains multi-functional unit providing the rotary movement of the
drillstring. As a result of the importance of the TD in oil drilling industry, many oil
equipment producers have launched their project for developing new generation of
the automated TDs. Such projects may lead to new top drive motors that can generate
high torques at low speed automated based on the torque on bit real-time variation.
Modeling of Torsional Vibrations Dynamic … 173
Fig. 1 Descriptive schema of rotary drilling system
3 Mathematical Model
The mathematical model of the drilling system can be derived by representing the
behavior of the system as the behavior of torsional pendulum, for which the pipes
behaves as torsional springs and the collars as a rigid body, we suppose that the top
drives rotates at constant speed (Fig. 2).
174 R. Riane et al.
Fig. 2 Representation of the mechanical part under torsional behavior of the drill-string
For simplification of this study, we suppose that the interaction between the drill-
string and the formation is in the bit, thus, no lateral of axial vibrations exist. that no
lateral or The dry frictional model is considered [1, 8, 9], the Interaction between the
bit and the formation is a mixture of cutting process and frictional forces [10–12].
The equation of motion can be given as follows.
3.1 The Dynamic of BHA
The BHA dynamic is given by the following equation [13].
Jb ϕ̈b = k(ϕt − ϕb ) − Cb ϕ̇b − T ob(ϕ̇b ) (1)
Let φ = ϕt − ϕb and, ϕ̇b = b , then
Jb Ω̇b = kφ − Cb b − T ob(b ) (2)

where: (ϕb ) (b ) (Jb ) (Cb ) are, the angular displacement, angular velocity, equivalent
of mass moment of inertia, equivalent viscous damping coefficient at the bottom of
the drill-string respectively, (ϕt ) is the angular displacement at the top of the drill-
string, (k) is the torsional stiffness coefficient, and Tob is a nonlinear function which
will be referred to the torque on bit.
3.2 The Dynamic of Drillstring
The components that defines the mechanical dynamic of the drillstring are: the
gearbox characterized by an (n:1) ratio, and the DC excited electric motor. The
equations that describe this behavior are given as follows.
3.2.1 Mechanical Behavior

Jt t = T − Ct t − kφ (3)
where (Ω t ) (J t ) (C t ) are the angular velocity, equivalent of mass moment of inertia,

equivalent viscous damping coefficient at the top of the drill-string respectively, T
is the torque delivered by the motor to the system multiplied by the gearbox ratio
(n):T = nTm .
3.2.2 Electrical Equation

di
v=l + ri + vcem (4)
dt
where (l), (r), (i) and (v) are respectively defined as motor current, motor resistance,
motor inductance and motor input voltage, (vcem ) is the counter-electromotive force
(back-emf), (K) is the motor constant multiplied by the gearbox ratio such as K =
n Km .
The motor speed and current are connected to counter-electromotive force, and
the torque, by the coefficient K respectively.
vcem = K t , And, T = K i
Finally, we can find
Jt Ω̇t = K i − Ct t − kφ (5)
di
l = v − ri − K t (6)
dt
176 R. Riane et al.
As explained earlier, the Tob is of frictional contact between the bit and the rock
as well as the cutting operation [12].
T ob = Tc + T f (7)
By exploring conclusion in [14], we can say that the cutting torque can be written
as
1 2
Tc = R εd (8)
2 bit
ε is the amount of energy required to cut a unit volume of rock, and d the cut depth,
and (Rbit ) is the radius of the bit.
In [6, 10, 12] the frictional torque is considered continuous static function,
however, in this study it has been taken as a dynamic discontinuous function for
dry friction. The discontinuity between stick and slip phase caused the model to
be more complex especially for low velocity values of the slip phase [15]. During
the slip phase, the relative movement is almost zero, thus, the friction can be seen
as constraints that preserve the initial condition of zero velocity between rubbing
surfaces and the rock. Coulomb model, shown in Fig. 3, has been lengthily used
even it reveals discontinuity problem, consequently KARNOOP [16] proposed a
zero velocity constraints in his model in which where a transition interval from stick
to the slip phases is set up, the Coulomb model is still valide outside that interval
[17]. The Karnopp model can be written as (9).
⎧
⎨ Te i f b ≺ Dv and Te ≺ Ts
T f = Ts sign(Te ) i f b ≺ Dv and Te Ts (9)
⎩
Td sign(ϕb ) i f b Dv
Fig. 3 PI Observer diagram

for unknown input torque
T e: is the external torque that must be applied to conquer the static friction torques
T s . T d is the dynamic friction torque, Dv is the zero-velocity band, where:
1
Ts(d) = W ob · g · μs(d) · Rbit (10)
2
μs(d): is the static (or dynamic) dry frictional coefficient, Wob is the Weight on Bit.
In conclusion, the mathematical model of the entire system can be written as
⎧
⎪
⎪ φ = t − b
⎪
⎨ Ω̇ = − k φ − Ct + n K i
t Jt Jt t Jt
(11)
⎪
⎪ Ω̇ = k
φ − Cb
− 1
T ob(b , W ob)
⎪
⎩ di
b Jb Jb
b J b
dt
= − l t − l i + l v
nK r 1
4 PI Observer Design
The PI observer is a system in which the added terms are proportional and integral of
unknown variable to be estimated so that the robustness and performance is improved.
In this study, the unknown variable is the input torque which represents the state and
the unknown nonlinear torque on bit as shown in Fig. 5, [5, 18].
The mathematical model given in (11) can be rewritten the state space model in
(12), it represents the torsional behavior of the drillstring.

Ẋ = AX + Bu + Ed
(12)
Y = CX
X = ( Ωt Ωb i)t is the state vector, Y = (Ωt i)t is the output vector, u = v is the
known input (voltage), d = Tob is an unknown input torque on bit). A, B, C, and E
are known matrices with appropriate dimensions.
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0 1 −1 0 0 0
⎜− k −C 0 k ⎟ ⎜0⎟ ⎜ 0 ⎟
⎜ Jt Jt ⎟ ⎜ ⎟
A=⎜ k Jt
⎟, B = ⎜ ⎟
⎝ 0 ⎠, E = ⎜ ⎟
⎝ Jb 0 −Cb 0 ⎠ ⎝ − J1b ⎠
1
0 − Kl 0 − rl l 0

0100
C=
0001
The PI observer can designed and analyzed based on the new augmented system
given by (13).
178 R. Riane et al.

Ż = A z Z + Bz u
(13)
Y = Cz Z

T T AE
where Z = X d , Bz = B 0 , C z = C 0 , and A z =
0 0
By checking the observability (Az, C z) , the observer is proposed for the system
(12) as shown in (14).
⎧
⎨ Ẋ = A X̃ + Bu + E d̃ + K p Y − C X̃
(14)
⎩ d̃ = K I Y − C X̃
K P and K I are respectively proportional and integral gains, the observer in (14) can
now be rewritten in the following form [17].

Ż = A z Z̃ + Bz u + K z Y − C z Z̃
(15)
Ỹ = C Z Z̃
T
where Z̃ = X̃ d̃ , and K z = K p K I
The error of estimation of the unknown input torque is given as:
e = Z − Z̃
Thus
ė = Ż − Z̃˙ = (A z − K z C z )e (16)
It is noticeable that the error in (16) converges asymptotically to zero if we prove

that the AO = Az – Kz Cz is Hurwitz. Because the observability (Az, Cz) is proven,
we can use many techniques to find the gain Kz that forces the observer error toward
zero, in this study, pole placement technique has been used.
5 Results and Discussion
The effectiveness and robustness of the proposed observer have examined by simu-
lation of the model with different scenarios, they includes random variation of Wob
and voltage of the motor, then estimate the torque on bit for these each variation. The
simulations have been carried out in LabView environment (S/N: M76X07294), all
model parameters are given in Tables 1 and 2.
Practically, to mitigate torsional vibrations the drillers manipulate the drilling
parameters such as the Wob, the top drive torque, and the mud viscosity. It is based
Table 1 Numerical values of the drilling system

Parameter Description Value Unit
Jt Mass moment of inertia at the top of the drill-string 1000 kgm2
Jb Mass moment of inertia at the bottom of the drill-string 127 kgm2
Ct Viscous damping coeffi-cientat the top of the drill-string 51 Nms/rad
Cb Viscous damping coeffi-cientat the bottom of the drill-string 40 Nms/rad
k Drill-string stiffness 481 Nm/rad
r Motor resistance 10 m
l Motor inductance 5 mH
Km Motor constant 6 Nm/A
n Gearbox ratio 7.20 –
Table 2 Numerical values of the Bit-Rock interaction

Parameter Description Value Unit
e Intrinsic specific energy of the rock 130 MJ/m3
d Depth of cut 4 mm/rev
μs Static dry frictional coefficient 0,6 –
μd Dynamic dry frictional coefficient 0,4 –
Rbit Bit radius 0.10 m
on their recommendations; the main drawback of such practice is that the time delay
between the beginning of the oscillation phenomenon and the time of observing it
and manipulating parameters and its effect [19, 20]. In order to show the effectiveness
of the proposed observer for minimizing this time delay, the same methodology used
in the field is considered in this simulation, the only difference is that we replace the
human observation by PI observer.
5.1 Weight on Bit and Top Drive Speed Variations
It has been recommended by drillers practice that severity of torsional vibrations can
be minimized by reducing the Weight on bit (WOB) and increasing the Top drive
angular velocity (by increasing the voltage of DC motor), this recommendation has
been simulated as given by Figs. 4 and 5.
180 R. Riane et al.
Fig. 4 First scenario:

increase of DC motor
voltage with Wob = 15T
Fig. 5 Second scenario:

Wob decreasing with
constant top drive angular
velocity (v = 125 V)
5.2 Drilling Fluid Viscosity Variation
The third scenario for torsional vibrations mitigation is by manipulating the drilling
fluid viscosity; the increase of this characteristic leads to amplification of the damping
coefficient in the bottom hole at the bit. Figure 6 demonstrates the effect of increasing
the equivalent damping coefficients C b and C t.
Fig. 6 Third scenario:

Variation of drilling fluid
viscosity with Wob = 15T
and (v = 125 V)
As the results of simulation that indicate the same dynamic behavior of field
practical scenario, we can say that the model is representative and valid (Fig. 7).
For the second part of the simulation, we design the PI observer to estimate the
unknown Tob as well as the angular velocity of the BHA. Pole placement technique
has been considered for gain calculation for which the optimal Eigen values are found
at (-10-50-70-100-130). Figure 8 demonstrate the effectiveness of the PI observer
estimation for BHA angular velocity, however, Fig. 9 is for Tob estimation. The Wob
and Motor voltage (proportional to the Top drive speed) used for this scenario are
shown in Fig. 7.
Figure 10 represents the Wob under perturbations; the latter can be caused by
many factors. One of them is the presence of axial vibration stimulated by the slip
Fig. 7 weight-on-bit and

motor voltage used with PI
observer model
182 R. Riane et al.
Fig. 8 The BHA angular

velocity and its estimation,
and estimation error (on
bottom)
Fig. 9 Torque-on-bit and its

estimation, estimation error
(on bottom)
Fig. 10 The weight-on-bit

variation under perturbation
phase of torsional vibrations. The designed PI observer has tested under such types
of disturbances. The Eigen values has been recalculated to improve the estimation
performance of our observer under structured and unstructured perturbations, the
optimal values are (−10 −100 −110 −120 −130) (Figs. 11 and 12).
Fig. 11 BHA angular

velocity and its estimation
under perturbations, and
estimation error (bottom)
Fig. 12 Torque-on-bit and

its estimation under
perturbation, and estimation
error (on bottom)
184 R. Riane et al.
6 Conclusion
This study proposed the use of PI observer to estimate the nonlinear torsional dynamic
of drillstring in petroleum wells. The mode that describes the behavior of the system
under these vibrations has been detailed. The bit-rock interaction term is also used
to characterize the frictional forces during drilling process, even though it may vary
instantly due the change of rock mechanical characteristics. The proposed PI observer
is designed to estimate the instantaneous angular velocity variation of the BHA,
as well as the variation of unknown disturbances in the Torque on bit, the good
estimation is used to define the compensation value to be applied in real-time to
suppress (or mitigate) the torsional vibrations and protect drilling equipment.
References
1. Riane R (2015) Contribution to the adaptive observer synthesis with unknown input of rotary
drilling system: simulation under LabView Environment. Magister thesis, University M’hamed
Bougara of Boumerdes
2. Mihajlović N (2005) Torsional and lateral vibrations in flexible rotor systems with friction.
Doctoral thesis, Technical University of Eindhoven
3. Abdulgalil F, Siguerdidjane H (2004) Nonlinear friction compensation design for suppressing
stick-slip oscillations in oil-well drill-strings. In: 7th IFAC DYCOPS, Massachusets, USA.
https://doi.org/10.1016/S1474-6670(17)31903-1
4. Doghmane MZ, Kidouche M (2018) Decentralized controller Robustness improvement using
longitudinal overlapping decomposition-application to web winding system. Elektronika ir
Elektronika 24(5):10–18. https://doi.org/10.5755/j01.eie.24.5.21837
5. Koenig D, Mammar S (2002) Design of Proportional-Integral Observer for Unknown Input
Descriptor Systems. IEEE Trans Autom Control 47(12):2057–2062. https://doi.org/10.1109/
TAC.2002.805675
6. Detournaya E, Richard T, Shepherd M (2008) Drilling response of drag bits: theory and
experiment. Int J Rock Mech Min Sci 45:1347–1360. https://doi.org/10.1016/j.ijrmms.2008.
01.010
7. Detournay E, Defourny P (1992) A phenomenological model for the drilling action of drag
bits. Int J Rock Mech Min Sci Geomech Abstr 29(1):13–23. https://doi.org/10.1016/0148-906
2(92)91041-3
8. Navarro-López EM, Suárez R (2004) Modelling and analysis of stick-slip behaviour in a
drill-string under dry friction. In: The congress of the Mexican Association of Automatic
Control
9. Li L, Zhang Q-Z, Rasol N (2011) Time-varying sliding mode adaptive control for rotary drilling
system. J Comput 6(3):564–570. https://doi.org/10.4304/jcp.6.3
10. Richard T, Germay C, Detournay E (2000) Self-excited stick–slip oscillations of drill bits. C
R Mecanique 332:619–626. https://doi.org/10.1016/j.crme.2004.01.016
11. Besselink B, van de Wouw N, Nijmeijer H (2011) Model-based analysis and control of axial
and torsional stick-sliposcillations in drilling systems. In: IEEE international conference on
control applications (CCA). https://doi.org/10.1109/CCA.2011.6044505
12. Bayliss MT, Panchal N, Whidborne JF (2012) Rotary steerable directional drilling stick-slip
mitigation control. In: The proceedings of the 2012 IFAC workshop on automatic control in
offshore oil and gas production, Norwegian University of Science and Technology, Trondheim,
Norway, May 31–June 1, 2012. https://doi.org/10.3182/20120531-2-no-4020.00001
13. Riane R, Kidouche M, Doghmane MZ (2018) Kalman filter for unknown input estimation
application to a rotary drilling system modeled with simscape simulink environment. In: Inter-
national conference on technological advances in electrical engineering, 10–12 December,
2018, Skikda, Algeria
14. Detournay E, Defourny P ()1992 A phenomenological model for the drilling action of drag
bits. Int J Rock Mech Min Sci Geomech Abstr 29(1):13–23. https://doi.org/10.1016/0148-906
2(92)91041-3
15. Bousaada I, Saldivar B, Mounier H, Mondié S, Cela A, Niculescu S-L (2016) Delay system
modeling of rotary drilling vibrations. In: E Witrant et al (eds) Recent results on time-
delay systems, advances in delays and dynamics, 5th edn. Springer International Publishing
Switzerland, pp 23–43. https://doi.org/10.1007/978-3-319-26369-4_2
16. Karnopp D (1985) Computer simulation of stick-slip friction in mechanical dynamic systems,
asme journal of dynamics systems. ASME J Dyn Syst Meas Control 107(1):100–103. https://
doi.org/10.1115/1.3140698
17. Iamandei A, Miloiu G (2013) Motor drives of modern drillingand servicing rigs for oil and
gas wells. Springer Science+Business Media Dordrecht. https://doi.org/10.1007/978-94-007-
6558-0_50
18. Doghmane MZ (2019) Conception de commande décentralisée des systèmes complexes en
utilisant les stratégies de décomposition et optimisation par BMI. Doctoral thesis, University
M’hamed Bougara of Boumerdes, Algeria
19. Vromen TGM (2015) Control of stick-slip vibrations in drilling systems. Technische Univer-
siteit Eindhoven, Eindhoven
20. Dan S, Roar N, Vahid A (2013) Real-time optimization of rate of penetration during drilling
operation. In: 10th IEEE international conference on control and automation (ICCA) Hangzhou,
China, June, pp 12–14. https://doi.org/10.1109/icca.2013.6564893
Power System Generator Coherency
Identification for Large Disturbances
by Koopman Modes Analysis
Zahra Jlassi, Khadija Ben Kilani, Mohamed Elleuch, and Lamine Mili
Abstract This paper presents an effects analysis of large disturbances on power

system generator coherency. The analysis is based on a comprehensive parametric
study combining fault parameters with system operating conditions. Generator
coherency is identified via nonlinear Koopman modal analysis technique applied on
generators rotor speed dynamics following large disturbances. The Koopman oper-
ator captures the highly nonlinear spatiotemporal dynamics that cannot be assessed
with standard linear coherency identification techniques. Faults are short-circuits of
variable durations and locations. System operating parameters include loading levels
and on line generators. The study methodology is demonstrated on the Tunisian elec-
tric network comprising 153-buses 26-machines with a comparison to the conven-
tional slow coherency method. The results show that large disturbance location and
system loading levels, alter the coherent grouping determined by weak connections
topology. Slow coherent areas may degenerate into smaller coherent groups under
large disturbances. Such information is particularly useful for emergency control
where intentional islanding actions ought to be taken to mitigate large blackouts.
Keywords Electric power system · Coherency · Koopman modes analysis · Large

disturbances
Z. Jlassi · K. Ben Kilani · M. Elleuch (B)

University of Tunis El Manar, ENIT-LSE. B.P 37- le Belvédère 1002, Tunis, Tunisia
Z. Jlassi
K. Ben Kilani
L. Mili
Department of Electrical and Computer Engineering, Virginia Polytechnic Institute and State
University, Blacksburg, Virginia, USA

https://doi.org/10.1007/978-981-15-6403-1_13
188 Z. Jlassi et al.
1 Introduction
Unit coherency of power system generators is a property describing similarity in their

dynamic responses following disturbances. Synchronous generators are coherent if
they swing together in phase and frequency with constant angular difference over a
given time interval, and within a specific tolerance. Coherency importance emanates
from its dynamic equivalency implications. Its applications range from model aggre-
gation needed in off line system security assessment and control design, to on line
security assessment, and emergency control. In recent works, the coherency concept
has been utilized to develop suitable corrective actions for intentional power grid
islanding [1–4].
One important aspect in coherency identification is the set of disturbances consid-
ered and the identification technique. In the literature, many generator coherency
determination techniques have been proposed and advanced techniques are still being
developed. In [2, 5–8], the authors used data of generators swing curves or rotor angle
evolution with time following the disturbance. From these time responses, coherent
generators can be identified using a clustering algorithm with specific coherency
tolerance value. This technique is able to capture large signal dynamics, but is
computationally burdensome and its performance become poor as the power system
size increases. The large number handled of generators makes the correspondence
between modes and machines invisible and hard to study. In [9], linear time simula-
tion has been used to identify the coherent machines groups. Linear time simulations
are advantageous over nonlinear in their time frames. In [10, 11] the authors used
the slow coherency method to determine coherent generators groups of a linearized
model by selecting the slowest modes. The disadvantage of this method is the linear
approximation of modes and the need of power system modeling.
Large disturbance excite various high nonlinear low frequency oscillations and
would pull the operating point of the power system far away from its stable equi-
librium [12]. Post disturbance oscillations are a complex phenomenon involving all
power system generators which may swing in phase or oscillate against each other.
Loss of synchronism may result among machines in very few seconds following large
disturbances. Power system stability is affected, power transfers may be limited and
potentially widespread blackouts can be developed. Hence, the linear approxima-
tion may be invalid for large disturbances and not effective for emergency control
and system partitioning under severe disturbances. Recently, a mathematical method
of decomposing nonlinear dynamics into modes called Koopman Modes Analysis
(KMA) has gained an increased attention as recent technique applied to electric
power systems [4, 13, 14]. This method keeps all system nonlinear dynamics and
doesn’t employ explicit system analytical models. The Koopman operator captures
the highly nonlinear spatiotemporal dynamics that cannot be assessed with standard
linear coherency identification techniques.
The presented work aims at conducting an effects analysis of large disturbances on
generator coherency. Generator coherency is identified via nonlinear Koopman modal
Power System Generator Coherency Identification … 189
analysis technique applied on generators rotor speed dynamics. Faults are short-
circuits of variable durations and locations. System operating parameters include
loading levels and on line generators. The study is demonstrated on the Tunisian
electric network with a comparison to the conventional slow coherency method.
2 Koopman Modes Analysis
2.1 Koopman Modes Definition
For a given certain measurement, a Koopman modes set is defined. Consider the
dynamical discrete-time system evolving on a given finite-dimensional manifold M:
xk+1 = f (xk ) (1)
where k is an integer index and f is a map from the manifold M to itself. The linear
Koopman operator U acts on scalar-valued functions evolving on M as follow: for a
scalar-valued function g : M → R, the operator U maps g into a novel function Ug
given by:
U g(x) = g( f (x)) (2)
The idea is the analysis of nonlinear dynamics governed by (1) via the eigen-
functions and the eigenvalues of the operator U. from available data collected either
experimentally or numerically. To this end, let λ j ∈ C denote the Koopman operator
eigenvalues and ϕ j : M → R denote its eigenfunctions,
U ϕ j (x) = λ j ϕ j (x) j = 1, 2, . . . (3)
Consider x ∈ M and a vector-valued observable g : M → R p . For instance, if

x contains the full information at a particular time about a flow field, g(x) denote
a vector of any interest quantities, such as generator frequency measurements. If
all the p components of g lie within the Koopman eigenfunctions span, then the
vector-valued g is expanded in terms of these eigenfunctions as:
∞

g(x) = ϕ j (x)v j (4)
j=1
We typically think of (4) as expanding g(x) as a linear combination of Koopman

vectors v j . But, alternatively we may think of (4) as expanding g(x) as a linear
combination of Koopman operator eigenfunctions ϕ j , where v j are the vector coef-
ficients in this expansion. We will refer in this paper to the eigenfunctions ϕ j as the
Koopman eigenfunctions and the corresponding v j vectors in (4) as Koopman modes

of f , corresponding to the study observable g.
Each Koopman eigenvalue λ j characterizes the temporal behavior of each corre-
sponding Koopman mode v j : the magnitude of the eigenvalue determines its growth
rate and the phase determines its frequency.
2.2 Koopman Modes Computation from Snapshots
The commonly used Arnoldi algorithm actually produces approximations to the

Koopman operator eigenvalues and their corresponding modes [4, 13, 14]. The Ritz
values λ̃ j and Ritz vectors ṽ j behave in exactly the same manner as the Koopman
operator eigenvalues λ j and modes v j . Consider the sequence {g(x0 ), . . . , g(xm−1 )}
where g(x j ) is a vector measuring the observables at a certain time t j . The empirical
Ritz vectors ṽ j and values λ̃ j of this sequence are defined by this following algorithm:
(i) Define K by:

K = g(x0 )g(x1 )g(x2 ) . . . g(xm−1 ) (5)
(ii) Find the constants c j where a residual r is defined as:

m−1

r = g(xm ) − c j g x j = g(xm ) − K C,
j=0
r ⊥span{g(x0 ), . . . , g(xm−1 )} (6)
(iii) Define the companion matrix by:

⎡ ⎤
0 0 ... 0 c0
⎢1 0 ... 0 c1 ⎥
⎢ ⎥
⎢ ⎥
C = ⎢0 1 0 c2 ⎥ (7)
⎢. .. .. ⎥
⎣ .. . . ⎦
0 0 . . . 1 cm−1
(iv) Find the companion matrix eigenvalues:
C = T −1 T, = diag(λ̃1 , . . . , λ̃m ) (8)
where T is the Vandermonde matrix defined as:

⎡ ⎤
1 λ̃1 λ̃21 ... λ̃m−1
1
⎢1 λ̃2 λ̃22 ... λ̃m−1 ⎥
⎢ 2 ⎥
⎢ λ̃3 λ̃23 ... m−1 ⎥
λ̃3 ⎥
T = ⎢1 (9)
⎢. .. .. .. .. ⎥
⎣ .. . . . . ⎦
1 λ̃m λ̃2m . . . λ̃m−1
m
(v) Define the Ritz vectors ṽ j to be the columns of:
V = K T −1 (10)
The eigenvalues of the defined companion matrix C are approximations to

Koopman eigenvalues, called Ritz values. The corresponding eigenvectors given
by (10), called Ritz vectors, are approximations to Koopman modes. The Koopman
mode growth rate «GR» assesses the sampled dynamics damping. It is defined for a
given mode j as the modulus of the Koopman eigenvalue λ̃ j given by:
GR = λ̃ j (11)
Largest growth rates of Koopman modes denote its smallest damping ratios [4,
13, 14]. The frequency of each Koopman mode is computed by (12), where f s is the
chosen sampling frequency of sampled observables data:

f K M = I m ln λ̃ j ∗ f s /2π (12)
The norm of a Koopman Mode j defined by (13) quantifies its contribution

magnitude in the dynamics:

ṽ j = ṽ Tj ṽ j (13)
3 Coherency in Power Systems
Power system generators coherency is a property describing similarity in their

dynamic responses following disturbances. Two synchronous machines are consid-
ered to be coherent, if they swing together in phase and frequency with constant
angular difference over a given time interval, and within a specified tolerance [4].
The coherency condition in general simplifies to:
δi (t) − δj (t) ≤ constant (14)
where δi (t) is the i th generator rotor angle and δj (t) is the j th generator rotor angle.
3.1 Conventional Slow Coherency Method
Slow coherency is coherency arising from the slower inter-area modes, which are
oscillatory modes due to groups of machines oscillating against each other across
power transfer interfaces. From the slow coherency theory, the eigen-subspace of the
inter-area modes can be used to identify the slow coherent machines [10]. The weak
coupling and the singular perturbation theories may be used to separate variables
into slow and fast ones [11]. The slow variables are the interarea oscillatory modes
variables. The local modes of oscillations within the areas described the fast variables.
Generators coherency identification based on this linear technique is independent of
applied disturbance.
3.2 Coherency in the Koopman Modes
The concept of coherency in the context of Koopman modes has been defined
in various works [4, 13, 14]. This concept is applied on the generators dynamic
responses following arbitrary disturbances. For a given oscillatory KM v j , or Mode
j, to identify coherent generators swing dynamics in frequency and phase, it is quite
sufficient to check the initial phases α ji of mode j defined by (15) [4, 14]. A set of
oscillatory observables I ⊆ [1 . . . p] is a coherent group with respect to Mode j, if
all i ∈ I have similar initial phases behaviors [4].

tan α ji = I m ϕ j (x0 )v j i /Re ϕ j (x0 )v j i (15)

The notation Re ϕ j (x0 )v j i denotes the i th component of the vector real part
of ϕ j (x0 )v j . This vector stands for initial amplitudes of the modal dynamics. The
imaginary part vector I m[ϕ j (x0 )v j ] affects the initial phases of the modal dynamics.
4 Study Case
4.1 System Description
The proposed technique is conducted on the Tunisian power system comprising

153-bus 26-machines for load peaks and average load cases and comprising 137-bus
10-machines for load dips case. Synchronous generators are equipped with auto-
matic voltage and speed regulators. They are geographically grouped into three main
groups: North, Center and South. The North contains 12 generators for load peaks
and average load cases and 7 generators for load dips case.
Table 1 Study cases data

Loading level Disturbance
Location/Duration (ms)
North Center South
50 150 250 50 150 250 50 250 250
Load peaks x x x x x x x x x
Average load x x x x x x x x x
Load dips x x x x x x x x x
The Center contains 7 machines for load peaks and average load cases and only
two machines for load dips case. The South contains 7 generators for load peaks and
average load cases and only one generator for load dips case.
4.2 Study Cases
The KM analysis is performed on sampled generators angular speeds dynamics

following a short-circuit fault. Generators coherency identification based on
Koopman modes is performed and compared to the conventional slow coherency
method. Three loading levels are studied: an average load (3000 MW), a load peaks
(4200 MW) and a load dips (1600 MW). For each loading level, three fault durations
(50 ms, 150 ms and 250 ms) and three locations of applied fault are tested: in the
North (in RadesII bus), in the Center (in SousseC bus) and in the South (in Gannouch
bus). The different study cases are defined in Table 1. The simulations are obtained
using the software PSS/E: Power System Simulator for Engineering.
4.3 Simulation Results
Table 2 summarizes the slow modes and the KM analysis results for three dominant
Koopman modes. For all fault parameters and system loading levels, the modes are
oscillatory with low frequencies, corresponding to inter-are modes. For the standard
linear modal analysis of the system, one inter-area oscillatory mode is indicated,
independently of the fault. Intrinsic linear modes are captured by the KMA method
and nonlinear intrinsic modes of the power system are identified as shown in Table 2.
Table 2 Frequencies results of Koopman and linear modes for the different study cases
Fault parameter Interarea mode frequency (Hz)
Loading level Location Duration (ms) Koopman mode Linear mode
Average North 50 0.703-0.755-0.649 0.706
150 0.698-0.749-0.647
250 0.695-0.745-0.640
Center 50 0.640-0.727-0.671 0.706
150 0.767-0.728-0.668
250 0.772-0.721-0.658
South 50 0.686-0.694-0.766 0.706
150 0.709-0.655-0.767
250 0.710-0.656-0.767
Load peaks North 50 0.726-0.747-0.652 0.726
150 0.728-0.760-0.659
250 0.723-0.765-0.663
Center 50 0.776-0.730-0.660 0.726
150 0.773-0.733-0.661
250 0.764-0.737-0.664
South 50 0.728-0.763-0.659 0.726
150 0.718-0.751-0.651
250 0.718-0.758-0.590
Load dips North 50 0.827-0.771-0.721 0.780
150 0.830-0.774-0.721
250 0.831-0.777-0.723
Center 50 0.830-0.776-0.725 0.780
150 0.835-0.782-0.728
250 0.838-0.785-0.732
South 50 0.828-0.776-0.725 0.780
150 0.829-0.776-0.724
250 0.833-0.779-0.726
5 Faults Effects Analysis
Based on phase-coherency results in nonlinear Koopman modes, coherent genera-

tors groups are defined by applying K-means clustering algorithm. Combining the
grouping results of dominant Koopman interarea modes, a fine partition into coherent
machines of a power system can be determined. Coherency is identified for all study
cases. The effect of each case parameter is studied.
5.1 Fault Duration
For the average load case, the simulation results of generators angular speeds
following northern short-circuit with various durations (50, 150 and 250 ms) is
depicted in Fig. 1. For each fault duration case, northern generators have coherent
dynamics, central generators have other coherent dynamics and southern generators
have another coherent dynamics as shown in Fig. 1. The three coherent motions are
depicted by combining the coherent generators partitions for dominant Koopman
modes given in Fig. 2.
Fig. 1 Generators responses for average load case following northern fault with various durations.
a Northern generators. b Central generators. c Southern generators
(a)
Mode of frequency 0.703 Hz Mode of frequency 0.755 Hz Mode of frequency 0.649 Hz
4 4 4
North
North
North
Phase (rad)
Center
Phase (rad)
Phase (rad)
Center Center
0 0 0
South
South South
-4 -4 -4
0 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Amplitude Amplitude Amplitude
(b)
Mode of frequency 0.749 Hz Mode of frequency 0.647 Hz
Mode of frequency 0.698 Hz
4 4 4
North
North North
Phase (rad)
Phase (rad)
Phase (rad)
Center
Center Center
0 0 0
South
South South
-4 -4 -4
0 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Amplitude
Amplitude Amplitude
(c)

4 4
4 North
North North
Phase (rad)
Phase (rad)
Center
Phase (rad)
Center Center
0 0 0
South
South South
-4 -4 -4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fig. 2 Coherency in Koopman modes for average load case following northern fault with various
durations. a Duration of 50 ms. b Duration of 150 ms. c Duration of 250 ms
5.2 Fault Location
For each loading level, the effects of short-circuit location on generators coherency
is analysed. Results led to the same conclusion that depending on fault location,
the power system may exhibit distinct generator coherent groups. For example, for
load peaks study case, a fault of duration 150 ms is studied for three cases of loca-
tions: in the north, in the center and in the south. The coherency results in dominant
Koopman modes for this study case are given in Fig. 3. For northern fault location,
the dominant Koopman mode of frequency 0.830 Hz captures three coherent groups:
northern generators group, central generators group and southern generators group.
(a)
4 4 4
North North
North
Phase (rad)
Phase (rad)
Phase (rad)
Center Center
Center
0 0 0
South South
South
-4 -4 -4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(b)
4 4 4
North North
North
Phase (rad)
Phase (rad)
Phase (rad)
Center Center
Center
0 0 0
South South
South
-4 -4 -4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(c)
4 4 4
North North
North
Phase (rad)
Phase (rad)
Phase (rad)
Center Center
Center
0 0 0
South South
South
-4 -4 -4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fig. 3 Coherency in Koopman modes for load dips case following a fault of duration 150 ms with
various locations. a In the North. b In the Center. c In the South
For the Koopman modes of frequencies 0.774 Hz and 0.721 Hz, two coherent groups
are captured as depicted in Fig. 3. The northern generators form a first group, the
central and southern generators form a second group. By combining the coherent
generators partitions for all dominant Koopman modes, a partition of the Tunisian
test system with three coherent groups is obtained: North, Center and South. The
same partition is captured by dominant Koopman modes for central fault location
as given in Fig. 3. For southern fault location, the three dominant Koopman modes
capture two coherent motions as shown in Fig. 3. A first coherent motion for southern
generators is captured, a second coherent motion for northern and central generators
is captured. Thus, for this test fault location, the power system is decomposed into
two coherent zones instead of three zones.
5.3 System Loading
For the average load case, following a central fault of duration 250 ms, the coherency
in identified dominant Koopman modes are given in Fig. 4. For mode of frequency
(a)
North
4 4 4
North
Center North
Phase (rad)
Phase (rad)
Phase (rad)
Center
Southern- Center
0 0 0
South
Eastern South
Southern-
-4 -4 -4
0 0.5 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Western
(b)
4 4 4
North North
North
Phase (rad)
Phase (rad)
Phase (rad)
Center Center
Center
0 0 0
South South
South
-4 -4 -4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(c)

4 4 4
North North
North
Phase (rad)
Phase (rad)
Phase (rad)
Center Center
Center
0 0 0
South South
South
-4 -4 -4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fig. 4 Coherency in Koopman modes following a central fault with duration 250 ms for various
system loading levels. a Average load. b Load peaks. c Load dips
0.772 Hz, two coherent motions are captured. Northern and southern-western gener-
ators show a first coherent motion. Central and southern-eastern generators show
a second coherent motion. The mode of frequency 0.721 Hz captures other two
coherent groups: southern and central generators group and northern generators
groups. The mode of frequency 0.658 Hz captures three coherent groups: north,
center and south. Combining the grouping results of these three Koopman modes, a
partition into four coherent zones is determined: northern generators group, central
generators group, southern-eastern generators group and southern-western genera-
tors group. Nonetheless, following the same applied central fault of duration 250 ms,
for load dips and load peaks cases, a partition into three coherent zones is captured
by combining the grouping results of dominant Koopman modes given in Fig. 4.
The two Koopman modes of frequencies (0.737 Hz and 0.664 Hz), (0.785 Hz and
0.732 Hz) for load peaks case and load dips case respectively, capture two coherent
groups: northern and central generators group and southern generators group. The
Koopman mode of frequency 0.838 Hz for load dips case captures three coherent
groups: north, center and south. Southern and central generators group and northern
generators group are captured by mode of frequency 0.764 Hz for load peaks case.
Thus depending on loading level, different coherent dynamics may be captured.
6 Coherency Identification Comparison
Based on linear modal analysis of the study Tunisian power system, generators
slow coherency for each loading level case is identified. Figure 5 defines the mode
shapes for each case. The results show that for any system operating condition, the
northern and central generators show coherent dynamics and the southern generators
have other coherent dynamics. Therefore, a partition with two disjoint parts (north
and south) is defined based on slow coherency method as depicted in Fig. 6. But
based on Koopman modes analysis, depending on fault location and power system
loading level, the Tunisian power grid may be decomposed into three or four coherent
machines zones instead of two topological zones as shown in Fig. 6. The northern
slow coherent generators zone is decomposed into two coherent groups for load
peaks case and average load case under any applied fault in the north, in the center
or in the south. The southern slow coherent generators zone is decomposed into two
coherent groups (southern-eastern and southern-western) for average load case under
central fault. Only, for load dips case following a southern short-circuit fault, the two
topological slow coherent zones (northern and southern) are not decomposed into
other coherent groups. Thus, coherency property based on linear and topological
approximations may be invalid under large disturbances. Several generators may
switch from a group to another. A slow coherent area may degenerate into smaller
coherent zones.
90 1 90 1
120 60 120 0.8
60
0.8
0.6 0.6
150 30 150 30
0.4 0.4
0.2 0.2
180 0 180 0
210 330 210 330
240 300 240 300

270 270
Average load (0.706 Hz) Load peaks (0.726 Hz)

90 1
120 60
0.8
0.6
150
0.4
30
North
0.2
180 0
Center
South
210 330
240 300
270
Load dips (0.780 Hz)
Fig. 5 Coherency in the Tunisian slow mode for various loading levels
7 Conclusion
This paper proposes a comprehensive parametric study combining fault parameters

with power system loading levels to analyze their effects on generators coherency.
Faults are short-circuits of variable durations and locations. Coherency is identified
via the nonlinear Koopman modes analysis method to capture the highly nonlinear
spatiotemporal dynamics that cannot be assessed with standard linear coherency
identification techniques. The nonlinear modal analysis is applied on generators
rotor speed dynamics following large disturbances. The study is demonstrated on a
large scale industrial electrical network: the Tunisian power system. A comparison
of the nonlinear KMA based coherency identification method is carried out with the
conventional slow coherency method. The results have shown that the tested power
system has exhibited distinct generator coherent groups, emerging under different
severe system faults depending on their location and the power system loading level.
Predetermined Tunisian slow coherent areas based on weak connections topology
and linear approximations may be decomposed into two coherent groups under large
North
Center
Southern-
Southern- Eastern
Western
South
Legend
Gas Turbine
generator
Thermal Turbine
generator
Combined Cycle
Turbine generator
Groups based on slow coherency & KMA for load dips case
following southern fault.
Groups based on KMA for load peaks case following northern,
central and southern faults & KMA for load dips case following
northern and central faults & KMA for average load case
following northern and southern faults.
Groups based on KMA for average load case following central
fault.
Fig. 6 Coherent generators groups of the Tunisian power system under study cases
disturbances. Such information is particularly useful for emergency control where

intentional islanding actions ought to be taken to mitigate large blackouts.
References
1. Lin ZZ, Wen FS, Zhao JH, Xue Y (2016) Controlled islanding schemes for interconnected
power systems based on coherent generator group identification and wide-area measurements.
J. Mod. Power Syst Clean Energy 4(3):440–453
2. Koochi MHR, Esmaeili S, Dehghanian P (2018) Coherency detection and network partitioning
supported by wide area measurement system. In: 2018 IEEE Texas power and energy conference
(TPEC), College Station, TX, USA, IEEE, Mar 2018
3. Ab Salam AN, Hasmaini M, Dahlan NY, Raza S (2017) Performance of multiple passive
islanding detection technique for synchronous type of DG. J Electric Syst 13(3):568–578
4. Raak F, Susuki Y, Hikihara T (2015) Data-driven partitioning of power networks via nonlinear
Koopman mode analysis. IEEE Trans Power Syst 31(4):2799–2808
5. Khalil AM, Iravani R (2016) A dynamic coherency identification method based on frequency
deviation signals. IEEE Trans Power Syst 31(3):1779–1787
6. Yang S, Zhang B, Hojo M (2018) A dynamic generator coherency identification method based
on phase trajectory vector. In: 2018 IEEE innovative smart grid technologies-Asia (ISGT Asia),
Singapore, IEEE, Sept 2018
7. Lin Z, Wen F, Ding Y, Xue Y (2018) Data-driven coherency identification for generators based
on spectral clustering. IEEE Trans Ind Inf 14(3):1275–1285
8. Ul Banna H, Iqbal T, Khan A, Zahra Z (2018) Generators coherency identification using relative
correlation based clustering. In: 2018 International conference on engineering and emerging
technologies (ICEET), Lahore, Pakistan, IEEE, Apr 2018
9. Kyriakidis T, Cherkaoui R, Kayal M (2013) Generator coherency identification algorithm using
modal and time-domain information. In: EUROCON, Zagreb, Croatia, IEEE
10. Verdejo H, Montes G, Olgui X (2014) Identification of coherent machines using modal analysis
for the reduction of multimachine systems. Latin Am Trans IEEE 12(3):416–422
11. Stadler J, Renner H, Köck K (2015) An inter-area oscillation based approach for coherency
identification in power systems. In: 2014 power systems computation conference, Wroclaw,
Poland, IEEE, Feb 2015
12. Zaid MM, Malik MU, Bhatti MS, Razzaq H, Aslam MU (2017) Detection and classification
of short and long duration disturbances in power system. J Electric Syst 13(4):779–789
13. Susuki Y, Mezic I, Raak F, Hikihara T (2016) Applied Koopman operator theory for power
systems technology. Nonlinear Theory Appl IEICE 7(4):430–459
14. Susuki Y, Mezic I (2011) Nonlinear Koopman modes and coherency identification of coupled
swing dynamics. IEEE Trans Power Syst 26(4):1894–1904
Analysis and Experimental Validation
of Single Phase Series Resonance Inverter
A 3–7 KHz Full-Bridge Inverter
Alla Eddine Toubal Maamar, M’hamed Helaimi, and Rachid Taleb
Abstract In this paper, analysis and modelling of a single-phase cascaded full-

bridge resonance inverter are considered. The control of proposed inverter by
full-wave phase method at high frequency is implemented. In the first step,
MATLAB/Simulink environments are used to simulate the model and show obtained
results of waveforms. The second step is the experimental validation of simulation
results by a realization of the inverter and generated signals. The Root-Mean-Squared
voltage of a capacitor waveform between 3 and 7 kHz frequency with simulation and
realization are presented, comparatively, for a comparison. The results obtained are
satisfactory and show the efficiencies of analysis and the proposed model of the
full-bridge resonance inverter.
Keywords Power converters · Cascaded full bridge · High-frequency control ·

Resonant inverters · MATLAB/Simulink
1 Introduction
In recent years, cascaded full-bridge inverter have received more attention and have
drawn great research interest for several applications, some examples include high
power induction motor, resonant DC-AC converters, microgrid with solar panels and
storage batteries [1–3].
Resonance DC-AC inverter is a type of electric power converter, which contains
a resistor (R), inductance (L) and capacitor (C), there has been a great increase in the
use of this converter in a number of applications, some examples include: induction
heating systems, radio transmitter, heat treatment of metals in industrial processes
by melting and welding [3, 4].
A. E. Toubal Maamar (B) · M. Helaimi · R. Taleb

Laboratoire Génie Electrique et Energies Renouvelables (LGEER), Electrical Engineering
Department, Hassiba Benbouali University of Chlef, BP. 78C, 02180 Ouled Fares, Chlef, Algeria

https://doi.org/10.1007/978-981-15-6403-1_14
204 A. E. Toubal Maamar et al.
A lot of resonant converters have been developed, as (SRI) Series Resonant

Inverter, (PRI) Parallel Resonant Inverter, (ZVS) Zero Voltage Switching Reso-
nant Converter, (ZCS) Zero Current Switching Resonant Converter, Two Quadrant
ZVS Resonant Converter, (CERC) Class E Resonant Converter, Resonant DC-Link
Inverter. The series resonant inverter is being more popular because it simplified
topology, operation and control [5].
Many projects used the Arduino technologies for their advantages, as the design
of digital filter, power measurement, control of motor for robotic systems [6–8].
In this work, the control of electronic switches is ensured by Arduino Mega 2560
Microcontroller, and we interest to control the output power by varying the frequency
with a fixed duty cycle.
The paper is organized as follows. In Sect. 2, Analysis of the inverter operation
and simulation of a series resonant inverter are discussed. Some experimental results
using a prototype are given in Sect. 3. Results comparison, simulation and experi-
mental, are presented in Sect. 4. Finally, conclusions and future trends of work are
presented in Sect. 5.
2 Analysis and Simulation of a Series Resonant Inverter
A. Analysis of the Inverter Operation
A full-bridge inverter has two arms, and each arm with two electronic switches, with
series RLC Load, Resistor, Inductor, and Capacitor, the inverter powered by a DC
power supply, the topology shown in Fig. 1.
The operation of the switches is periodic with adjustable period T, and the
switching devices must be controlled in a complementary mode to avoid the short-
circuiting of the DC power supply, a time delay is inserted between the signals
to avoid short-circuit, so depending on the states of the switches S1, S2, S3, S4,
four operating sequences of the full-bridge inverter can be distinguished during a
switching period, and we work with two sequences, positive and negative. Figures 2,
3, 4, and 5, shows the operation sequences of the inverter.
Fig. 1 The structure of a

full-bridge series resonance
inverter S1 S3
+ R L C
UE - i
S2 UC S4
CONTROL
Analysis and Experimental Validation … 205
(a) Sequence 1: The switch S1, S3 was closed and switch S2, S4 was opened, the
output voltage URLC = 0.
(b) Sequence 2: The switch S1, S4 was closed and switch S2, S3 was opened, the
output voltage URLC = +UE .
(c) Sequence 3: The switch S1, S3 was opened and switch S2, S4 was closed, the
output voltage URLC = 0.
(d) Sequence 4: The switch S1, S4 was opened and switch S2, S3 was closed, the
output voltage URLC = −UE
B. Simulation of a Series Resonant Inverter
Simulation of the electrical circuit of the system is done in MATLAB environment

(SIM/POWER/SYSTEMS). The simulated circuit is a MOSFET based RLC series
Fig. 2 Sequence 1 of the

inverter operation
1
S1 S3
+ R L C
UE - i
S2 UC S4
CONTROL

inverter operation
S1 S3
+ R L C
UE - i
S2 2 UC S4
CONTROL

inverter operation
S1 S3
+ R L C
UE - i
S2 UC S4
3
CONTROL
S1 4 S3
+ R L C
UE - i
S2 UC S4
CONTROL
Fig. 5 Sequence 4 of the inverter operation
resonance inverter with four switches. Figures 6 and 7, shows the model of inverter
in MATLAB environement and the control block.
To have a similar model to a real system, and identification, we have measured
the electronic components values, resistor, inductor, and capacitor. RRESISTOR =
1004 [Ohm]; RINDUCTOR = 200 [Ohm]; RTOTAL = RRESISTOR + RINDUCTOR = 1204
[Ohm]; L = 310e−3 [H]; C = 3.13e−9 [F], and use these values in simulation.
Figures 8, 9, 10, 11, and 12, shows the simulation results, a permanent and tempo-
rary regime is obviously observed, a temporary regime caused by the non-linearity of
electronic components, saturation and hysteresis. From the figures, we can see that
Fig. 10 is a very special one, the wavelength is increasing and regular because the
control frequency is near to the circuit resonance frequency. The resonance frequency
of a circuit is determined by the following relationship:
Fig. 6 The Resonance inverter using MATLAB

Fig. 7 The control block of resonance inverter
Fig. 8 The output 25

waveforms, MATLAB, with 20
f = 3 kHz, D = 50%
15
10
5
Voltage [v]
-5
-10
-15
-20
-25
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
-3
x 10
Time [s]

waveforms, MATLAB, with
f = 4 kHz, D = 50% 30
20
10
Voltage [v]
0
-10
-20
-30
-40
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
-3
Time [s] x 10

f = 5 kHz, D = 50%
60
40
20
Voltage [v]
-20
-40
-60
-80
-100
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
-3
x 10
Time [s]
1
f0 = √
2π LC
by numerical application with the values previously declared, RTOTAL = 1204 [];
L = 310e−3 [H]; C = 3.13e−9 [F], the resonance frequency f0 = 5.1 kHz.
1
f0 = √ = 5.1120e+03
2π 310e−3 × 3.13e−9

f = 6 kHz, D = 50%
30
20
Voltage [v]
10
-10
-20
-30
-40
-50
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
-3
Time [s] x 10

waveforms, MATLAB, with
f = 7 kHz, D = 50% 20
10
Voltage [v]
-10
-20
-30
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
-3
Time [s] x 10
The experimental prototype of the system is shown in Fig. 13. The full-bridge inverter
consists of four MOSFET switches IRFZ44N. The MOSFET drivers use TLP 250
photocouplers. The load consists of a resistor, inductor and capacitor in series.
The generalized control signals have been implemented using the programmable
board (ARDUINO Mega 2560 Microcontroller) and PC with open source software
(Arduino IDE).
Figures 14, 15, 16, 17 and 18, shows the output voltages of RLC Load and the
capacitor voltage, with different frequency from 3 to 7 [KHz] and fixed duty cycle
D = 50%.
Fig. 13 The Experimental

prototype of a series resonant
inverter. A: Pc + ARDUINO
IDE software. B: Inverter
(driver circuits + MOSFET
switches) + ARDUINO
Mega. C: Series RLC
(Resistor, Inductor,
Capacitor) circuit. D: Digital
oscilloscope
Fig. 14 The output

waveforms, with f = 3 kHz,
D = 50%
Fig. 15 The output

D = 50%
Fig. 16 The output

D = 50%
Fig. 17 The output

D = 50%
Fig. 18 The output

D = 50%
The first step is the realization of the Experimental prototype, and next step is the
implementation of the digital control through ARDUINO Mega 2560 and calculates
the RMS values of capacitor voltage with the variation of frequency. To evaluate the
obtained results, the comparison is necessary, and this is what we will present in the
following section.
4 Comparison and Discussions
After the implementation of the resonance inverter and calculates the RMS value
with the variation of frequency, the collected data is shown in the following Table 1.
From presented data, two curves (simulation with MATLAB and experimental)
are generated, shown in Fig. 19.
From these two comparison curves, it is observed that the peak value of the curves
indicates the resonance frequency, also there is a shift between the two curves (simu-
lation and experimental) because of the perturbations of electrical equipment and
electronic components, but generally the obtained results show the good concor-
dance existing between the simulation model and the real system, and it is important
to note that the output power of the resonant inverter can be controlled by adjusting
the frequency of the switches or adjusting the duty cycle (D), some testing results
with variable duty cycle are presented in Table 2. Figures 20, 21 and 22, shows the
obtained results with variable frequency and duty cycle.
5 Conclusion
The simulation and realization of a single-phase cascaded Full-bridge resonance

inverter are discussed in this paper. The effectiveness of the system is verified by
obtained results, simulation and experimental. The resonance is a very attractive
physical phenomenon used in many industrial applications, with the periodic control
there is energy exchange between the inductor and capacitor, if the inductive reac-
tance equals that of the capacitive reactance (XL = XC), then a physical phenomenon
obtained, with a generation of higher output voltages. This work opens new ways
for future research with other topologies and other control cards. For an effective
control system for the resonant inverter, the uses of closed-loop circuit to adjusting
the duty cycle and frequency is very important.
Table 1 RMS voltage in

Frequency [KHz] UCRMS [V] UCRMS [V]
capacitor, simulation and
experimental MATLAB
reaisation
3 13.563 13.74
3.1 14.033 14.22
3.2 14.535 14.76
3.3 15.072 15.36
3.4 15.712 16.04
3.5 16.425 16.81
3.6 17.242 17.67
3.7 18.153 18.66
3.8 19.195 19.8
3.9 20.42 21.11
4 21.810 22.64
4.1 23.56 24.45
4.2 25.56 26.6
4.3 28.27 29.19
4.4 31.17 32.35
4.5 34.92 36.26
4.6 39.84 40.13
4.7 45.29 47.21
4.8 52.24 53.59
4.9 58.92 62.8
5 64.01 69.89
5.1 64.87 72.41
5.2 61.27 68.42
5.3 54.88 60.25
5.4 48.09 51.41
5.5 41.48 43.66
5.6 36.26 37.37
5.7 31.75 33.38
5.8 27.92 28.36
5.9 24.96 25.14
6 22.440 22.5
6.1 20.06 20.31
6.2 18.554 18.47
6.3 16.887 16.91
6.4 15.630 15.57
6.5 14.496 14.41
(continued)
Table 1 (continued)
Frequency [KHz] UCRMS [V] UCRMS [V]
experimental MATLAB
6.6 13.453 13.39
6.7 12.522 12.5
6.8 11.740 11.71
6.9 10.977 11
7 10.347 10.37
Fig. 19 The comparison of 80

resonance frequency results, Experimental
MATLAB and experimental 70 MATLAB
60
RMS Voltage [v]
50
40
30
20
10
0
2 3 4 5 6 7 8
Frequency [KHz]
Table 2 RMS voltage with

Frequency D = 25% D = 50% D = 75%
variable duty cycle
[KHz]
5.0 UCRMS = UCRMS = UCRMS =
45.37 [v] 64.01 [v] 45.22 [v]
46.27 [v] 64.87 [v] 46.13 [v]
43.56 [v] 61.27 [v] 43.42 [v]
Fig. 20 The output

waveforms, with f =
5.1 kHz, D = 25%
Fig. 21 The output

waveforms, with f =
5.1 kHz, D = 50%
Fig. 22 The output

waveforms, with f =
5.1 kHz, D = 75%
References
1. Taleb R, Benyoucef D, Helaimi M, Boudjemaa Z, Saidi H (2015) Cascaded H-bridge asym-

metrical seven-level inverter using THIPWM for high power induction motor. Energy Procedia
74:844–853
2. Helaimi M, Benghanem M, Beldamani B (2012) An improved PIλ controller for resonant inverter
induction heating systems under load and line variations. Stud Inform Control 21(4):423–430
3. Abdar HM, Chakraverty A, Moore DH, Murray JM, Loparo KA (2012) Design and imple-
mentation a specific grid-tie inverter for an agent-based microgrid. In: 2012 IEEE energytech
conference
4. Feng W, Mattavelli P, Lee FC (2013) Pulsewidth locked loop (PWLL) for automatic reso-
nant frequency tracking in LLC DC–DC transformer (LLC-DCX). IEEE Trans Power Electron
28(4):1862–1869
5. Kazimierczuk MK, Czarkowski D (2011) Resonant power converters. Wiley, New York
6. Toubal Maamar AE, Helaimi M, Taleb R, Chabni F (2018) Analysis and implementation of
half-bridge series resonant inverter using Arduino. In: ICCEE’18 international conference on
communications and electrical engineering, El-Oued, Algeria
7. Adel Z, Hamou AA, Abdellatif S (2018) Design of real-time PID tracking controller using
Arduino Mega 2560 for a permanent magnet DC motor under real disturbances. In: 2018 inter-
national conference on electrical sciences and technologies in Maghreb (CISTEM), Algiers, pp
1–5
8. Fannakh M, Elhafyani ML, Zouggar S (2019) Hardware implementation of the fuzzy logic
MPPT in an Arduino card using a Simulink support package for PV application. IET Renew
Power Gener 13(3):510–518
Optimal Decentralized State Control
of Multi-machine Power System Based
on Loop Multi-overlapping
Decomposition Strategy
M. Z. Doghmane and M. Kidouche
Abstract Most of large scale systems are described by their complex mathematical
model’s structures which caused by the overlapping between its inputs and outputs
variables. In many cases, the complexity of the model may lead to lose some part
of the input-output signals. Besides, it obscures study and analysis of such type of
systems for researchers. Many studies have been discussed the design of simpler
controller based on the mathematical structure of the systems’ model, wherein the
aim was to decrease control complexity. The main objective of this manuscript is to
benefit from the multi-overlapping structure of complex systems in order to design a
simpler controller using loop approach; the latter allows highlighting the advantages
of applying the overlapping decomposition strategy in order to improve the robustness
of the designed controller. Therefore, it can be considered as an efficient strategy
suitable for such type of large scale systems. A multi-machine power system with
loop structure is considered in order to demonstrate the usefulness of the proposed
design.
Keywords Large scale systems · Multi-overlapping decompositin · Loop

approach · Multi-machine power system
1 Introduction
The study of large scale systems is one of the biggest challenges facing researchers
in the last few years, the mains reason for that is the complexity that exist in the
mathematical model and also in its analysis. Many studies have been conducted in
order to overcome such kind of difficulties [1–5] (Bochen and Stankovic [6]; Zecvic
and Siljack [7]); these researchers have developed decomposition technique named
“Overlapping” in order to decouple the original system into subsystems with no
M. Z. Doghmane (B) · M. Kidouche

Laboratory of Applied Automatic (LAA), Department of Automation, University M’hamed
Bougara of Boumerdes, Boumerdes, Algeria

https://doi.org/10.1007/978-981-15-6403-1_15
218 M. Z. Doghmane and M. Kidouche
overlapped variables between them. Despite its usefulness, there are many neces-
sary and sufficient mathematical conditions that should be verified in order to apply
such technique. Lately, Francisco et al. [8] has extended the application of this tech-
nique for systems composed of more than two subsystems namely multi-overlapping
decomposition, it has been used to overcome the difficulties of controlling large scale
systems globally with convex optimization constraints. Multi-machine power system
is very known industrial system for which many control strategies have been applied;
the structure of the nonlinearity of its model and the interactions between its vari-
ables present a challenge for researchers. Many papers have treated the described
problematic [9]. One of the main reasons of studying this system is that its math-
ematical model verifies the necessary and sufficient conditions, thus it can be used
to illustrate the construction of the decomposed subsystems by using an overlapped
original system.
Our objective in this paper is to design an optimal decentralized state feedback
controller based on multi-overlapping decomposition strategy in order to reduce the
complexity of the design with preserving the global stability of the original system.
2 Control Structure and Complex Systems
A. Centralized Controller
In this structure, the control system uses all input variables to generate one control
law for all the outputs, this means that all the states and their interconnections are
considered as one single dynamic system. In the case of large-scale systems, the
order of the multivariable controller becomes high, which complicated its real time
implementation.
The feedback control law has been used to improve performance as well as reject
external disturbances. Control designs, such as PID, loop shaping, gain-scheduling
[10], multivariable control [11] are industrially used with centralized structure.
Unlike the centralized configuration, in the decentralized configuration scheme, each
manipulation variable is connected to one controlled variable.
B. Decentralized Controller
In decentralized control of large scale systems, the interactions terms between states
and control inputs are the main obstacle, these terms are either neglected or analyzed
for design simplification reasons [6].
The main contribution in such design is how to find a controller that reduces the
influence of the interactions terms without losing the dynamic of the original system
from one side, and ensures the stability of the total original system from the other
side.
Optimal Decentralized State Control of Multi-machine … 219
C. Advantages of Decentralized Controller
The advantage of the decentralized structure is that there are no any measurements
of the coupling quantities even if it is required to identify the active quantities of
coupling for each subsystem. Even though, the obtained controller is robust against
external parametric variation in a limited interval [6]. Thus, we can set the advantages
of decentralized controller, in comparison to centralized controller, for large scale
system as follows
• More robust control system with high degree of stability for the decomposed
system.
• Less complication for control implementation, the frequency of regulation interval
is wider.
3 Multi-overlapping Decomposition Strategy
In literature, there are many definition of Large-scale dynamic systems, the most
used one is that any system that is represented by large number of variables that
are strongly interconnected can be considered as complex systems [3]. Stability
analysis of the complex systems has also been very interesting research topic in the
last decades (Karacanias and Wilson [12]; Wu [13]; Doghmane and Kidouche [14,
15]). The analysis and control of such systems can be achieved by either studying
the whole mathematical model of the system which is very difficult because of the
interaction terms, or by decomposing the model into less complicated subsystems.
A. Decomposition principle
Complex systems’ dynamic can be composed of overlapped subsystems that share

common variables [6]. This structure can be expanded into larger state space represen-
tation in which the subsystems are separated from each other; consequently, higher
order system is obtained. The latter includes the necessary information of the orig-
inal system; the expanded system is then contracted by applying standard disjoint
decomposition strategy [16]. Consider the large scale system given by (1)

ẋ(t) = A(t)x(t) + B(t)u(t)
, (1)
y(t) = C(t)x(t)
where
⎧ T T T T
⎪
⎨ u = u 1 , u 2 , u 3 , u i ∈ i , i = 1, 2, 3
m
T T
x = x1 , x2 , x3 , xi ∈ ni , i = 1, 2, 3 .
T T (2)
⎪
⎩ T
y = y1T , y2T , y3T , yi ∈ li , i = 1, 2, 3
and consider the system given by (2)


ẋe = Ae xe + Be u e
, (3)
ye = Ce xe
xe ∈ n e , u e ∈ m e and ye ∈ le , are, state, input and output vectors of system (3)
respectively. The decomposition principle is based on the idea of finding a full-rank
transformations matrices: T : n → n e , N : m → m e , S : l → le , for any
initial state xe ∈ n e , of the system (1) and any input u e (t) ∈ m e , so that the system
in (3) can be expansion of the system (1) (or system (1) is contraction of system (3)),
which means

xe0 = T x0 xe (t; xe0 ; u e ) = T x(t; x0 ; u)
⇒ , ∀t ≥ 0 (4)
u(t) = N u e (t) ye (t, xe ) = Sy(t; x)
If condition in (4) is satisfied, then we can say that system (3) is an expansion of
system (1) or system (1) is contraction of (3) [8].
B. Restrictions on Decomposition Principle
Restriction on applying decomposition principle can be summarized by the necessary

and sufficient conditions in Theorem 1.
Theorem 1 The system in (3) is an expansion of the system in (1) if and only if there
exist full-rank matrices T ∈ n e × n , N ∈ m × m e , S ∈ le × l such that
⎧
⎨ T A = Ae T
T B N = Be , (5)
⎩
SC = Ce T
The two systems are related by (6)

⎧
⎨ Ae = T AT I + E A
B = T B N + EB (6)
⎩ e
Ce = SC T I + E C
T I is the pseudo-inverse of T, satisfying (7)
T I T = In

−1 I
TI = TIT T . (7)
E A ∈ n e ×n e , E B ∈ n e ×m e , and E C ∈ le ×n e , are the complementary matrices [17].

They must satisfy conditions given by (8) so that Theorem 1 is verified [16]. Consider
that the matrices in (1) and (3) satisfy (6); then theorem is verified if and only if (8)
is verified [18].
E A T = 0, E B = 0, E C T = 0. (8)
C. Contraction
Contraction is an equivalent principle of expansion, reversibly, the system (1) is

contraction of system (3) if, for any initial state x0 of (4), there is a contraction
matrix T with full row rank such that
xe0 = T I x0 ⇒ x(t, x0 ) = T xe (t, xe0 ), ∀t ≥ 0. (9)
A complex system has an overlapping structure if at least two of its subsystems

share some common states, inputs, and outputs. Consider the system (1) where A, B
has the form given in (10).
(10)
The system (10) can be now expanded into a larger system that preserves all
necessary dynamic of the original system; the system can be then decomposed into
subsystems controlled separately, the obtained controller is contracted for implemen-
tation in the original system [1, 2]. The state vector x in (1) is composed of three

T
vectors x1 , x2 and x3 with dimensions n 1 , n 2 and n 3 , so that x = x1T
, x2T , x3T and
n = n 1 + n 2 + n 3 , the dimensions of sub-matrices Ai j are n i × n j , i, j = 1, 2, 3
[2]. The structure of the gain matrix of the controller always pursues the structure of
the mathematical model of the original system. Consider the matrices T : n → n e ,
N : m → m e , such that
{T A = Ae T, T B N = Be , (11)
Our goal is derive the overlapping dynamic matrices (10) into matrices structure
given by (12).
⎡ ⎤ ⎡ ⎤
A11 A12 0 A13 B11 B12
⎢ A21 A22 0 A23 ⎥ ⎢ B21 B22 ⎥
Ae = ⎢
⎣ A21
⎥, Be = ⎢ ⎥,
0 A22 A23 ⎦ ⎣ B21 B22 ⎦
A31 0 A32 A33 B B32
⎡ ⎤ 31 (12)
0 A12 −A12 0
⎢ 0 A22 −A22 0 ⎥
E A = 21 ⎢
⎣ 0 −A22
⎥
A22 0 ⎦
0 −A32 A32 0
where (12) can be written as


Ae = T AT I + E A
. (13)
Be = T B N + E B
By rearranging the components of the state x into two overlapping subsys-

T
T
tems xe1 = x1T , x2T , and xe2 = x2T , x3T , which compose a new vector
T T
T
xe = xe1 , xe2 , and persuade the decomposition of A as shown by dashed lines in
(12). T is an n e × n matrix defined by (14).
(14)
The expanded system Se can now be conducted into two interconnected subsys-
tems Se1 and Se2 defined by (15).

Se1 : ẋe1 = Ae11 xe1 + Ae12 xe2 + Be11 u e1 + Be12 u e2
, (15)
Se2 : ẋe2 = Ae21 xe1 + Ae22 xe2 + Be21 u e1 + Be22 u e2
Ae12 , Ae21 : are the interconnection matrices. System in (15) indicates that the
decoupled matrices of the expanded matrix Ae correspond to the decomposed part of
the original matrix A [19, 3]. The class of expanded systems Se which corresponds to
the system S can be found by selecting the appropriate transformation matrix T and
its complementary matrices. E A , E B must satisfy the condition {E A T = 0, E B = 0.
The matrix in (12) has identity diagonal blocks, thus, the resultant expanded system
contains two identical subsystems [17]. System in (15) can now be stabilized by the
control law given in (16) [4].
⎧ ⎧
⎪
⎪ ⎨ ż e1 = Fe1 z e1 + G 1e1 G 1e2 ye1
⎪
⎪ 1 1 1
⎪
⎪ e1 : He1 K 11 K 12
⎪
⎨ ⎩ u e1 = 1 z e1 + 1 1 ye1
He2 2K 21 2K 22

⎧ , (16)
⎪
⎪ ⎨ ż e2 = Fe2 z e2 + G e1 G e2 ye2
⎪
⎪ 2 2 2
⎪
⎪ e2 : He1 K 11 K 12
⎪
⎩ ⎩ u e2 = z e1 + ye2
2 2 2
He2 K 21 K 22
(17)
The controller in (16) is contracted to decentralized controller in (17) using

Theorem 1 for implementation on the system (1). The necessary and sufficient
conditions for this transformation are all verified and respected.
4 Multi-overlapping Decomposition
Multi-overlapping decomposition strategy has been widely used to solve many prob-
lems in different domains such as: applied mathematics, structures’ construction,
robotics, and electric power systems. The mathematical framework of decompo-
sition is essentially based on expansion-contraction principle detailed before. The
principle is to expand the state space model in away the common variables appear as
disjoint, then design control law for each subsystem alone and contract the obtained
controllers to the original system [8]. Satisfaction of the inclusion conditions is
essential for transferring properties of the expanded system to the original one. In
loop topology, the common variable are shared by all the subsystems, in the case
of complex system composed of three subsystems, each two subsystems share a
common overlapped variables different from the other pairs [6]. An application of
designing a loop structure controller of system based on a multi-machine power
system is given in the next subsection.
A. Loop structure of Multi-overlapping system
The system in (1) can be rewritten in the form of (18)

ẋi = Aii xi + Bii u i
Si : i = 1, 2, . . . , N . (18)
yi = Cii xi
The structure of matrix A can be longitudinal, radial or loop. In this study, the
loop structure is considered in order to apply the proposed decomposition strategy to
multi-machine power system [12]. The following steps are followed to achieve the
objective of this paper
First, the transformation matrices are given by (19)

T = blockdiag [I11 , I11 ]T , . . . , [I N N , I N N ]T

, (19)
N = 0.5 blockdiag ([I11 , I11 ], . . . , [I N N , I N N ])
where E A is proposed as in (20).

⎡ ⎤
A11 −A11 A12 −A12 . . . −A1N A1N
⎢ −A −A12 . . . −A1N ⎥
⎢ 11 A11 A12 A1N ⎥
⎢ ⎥
⎢ −A 21 A21 A22 −A22 ... 0 0 ⎥
⎢ ⎥
E A = 0.5⎢ −A21
⎢ .
A21 −A22 A22 ... 0 0 ⎥,
⎥ (20)
⎢ . .. .. .. .. .. ⎥
⎢ . . . . ··· . . ⎥
⎢ ⎥
⎣ AN1 −A N 1 0 0 · · · AN N −A N N ⎦
AN1 −A N 1 0 0 · · · −A N N AN N
Thus, the expanded matrix is found as in (21).

(21)
Second, the permutation matrices (22) are applied to (21).

⎡ ⎤
0 · · · 0 I11
2N −1 ⎢ I11 · · · 0 0 ⎥
⎢ ⎥
T A = T A12 T A23 · · · T A(2N −1)(2N ) = T Ak(k+1) = ⎢ .. . . .. .. ⎥ (22)
k=1 ⎣ . . . . ⎦
0 . . . IN N 0
So that, the expanded system becomes:
(23)
Theorem 1 discussed early in this paper, has been used to decompose the matrices
of the expanded system (23), they have multi-overlapping structure with loop
topology. For each subsystem Sei , a local control law given by (24) is proposed
[8].

ui K ii K i j xi
=− , (24)
uj K ji K j j xj
The gain matrix for this control law is written as

K ii K i j
K e = blockdiag . . . , ,... , (25)
K ji K j j
The conditions of contraction are verified in [10, 11, 15], and its contracted
controller is found as:
p V p .
K = QpK (26)
B. Loop Structure of Multi-machine power system

A power system, composed of N machines, is seen as complex system with N

subsystems with nonlinear interconnection terms.
Let xi = [δi (t) ωi (t) δmi (t) δei (t)]T be the state vector of the ith
machine with i = 1, . . . , N . Its dynamic is described by (27) [9].

ẋi = Ai xi (t) + Bi u i (t) + h i (x, t)
, (27)
yi = Ci xi (t)
N
h i (x, t) = j=1, j=i pi j G i j gi j x i , x j is the nonlinear interaction terms between

machine i and all other machines. The dynamic matrices of the ith subsystems are
given by (28).
⎡ ⎤ ⎡ ⎤
0 1 0 0 0
⎢ 0 − Di ω0
0 ⎥ ⎢ ω0 Eqi Eq j Bi j ⎥
⎢ 2Hi 2Hi ⎥ ⎢ − 2Hi ⎥
Ai = ⎢ − 1
− K mi ⎥, G i j = ⎢ ⎥ (28)
⎣0 0 Tm i Tm i ⎦ ⎣ 0 ⎦
0 − Te KReii ω0 0 − 1
Te 0
i i
The
interactions

between

the subsystems are defined as gi j xi , x j =
sin δi (t) − δ jt − sin δi0 − δ j0 .
Parameters of system (27) are described by (29)
⎧
⎨ δi (t) = δi (t) − δi0
Pmi (t) = Pmi (t) − Pmi0 , (29)
⎩
X ei (t) = X ei (t) − X ei0
u i (t) is the control vector of the ith subsystem, yi (t) is the output vector of the ith
subsystem; δi (t) is the rotor angle for the ith machine (rad); ωi (t) is the relative speed
for the ith machine (rad); Pmi (t) is the per unit mechanical power for ith machine
[9] (Fig. 1).
Fig. 1 Three-machine
power system diagram
5 State Feedback Control
Using the developed algorithm, solve for K i matrix for each subsystem; by applying
the expansion–contraction theorem, the control design can be expressed by finding
matrices Fe , G e , He and K e of the controller e that minimizes the cost function Je
such that the closed loop system in (30)

ẋ Ae + Be K e Ce Be He xe
Sec : e = , (30)
ż e G e Ce Fe ze
is asymptotically stable [5, 13, 14], with
Je (Fe , G e , He , K e , xe0 , z e0 ) ≤ Je F e , G e , H e , K e , xe0 , z e0 , ∀ F e , G e , H e , K e ∈ e

(31)
Je F e , G e , H e , K e , xe0 , z e0 represents the cost function value for the controller

T T T
(26) in the closed-loop system (30) with initial state xe0 z e0 . The conditions
for controller design and system stability are achieved by transforming (30) with its
performance function into LQR standard form [7, 20]. Consider the system described
by (32).
υ
x̂˙e = Ae x̂e + i=1

B ei û ei
Se :
, i = 1, 2, . . . , υ (32)
ŷei = C ei x̂e

K ei (i = 1, 2, . . . υ) are gain matrices for the system S e that minimize the cost
function

∞ υ

J e = Je = x̂eT Q e x̂e + û eiT R ei û ei dt
0 i=1
The semi-decentralized controllers can be now found by solving recursively the

four parts of (33) [21].
⎧
⎪
⎪ T P + P + Q + C T K T R K C = 0
⎪
⎨ ∇ J = B T P LC T + R K C LC T = 0
Ki i i
i i i

i −1 . (33)
⎪
⎪ K i = −RiT BiT P LCiT Ci LCiT
⎪
⎩ L + LT + X 0 = 0
As shown in Figs. 2, 3, 4, 5 and 6, the input reference for multi-machine power

system are tracked,
• For state responses, the controller tracks the input for machine 1 (Fig. 2), machine
2 (Fig. 3), and machine 3 (Fig. 4). The cost function values are found from solving
second equation in (33) J01 = 1.54 × 103 , J02 = 1.5 × 102 , and J03 = 0.76 × 103
for machine 1, 2, and 3 respectively.
• For the errors of overlapped zone, the quadratic error for machine 1 is e1 = 2.7%,
and for machine 2 is e2 = 1.68%, and for machine 3 is e3 = 4.05%, the overlapped
response error of state 3 in very high (Fig. 6).
Using another input reference, the desired output feedback control signals for
multi-machine power system are reached as it is shown in Fig. 7, where
Fig. 2 State response for multi-power machine system after decomposition—Machine 1

Fig. 5 State response of overlapped system
Fig. 6 Error of responses of overlapped system

Fig. 8 Error of responses for multi-power machine system after decomposition—Machine 3
• For state responses, the controller tracks the input for machine 1, machine 2
(Fig. 7), and machine 3. The cost function values J01 = 2.35 × 102 , J02 =
1.25 × 103 , and J03 = 1.73 × 103 for machine 1, 2, and 3 respectively.
• For the errors, the quadratic error for machine 1 e1 = 2.64%, and for machine 2
is e2 = 3.04% (Fig. 8), and for machine 3 is e3 = 5.75%.
7 Conclusion
In this paper, a decomposition strategy based on multi-overlapping structure with loop

topology has been studied for complex multi-machine power system. The mathemat-
ical background of the proposed strategy has been detailed, theorem of inclusion-
contraction principle has been considered for designing decentralized controller
based on the decomposition, and for global stability analysis of the contracted system.
Necessary and sufficient conditions for application of such techniques for complex
systems with loop topology have been checked. Furthermore, power system with
multi-overlapping loop topology is considered to demonstrate the effectiveness of

the algorithm; optimal decentralized state feedback controller has been designed
for the studied system. Based on the obtained results, it has been established that
multi-overlapping decentralized controller could be advantageous in comparison to
other classical techniques, because it reduces the complexity of interaction terms and
controller implementation simplification. Moreover, this method has been improved
by introducing an optimization technique. The advantages of the proposed approach
can be showed by comparing obtained results with recent controller based on new
approaches such as sliding mode. Several other types of topologies can be discussed
for future scientific works.
References
1. Iftar A (1991) Decentralized optimal control with overlapping decompositions. In: IEEE inter-
national conference on system engineering, Dayton, OH, USA, pp 299–302. https://doi.org/
10.1109/icsyse.1991.161138
2. Iftar A (1993) Decentralized estimation and control with overlapping input-state and output
decomposition. IFAC Proc 23(8):99–104. https://doi.org/10.1016/S1474-6670(17)52078-9
3. Ikeda M, Siljack DD, White DE (1981) Decentralized control with overlapping information
sets. J Optim Theory Appl 34:279–310. https://doi.org/10.1007/bf00935477
4. Ikeda M, Siljack DD (1986) Overlapping decentralized control with input, state, and output
inclusion. Control Theory Adv Technol 2(2):155–172
5. Stankovic S, Stanojevic M, Siljak DD (2000) Decentralized overlapping control of a platoon
of vehicles. IEEE Trans Control Sys Technol 8:816–832. https://doi.org/10.1109/87.865854
6. Bochen X, Stankoviç S (2005) Decomposition and decentralized control of systems with multi
overlapping structure. Automatica 41:1765–1772. https://doi.org/10.1016/j.automatica.2005.
01.020
7. Zecevic AI, Šiljak DD (2005) A new approach to control design with overlapping informa-
tion structure constraints. Automatica 41:265–272. https://doi.org/10.1016/j.automatica.2004.
09.011
8. Francisco PQ, Rodellar J, Rosell JM (2010) Sequential design of multi-overlapping controllers
for longitudinal multi-overlapping systems. Appl Math Comput 207(3):1170–1183. https://doi.
org/10.1016/j.amc.2010.01.130
9. Tlili AS, Baiek NB (2009) Decentralized observer based guaranteed cost control for nonlinear
interconnected systems. Int J Control Autom 2(2):29–45
10. Koç H, Knittel D, Mathelin MD, Abba G (2000) Robust gain-scheduled control of winding
systems. In: IEEE conference decision and control, Sidney, Australia. https://doi.org/10.1109/
cdc.2000.912360
11. Koç H, Knittel D, Mathelin MD, Abba G (2002) Modeling and robust control of winding
systems for elastic webs. IEEE Trans Control Syst Technol 10(2). https://doi.org/10.1109/87.
987065
12. Karcanias N, Wilson DR (1989) Decentralized diagonal dynamic stabilization of linear
multivariable systems. Linear Algebra Appl 121:455–474. https://doi.org/10.1016/0024-379
5(89)90716-7
13. Wu HS (1999) Decentralized stabilizing state feedback controllers for a class of large-scale
systems including state delays in the interconnections. J Optim Theory Appl 100:59–87. https://
doi.org/10.1023/A:1021764814535
14. Doghmane MZ (2011) Optimal decentralized control design with overlapping structure.
Magister Thesis, University M’hamed Bougara of Boumerdes, Algeria
15. Doghmane MZ (2019) Conception de commande décentralisée des systems complexes en

utilisant les stratégies de décomposition et optimisation par BMI. PhD Thesis, University
M’hamed Bougara of Boumerdes, Algeria. http://dlibrary.univ-boumerdes.dz:8080/handle/123
456789/5432
16. Doghmane MZ, Kidouche M (2018) Decentralized controller robustness improvement using
longitudinal overlapping decomposition-application to web winding system. Elektronika ir
Elektrotechnika 24(5):10–19. https://doi.org/10.5755/j01.eie.24.5.21837
17. Doghmane MZ, Kidouche M, Habbi H et al (2015) A new decomposition strategy approach
applied for a multi-stage printing system control optimization. In: 4th international conference
on electrical engineering (ICEE), Boumerdes Algeria, May 2015. https://doi.org/10.1109/intee.
2015.7416751
18. Garcia CE, Prett DM, Morari M (1989) Model predictive control: theory and practice, a survey.
Automatica 25:335–348. https://doi.org/10.1016/0005-1098(89)90002-2
19. Iftar A, Ozguner V (1999) Contractible controller design and optimal control with state and
input inclusion. Automatica 26(03):593–597. https://doi.org/10.1016/0005-1098(90)90031-C
20. Chen J, Hailing H, Tongguang Y (2019) Robust decentralized H∞ control for a multi-motor
web-winding system. IEEE Access 7:41241–41249. https://doi.org/10.1109/ACCESS.2019.
2906223
21. Hailiang H, Xiaohong N, Chen J, Dengfeng X (2018) Decentralized coordinated control of
elastic web winding systems without tension sensor. ISA Trans 80:350–359. https://doi.org/
10.1016/j.isatra.2018.06.006
Large Synchronverter Integration
in Power Electrical System: Impacts
on SCR and CCT
Raouia Aouini, Khadija Ben Kilani, Mohamed Elleuch,

and Quoc Tuan TRAN
Abstract This paper investigates the impacts of large synchronverters (SV) integra-
tion on transient stability and strength of electric power systems. Large Synchron-
verters account for large amounts of electrical energy injected into the power grid
through power converters, as in PV and off shore wind power sources. Potential
stability problems may occur when large power converters are connected to weak
power systems (PES). An SV is an inverter that mimics a synchronous generator
(SG). Methodologically, the SV performances are contrasted to the standard SG of
like parameters in terms of system strength and transient stability. These criteria
are quantified respectively by the Short-circuit Ratio (SCR) and the critical clearing
time (CCT). The penetration ratio of SV generation is increased to the detriment
of SG power. Due to their current controllers’ limitations, the SV participation in
short-circuit currents is reduced to nominal values, diminishing the network strength.
Although SVs have similar properties as SGs and have variable virtual inertia, this
study reveals many limitations that must be taken into account when planning their
integration into the network. In opposition, SGs enhance the transient stability margin
even in weaker systems. Hence, strong buses in a PES indicated by large SCR and
CCT feature larger amounts of SG power compared to the power generated by SVs.
Keywords Synchronverter · Network strength · Inertia · SCR · CCT
R. Aouini · K. Ben Kilani · M. Elleuch (B)

Université de Tunis El Manar, ENIT, L.S.E, LR11ES15, BP 37-1002 Tunis le Belvédère, Tunisie
R. Aouini
K. Ben Kilani
Q. T. TRAN
INSTN, CEA–INES, Saclay, France

https://doi.org/10.1007/978-981-15-6403-1_16
234 R. Aouini et al.
1 Introduction
Electric power systems are undergoing significant changes in the generation mix
and source types. Power electronic converters have evolved in number and capacity,
being an essential equipment in renewable energy sources (RES), such as wind farms,
solar PV plants, storage batteries and HVDC interconnection. Although, mostly are
equipped with fast-acting controls, many stability issues may arise when large power
converters are connected to weak power grids [1]. System strength is one of the
important concerns in the integration of RES.
Technically, a weak AC system may be evaluated using several measures: low
ratio of inductance over resistance, high impedance, and low inertia [2]. The short-
circuit ratio (SCR) at the point of interconnection (POI) has evolved as an indicator
of system strength. It is defined as the ratio of system MVA short-circuit capacity
(Scc ), to the MW rated power of the interconnected device [3], which could absorb or
inject power at the POI. The POI bus is considered strong if its SCR is above three. As
in indicator of system strength, the SCR has originally been used for HVDC systems
[3], then extended to RES featuring similar architecture and control as HVDC devices
[4, 5].
As a stability indicator, the strength of a power system at the POI affects its ability
to maintain its voltage and power quality under variable generation levels of RES.
Connecting RES to weaker portions of the grid may tremendously harm voltage
stability and quality at the POI [6], in particular transient stability. For instance, a
major fault occurring in an interconnected PES through weak transmission lines may
lead to transient instability. The criteria used for transient stability evaluation is the
Critical Clearing Time. This is because, in AC power grids integrating RES, large
disturbances are mainly compensated by large conventional synchronous generators.
However, as electronic power converters lack inertia, their increasing number dimin-
ishes the overall system inertia, which affects the dynamic and transient stability of
the power grid [7].
Consequently, power system operators have set rules requiring that RES actively
take part in power regulation of the grid, just like conventional SGs. Based on this idea,
the concept of Synchronverters was proposed. The aim is to render grid-connected
inverters emulate the essential behavior of SGs, including the droop mechanism
and inertial properties [8]. As a result, the SV should have the ability of providing
grid support by automatically adjusting its active and reactive powers to comply with
frequency and voltage grid requirements [11]. Meanwhile, the inertial characteristics
emulated by the SV contribute to the total inertia of the grid, enhancing thereafter
the system strength. Hence, the SV provides a promising technology for various
applications, such as HVDC transmission [9], STATCOM [10], and wind power
systems. One important advantage of the SV is that some system parameters, such
as inertia, can be suitably chosen to improve the system dynamic performances.
In comparison, SGs inertia are constant, and prime movers may have large time
constants in droop control loops.
Large Synchronverter Integration in Power … 235
In this paper, we investigate the transient stability of a four-bus power system,

considering five configurations of SV integrations. The SV parameters are adapted
to the conventional SG parameters. Dynamic responses of the SV and the SG
are compared under different operating conditions. The SV benefits its operation
frequency and voltage drooping mechanisms for load sharing. For each configura-
tion, strength at POI is evaluated by its SCR. Also, transient stability of the systems
is assessed in terms of the CCT, for different levels of the SCR at the load bus.
2 Overview of the Synchronveretr Technology
In this Section, the synchronverter concept is briefly reviewed. The synchronous

generator (SG) is the dominant type of generator in conventional electric power
systems [12]. SGs offer many advantages such as (i) supplying inertial response out
of the kinetic energy stored in the rotating masses; (ii) they can supply and absorb
reactive power; (iii) they have simple control structures with guaranteed grid stabi-
lizing responses. Based on these proprieties, an inverter based on the synchronverter
controls is developed to mimic a SG [8].
A synchronverter consists of a power part, which is the same as a conventional
power electronic converter depicted in Fig. 1a, and an electronic part, which consists
of the measurement and the control circuits. The power part of the SV is the standard
hardware of a three-phase pulse width modulation (PWM) inverter, with an LCL
filter at the output of each phase. Following the terminology defined in [8], the three
inductors with parameters Ls and Rs play the role of stator coils in a SG. The input
of the power part is the value of the desired synchronous internal voltage e from the
electronic part. The three phase inverter then generates a high-frequency switching
voltage signal eabc = [ea eb ec ]T which average value over a switching period is e.
The basic mathematical model of an SV is given in Eqs. (1–6).
dω
2H = (Pm − Pe )/ω (1)
dt
dω
θ= (2)
dt
Pm = Pr e f + D p (ωn − ω) (3)
Pe = ωE m (i a sin θ + i b sin(θ − 2π/3) + i c sin(θ + 2π/3)) (4)
1
Em = (Dq (Ur e f − Um ) + Q r e f − Q e ) (5)
Ks
Q e = −ωE m (i a cos θ + i b cos(θ − 2π/3) + i c cos(θ + 2π/3)) (6)

a)
b)
Fig. 1 Organization of the synchronverter. a the power part of SV, b the electronic part of SV
where H is the inertia constant, Pref and Qref are the given values of active and reactive
powers, Pe and Qe are the actual output values of active power and reactive power,
ωn and ω are the rated and actual values of electric angular velocity, θ is the rotor
angle, Dp and Dq are the droop coefficients of active and reactive loops, Um and Uref
are the actual and given values of grid voltage amplitude, Em is the internal potential
amplitude of the SV, K is the integral coefficient.
The electronic part of a three-phase SV, as shown in Fig. 1b, includes the mathe-
matical model (1–6) of a three-phase round-rotor synchronous machine as the core.
The SV controller strategy processes the electric measurements (vabc and iabc ) as well
as the Pref and the Qref which purpose is to generate e, using the following expression,
eabc = ωE m [sin θ sin(θ − 2π/3) sin(θ + 2π/3)]T (7)

The latter expression is passed through a PWM generation block to generate six
pulses to drive the power semiconductors in Fig. 1a.
The DC-DC converter and the PV panel or any other renewable source, presented
in Fig. 1a, can be represented by one equivalent constant voltage source noted VDC .
It is noted that there is a difference in hardware between a SV and SG: some modest
energy storage is required on the DC bus of the inverter to mimic the effect of rotor
inertia.
The mechanical equation of the SV rotor is expressed by (1). The SV control can
be divided into active loop control and reactive loop control as described by (3) and
(5) respectively. Its active loop includes active frequency droop control and iner-
tial response control, which mainly realizes the function of independent frequency
response. The reactive loop consists of reactive power-voltage droop control and end-
voltage closed-loop control, which achieve automatic voltage regulation and voltage
amplitude control of the SV [8].
3 Short-Circuit Ratio and System Strength
The Short-circuit Ratio is an indicator of the system strength, which is defined as the
ratio of system MVA short-circuit power to the MW rating power of the intercon-
nected device [3]. The strength of a power system at the POI of the device is viewed
as the ability of the system at the point to maintain its voltage stability and quality.
In this section, the relationship between the SCR and the system strength at the POI
of RES is explained.
For an HVDC interconnection, the SCR is used to quantify the strength at the POI
and its value determines how strong the AC system to support the stable operation
of HVDC converter. Moreover, the RES can either be on grid or off grid. When
connected to the grid, RES are typically interfaced to the distribution, far from the
main grid. The points of interconnection of these RES are generally weak, and voltage
stability problems are likely to occur [5, 6]. Therefore, it is necessary to measure
the system strength at all possible interconnection points. Like HVDC devices, the
commonly used SCR has been extended to quantify the strength at the POI of RES.
The SCR at bus i can be expressed as:
Sac,i |Vi |2 1
SC Ri = = . (8)
Pd,i Pd,i |Z i |
where Sac,i is the short-circuit capacity of the system at bus i; Pd,i is the MW rated
power of the RES or the load connected to bus i; |Vi | and |Z i | are respectively the
Thevenin voltage and impedance viewed at bus i.
From the expression of the SCR, it is clear the system strength is highly dependent
on the ratio of short-circuit capacity to the power of the interconnected device at a
specific point. The short-circuit capacity may also be expressed as:
Sac = |V0 |.|Icc | (9)
where |V0 | and |Icc | are the pre-fault bus voltage magnitude and fault current,
respectively.
4 The Testedfour Area Power System
The synchronverter control strategy is applied to the system described in Fig. 2.

The system is built based on MATLAB/Simulink toolbox. A series of time domain
simulations are performed to investigate the transient stability of the tested system,
considering five configurations. The impact of the SV integration on load point
strength and on its transient stability is assessed in terms of the CCT and SCR.
The proposed tests consist of four control areas designated as Area 1, Area 2, Area
3 and Area 4, feeding a common load area at POI, via AC lines as in Fig. 2. Two
different generation sources are studied: synchronverters and synchronous gener-
ators. SVs are modeled by (1–7), SGs of the same power rating are equipped
with power frequency and voltage reactive power controls described in (3) and (5),
respectively.
The study cases are defined as follows:
Fig. 2 Four areas and four SV/SG interconnected generation

– Case 1: Areas 1, 2, 3 and 4 are SVs designated as SV1, SV2, SV3 and SV4. This
case is named 4SV with SV integration ratio equals to 100%.
– Case 2: SV in Area 1 has been replaced byan equivalent SG (marked as SG1 in
Fig. 2). Areas 2, 3 and 4 arekept as SVs.
– Case 3: Areas 1 and 2 are SGs named SG1 and SG2, respectively.
This case is called 2SG-2SV with SV power integration ratio equal to 50%.
– Case 4: Areas 1, 2 and 3 are SGs. Area 4 is an SV. This case is designed as
3SG-1SVwithSV power integration ratio equals 25%.
– Case 5 named 4SG: Area 1, 2, 3 and 4 are SGs designated SG1, SG2, SG3 and
SG4with no SV power integration.
The generation sources are chosen to have identical generator, with voltage, power
and frequency ratings given respectively by V = 100 kV, Sn = 200 MVA, f = 50 Hz.
The nominal current of the SV and the SG is expressed by:
Sn 200
In = √ =√ = 1.15 kA (10)
3Vn 3.100
For the SV parameters, they are firstly adapted to match standard SG parameters
[9, 12]: a 5% governor droop (1/Dp-SV and 1/Dp-SG ), inertia constant H = 4 s, and3%
voltage regulation gain (1/DQ-SV and 1/DQ-SG ).
The main advantage of the SVs is to feature the structure of classic SGs, which
controls are well known. The SV operation is verified on the test power system
shown in Fig. 2. Dynamic and transient performances of SV are compared with SG
of same capacity under different faults.
A. Load share of the SV versus SG
An important mechanism for SVs and SGs for even load sharing is to adjust the
real power delivered to the grid according to the grid frequency. In this Section, the
dynamic performances of the cases 4SV and 3SV-1SG are compared. The responses
of these cases to a load change at t = 10 s are shown in Figs. 3 and 4. In Fig. 3, SV1,
SV2, SV3 and SV1 responses of the 4SV system are presented. Figure 4 depicts
the 3SV-1SG responses. The active powers, the speeds, the voltages and the reactive
powers dynamics are similar.
These results confirm that the SV performances are nearly identical to SG ones, for
steady state and small signal stability. The SV and the SG coupled via an AC lines as
shown in Fig. 2 are set in a parallel arrangement and are initially at nominal frequency
f0 = 1 p.u with power outputs, respectively, PSV1 = PSG1 = 0.8 p.u (Figs. 3a, b and 4a,
b show, a slow down of the speed response to an active load increase ΔPL = +0.8 p.u.
Fig. 3 Responses of the 4SV system toa +0.8 p.u step in the load. a Responses of the SVs active
powers in (p.u), b Responses of the SV speeds in (p.u), c Responses of the SVs reactive powers in
(p.u), d Responses of the SV voltages in (p.u)
The governors increase their power output until they reach a new common operating
frequency f . The frequency of the SG followed the SV frequency, and the real and
reactive powers tracked their set points. For example, for the 3SV-1SG system, the
amount of load picked up by each unit depends on the droop characteristics, such
that:
PL 0.8
f = f − f0 = = = 0.01 p.u. (11)
D SG1 + D SV 2 + D SV 3 + D SV 4 4 × 20
The SG1, SV2, SV3 and SV4 increased their real powers output by 0.2 p.u
PSG1 = f × D P-SG1 = 0.01 × 20 = 0.2 p.u (12)
PSV 2 = PSV 3 = PSV 4 = f × D P-SV = 0.01 × 20 = 0.2 p.u (13)
The amount of load increase corresponds to a 0.1% drop of the frequency. Hence,
PSV
PSG1
= DDP-SG1
P-SV
, where DP-SV and DP-SG1 are respectively, the static droop of SV and
SG, which are equal to 1/5%.
Fig. 4 Responses of the SV1 and the SG1 to +0.8 p.u step in the load. a Responses of the SVs
active powers in (p.u), b Responses of the SV speeds in (p.u), c Responses of the SVs reactive
powers in (p.u), d Responses of the SV voltages in (p.u)
Due to the load increase, the local terminal voltage Vabc decreased by about 0.6%
from nominal (as in Figs. 3d and 4d). In response, the reactive power output of SV1
and SG1increased by about 0.195 p.u (see Figs. 3c and 4c), depending on the droop
control value
VSG1 0.992 − 0.998
Q SG1 = − =− = 0.195 p.u. (14)
Dq-SG1 0.03
B. Evaluation of the SCR and the CCT for different ratio of SV power penetration
In this Section, the impact of large SVs integration is evaluated in terms transient
stability, and system strength at the POI, using Fig. 2. Transient stability is quantified
by the Critical Clearing Time (CCT), which is the maximum time duration that a
short- circuit may be applied without losing the system capacity to recover to a steady-
state stable operation. To evaluate the CCT, a three phase short-circuit was simulated
near the load. The power at the POI may be considered as a load, or RES power
injection. Different levels of SCRs are used to investigate the impact of the system
strength on transient stability. To do so, the transmission lines lengths (L) of the test
system in Fig. 2 are varied: (400, 300, 200 and 50 km) to change the SCR values at
POI. Tables 1, 2, 3 and 4 and Fig. 6 depict the CCT and the SCR for SV integration
cases and for different transmission lines lengths. The following methodology is
adopted. First, the test power system in Fig. 2 is replaced by an equivalent Thevenin
generator Eth in series with its impedance Zth (Fig. 5), as viewed from the terminals
a and b (at the POI) with all the generation sources (SV and SG).
The SCR at POI is evaluated as follows:
• Base Case: the system in Fig. 5 is in normal operating condition, and governed
by:
Table 1 CCTs, Current default Icc , Eth , Zth and the SCRs at POI for different ratios of SV integration
with a line length L = 400 km
Length 400 km

Cases CCT (s) I cc (p.u) E th (p.u) Z th (p.u) SCR
4SG 2.1 2.9 0.95 0.32 1.09
3SG-1SV 1.75 s 2.12 0.95 0.44 0.80
2SV-2SG 0.9 s 1.65 0.949 0.57 0.64
3SV-1SG 0.75 s 1.3 0.948 0.72 0.49
4SV 0.6 s 1.27 0.948 0.74 0.48
Table 2 CCTs, Current default Icc , Eth , Zth and the SCRs at POI for different ratio of SV integration
Length 300 km

4SG 2.5 3.53 0.965 0.27 1.52
3SG-1SV 1.85 2.82 0.965 0.34 1.08
2SV-2SG 1 1.72 0.964 0.56 0.84
3SV-1SG 0.8 1.55 0.964 0.62 0.59
4SV 0.75 1.52 0.964 0.63 0.58
Length 200 km

4SG 2.8 4.6 0.97 0.21 1.77
3SG-1SV 2.3 3.7 0.97 0.26 1.42
2SV-2SG 1.1 2.4 0.969 0.4 0.92
3SV-1SG 0.9 2 0.969 0.48 0.77
4SV 0.8 1.90 0.969 0.51 0.73
Length 50 km

4SG 3.4 10.6 0.99 0.09 4.18
3SG-1SV 2.8 7.76 0.99 0.12 3.06
2SV-2SG 2.2 4.6 0.989 0.21 1.81
3SV-1SG 1.75 3.5 0.989 0.28 1.38
4SV 1.45 3.4 0.989 0.29 1.34
Fig. 5 Thevenin equivalent

circuit as seen from POI
V L = E th − Z th .I L (15)

from which I L and V L have been measured.
• Contingency Case: athree phase short-circuit fault is applied at the POI bus, and
the fault current I cc is measured for different cases as summarized in Tables 1,
2, 3 and 4. The fault current is given by:
E th
I cc = (16)
Z th
Then, the Thevenin voltage is computed as:
V L .I cc
E th = (17)
I cc − I L
Finally, the SCRs in Tables 1, 2, 3 and 4, are given by:

sac, poi
scr poi = (18)
pd, poi
Fig. 6 The CCTs and the SCRs of different configurations and line length. a CCTs, b SCRs
where Sac,POI is the short-circuit capacity and Pd,POI is the load power at POI. As
indicated in Eq. (9), Sac,POI is calculated by:

Sac,P O I = E th . I cc (19)
The load power Pd, POI is equal to 500 MW which may also be considered as the
RES power to be injected at the POI.
C. The impact of the SVs integrations on the CCT and the SCR
First, it can be observed from Tables 1, 2, 3 and 4 (column 2) that the best CCT is
for case 4SG. The CCT decreased with increasing ratios of SV integration, and with
the decoupling between generators by increasing their tie lines. Figure 6a illustrates
the worst CCTs for the case with full SV integration and no SG.
Second, it is confirmed that the strength of the grid at POI is well reflected by its
SCR and its fault current Icc . In fact, the I cc which is almost proportional to the
SCR, has higher values when the more power is generated by conventional SG with
less SV integration. It is obvious that SGs supply fault currents up to several times
their rated current (Table 4), whereas converters have to reduce their fault current
up to almost their nominal values [13], which results in low participation in the
fault current Icc . This leads to the decrease of SCR, and network strength reduction.
Tables 1, 2, 3 and 4 and Fig. 6b confirm the increase of Icc , SCR, and the CCT with the
generation sources proximity to the POI. The best CCT and SCR values are obtained
for the shortest tie-line (50
km).
As shown in column 3 of Tables 1, 2, 3 and 4, the
fault current magnitudes I cc are reduced as the line length increased, specifically,
as the Thevenin impedance magnitude Z th increased (as in column 5).
In addition, as it can be seen from Fig. 6b and Table 4 (column 3 and 5), with
a line length L = 50 km, higher fault current values are depicted, because of the
reduction of the Thevenin equivalent impedance
Z th seen from the POI bus which is considered strong. A higher fault current
value reflects a strong response of the generation in a PES, which produces, at the
fault location, a higher current and a voltage drop. Nonetheless, in Table 1 with a
Table 5 The CCTs for different inertia constant

Constant inertia H (s) CCT(s)
0.5 2 4 10 15 20
3SG-1SV 1.5 2.1 2.4 2.5 2.8 Unstable
2SV-2SG 0.95 1 1.2 1.8 2 Unstable
3SV-1SG 0.75 0.8 1 1.15 1.5 Unstable
4SV 0.22 0.45 0.98 1.1 1.4 Unstable

line length L = 400 km, the fault current levels I cc decreased when the POI bus is
considered weak, since the SCR is smaller than 2. Similar observations for the CCT
can be made based on data in Tables 1, 2, 3 and 4 and Fig. 6a, where the CCT tends
to overestimate the system strength in terms of transient stability. The lower CCT
is observed with weaker POI bus; the higher CCT indicates strong POI bus. Thus,
strong buses in a PES indicated by large SCR and CCT values feature higher power
quantity from SGs and less from SVs. The higher values of fault current at theses
buses tend to maintain stable operation during disturbances.
D. The effect of the SV inertia constant on the CCT
One important advantage of SVs over SGs is the possibility to vary their parameters,
like the inertia constant, to improve power system stability. In this section, the inertia
constant H of cases with SV integration is varied to enhance the transient stability in
term of CCT. The SG inertia constant H is fixed to 4 s, while the SV inertia constant
is increased up to 20 s. The results summarized in Table 5, confirm the improvement
of transient stability with increased inertia, limited by a threshold value, beyond
which the system becomes instable. For the system strength, the inertia constant has
no significant effect on the SCR, since it doesn’t affect the fault current Icc . This is
another limitation of SV performance.
6 Conclusion
This paper dealt with a comparative analysis of a SV/SG in four-areas network. The
performance criteria used are CCT and SCR. With appropriate controls, it is shown
that SVs are able to recover many performances of SGs in terms of frequency and
voltage droop control.
In addition, thanks to its variable virtual inertia, SV has improved the transient
stability of the tested network. However, the SV has some limitations highlighted
along this work which consisted of:
– Low participation in the fault current leading to low SCR and consequently a
reduction of the network strength;
– Maximum value of the inertia constant beyond which the system loses its stability;
– The increase of the integration ratio of SVs power has a negative impact on
network performances (by reducing the SCR and CCT) as opposed to SGs.
References
1. Urdal H, Ierna R, Zhu J, Ivanov C, Dahresobh A, Rostom D (2015) System strength

considerations in a converter dominated power system. IET, Renew. Power Gen 9(1):10–17
2. Gavrilovic A (1991) AC/DC system strength as indicated by short-circuit ratios. In: Inter-
national conference on AC and DC power transmission, September 17–20, 1991, pp
27–32
3. Jia C, Wen J, Bie X (2012) Study of the effect of AC system strength on the HVDC startup char-
acteristics. In: International conference on sustainable power generation and supply, September
8–9, 2012
4. D. Wu, G. Li, M. Javadi (2018) Assessing impact of renewable energy integration on system
strength using site-dependent short-circuit ratio. IEEE Trans Sustain Energy 9:1072–1080
5. Zhang Y, Huang SH, Schmall J, Conto J, Billo J, Rehman E (2014) Evaluating system strength
for large-scale wind plant integration. In: IEEE power and energy society general meeting, July
2014, pp 1–5
6. Huang SH, Schmall J, Conto J, Zhang Y, Li Y, Billo J (2015) Voltage stability of large-scale
wind plants integrated in weak networks: an ERCOT case study. IEEE power and energy society
general meeting, July 2015, pp 1–5
7. Nkechi AI, Howlader AM, Yona A (2018) Integration of photovoltaic energy to the grid, using
the virtual synchronous generator control technique. J Energy Power Eng 12:329–339
8. Zhong Q-C, Weiss G (2011) Synchronverters: inverters that mimic synchronous generators.
IEEE Trans Ind Electron 58(4):1259–1267
9. Aouini R, Marinescu B, Ben Kilani K, Elleuch M (2016) Synchronverter-based emulation and
control of HVDC transmission. IEEE Trans Power Syst (1):278–286
10. Aouini R, Nevzi I, Ben-Kilani K, Elleuch M (2016) Exploitation of synchronverter control to
improve the integration of renewable sources to the grid control. J Electr Syst 12–3:515–528
11. Nguyen P-L, Zhong Q-C, Blaabjerg F, Guerrero J-M (2012) Synchronverter-based OPERA-
TION OF STATCOM to mimic synchronous condensers. In: 2012 7th IEEE conference on
industrial electronics and applications (ICIEA)
12. Kundur P (1994) Power system stability and control. Mc Graw-HillInc
13. Callavik M, Blomberg A, Häfner J, Jacobson B (2012) The hybrid HVDC breaker—aninno-
vation breakthrough enabling reliable HVDC grids, Technical Paper, ABBGrid Systems
Immersion and Invariance Based
Adaptive Dynamic Surface Control
for Parametric Strict-Feedback
Nonlinear Systems
Y. Soukkou, S. Labiod, M. Tadjine, Q. M. Zhu, and M. Nibouche
Abstract This paper presents an immersion and invariance based adaptive dynamic
control scheme for a class of parametric strict-feedback nonlinear systems. By using
dynamic surface control, the problem of explosion of complexity inherent in the
conventional backstepping control method is avoided. The uniformly ultimately
boundedness (UUB) of all signals in the closed-loop system is performed based
on the Lyapunov stability analysis theory. Simulation results for the control of a
one-link manipulator actuated by a brush DC motor are provided to demonstrate the
effectiveness of the proposed adaptive control method.
Keywords Adaptive control · Immersion and invariance · Dynamic surface

control · Lyapunov stability theory · One-link manipulator actuated by a brush DC
motor
1 Introduction
Over the past few decades, adaptive backstepping control method is one of the most
popular and effective control approaches for a class of parametric strict-feedback
nonlinear systems, which has received much attention [1–5]. Nevertheless, the main
drawback in this approach is the overparametrization problem. Tuning functions
Y. Soukkou · M. Tadjine
LCP, Department of Automatic Control, National Polytechnic School (ENP), 10, Av. Hassen
Badi, BP. 182, Algiers, Algeria
Y. Soukkou (B)
Research Center in Industrial Technologies CRTI, P. O. Box. 64, 16014 Cheraga, Algiers, Algeria
S. Labiod
LAJ, Faculty of Science and Technology, University of Jijel, BP. 98, Ouled Aissa, Jijel, Algeria
Q. M. Zhu · M. Nibouche
Department of Engineering Design and Mathematics, University of the West of England,
Frenchay Campus, Coldharbour Lane, Bristol BS16 1QY, UK

https://doi.org/10.1007/978-981-15-6403-1_17
248 Y. Soukkou et al.
based adaptive backstepping control design is proposed to avoid the overparametriza-

tion problem [2–5]. However, adaptive backstepping control and tuning functions
based adaptive backstepping control designs suffer from the problem of explo-
sion of complexity, which is caused by the repeated derivations of virtual control
inputs [6–8]. Recently, a dynamic surface control has been proposed to eliminate
the problem of explosion of complexity by introducing a first-order low-pass filter
at each step of the conventional backstepping control method [6–8]. In [6, 9], the
dynamic surface control method has extended to adaptive control. For parametric
strict-feedback nonlinear systems, adaptive dynamic surface control has been widely
applied [10–13].
In recent years, immersion and invariance based adaptive control of nonlinear
systems was proposed by Astolfi and Ortega [14]. It has been established for control-
ling of class of parametric strict-feedback nonlinear systems [15, 16]. It has been also
a great attention in the few recent years, which has been widely used to the control
of a class of parametric strict-feedback nonlinear systems [17, 18].
During the last few years, the control methodologies of the one-link manipulator
actuated by a brush DC motor have drawn considerable interest [11, 19–22]. Several
nonlinear control methodologies and adaptive approaches have been also studied and
applied to electromechanical systems.
This paper focuses on an immersion and invariance based adaptive dynamic
surface control for a class of parametric strict-feedback nonlinear systems. Immer-
sion and invariance based adaptive dynamic surface control is introduced to avoid
the problem of explosion of complexity. The uniformly ultimately boundedness
(UUB) of all signals in the closed-loop system is proven based on the Lyapunov
stability theory. Simulation results for a one-link manipulator actuated by a brush
DC motor are provided to demonstrate the effectiveness of the proposed adaptive
control scheme.
The remaining of the paper is organized as follows. The problem formulation
and preliminaries are presented in Sect. 2. Section 3 is devoted to the design of
immersion and invariance based adaptive dynamic surface control. Stability analysis
is discussed in Sect. 4. Simulation results are included in Sect. 5. Conclusions are
given in Sect. 6.
2 Problem Formulation and Preliminaries
We will consider the following nonlinear system in parametric strict-feedback form:
ẋ1 = g1 (x1 )x2 + ϕ1T (x1 )θ + ψ1 (x1 )

ẋi = gi (x̄i )xi+1 + ϕiT (x̄i )θ + ψi (x̄i ), i = 2, · · · , n − 1
ẋn = gn (x)u + ϕnT (x)θ + ψn (x) (1)
Immersion and Invariance Based Adaptive Dynamic Surface Control … 249
T
where, x = x1 x2 · · · xn ∈ Rn and u ∈ R are system states and the control
T
input. x̄i = x1 x2 · · · xi and θ ∈ R p are unknown constant parameter vector.
The nonlinear functions ϕi , ψi and gi = 0 are known and continuous. The control
T
objective is to design the controller u such that the output x1 tracks the desired
trajectory x1d and to ensure the uniformly ultimately boundedness (UUB) of all
signals in the closed-loop system.
Assumption 1 There exists a positive constant g0 such that |gi (x̄i )| ≥ g0 , i =

1, · · · , n.
Assumption 2 The desired trajectory x1d and its derivatives ẋ1d and ẍ1d are known,
continuous and bounded.
3 Immersion and Invariance Based Adaptive Dynamic

Surface Control
The design procedures of the proposed adaptive control consisted of a general two
steps, while the first step is to design an estimator and the second step is to design a
control law.
A. Estimator Design
We define the estimation errors as [16]:

θi = θ i − θ + βi (x̄i ) (2)

where, θ i are the estimator states and βi : Ri → R p are functions yet to be specified.
Then, the estimation dynamics are given by [16]:
∂βi
i i ∂βi
θ̃˙i = θ̂˙i + ẋk = θ̂˙i + gk xk+1 + ϕkT θ̂i + βi −
θi + ψk (3)
k=1
∂ xk k=1
∂ xk
where, xn+1 = u. Selecting the adaptive laws θ̂i as:
i ∂βi
θ̂˙i = −

gk xk+1 + ϕkT θ i + βi + ψk (4)

k=1
∂ xk
yields the error dynamics:

i
˙θ̃ = − ∂βi ϕ T θi (5)

i
k=1
∂ xk k
where, the functions βi are selected as [16]:
xi
βi = i ϕi (χ )dχ + εi (6)
0
where, i > 0 are constants and εi are functions with ε1 = 0.
Assumption 3 There exist functions εi satisfying the partial differential (matrix)

inequality [16]:
Fi + FiT ≥ 0 (7)
for, i = 2, · · · , n − 1, where,
⎛ x ⎞

i−1 i
∂ ⎝ ∂εi T
Fi = i ϕi (χ )dχ ⎠ϕkT + ϕ (8)
k=1
∂ xk ∂ xi i
0
Remark 1 In the special case when ϕi (·) is a function of xi only, the partial differential
(matrix) inequality (7) admits the trivial solution εi = 0 for, i = 2, · · · , n [16].
Lemma 1 Consider the system (1), where the functions βi are given by (6) and func-
tions εi exist which satisfy (7). Then, the system (1) has a uniformly globally stable
θi ∈ L ∞ (bounded) and ϕiT
equilibrium at the origin, θi ∈ L 2 (square integrable). In
addition, ϕi and its time derivative ϕ̇i are bounded, then ϕiT
θi converges to zero.
Proof We consider the Lyapunov function W as:
1 T −1
n
W =
θ θi (9)
2 i=1 i i
The derivative of the Lyapunov function Ẇ becomes:
1 ˙ T −1
n n
Ẇ = θ̃i i θiT i−1 θ̃˙i = −
θi +
θiT i−1 i ϕi ϕiT + Fi + FiT
θi
2 i=1 i=1

n
n
T 2
≤−
θiT ϕi ϕiT
θi ≤ − ϕi
θi (10)
i=1 i=1
By integrating of inequality (10) over [0, ∞], we obtain:
∞
n
T 2
ϕi
θi dτ ≤ W (0) − W (∞) < ∞ (11)
0 i=1
As a result, W ∈ L ∞ , θi ∈ L ∞ and ϕiT

θi ∈ L 2 . Furthermore, ϕi ∈ L ∞ and
ϕ̇i ∈ L ∞ . By using Barbala’s lemma [23], we get limt→∞ ϕiT
θi = 0.
B. Controller Design
The controller design presented here includes adaptive dynamic surface control.
Step 1(i = 1): Define the first surface error as S1 = x1 − x1d , and the time
derivative of S1 is defined as:
Ṡ1 = g1 x2 + ϕ1T θ + ψ1 − ẋ1d (12)
The virtual control x̄2 is chosen as:
1 T

x̄2 = −ϕ1 θ 1 + β1 − ψ1 + ẋ1d − K 1 S1 − g1 S1 (13)
g1
where, K 1 > 0. We introduce a new variable x2d and let x̄2 pass through a first order
filter with time constant τ2 to obtain x2d as:
τ2 ẋ2d + x2d = x̄2 , x2d (0) = x̄2 (0) (14)

1 1 T

ẋ2d = −x2d + −ϕ1 θ 1 + β1 − ψ1 + ẋ1d − K 1 S1 − g1 S1 (15)
τ2 g1
Step i(i = 2, · · · , n − 1): Define the ith surface error Si = xi − xid , and it’s time
derivative is defined as:
Ṡi = gi xi+1 + ϕiT θ + ψi − ẋid (16)
The virtual controllers x̄i+1 are chosen as:
1 T

x̄i+1 = −ϕi θ i + βi − ψi + ẋid − K i Si − gi Si (17)
gi
where, K i > 0. We introduce a new variable x(i+1)d and let x̄i+1 pass through a first
order filter with time constant τi+1 to obtain x(i+1)d as:
τi+1 ẋ(i+1)d + x(i+1)d = x̄i+1 , x(i+1)d (0) = x̄i+1 (0) (18)

1 1 T

ẋ(i+1)d = −x(i+1)d + −ϕi θ i + βi − ψi + ẋid − K i Si − gi Si
τi+1 gi
(19)
Step n: Define the nth surface error Sn = xn − xnd , and its time derivative is
defined as:
Ṡn = gn u + ϕnT θ + ψn − ẋnd (20)
The actual control input u is given as follow:
1 T

u= −ϕn θ n + βn − ψn + ẋnd − K n Sn (21)
gn
where, K n > 0.
4 Stability Analysis
Define the boundary layer error as [9, 10, 12, 13]:
yi = xid − x̄i , i = 2, · · · , n (22)
and the parameter estimation errors as [16]:

θi = θ i − θ + βi , i = 1, 2, · · · , n (23)
Then, the closed-loop dynamics can be expressed in terms of the surface errors
Si , the boundary layer errors yi , and the parameter estimation errors
θi .
The surface errors dynamics are expressed, for i = 1, as:
Ṡ1 = ẋ1 − ẋ1d = g1 x2 + ϕ1T θ + ψ1 − ẋ1d

= g1 S2 + g1 x2d + ϕ1T θ + ψ1 − ẋ1d
= g1 S2 + g1 y2 + g1 x̄2 + ϕ1T θ + ψ1 − ẋ1d
= g1 S2 + g1 y2 − K 1 S1 − ϕ1T θ1 − g12 S1 (24)
For i = 2, · · · , n − 1:
Ṡi = ẋi − ẋid = gi xi+1 + ϕiT θ + ψi − ẋid

= gi Si+1 + gi x(i+1)d + ϕiT θ + ψi − ẋid
= gi Si+1 + gi yi+1 + gi x̄i+1 + ϕiT θ + ψi − ẋid
= gi Si+1 + gi yi+1 − K i Si − ϕiT
θi − gi2 Si (25)
For i = n:
Ṡn = ẋn − ẋnd = gn u + ϕnT θ + ψn − ẋnd = −K n Sn − ϕnT

θn (26)
The boundary layer errors dynamics are expressed, for i = 2, as:

1 1
ẏ2 = ẋ2d − x̄˙2 = (−x2d + x̄2 ) − x̄˙2 = − y2 − x̄˙2 (27)
τ2 τ2
For i = 3, · · · , n:
1 1
ẏi = ẋid − x̄˙i = (−xid + x̄i ) − x̄˙i = − yi − x̄˙i (28)
τi τi
We consider the following Lyapunov function as:
1 2 1 1 T −1
Vi S = S , Vi y = yi2 , Viθ =
θ θi (29)
2 i 2 2 i i
where,

n
n
n
V = Vi S + Vi y + Viθ (30)
i=1 i=2 i=1
Then, for i = 1, · · · , n − 1, by some simple computations, one has,
V̇i S = Si Ṡi = gi Si Si+1 + gi Si yi+1 − K i Si2 − Si ϕiT

θi − gi2 Si2 (31)
Applying of the following Young’s inequalities:
1 2 2 1 2
gi Si Si+1 ≤ g S + Si+1 (32)
2 i i 2
1 2 2 1 2
gi Si yi+1 ≤ g S + yi+1 (33)
2 i i 2
The following Young’s inequalities hold:
1 2 1 2
V̇i S ≤ −K i Si2 + Si+1 + yi+1 − Si ϕiT
θi (34)
2 2
and, for i = n:
V̇nS = Sn Ṡn = −K n Sn2 − Sn ϕnT

θn (35)
one has, ẋ1 = Ṡ1 + ẋ1d and ẋi = Ṡi + ẏi + x̄˙i , then from:
T
˙x̄2 = 1 − ∂ϕ1 ẋ1 θ 1 + β1 − ϕ1T θ̂˙1 + β̇1

g1 ∂ x1

∂ψ1 ∂g1
− ẋ1 + ẍ1d − K 1 Ṡ1 − g1 Ṡ1 − ẋ1 S1 (36)
∂ x1 ∂ x1
and by induction, it is easy from:

⎛ ⎛ ⎞

i−1
∂ϕi−1
T
θ̂˙i−1 + β̇i−1
1
x̄˙i = ⎝−⎝ ẋ j ⎠ θ i−1 + βi−1 − ϕi−1

T
gi−1 j=1
∂x j
⎞ ⎛ ⎞

i−1
∂ψi−1
i−1
∂g
ẋ j + ẍ(i−1)d − K i−1 Ṡi−1 ⎠ − gi−1 Ṡi−1 − ⎝ ẋ j ⎠ Si−1
i−1
−
j=1
∂ x j j=1
∂ x j
(37)
one has,
T
∂ϕ1
ẋ1 θ 1 + β1 − ϕ1T θ̂˙1 + β̇1
1 1
ẏ2 = − y2 − −
τ2 g1 ∂ x1

∂ψ1 ∂g1
− ẋ1 + ẍ1d − K 1 Ṡ1 + g1 Ṡ1 + ẋ1 S1 (38)
∂ x1 ∂ x1
Since all terms in (38) can be dominated by some continuous functions, it follows
that:

1 1

ẏ2 + y2 ≤ ẏ2 + y2 ≤ B2 S1 , S2 , y2 , θ 1 , β1 , x1d , ẋ1d , ẍ1d (39)

τ2 τ2
where, B2 (.) is a continuous function. Thus:
1 2
y2 ẏ2 + y ≤ B2 (.)|y2 | (40)
τ2 2
By using Young’s inequality, it gives:
1 2 y22 B 2 (.)
y2 ẏ2 ≤ − y2 + + 2 (41)
τ2 2 2
as similarly for i = 3, · · · , n, we obtain:
1 2 yn2 B 2 (.)
yn ẏn ≤ − yn + + n (42)
τn 2 2

where, Bn (.) = Bn S1 · · · Sn , y2 · · · yn , θ 1 · · · θ n , β1 · · · βn , x1d , ẋ1d , ẍ1d is a
continuous function. From (41) and (42), we can write:
1 2 yi2 B 2 (.)
V̇i y ≤ − yi + + i , i = 2, · · · , n (43)
τi 2 2
From (10), one has,

2
V̇iθ ≤ − ϕiT
θi (44)
Therefore, the derivative of the Lyapunov function V̇ becomes:

n
n
n
V̇ = V̇i S + V̇i y + V̇iθ
i=1 i=2 i=1
n−1

1 2 1 2
≤ −K i Si + Si+1 + yi+1 − Si ϕi θi − K n Sn2 − Sn ϕnT
2 T
θn
i=1
2 2
n
1 y2 B 2 (.)
n
T 2
+ − yi2 + i + i − ϕi
θi (45)
i=2
τi 2 2 i=1
Applying of the following inequality:
1 2 1 T 2
−Si ϕiT
θi ≤ S + ϕi θi (46)
2 i 2
We obtain:

1 2
n−1
V̇ ≤ − K 1 − S1 − (K i − 1)Si2 − (K n − 1)Sn2
2 i=2
n
1 T 2
n
1
− − 1 yi2 + ϕ − ϕ θi (47)
i=2
τi 2 i=1 i

n
where, ϕ = Mi
2
and Mi are the maximums of Bi (.). By choosing the design
i=2
parameters such that, K 1 > 21 , K i > 1, K n > 1 and τi < 1, it can be concluded that,
V , Si , yi and ϕiT θi are bounded. Furthermore, all signals in the closed-loop system,

i.e., xi , xid , ẋid , x̄2 , · · · , x̄i+1 , u and θ i + βi are also bounded.
Remark 2 To robustify the adaptive laws, we introduce a σ-modification term into

the adaptive laws (4). Then, we choose the adaptive laws for θ i as:
∂βi
i
θ̂˙i = −

gk xk+1 + ϕkT θ i + βi + ψk − i σθi θ i + βi (48)

k=1
∂ xk
where, σθi are small design positive constants. The estimation dynamics are given
by:
i

˙θ̃ = − ∂βi ϕ T

i k θi − i σθi θ i + βi (49)
k=1
∂ xk
one has, V̇i S , V̇nS and V̇i y are expressed in (34), (35), (43), and,
2

V̇iθ ≤ − ϕiT
θi − σθi
θiT θ i + βi (50)
By using of inequality (46), the derivative of the Lyapunov function V̇ becomes:

n
1 2
n−1
1
V̇ ≤ − K 1 − S1 − (K i − 1)Si2 − (K n − 1)Sn2 − − 1 yi2
2 i=2 i=2
τi
1 T 2
n n

+ϕ− ϕi
θi − σθi
θiT θ i + βi (51)
2 i=1 i=1
From Young’s inequality, it is easy to show that:

1 T 1
θiT θ i + βi ≤ −
− θi
θi + θ 2 (52)
2 2
Substituting (52) into (51) yields:
n
1 2
n−1
1
V̇ ≤ − K 1 − S1 − (K i − 1)Si2 − (K n − 1)Sn2 − − 1 yi2
2 i=2 i=2
τi
1 T 2 σθi T σθ
n n n
+ϕ− ϕi
θi −
θi
θi + i
θ 2
2 i=1 i=1
2 i=1
2
≤ −π V + μ (53)

where, π = min 2 K 1 − 21 , 2(K i − 1), 2(K n − 1), 2 τ1i − 1 , i σθi and μ =
n
σθi
ϕ+ 2
θ 2 .
i=1
Theorem 1 Consider the closed-loop system composed of plant described by (1).

Suppose that Assumptions 1 and 2 are satisfied. Then, the virtual controllers (13) and
(17), the actual control input (21) and the parameter adaptive laws (48) guarantee
that all signals in the closed-loop system are uniformly ultimately bounded (UUB)
and the surface errors converge to a sufficiently small neighborhood of the origin by
appropriately adjusting the design parameters.
Proof Multiplying both sides in (53) by eπt yields:

d
V (t)eπt ≤ μeπt (54)
dt
Moreover, by integrating (54) over [0, t], we obtain:
μ μ −πt
0 ≤ V (t) ≤ + V (0) − e (55)
π π
μ
Since π
> 0, it can be obtained that:
μ
0 ≤ V (t) ≤ V (0)e−πt + (56)
π
Therefore, we know that, Si , yi , ϕiT

θi and θi are UUB. Furthermore, all signals in

the closed-loop system, i.e., ϕi , ϕ̇i , xi , xid , ẋid , x̄2 , · · · , x̄i+1 , u and θ
i + βi are
n
also UUB. In addition, from (30) and (56), it follows that: S = Si2 ≤
i=1
√
−0.5πt
2V (0)e + 2μ π . Accordingly, when t → ∞, it is easy to show that:

S ≤ 2μ π . This completes the proof.
The dynamics of a one-link manipulator actuated by a brush DC motor are described

as follows [20–22]:
D q̈ + B q̇ + N sin(q) = I
M I˙ + H I + K m q̇ = V (57)
Denote, x1 = q, x2 = q̇ and x3 = I , then, (57) can be rewritten as:
ẋ1 = x2
ẋ2 = f 2 (x̄2 ) + b2 x3
ẋ3 = f 3 (x) + b3 u (58)
where, f 2 (x̄2 ) = − ND sin(x1 ) − DB x2 , b2 = D1 , f 3 (x) = − KMm x2 − M

H
x3 , b3 = M1
and u = V . The parameters of the one-link manipulator actuated by a brush DC
motor are given by [11, 19, 22]: D = 1, B = 1, M = 0.05, H = 0.5, N = 10 and
K m = 10. The values of unknown constant parameters are assumed that: θ2 = ND and
θ3 = KMm . The objective of this simulation is to design the proposed adaptive controller
u in such a way that the link angular position q tracks the reference trajectory [20]:
π −0.1t 2
x1d = 2 sin(t) 1 − e .
We define the signals x2d and x3d by:
1
ẋ2d = (−x2d + ẋ1d − K 1 S1 − S1 ) (59)
τ2
B
1 1
ẋ3d = −x3d + sin(x1 ) θ 2 + β2 + x2 + ẋ2d − K 2 S2 − b2 S2 (60)

τ3 b2 D
The actual control input u is given by:

H
1
u= x2 θ 3 + β3 + x3 + ẋ3d − K 3 S3 (61)
b3 M
The adaptive laws are defined as:

B
∂β2 ∂β2
θ̂˙2 = −

x2 − b2 x3 − sin(x1 ) θ 2 + β2 − x2
∂ x1 ∂ x2 D

− 2 σθ2 θ 2 + β2 (62)

˙θ̂ = − ∂β3 x − ∂β3 b x − sin(x ) θ + β − B x

3 2 2 3 1 3 3 2
∂ x1 ∂ x2 D
H
∂β3
− b3 u − x2 θ 3 + β3 − x3 − 3 σθ3 θ 3 + β3 (63)
∂ x3 M
where, β2 = −2 x2 sin(x1 ) and β3 = −3 x2 x3 . The initial conditions are chosen
T
as: x(0) = 0 0 0 , θ 2 (0) = θ 3 (0) = 0 and x2d (0) = x3d (0) = 0. The control
parameters are chosen as: K 1 = K 2 = 50, K 3 = 500, 2 = 3 = 0.5, σθ2 = σθ3 =
0.01 and τ2 = τ3 = 10−3 . The simulation results are shown in Figs. 1, 2, 3 and 4.
From these Figures, we can obviously see that all states asymptotically converge
to their desired values, the ultimate surface errors converge to zero, and the conver-
gence of the parameter estimates to their true values is guaranteed. We can conclude
then that the proposed adaptive control method provides better trajectory tracking
performances.
6 Conclusion
In this paper, a new design of immersion and invariance based adaptive dynamic
surface control for parametric strict-feedback nonlinear systems is proposed. The
proposed adaptive control method is designed to avoid the explosion of complexity
2
x1d
1.5
Angular position [rad]

x1
1
0.5
0
-0.5
-1
-1.5
-2
0 5 10 15 20 25
Time [sec]
2
Angular velocity [rad/sec]
x2d
1.5
x2
1
0.5
0
-0.5
-1
-1.5
-2
0 5 10 15 20 25
Time [sec]
Motor armature current [A]
10
x3d
8
6 x3
4
2
0
-2
-4
-6
-8
-10 0 5 10 15 20 25
Time [sec]
Fig. 1 Trajectories of the output variables
0.3
S1
0.25
0.2 S2
Surface errors
0.15 S3
0.1
0.05
0
-0.05
-0.1
-0.15
-0.2
0 5 10 15 20 25
Time [sec]
Fig. 2 Trajectories of the surface errors

12
10
Parameter estimate 8
4 θ1
θ̂1
2
0
0 5 10 15 20 25
Time [sec]
250
Parameter estimate
200
150
100
θ2
50 θ̂2
0
0 5 10 15 20 25
Time [sec]
Fig. 3 Trajectories of the parameter estimates
25
20 u
Input control voltage [V]
15
10
5
0
-5
-10
-15
-20
-25
0 5 10 15 20 25
Time [sec]
Fig. 4 Trajectory of the control input
problem. Based on the Lyapunov stability theory, it has been proven that the proposed
adaptive control method guarantee the uniformly ultimately boundedness (UUB) of
all signals in the closed-loop system and the surface errors converge to a small
neighborhood of the origin. The simulation results clearly show the effectiveness of
the proposed adaptive control scheme.
References
1. Kanellakopoulos I, Kokotović PV, Morse AS (1991) Systematic design of adaptive controllers

for feedback linearizable systems. IEEE Trans Autom Control 36(11):1241–1253
2. Krstić M, Kanellakopoulos I, Kokotović PV (1995) Nonlinear and adaptive control design.
John Wiley & Sons, New York
3. Krstić M, Kanellakopoulos I, Kokotović PV (1992) Adaptive nonlinear control without
overparameterization. Syst Control Lett 19(3):177–185
4. Soukkou Y, Labiod S (2015) Adaptive backstepping control using combined direct and indirect
adaptation for a single-link flexible-joint robot. Int J Ind Electron Driv 2(1):11–19
5. Zhou J, Wen C (2008) Adaptive backstepping control of uncertain systems: nonsmooth
nonlinearities, interactions or time-variations. Springer-Verlag, Berlin/Heidelberg
6. Hedrick JK, Yip PP (2000) Multiple sliding surface control: theory and application. J Dyn Syst
Meas Contr 122(4):586–593
7. Swaroop D, Hedrick JK, Yip PP, Gerdes JC (2000) Dynamic surface control for a class of
nonlinear systems. IEEE Trans Autom Control 45(10):1893–1899
8. Swaroop D, Gerdes JC, Yip PP, Hedrick JK (1997) Dynamic surface control of nonlinear
systems. In: Proceedings of the American control conference, Albuquerque, New Mexico, pp
3028–3034
9. Yip PP, Hedrick JK (1998) Adaptive dynamic surface control: a simplified algorithm for
adaptive backstepping control of nonlinear systems. Int J Control 71(5):959–979
10. Hou MZ, Duan GR (2011) Robust adaptive dynamic surface control of uncertain nonlinear
systems. Int J Control Autom Syst 9(1):161–168
11. Li TS, Wang D, Feng G, Tong SC (2010) A DSC approach to robust adaptive NN tracking
control for strict-feedback nonlinear systems. IEEE Trans Syst Man Cybern Part B Cybern
40(3):915–927
12. Soukkou Y, Labiod S, Tadjine M (2018) Composite adaptive dynamic surface control of
nonlinear systems in parametric strict-feedback form. Trans Inst Meas Control 40(4):1127–
1135
13. Soukkou Y, Labiod S (2017) Adaptive backstepping control using combined direct and indirect
σ-modification adaptation. Lecture Notes in Electrical Engineering, vol 411. Springer, Cham,
pp 17–30
14. Astolfi A, Ortega R (2003) Immersion and invariance: a new tool for stabilization and adaptive
control of nonlinear systems. IEEE Trans Autom Control 48(4):590–606
15. Astolfi A, Karagiannis D, Ortega R (2008) Nonlinear and adaptive control with applications.
Springer-Verlag, London
16. Karagiannis D, Astolfi A (2008) Nonlinear adaptive control of systems in feedback form: an
alternative to adaptive backstepping. Syst Control Lett 57(9):733–739
17. Han C, Liu Z, Yi J (2019) Immersion and invariance adaptive finite-time control of air-breathing
hypersonic vehicles. Proc Inst Mech Eng Part G J Aerosp Eng 233(7):2626–2641
18. Han C, Liu Z, Yi J (2018) Immersion and invariance adaptive control with σ-modification for
uncertain nonlinear systems. J Franklin Inst 355(5):2091–2111
19. Bechlioulis CP, Rovithakis GA (2013) Reinforcing robustness of adaptive dynamic surface
control. Int J Adapt Control Signal Process 27(4):323–339
20. Carroll JJ, Dawson DM (1995) Integrator backstepping techniques for the tracking control of
permanent magnet brush DC motors. IEEE Trans Ind Appl 31(2):248–255
21. Dawson DM, Carroll JJ, Schneider M (1994) Integrator backstepping control of a brush DC
motor turning a robotic load. IEEE Trans Control Syst Technol 2(3):233–244
22. Sun G, Wang D, Li X, Peng Z (2013) A DSC approach to adaptive neural network tracking
control for pure-feedback nonlinear systems. Appl Math Comput 219(11):6224–6235
23. Slotine JJE, Li W (1991) Applied nonlinear control. Prentice Hall, Englewood Cliffs, New
Jersey
Design of an Adaptive Fuzzy
Backstepping Synergetic Control Scheme
for a Class of Strict-Feedback Nonlinear
Systems
Aissa Rebai, Kamel Guesmi, and Mohamed Bougrine
Abstract The aim of this paper is to construct an adaptive fuzzy controller for strict
feedback nonlinear uncertain disturbed systems. The backstepping and synergetic
control techniques are combined to develop a robust controller for the considered
class of systems. In this paper, fuzzy systems are utilized to estimate the unknown
behaviors of the system. The proposed approach guarantees that all signals in the
closed-loop adaptive control system are semi-globally uniformly ultimately bounded,
and the tracking error eventually converges to a small neighbourhood of the origin.
Numerical simulations are provided to show the effectiveness of the presented control
technique.
Keywords Nonlinear systems · Backstepping control · Synergetic control · Fuzzy

systems · Adaptive control
1 Introduction
During the last decades, the control of nonlinear systems has been widely studied
to deal with nonlinear characteristics existing in practical systems. In recent years,
various advanced control approaches have been proposed for realistic applications in
the literature such as sliding mode control [1–5], backstepping control [6–8], fuzzy
logic and neural network control [9–12].
A. Rebai
Département Maintenance et Sécurité Industrielle, Institut Algérien de Pétrole, Hassi Messaoud,
Algérie
K. Guesmi (B)
CReSTIC, IUT de Troyes, 09 rue de Quebec, 10026 Troyes, France
M. Bougrine
LACoSERE Lab, Amar Telidji, University of Laghouat, Laghouat, Algeria

https://doi.org/10.1007/978-981-15-6403-1_18
264 A. Rebai et al.
During the last years, an increased attention is paid to backstepping-based control

of nonlinear systems due to its efficiency [13, 14]. To cope with the highly uncertain
and complex nonlinear systems, neural networks and fuzzy logic theories are used
to approximate the nonlinear behaviors and to deal with the tracking and regula-
tion control problems. Up to now, various interesting results have been presented
in the literature [15–20]. In [15] a fuzzy adaptive backstepping controller is devel-
oped for uncertain nonlinear systems with triangular form using the dynamic surface
control technique. This technique is used to construct an adaptive fuzzy backstep-
ping controller for input-delay nonlinear systems [16]. In [17], the authors proposed
a fuzzy adaptive output feedback control scheme for switched nonlinear systems
with unknown nonlinearities and immeasurable states. The authors in [18] addressed
the problem of robust adaptive fuzzy funnel control for nonlinear strict-feedback
systems with dynamic uncertainties. Furthermore, in [19, 20], the backstepping and
small gain approaches are combined to design adaptive fuzzy controller for uncertain
nonlinear systems with unmodeled dynamics. In spite of the good tracking control
performances, the controller robustness remains questionable.
Since its introduction by the Russian researcher Kolesnikov [21], synergetic
control theory has been used in several fields such as power systems [22], electrical
systems [23], robotics [24], and other nonlinear systems [25–27]. Compared to sliding
mode control approach, the synergetic control theory presents better control perfor-
mance outside the manifold, and the chattering phenomenon is avoided. Despite the
application of synergistic control theory in several research axes, the use of this theory
for the development of robust control approaches has always been a challenging yet
a rewarding task, which motivate this study.
In this work, the backstepping and synergetic control theory are combined to
design a new robust adaptive fuzzy backstepping synergetic controller for a class
of strict-feedback nonlinear systems. The key contributions of this work lie in
the following: (i) a systematical approach to control a class of non-linear strict-
feedback systems with external disturbances and parameters uncertainties; (ii) the
synergetic control theory to improve the tracking control performances; (iii) fuzzy
systems to approximate the nonlinear unknown dynamics and to overcome the design
complexity; (iv) a fuzzy system to estimate the external disturbance and to ensure
the controller robustness.
The outline of this paper is summarized as follows: the problem formulation and
some assumptions are given in Sect. 2. The proposed adaptive fuzzy backstepping
synergetic control scheme is presented in Sects. 3 and 4. Finally, some simulation
results are given in Sect. 5.
We consider the class of uncertain strict-feedback nonlinear SISO systems repre-

sented by:
Design of an Adaptive Fuzzy Backstepping Synergetic Control … 265
⎧
⎨ ẋi = xi+1 , i = 1, . . . , n − 1
ẋ = [ f (x) + f (x)] + [g(x) + g(x)]u + d(t) (1)
⎩ n
y = x1
where x = [x1 , x2 , . . . , xn ]T ∈ R n , u ∈ R and y ∈ R are the state vector, the

control input, and the system output, respectively. f (x) and g(x) are two unknown
non-linear continuous functions to be estimated using fuzzy systems. d(t), f (x)
and g(x) represent the unknown external disturbance and the unknown bounded
uncertainties, respectively.
Assumption 1 For controllability requirement, the function g(x) must verify gx / =

0 and must be strictly positive or strictly negative. Without loss of generality, we
assume that g(x) > 0. For all x ∈ R n , there exists some unknown constants g1 and
g2 such that:
0 < g1 < g(x) < g2 (2)
Assumption 2 The desired reference yd and its n first derivatives exist, are
continuous and bounded. i.e.

(i)
yd ≤ ym i , i = 1, . . . , n (3)
Assumption 3 The unknown quantities f (x), g(x), and d(t) are bounded :
⎧
⎨ ||d(t)|| ≤ dm
|| f (x)|| ≤ f m (4)
⎩
||g(x)|| ≤ gm
Assumption 4 There exists a large positive constant u m such that
||u(t)|| ≤ u m (5)
The system (1) can be transformed to the following form:

⎧
⎨ ẋi = xi+1 , i = 1, . . . , n − 1
ẋ = f (x) + g(x)u + d̄(t) (6)
⎩ n
y = x1
where d̄(t) is considered as a global disturbance and is given by
d̄(t) = f (x) + g(x)u(t) + d(t) (7)
From Assumptions 3 and 4, we have

266 A. Rebai et al.

d̄(t) ≤ f m + gm u m + dm (8)
The control objective of this work is to design an adaptive fuzzy sliding mode
control law to ensure:
• all signals in the closed-loop system are semi-globally uniformly ultimately
bounded.
• the desired trajectory yd (t) can be tracked by the system output y(t) with a
neglected tracking error.
In this paper, fuzzy logic systems (FLS) are used to identify the unknown nonlinear
continuous functions f, g and d̄ defined on some compacts. Consider the following
form of fuzzy rules:
j j j
Ri I f x1 is F1 and x2 is F2 . . . and xn is Fn ,
(9)
then ω is B j ( j = 1, . . . , N )
where x = [x1 , x2 , . . . , xn ] ∈ R n and ω ∈ R are the FLS input and output, respec-
tively. F ji and B j are the fuzzy sets associated with the membership functions μ F jl (xi )
and μ B j (y), respectively. Since the strategy of singleton output fuzzification, center
average defuzzifier and product inference is used, the fuzzy system output can be
expressed as:
N
n
ȳll=1 i=1 μ Fil (x i )
ω(x) = (10)
N n
l=1 i=1 μ Fil (x i )
where ω̄l = maxω∈R μ B j (ω), μ Fil (xi ) is the membership function value at xi for the
fuzzy rule l. The FLS (10) can be reformulated as follows:
ω(x) = θ T φ(x) (11)

T
where θ = ω̄1 , ω̄2 , . . . , ω̄ N is ideal constant weight vector, and φ(x) =
1 T
φ (x), φ (x), . . . , φ (x) is the basis function vector defined by:
2 N
n
i=1 μ F l (x i )
φ(x) = i (12)
N n
If all membership functions are chosen Gaussian form, the following lemma
points out that the FLS (12) can approximate any nonlinear continuous function over
a compact set in any accuracy.
Lemma 1 Wang [28]. Let f (x) be a continuous nonlinear function defined on a

compact set ∈ R n . Then, there exists an FLS of form (12) such that

sup f (x) − θ T φ(x) ≤ ζ (13)
x∈R n
where ζ is a given positive constant.
3 Adaptive Fuzzy Backstepping Synergetic Controller

Design
The backstepping design procedure consists of n steps [29]. For this procedure, the
system (6) is regarded as a series of subsystems. In each design step, we use a virtual
control variable αi to stabilize a subsystem using an appropriate Lyapunov function.
Then, we calculate the derivative of the tracking error ėi . The detailed design is
described step-by-step as follows.
Step 01:
The tracking errors e1 and e2 are defined as
e1 = x1 − yd
e2 = x2 − α1 (14)
The Lyapunov function is chosen as
1 2
V1 = e (15)
2 1
The derivative of V1 is
V̇1 = e1 ė1 = (x1 − yd )(α1 − ẏd ) (16)
Then, the virtual control variable α1 and the derivative of e1 are given by
α1 = −k1 e1 + ẏd
ė1 = e2 − k1 e1 (17)
where k1 is a positive constant.

Step 02:
The tracking errors e2 and e3 are
e2 = x2 − α1
e3 = x3 − α2 (18)
The Lyapunov function is chosen as

268 A. Rebai et al.
1 2
V2 = e1 + e22 (19)
2
The derivative of V2 is
V̇2 = e1 ė1 + e2 ė2 = e1 (e2 − k1 e1 ) + e2 (α2 − α̇1 ) (20)
Then, the virtual control variable α2 and the derivative of e2 are obtained as
α2 = −e1 − k2 e2 + α̇1
(21)
ė2 = e3 − e1 − k2 e2
where k2 is a positive constant.

Step i, i ≤ n − 1:
The tracking errors ei and ei+1 are
ei = xi − αi−1
ei+1 = xi+1 − αi (22)
The Lyapunov function is selected as
1 2
i
Vi = e (23)
2 k=1 k
Then, the virtual control variable αi and the derivative of ei are obtained as
αi = −ei−1 − ki ei + α̇i−1
(24)
ėi = ei+1 − ei−1 − ki ei
where ki is a positive constant.

Step n:
The design procedure of a synergetic controller begins by selecting a macro-
variable σ which is a function of the system state variables [30]. The control law
forces the trajectories to operate on the attractor σ (x) = 0. The desired dynamic
evolution of σ is generally chosen such that
τ σ̇ + σ = 0, τ > 0 (25)
For our controller design, the tracking error at step n and the macro-variable σ
are chosen as
en = xn − αn−1
(26)
σ = cen−1 + en
where c is a positive constant.

Then, the control input u is obtained as
1
u= −c(en − en−1 − kn−1 en−1 ) − f (x) − d̄ + α̇n−1
g(x)

1
−en−1 − σ (27)
τ
To obtain the adaptive fuzzy backstepping synergetic controller (AFBSC), we

replace the unknown functions f (x), g(x) and d̄(t) by their fuzzy logic estimations.
Then, the proposed control law is expressed as
1
u= −c(en − en−2 − kn−1 en−1 ) − fˆ(x) − d̄ˆ + α̇n−1
ĝ(x)

1
−en−1 − σ (28)
τ
ˆ are given by
where fˆ(x), ĝ(x) and d̄(t)

fˆ x, θ f = θ Tf φ f (x) (29)

ĝ x, θg = θgT φg (x) (30)

d̄ˆ x, θd̄ = θd̄T φd̄ (x) (31)
Finally, the new AFBSC is summarized through Theorem 1.
Theorem 1 Consider the nonlinear strict-feedback system (1) with unknown

bounded uncertainties and external disturbances. The AFBS controller is designed
as described in (28). If the adapting laws are chosen as follows:
θ̇ f = γ1 σ φ f (x) (32)
θ̇g = γ2 σ φg (x)u (33)
θ̇d̄ = γ3 σ φd̄ (x) (34)
where γ1 , γ2 and γ3 are the positive learning rates, then the tracking error converges
asymptotically to a small neighbourhood of the origin and all signals in the closed
loop system are bounded.
270 A. Rebai et al.
4 Stability Analysis
For the stability analysis of the closed system, we consider the Lyapunov functional
candidate given by:
n−1
1 2 1 T 1 T 1 T
V (t) = e + σ + θ̃ f θ̃ f + θ̃g θ̃g + θ̃d̄ θ̃d̄
2
(35)
2 l=1 l γ1 γ2 γ3
where
θ̃ f = θ ∗f − θ f
θ̃g = θg∗ − θg
θ̃d̄ = θd̄∗ − θd̄
where θ ∗f , θg∗ and θd̄∗ are the optimal parameter vectors expressed, respectively, as

θ ∗f = arg min sup fˆ x|θ f − f (x, t) (36)
θ f ∈ f x∈R n

θg∗ = arg min sup |ĝ(x|θg ) − g(x, t)| (37)
θg ∈g x∈R n

θd̄∗ = arg min ˆ
sup |d̄(x|θd̄ ) − d̄(t)| (38)
θd̄ ∈d̄ x∈R n
where f , g and d̄ are constraint sets for θ f , θg and θd̄ , respectively.
The time derivative of V (t) is obtained as

n−1
1 T 1 1
V̇ = el ėl + σ σ̇ − θ̃ θ̇ f − θ̃gT θ̇g − θ̃d̄T θ̇d̄ (39)
l=1
γ1 f γ2 γ3
From (26) we have
σ̇ = cėn−1 + ėn (40)
Using (22), (24) and (26) in (40) we obtain
σ̇ = cėn−1 + f (x) + g(x)u + d̄(t) − α̇n−1

= cėn−1 + f (x) + g(x)u + d̄(t) − ĝ(x)u + cėn−1 + fˆ(x)

1 ˆ
+en−1 + σ + d̄
τ

= f (x) − fˆ(x) + g(x) − ĝ(x) u + d̄ − d̄ˆ − en−1
1
− σ (41)
τ
Substituting (41) in (39), it results in

n−1
V̇ = − kl el2 + en−1 en + σ f (x) − fˆ(x) + (g(x)
l=1

ˆ 1 1 1
−ĝ(x) u + d̄ − d̄ − en−1 − σ − θ̃ Tf θ̇ f − θ̃gT θ̇g
τ γ1 γ2
1
− θ̃d̄T θ̇d̄ (42)
γ3
The minimum approximation error is described as
ε = ε1 + ε2 + ε3
with

ε1 = f (x) − fˆ x, θ f (43)

ε2 = g(x) − ĝ x, θg u (44)

ε3 = d̄ − d̄ˆ θd̄ (45)
From (43)–(45), it can be written

f (x) − fˆ x, θ f = θ̃ Tf φ f (x) + ε1 (46)

g(x) − ĝ x, θg u = θ̃gT φg (x)u + ε2 (47)

d̄ − d̄ˆ θd̄ = θ̃d̄T φd̄ + ε3 (48)
Using (46)–(48) in (42) gives that

n−1
1 2
V̇ = − kl el2 + en−1 en − σ en−1 − σ + σ θ̃ Tf φ f (x)
l=1
τ
1 1 1
+ θ̃gT φg (x)u + θ̃d̄T φd̄ + ε − θ̃ Tf θ̇ f − θ̃gT θ̇g − θ̃d̄T θ̇d̄ (49)
γ1 γ2 γ3
272 A. Rebai et al.
Substituting the adapting laws of (32)–(34) in (49) yields that

n−1
1 2
V̇ = − kl el2 + en−1 en − σ en−1 − σ + σε (50)
l=1
τ
Replacing (26) in (50) results that

n−1
1 2
V̇ = − kl el2 − cen−1
2
− σ + σε < 0 (51)
l=1
τ
The inequality (51) is the optimum result that we can obtain because ε is the
minimum approximation error. As a result, we can conclude that all the signals
of the closed-loop adaptive control system are semi-globally uniformly ultimately
bounded.
5 Numerical Simulations
Consider the following inverted pendulum system

⎧
⎪
⎪ ẋ1 =
x2
⎪
⎨ mlx 2 cosx1 sinx1 cosx1
gsinx1 − 2m c +m
ẋ2 = + 4 mm(cosx
c +m u + d(t) (52)
⎪
⎪
m(cosx )2
l 43 − m c +m1 l 3 − m c +m1
)2
⎪
⎩
y = x1
where x1 and x2 are the swing angle and swing rate respectively. g = 9.8 m/s2 , mc = 1
kg, m = 0.1 kg, l = 0.5 m are the gravity, mass of the cart, mass of the pendulum and
length of pendulum, respectively. u(t) represents the control input. The disturbance
in this example is chosen as dt = 0.4 ∗ sin(π/100t). The membership functions of
the FLS are given as follows:

(xi + π/6 − ( j − 1)π/12) 2
μ F j (xi ) = exp − ,
i π/24
i = 1, 2; j = 1, . . . , 5
The parameters of design are chosen as: c = 100, k1 = 10 and τ = 0.01. The
initial conditions are defined as [x1 (0), x2 (0)]T = [π/60, 0]T . The desired signal is
defined as: yd = 0.1 sin(t).
The simulation results are shown on Figs. 1, 2, 3, 4 and 5. Figure 1 plots the
responses of output y(t) and the desired reference yd (t). Figure 2 gives the trajectories
of the states x1 (t) and x2 (t). It can be seen that the closed loop system is stable.
0.15
Desired
Approach in [31]
Proposed approach
0.1
0.08
0.06
0.05
System output y(t)
0.04
0.02
0
0
0 0.2 0.4 0.6 0.8
-0.05
-0.1
-0.15
0 1 2 3 4 5 6 7 8 9 10
Time (sec)
Fig. 1 Trajectories of output y(t) and the desired reference yd (t)
0.5
x (t)
1
x (t)
2
0.4
0.3
0.2
System state x(t)
0.1
-0.1
-0.2
-0.3
-0.4
-0.5
0 1 2 3 4 5 6 7 8 9 10
Time (sec)
Fig. 2 Trajectories of states x1 (t) and x2 (t)

274 A. Rebai et al.
Fig. 3 Trajectory of control input u(t)
0.06
Approch in [31]
Proposed approach
0.05
0.04
Tracking error e(t)
0.03
0.02
0.01
-0.01
0 1 2 3 4 5 6 7 8 9 10
Time (sec)
Fig. 4 Trajectory of tracking error e(t)

Fig. 5 Trajectory of macro-variable σ (t)
Figures 3, 4 and 5 show the control signal u(t), the response error e(t) and the
macrovariable σ (t), respectively. From these figures, it can be seen that all variables
of the considered system are bounded. It is clear from the obtained results that
the proposed approach presents good tracking performance with a setting time of
0.033 s compared to 0.467 s for the approach in [31]. The performance index is
I T AE = 0.2543 for the proposed approach and I T AE = 14.7487 for the approach
in [31]. According to the obtained simulation results, the effectiveness of the proposed
control approach is proven.
6 Conclusion
In this paper, an adaptive fuzzy control approach has been presented for a class
for strict-feedback nonlinear systems. Fuzzy logic systems are used to describe the
unknown nonlinear functions existing in the system. Furthermore, based on backstep-
ping and synergetic control strategies, the proposed adaptive fuzzy control scheme
guarantees that all signals of the closed-loop system remain bounded and the tracking
error converges asymptotically to the origin. Finally, simulation results have provided
the effectiveness the proposed control approach.
276 A. Rebai et al.
References
1. Ray PK, Paital SR, Mohanty A, Eddy FY, Gooi HB (2018) A robust power system stabilizer
for enhancement of stability in power system using adaptive fuzzy sliding mode control. Appl
Soft Comput 73:471–481
2. Xiuxiang C, Ting W, Yongkun Z, Wen Q, Xinghua Z (2018) An adaptive fuzzy sliding mode
control for angle tracking of human musculoskeletal arm model. Comput Electr Eng 72:214–
223
3. Zhao Y, Wang J, Yan F, Shen Y (2019) Adaptive sliding mode fault-tolerant control for type-2
fuzzy systems with distributed delays. Inf Sci 473:227–238
4. Nguyen SD, Vo HD, Seo T-I (2017) Nonlinear adaptive control based on fuzzy sliding mode
technique and fuzzy-based compensator. ISA Trans 70:309–321
5. Farahani M, Ganjefar S (2017) Intelligent power system stabilizer design using adaptive fuzzy
sliding mode controller. Neurocomputing 226:135–144
6. Baigzadehnoe B, Rahmani Z, Khosravi A, Rezaie B (2017) On position/force tracking control
problem of cooperative robot manipulators using adaptive fuzzy backstepping approach. ISA
Trans 70:432–446
7. Sadek U, Sarjas A, Chowdhury A, Svecko R (2017) Improved adaptive fuzzy backstepping
control of a magnetic levitation system based on Symbiotic Organism Search. Appl Soft Comput
56:19–33
8. Qiao J-F, Hou Y, Zhang L, Han H-G (2018) Adaptive fuzzy neural network control of wastewater
treatment process with multiobjective operation. Neurocomputing 275:383–393
9. Wu Y, Huang R, Li X, Liu S (2019) Adaptive neural network control of uncertain robotic
manipulators with external disturbance and timevarying output constraints. Neurocomputing
323:108–116
10. Jin L, Li S, Yu J, He J (2018) Robot manipulator control using neural networks: A survey.
Neurocomputing 285:23–34
11. de Jesús Rubio J (2018) Discrete time control based in neural networks for pendulums. Appl
Soft Comput 68:821–832
12. Sui S, Chen CLP, Tong S (2019) Fuzzy adaptive finite-time control design for non-triangular
stochastic nonlinear systems. IEEE Trans Fuzzy Syst 27(1):172–184
13. Wang W, Xie B, Zuo Z, Fan H (2018) Adaptive backstepping control of uncertain gear
transmission servosystems with asymmetric dead-zone nonlinearity. IEEE Trans Ind Electron
66(5):3752–3762
14. Roy TK, Mahmud MA, Oo AMT (2019) Robust adaptive backstepping excitation controller
design for higher-order models of synchronous generators in multimachine power systems.
IEEE Trans Power Syst 34(1):40–51
15. Peng J, Dubay R (2019) Adaptive fuzzy backstepping control for a class of uncertain nonlinear
strict-feedback systems based on dynamic surface control approach. Expert Syst Appl 120:239–
252
16. Zhou Q, Wu C, Jing X, Wang L (2016) Adaptive fuzzy backstepping dynamic surface control
for nonlinear input-delay systems. Neurocomputing 199:58–65
17. Hou Y, Tong S (2017) Command filter-based adaptive fuzzy backstepping control for a class
of switched nonlinear systems. Fuzzy Sets Syst 314:46–60
18. Wang H, Zou Y, Liu PX, Liu X (2018) Robust fuzzy adaptive funnel control of nonlinear
systems with dynamic uncertainties. Neurocomputing 314:299–309
19. Su H, Zhang T, Zhang W (2017) Fuzzy adaptive control for SISO nonlinear uncertain systems
based on backstepping and small-gain approach. Neurocomputing 238:212–226
20. Zhang X, Liu X, Li Y (2017) Adaptive fuzzy tracking control for nonlinear strict-feedback
systems with unmodeled dynamics via backstepping technique. Neurocomputing 235:182–191
21. Kolesnikov AA (2014) Introduction of synergetic control. In: American control conference,
Portland, OR, USA, June 4–6 2014, pp 3013–3016
22. Nechadi E, Harmas MN, Essounbouli N, Hamzaoui A (2016) Optimal synergetic control based
Bat Algorithm for DC-DC boost converter. IFAC-PapersOnLine 49(12):698–703
23. Wang C, Zhang D, Zhuang H, Lu B (2018) Coordinated synchronization control of multi-

motor system based on synergetic control theory. In: Chinese control and decision conference,
Shenyang, China, June 9–11 2018, pp 160–164
24. Sklyarov AA, Veselov GE, Sklyarov SA, Pohilina EE (2017) Synthesis of the synergetic control
law of the transport robotic platform. In: IEEE II international conference on control in technical
systems, St. Petersburg, Russia, October 25–27 2017, pp 285–288
25. Rastegar S, Araújo R, Sadati J, Mendes J (2017) A novel robust control scheme for LTV systems
using output integral discrete-time synergetic control theory. Eur J Control 34:39–48
26. Djennoune S, Bettayeb M (2013) Optimal synergetic control for fractional-order systems.
Automatica 49:2243–2249
27. Rebai A, Guesmi K, Hemici B (2016) Adaptive fuzzy synergetic control for nonlinear hysteretic
systems. Nonlinear Dyn 86:1445–1454
28. Wang LX (1994) Adaptive fuzzy systems and control: design and stability analysis. Prentice-
Hall, Englewood Cliffs, NJ
29. Zhou J, Wen C (2008) Adaptive backstepping control of uncertain systems: nonsmooth
nonlinearities, interactions or time-variations. Springer, Berlin, Heidelberg
30. Rastegar S, Araujo R, Sadati J (2018) Robust synergetic control design under inputs and states
constraints. Int J Control 91(3):639–657
31. Wang J, Rad AB, Chan PT (2001) Indirect adaptive fuzzy sliding mode control: part I: fuzzy
switching. Fuzzy Sets Syst 122(1):21–30
Improvement of the Stability
Performance of a Quad-Copter
Helicopter by a Neuro-Fuzzy Controller
Djalal Baladji, Kheireddine Lamamra, and Farida Batat
Abstract UAVs are quad-copter unmanned helicopters, They been the subject of
growing importance and are widely used in several fields; for this reason and to
improve their performance, in this article we have proposed an algorithm based on
the ANFIS. The genetic algorithms is used in order to optimize the ANFIS parameters
and thereby ensure its learning, which maintains the ability for self-organization and
self-learning. The proposed control scheme aims to implement good capabilities
such as the description of qualitative knowledge, a learning mechanism and direct
processing of quadcopter helicopter quantitative data. This control is adopted the
precision and time of adjustment. On the other hand, If the position and attitude
deviation becomes relatively smaller, The PID command will be used to limit this
error. Experimental results indicate that the proposed ANFIS control algorithm has
good performance in the flight process.
Keywords Quad-rotor helicopter · Stability performance · Neural network control

(ANFIS) · Fuzzy control · Flight angle deviation
1 Introduction
Recently, The quad-copter helicopter has attracted attention because of its very good
flexibility, its ability to perform various air missions even in difficult conditions. The
quad-copter helicopter can adopt a variety of flight attitudes, which is attributed to
D. Baladji (B) · K. Lamamra · F. Batat

Department of Electrical, Engineering, University of Oum El Bouaghi, Oum El Bouaghi, Algeria
K. Lamamra
F. Batat

https://doi.org/10.1007/978-981-15-6403-1_19
280 D. Baladji et al.
the efficient engine speed controller with about four propellers. In order to enable the
quad-copter helicopter to better complete its flight mission, flight stability perfor-
mance becomes particularly important. A neuron-fuzzy control algorithm (ANFIS)
is proposed in this article to ensure the stability performance of the quadcopter
helicopter.
In recent years, some current research has focused on the fuzzy and neural
network; for example, Qu et al. [8] has addressed the design of an autopilot for
an autonomous UAV using a draft genetic algorithm for an evolution on the rules
of fuzzy limb functions. Roopashree et al. [11] explained the design of a compact,
accurate and economical system. Fuzzy logic controllers (FLC) and fuzzy infer-
ence systems (FIS), which estimated the attitude of drones. Sabo and Cohen [13]
proposed a methodology for two-dimensional data management [1–3]. In order to
improve the system’s performance characterization, a recurrent online neural network
modeling for the dynamic uncertainty of the four-rotor unmanned aerial (QUAV) was
used to develop an under-operated sliding mode control based on a recurrent neural
network (Hwang 2012) [5, 21]. A new adaptive neural control scheme for stabilizing
quadcopter helicopters in the presence of sinusoidal turbulence has been proposed
(Boudjedir et al. 2012) [6, 8]. Due to the excellent performance of the fuzzy system
and the neural network, fuzzy control of the neural network has been used in several
areas; for example, He and Dong [2] conducted research for adaptive control of the
neural fuzzy network (NN) by learning [12, 17, 18].
The above control strategies, with the aim of suggesting a flexible control scheme
based on piloting errors, which uses the advantages of the ANFIS system, so that
regulation is effective on the four rotors of the four helicopter quad-copters.
2 Quad-Copter
Four-screw helicopters are four-engine quad-copters mounted on a cross, usually

made of carbon fibre, hence their name quad-copters. Examples are shown in Fig. 1.
In a quad-copter, the front and rear engines rotate clockwise while the right and
left engines rotate counter-clockwise. The propellers used are fixed-pitch. The pitch
is obtained by a difference in the rotational speed of the front and rear rotors. The
roll is obtained in a similar way with the speed difference of the lateral engines. Yaw
is achieved by increasing the speed of the front and rear engines while reducing the
speed of the side engines. There is also a version of X4 in which all rotors rotate
in the same direction. In the latter, two rotors are tilted horizontally to create a yaw
torque. The inclination can be fixed or variable and allows to control the movement
of the yaw. Since quadri-rotors are controlled by the difference in rotor speed, it is
important that the motor speed can be varied quickly. For this purpose, very light
blades and ratios should be used.
• The structure of the quad-copter is assumed to be rigid and symmetrical, which
means that the inertia matrix will be assumed to be diagonal.
Improvement of the Stability Performance of a Quad-Copter … 281
Fig. 1 Four-propeller helicopters
• Les hélices sont supposées rigides pour pouvoir négliger l’effet de leur déforma-
tion lors de la rotation.
• Le centre de masse et l’origine du repère lié à la structure coïncident.
• The lift and drag forces are proportional to the squares of the rotors’ rotational
speed, which is a very close approximation of aerodynamic behaviour.
To evaluate the mathematical model of the quad-copter, two benchmarks are used,
a fixed benchmark linked to the ground Rb and another mobile Rm. The passage
between the moving reference frame and the fixed reference frame is given by a
matrix called transformation matrix T which contains the orientation and position of
the moving reference frame with respect to the fixed reference frame.
The following axis convention is chosen: (Fig. 2)

Rξ
T =
0 1
with R the rotation matrix (describes the orientation of the moving object), ζ = [x y
z]T is the position vector. To determine the elements of the rotation matrix R, Euler
angles are used.
A. Euler’s Angles:
At the beginning the moving reference mark is coincident with the fixed reference
mark, after the moving reference
mark makes a rotational movement about the x-axis
by a roll angle − π2 < φ < π2 , followed by rotation about the y-axis at a pitch angle
π
− 2 < φ < π2 , followed by a rotation about the z axis of angle of (−π < φ < π ).
so we have the formula of the rotation matrix R :
R = Rotz (ψ) + Roty (θ ) + Rotx (φ)

Fig. 2 Géométrie du quadcopter
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
cψ −sψ 0 cθ 0 sθ 1 0 0
= ⎣ sψ cψ 0 ⎦ × ⎣ 0 1 0 ⎦ × ⎣ 0 cφ −sφ ⎦
0 0 1 −sθ 0 cθ 0 sφ cφ
⎡ ⎤
cψcθ sφsθ cψ − sψcφ cφsθ cψ + sψsφ
R = ⎣ sψcφ sφsθ sψ + cψcφ cφsθ sψ − cψsφ ⎦
−sθ sφcθ cφcθ
with : c = cos and s = sin

B. Angular speeds:
Rotation speeds 1, 2, 3 in the fixed reference frame are expressed as a function
of the rotation speeds ψ, θ, φ in the moving reference frame, we have:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 φ̇ 0 0
⎣ ⎦ ⎣ ⎦
= 2 = 0 + Rotx (φ) −1 ⎣ ⎦
θ̇ + (Rot y (θ )Rotx (φ)) −1 ⎣
0⎦
3 0 0 ψ̇
Indeed, the rotation in roll takes place when the marks are still confused. Then,
with regard to pitch, the vector representing rotation must be expressed in the fixed
reference frame: it is therefore multiplied by Rot x (φ)−1 . Similarly, the vector repre-
senting the yaw rotation must be expressed in the fixed reference frame which has
already undergone two rotations. This brings us to:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ . ⎤ ⎡ . ⎤
1 φ̇ 0 −ψsθ φ̇ − ψsθ
⎢ ⎥ ⎢ ⎥
= ⎣ 2 ⎦ = ⎣ 0 ⎦ + ⎣ θ̇ cφ ⎦ + ⎣ ψ̇sφcθ ⎦ = ⎣ φ̇cφ − ψ̇sφcθ ⎦
3 0 −θ̇ sφ ψ̇cφcθ ψ̇cφcθ − θ̇ sφ
⎡ ⎤ ⎡ ⎤
1 0 −sθ φ̇
= ⎣ 0 cφ sφcθ ⎦ × ⎣ θ̇ ⎦
0 −sφ cφcθ ψ̇
When the quadcopter makes small rotations, the following approximations can
be made:
cθ + cφ + cψ = 1 and sθ + sφ + sψ = 0
so the angular velocity will be:

= φ̇ θ̇ ψ̇
C. Linear speeds:
Linear speeds Vxb , Vyb , Vzb in the fixed reference mark as a function of linear speeds
⎡ b⎤ ⎡ m⎤
Vx Vx
Vxm , Vym , Vzm in the moving reference frame are given by: V = ⎣ Vyb ⎦ = R× ⎣ Vym ⎦
Vzb Vzm
3 ANFIS Commande (Adaptive Network-Based Fuzzy

Inference System)
ANFIS (Adaptive Network Based Fuzzy Inference System) it is a neuro-vague adap-

tive system of inference that consists in using a network neuron of type MLP in 5
layers for who every layer corresponds to the realisation of a stage of a system of
vague inference of type Takagi Sugeno. For simplicity, we assume that the vague
system of inference at two entries x and there, and in as an exit fr. Assume that the
usual base contains two vague rules of type Takagi-Sugeno.
Rule 1: SI x is A1 and y is B1 THEN f1 = p1 x + q1 y + r1
Rule 2: SI x is A2 and y is B2 THEN f2 = p2 x + q2 y + r2
L’ANFIS à une architecture posée par cinq couches comme représenté sur la
Fig. 3.
Fig. 3 Example ANFIS with 2 entries with 9 rules
The ANFIS system applies the learning mechanism of neural networks to fuzzy
inference techniques. Another term ANFIS is a fuzzy inference system (SIF) whose
parameters of the membership functions are adjusted using the backpropagation
learning algorithm, or in combination with another type of algorithms such as the
least square.
In the ANFIS architecture proposed in Fig. 3, the overall output can be expressed
as linear combinations of the consequent parameters. More precisely, the conclusion
(the output) in Fig. 3 can be rewritten as:
W1 W2
f= f1+ f2
w1 + w2 w1 + w2

= W 1 x p1 + W 1 y q1 + W 1 r 1 + W 2 x p2 + W 2 y q2 + W 2 r 2
The output is a linear function of the consequence parameters (p, q, r). ANFIS is
parametric representation two sets of parameters: S1 and S2
ANFIS uses a learning cycle of two passages:
• Passage forward: S1 is fixed and S2 is calculated using the least square error
algorithm (LSE). (The LSE is applied only once when starting to get the initial
values of the consequent parameters).
• Passback: S2 is fixed and S1 is calculated using the Retro-Propagation algorithm
Ek error measurement for the given Kiemme input learning:

n(L)
2
Ek = di − xL,i
i=1
4 Application of ANFIS on Quadcopter
The ANFIS controllers are taught along the three axes by the least squares method to
estimate the consequent parameters and the gradient descent algorithm to determine
the parameters of the premises (adjustment of the parameters related to the belonging
functions). The genetic algorithms is used here to choose the learning parameters of
the ANFIS. This is called “blended learning”. The ANFIS network used in this work
was programmed by the logic toolbox fuzzy Toolbox with which we built the fuzzy
rules (Si-Then) with their appropriate membership functions to four together, while
respecting the following steps:
1. Using the genetic algorithms to choose the ANFIS learning parameters
• The number of iterations = 20.
• The number of membership functions = 5.
• Error tolerance = 10−4 .
2. Start learning using the ANFIS parameters.
3. Stop when the tolerance is satisfactory.
4. Test the results with another data set.
The equivalent neural structure proposed to generate the vertical flight control is
presented in Fig. 4 we have the same structure for the other two axes: x and y to
generate the commands u2x and u2y respectively.
The characteristics of the ANFIS model used to control the XSF UAV in all three
axes are as follows:
Fig. 4 Neural structure of the proposed ANFIS controller to generate u3

Ɵme(s) Ɵme(s)
diplacement
follow (z)
diplacement diplacement
Ɵme(s) follow (y) follow (x)
Fig. 5 Reference and actual path of the straight connections (ANFIS command)
• Number of nodes: 75.

• Number of parameters of the premises: 30.
• Total number of parameters: 105.
• Number of learning data pairs: 25,001.
• Number of test data pairs: 25,001.
• Number of fuzzy rules: 25.
We can see from Fig. 5 that the trajectory of the XSF UAV follows the desired
trajectory (the reference) and this is shown in Fig. 6 whose errors tend towards zero.
From Fig. 7, we can see that the ANFIS method ensures the flight of our UAV
according to the desired trajectory (at equilibrium u 3 = mg and the commands
u 2x = u 2y = 0) but nevertheless it does not keep the readability of the rules.
4.1 Robustness Study
Figures 8 and 9 illustrate the simulation results for an opposite drag force of 2 N
and 4 N respectively, applied after take-off at time t = 8 s in the direction of vertical
flight. The zoom of the vertical flight shows very well the influence of the disturbance
at t = 8 s.
From Figs. 8 and 9, we can see that the variation of the error is proportional to the
variation of the disturbance: the more the tained force increases, the more the static
error increased with ocsillations.
From Figs. 10 and 11, we notice that the error of the ANFIS controller is much
smaller than that of the fuzzy PD controller but nevertheless the control by the ANFIS
controller presents ocsillations which increases with the increase in drag force which
can disturb the drone.
Ɵme(s)
Fig. 6 The commands u 3 , u 2x and u 2y for the next movement x, y and z

verƟcal flight (m)
Ɵme(s) Ɵme(s)
Ɵme(s) Ɵme(s)
Fig. 7 The influence of wind with a drag force, F_(trn) = 2 N (ANFIS command)
Ɵme(s) Ɵme(s)
Ɵme(s) Ɵme(s)
Fig. 8 The influence of wind with a drag force, F_(trn) = 2 N (ANFIS command)
Ɵme(s)
Fig. 9 The influence of wind with a drag force of 2 N on the error of the fuzzy PD and ANFIS
controller
Ɵme(s)
Fig. 10 The influence of wind with a drag force of 4 N on the error of the fuzzy PD and ANFIS
controller
diplacement follow (z)
diplacement follow (y) diplacement follow (x)
Fig. 11 Cone ascent
4.2 Generation of a Complex Trajectory (a Cone)
In this part, we generate a trajectory based on rational fractions in order to have

trajectories without decoupling from an initial point, with a zero speed, and reach
a well determined final point with the same zero speed. Next, we explicitly present
the expressions of the position variables that generate a cone, as well as a figure
illustrating its shape.
x r (t) = ρ(t)sin(α(t))
y r (t) = −ρ(t)sin(α(t))
t5
z r (t) = h dz 5
t5 t f − t
with: t f = 28 s and hdz = 5 m.
αr e f t 5
α= 5
t5 t f − t
with: αr e f = 4
t5
ρ = hr 5
t5 t f − t
diplacement follow (z)
diplacement follow (y) diplacement follow (x)
Fig. 12 Realization of a cone by the ANFIS controller
where hr = 2 m.
Figure 12 shows that our control system ensures the continuation of the complex
trajectory.
We notice that the errors tend towards zero so the desired position is reached.
5 Conclusion
It appears from this work that the use of an architecture called “ANFIS” is a simple
and effective way to obtain by learning a powerful controller whose behaviour can be
interpreted in the form of decision rules. We have found that the use of a simplified
structure allows us to obtain an efficient command law by online learning. The main
role of the genetic algorithm, is to choose the learning parameters of the ANFIS.
The controller results used (ANFIS) shows the validity of our approach for ordering
our system. So these techniques have been successfully applied to design control
algorithms that move the drone from an initial position to a desired equilibrium
position. The simulation results confirmed the controller performance used.
References
1. Ali MS, Gunasekaran N, Zhu Q (2017) State estimation of T-S fuzzy delayed neural networks
with Markovian jumping parameters using sampled-data control. Fuzzy Sets Syst 306:87–104
2. He W, Dong Y (2017) Adaptive fuzzy neural network control for a constrained robot using
impedance learning. IEEE Trans Neural Netw Learn Syst 29(4):1174–1186
3. Jin Z, Chen J, Sheng Y, Liu X (2017) Neural network based adaptive fuzzy PID-type sliding
mode attitude control for a reentry vehicle. Int J Control Autom Syst 15(1):404–415
4. Lin M-Y, Zheng C-D (2017) Novel stability conditions of fuzzy neural networks with mixed
delays under impulsive perturbations. Optik 131:869–884
5. Mi B, Liu D (2017) A fuzzy neural approach for vehicle guidance in real–time. Intell Autom
Soft Comput 23(1):13–19
6. Mushage BO, Chedjou JC, Kyamakya K (2017) Fuzzy neural network and observer-based
fault-tolerant adaptive nonlinear control of uncertain 5-DOF upper-limb exoskeleton robot for
passive rehabilitation. Nonlinear Dyn 87(3):2021–2037
7. Prasad M, Lin C-T, Li D-L, Hong C-T, Ding W-P, Chang J-Y (2015) Soft-boosted self-
constructing neural fuzzy inference network. IEEE Trans Syst Man Cybern Syst 47(3):584–588
8. Qu Y, Pandhiti S, Bullard KS, Potter WD, Fezer KF (2011) Development of a genetic
fuzzy controller for an unmanned aerial vehicle. In: International conference on industrial,
engineering and other applications of applied intelligent systems, pp 328–335
9. Raffo GV, Ortega MG, Rubio FR (2015) Robust nonlinear control for path tracking of a
quad-rotor helicopter. Asian J Control 17(1):142–156
10. Rodríguez-Canosa GR, Thomas S, Del Cerro J, Barrientos A, MacDonald B (2012) A real-time
method to detect and track moving objects (DATMO) from unmanned aerial vehicles (UAVs)
using a single camera. Remote Sens 4(4):1090–1111
11. Roopashree S, Deepika KM, Bhat S (2011) Design and implementation of fuzzy controller
for estimating the attitude of unmanned aerial vehicles. In: International joint conference on
advances in signal processing and information technology, pp 255–261
12. Ruan J, Chen X, Huang M, Zhang T (2017) Application of fuzzy neural networks for modeling
of biodegradation and biogas production in a full-scale internal circulation anaerobic reactor.
J Environ Sci Health Part A 52(1):7–14
13. Sabo C, Cohen K (2012) Fuzzy logic unmanned air vehicle motion planning. Adv Fuzzy Syst
2012:13
14. Sun L, Zuo Z (2015) Nonlinear adaptive trajectory tracking control for a quad-rotor with
parametric uncertainty. Proc Inst Mech Eng Part G J Aerosp Eng 229(9), 1709–1721
15. Sun W, Gao H, Kaynak O (2010) Finite frequency H_∞ control for vehicle active suspension
systems. IEEE Trans Control Syst Technol 19(2):416–422
16. Sun W, Gao H, Kaynak O (2014) Vibration isolation for active suspensions with performance
constraints and actuator saturation. IEEE ASME Trans Mechatron 20(2):675–683
17. Yao J, Deng W, Jiao Z (2015) Adaptive control of hydraulic actuators with LuGre model-based
friction compensation. IEEE Trans Ind Electron 62(10):6469–6477
18. Yao J, Jiao Z, Ma D, Yan L (2013) High-accuracy tracking control of hydraulic rotary actuators
with modeling uncertainties. IEEE ASME Trans Mechatron 19(2):633–641
19. Yao J, Jiao Z, Ma D (2014) Extended-state-observer-based output feedback nonlinear robust
control of hydraulic systems with backstepping. IEEE Trans Ind Electron 61(11):6285–6293
20. Yao J, Jiao Z, Ma D (2015) A practical nonlinear adaptive control of hydraulic servomechanisms
with periodic-like disturbances. IEEE ASME Trans Mechatron 20(6):2752–2760
21. Zhao D-J, Yang D-G (2016) Model-free control of quad-rotor vehicle via finite-time convergent
extended state observer. Int J Control Autom Syst 14(1):242–254
The Importance of Applying Artificial
Intelligence on Unmanned Aerial Vehicle
Amine Mohammed Taberkit, Ahmed Kechida, and Abdelmalek Bouguettaya
Abstract Unmanned Aerial Vehicles (UAVs) are used in several applications and
they are growing in popularity. Recent progress in unmanned aerial vehicles and
artificial intelligence constitutes a new chance for an autonomous operation and
flight. Nowadays, artificial intelligence and deep learning are driving the evolution
of UAVs and fueling their autonomous future. Computer vision achieved very impor-
tant progress in image classification and segmentation, and object detection, which
make it very attractive research field when it is applied on unmanned aerial vehicle.
Artificial intelligence is not only important and benefic, but can be rather, dangerous
and serious matter since the UAVs learns through algorithms, and use that for future
decision making. This work is a survey, where we present works, challenges and
dangerous part of using artificial intelligence on UAVs.
Keywords UAV · Machine learning · Artificial intelligence · System · Drone
1 Introduction
An unmanned aerial vehicle (UAV) is defined as an aircraft, without any presence

of a pilot on board. UAVs can be used to execute observation or detection (objects
and/or persons) missions [1] through automatic or remote control. Taking aerial
records have two main advantages, low cost and high movability. The reference
standards developed by the UAV community are based on several parameters such
as: flight altitude, endurance, speed, maximum payload and others [2]. The three main
components of an UAV system are: the aircraft with common or other sensor features,
Ground control station, and operator. There are a wide collection of UAV depending
on their shapes, functionalities, configurations, and characteristics. Their hardware
A. M. Taberkit (B) · A. Kechida · A. Bouguettaya

Research Center in Industrial Technologies CRTI, 16014 Cheraga, Algiers, Algeria
A. Bouguettaya

https://doi.org/10.1007/978-981-15-6403-1_20
294 A. M. Taberkit et al.
Fig. 1 Some quadcopter UAVs developed in CRTI research center
and software design can be modified with the task requirements. Quadcopter UAVs
are one of the most favored kinds of small unmanned aerial vehicles because of their
very simple mechanical construction and propulsion principle [3].
The civilian and military applications in several fields have got a drastic increase
during the last years [4]. Some examples encompass the inspection of power line
[5], wildlife conservation [6], building inspection [7], precision agriculture [8] and
military surveillance [9].
In our research center we developed many Quadcopter UAVs, two among them
were presented in the university Salah BOUBNIDER, Constantine 3 (Fig. 1).
2 Machine Learning
Machine Learning is defined as the capability allowing Artificial Intelligence (AI)

systems to learn through data. A suitable definition for what learning covers is as
follow: “A computer program is said to learn from experience E with respect to some
of class of tasks T and performance measure P if its performance at tasks in T, as
measured by P, improves with experience E” [4–10].
It can be defined as an evolving field of computational algorithms that are designed
to mimic the human intelligence by learning through the surrounding environment
[11]. This ability is the main key to develop successful machine learning.
Historically, the inception of machine learning can be traced to the seventeenth
century. Pascal and Leibniz [12] developed machines that can emulate ability of
add and subtract. Arthur Samuel, from IBM, invented the term “Machine Learning”
and demonstrated the computer ability of playing checkers [13]. In 1958, Rosenblatt
The Importance of Applying Artificial Intelligence … 295
developed the perceptron and it was one of the early neural networks [14]. A great
development was made in 1975 with the development of
Werbos multilayer perceptron (MLP) [15], and development of decision trees by
Quinlan in 1986 [16]. Several works based on Machine learning algorithms have
been proposed, including Adaboost [17] and random forests [18].
Deep learning (DL) in artificial intelligence (AI) has recently won a significant
interest. Deep learning is applied in many fields such as autonomous systems, facial
recognition, classification, and object detection. Among the most promising systems
that can exploit deep learning are unmanned aerial vehicles (UAVs).
Convolutional Neural Networks (CNNs) is class of deep neural networks, and
mostly applied to analyze visual imagery. CNN is considered as very good and
powerful tool in the classification and object detection [19]. The structure of a CNN
typically contains a feature extractor stage followed by a classifier [20].
Many object detectors have been proposed by the deep learning researchers,
including R-CNN [21], R-FCN [22], YOLO [23] and SSD [24]. Many auspicious
CNN architectures were recently proposed, such as CaffeNet [25] and GoogLeNet
[26].
In Fig. 2, we represent the evolution of some object detectors over the years.
CNN architecture includes many layers of different types [27]:
• Convolutional layers, as their name indicates, they compute the convolution of
the input image with the weights of the network [27].
• Pooling layers, the role of these layers is to diminish the size of the input layer
using some local non-linear operations.
• Normalization layers, the aim of using those layers is to improve generalization
of the CNN. Neurons used in these layers are sigmoid [28].
• Fully-connected layers, these layers are used in the last levels of the network
[28].
We distinguish three different categories: supervised, unsupervised, and reinforce-
ment learning (Fig. 3).
The reinforcement learning (RL) methods allow an agent to learn suitable actions
with little or without knowledge about its environment, and it can be used to adapt
to randomly changing environmental conditions [29, 30]. Reinforcement learning
is actually very used in Robotics [31]. Despite the fact that deep learning has
been extremely successful in many applications; it has remained limited to appli-
cations in which useful features can be created manually [32] or applications with
fully observed low dimensional state space. Volodymyr et al. [32] used deep neural
network to develop a new artificial agent, termed a deep Q-network, this agent can
learn successful policies from high-dimensional input. In this work they demon-
strated that a single architecture with very essential prior knowledge such as pixels
and the game score as inputs can successfully learn control policies at human-level
control. Alejandro et al., developed a deep reinforcement learning strategy for UAV
autonomous landing on a moving platform [33].
Compared to supervised learning, the amount of feedback the learning system
obtains in reinforcement learning is much less [34]. We can measure the learning
Fig. 2 Development of some objects detectors. Main developments in chronological order are:
R-CNN, SPPNet, Fast-RCNN, Faster RCNN, RFCN, FPN, Mask RCNN [66]
Fig. 3 Different fields and sub-fields of machine learning
performance by the number of correct answers, resulting in predictive accuracy. The

difficulty comes from the possibility that learning can generalize new unclassified
examples. In supervised and unsupervised learning, the data is usually considered
static, and it is not the case in RL, where we notice a stochastic nature, this issue
can be solved thought a constant interaction between evaluation and improvement
of policies, and the use of learning rate adaption schemes [34]. Additionally, in front
of some problems, it can be useful to provide agent with rewards for reaching inter-
mediate sub-goals [34]. Christiano and et al., reported that we need to communicate
complex goals to get sophisticated reinforcement learning systems, in their study
they show that’s possible to solve complex RL tasks without access to the reward
function [35].
The most universally used machine learning methods are supervised learning
methods, and we found them in several applications such as spam e-mail filtering,
face recognition and identification, and medical helping diagnosis [36]. In supervised
learning, the agent observes some input-output, and learns a map. The agent thought
knowledge of the input; achieve the output labeling [37]. If the output is a discrete
number of possible “classes”, this is called a classification problems, if the output is
continue it is called a regression [38]. In unsupervised learning, we haven’t labeled
output the agent focuses on observing patterns without labeled data. The aim is
to search pattern structure and features embedded with data [39]. Clustering is an
example of unsupervised learning. By eyeballing data, the agent will be able to
discover the existence of many apparent clusters, then classify each data into the
corresponding one [38].
3 Machine Learning Applied to UAVs for Autonomous

Flight
UAV can be used in many different services, in the military context, the research are
focused on enhancing the autonomy, with different kinds of missions [40]:
• Suppression of enemy air defense.
• Air-to-ground targeting scenario.
• Surveillance and reconnaissance.
• Avoidance of danger zones.
In the civilian context, UAVs could be used for purposes such as:
• Weather forecast/Disaster management [41, 42, 67].
• Urbane police surveillance.
• Agriculture production management [43].
• Border surveillance and inspection of infrastructures [44].
• Road Traffic Monitoring (RTM) [20].
• Emergency Response (ER) system [20].
Let us first cite what John Wyndham said: “The man and machine are naturally
complementary” [45]. One of the current issues in AI is to make the link between
heuristics used by human and programming. Another question is to know how to
use the large data collected from various drone sensors. Many questions are pushing
UAV research: how does a machine fly an aircraft like a human can, can a machine
really “think” enough to fly autonomously? [46]. Beside those questions, UAVs
have some challenges for control, real-time path planning and object recognition
beneath uncertain environments [37]. To solve these problems, many approaches
have been proposed such as negotiation approach, a heuristic approach, and graph
theory [47]. Machine learning is an attractive approach to overcome these challenges
for autonomous flight. It allows recognizing patterns or predicting from data [37].
Machine learning has contributed several fields of UAVs applications. Figure 4
shows the time-line of previous studies covered below [37].
Among the control strategies applied to an autonomous flight, we recognize the
parameter tuning and real time path planning and navigation.
To compensate the limitations of the Proportional, Integral Derivative (PID)
control systems technique, and operate in unpredictable and harsh environments,
reinforcement learning (RL) is an active and successful area of research [48]. In
reinforcements learning (RL) an agent is given a reward for every action it makes in
Fig. 4 Timeline of previous studies on machine learning applied on UAVs [37]
an environment with the objective to maximize the rewards. When we use reinforce-
ment learning, it is possible to develop some optimal control policies for a UAV, and
avoid making any assumptions about the aircraft dynamics [48].
Real-time path planning and navigation have been considered as elementary and
required in UAVs for autonomous flight [37]. Many researchers have exploited and
work on this area of research [49, 50, 68]. Loquerico et al. focused on civilian drones,
because controlling the UAVs can become difficult when we work in urban areas, in
those cases, the autonomous agent is not only expected to navigate while avoiding
collisions, but to interact safely with other agents such as pedestrians or cars [51].
In this research they proposed a model that learns to navigate by initiating cars and
bicycles and respecting the traffic rules.
Collision avoidance is also a very important research track, which can lead to new
and very interesting technologies. Lei et al. [52] proposed a restructured Q-value
learning algorithm based on reinforcement learning.
The autonomous navigation for large unmanned aerial vehicles (UAVs) is not
problem; it is the micro and small UAVs, which fly at low altitude in crowded envi-
ronments that meet many challenges in this area [53]. Ross et al., presented a small
quadrirotor that navigates autonomously at low altitude through natural forest envi-
ronments, They presented a MAV that autonomously fly at speeds of up to 1.5 m/s
and altitudes of up to 4 m above the ground. They developed model based on the
knowledge of a human pilot avowing collision with trees and they succeeded in
more than 680 trees [53]. However, this method is limited when we need to perform
longer flights and denser forests or any cluttered environments, The reactive method
is limited in case of narrow field view, and causes most of time failures. Junell et al.,
proposed a high level reinforcement learning algorithm for UAVs to avoid collisions
[54], they studied autonomous flight of UAV in unknown or uncertain environments,
by taking pictures from disaster site. This method is efficient to solve the conflict
between limited battery life and the big number of required iterations. However, it is
limited in finding an efficient rout to the destination, and learning the best recharging
point, in the aim of having the highest level of battery when it reaches the destination.
Nusrsultan.I and al, achieved experimental study of autonomous navigation by
using real-time model based reinforcement learning [29], which solved some limita-
tions of the previous method [54]. Moreover, deep learning contributed in many appli-
cations, such as construction of database for aerial image classification in emergency
response, and development of suitable CNN training strategy with low-computational
and low-cost low-power [55]. Dmitriy et al., worked also on emergency response,
but this time was about wildfires detection using unmanned aerial vehicles and they
compared several methods in term of speed and accuracy [56].
Cherian et al. proposed a semi-supervised machine learning (SSL) algorithm to
estimate the altitude of UAVs using top down aerial images [57]. The basic idea of
this study is to learn mapping between the texture information contained in an image
with a possible altitude. However this approach is only suited for low altitudes and
low speeds. Manjia Wu et al., proposed an approach to detect intruding drones in
sensitive areas, in real time using deep learning [58].
Most of cited studies are based on supervised learning, which control known
problems; it is suggested to aim researches on unsupervised learning methods
[37]. Reducing time spent on classification is another issue, Gonzalo-Martin et al.,
proposed a strategy to reduce the time spent on the classification trough another
method called “superpixel” segmentation. The results evince that while the classifi-
cation accuracy was identical to the results generated by pixel-based approach, the
time spent is dramatically decreased [28]. In the work of Gonzalo et al. [28], we
didn’t notice the tests of this method on unmanned aerial vehicle unlike the work of
Taro Suzuki et al. [59] who used this method in vegetation classification.
In classification approach based on CNN, training is achieved using big dataset
[29], which require high performance equipment for processing.
None of actual commercial drones, own sufficient control autonomy to achieve
missions without human skills, which makes the missions slow, dangerous and not
scalable [60].
In an interview with BBC, Bill Gates discussed AI and he told: “I am in the camp
that is concerned about super intelligence … That should be positive if we manage it
well. A few decades that thought the intelligence is strong enough to be a concern.”
Tesla’s Elon Musk believes that we should be very concerned when it comes to
AI. According to Stephen Hawking AI could be harmful to humans: “I think the
development of full artificial intelligence could spell the end of the human race”.
Legal regulation should give to autonomous flight, great attention to eliminate
the danger for people and things [61]. The existence of moving obstacles push the
researchers to equip the drone with additional sensors such as: “Sonars” and “Passive
infrared sensors” [61]. At the same level of concern, drones damages during flight
couldn’t be acceptable, and a check of components state and weather conditions
should be performed. A mistake and danger come from flights under bridges, inside
tunnels, or near high-voltage power lines, which could lead to GPS data errors, and
by result drone flight errors [62].
In the other hand, the scope of permissible self-help in defending the privacy
should be broad-ranging [63]. The worst scenario is when a drone escapes with
intrusive recordings that can be a major harm. Another issue concerns the images of
detectable individuals captured by aerial surveillance which should not be retained or
shared except if it contains criminal suspicion activity [2]. Froomkin and Colangelo,
suggested measures including forbidding weaponized robots, and mandating RFID
chips and serial numbers, in the aim of identifying the robot’s owner [63]. The United
States has had a monopoly over the use of drones, but cannot maintain that much
more [64]. These new weapons will not transform the international system, as did the
fast increase of nuclear weapons and ballistic missiles, they still be highly dangerous
and deadly [64]. In USA, the Federal Aviation Administration (FAA) predicted that
30,000 drones could be flying in sky in the next 20 years [65]. When the numbers
of UAVs increase, many accidents can happen in the sky, that is why several models
was proposed to make the process easy and safe [65].
4 Conclusion
UAVs are expected to be more used, as there is considerable demand in all sectors
(private and public). UAVs have the potential to be used in huge number of appli-
cations. Many solutions are suggested such as: adding sensors, enhancements of
data processing and others can expand their use. Artificial intelligence is the most
important solution, to boost their performances and allowing them to be autonomous.
However, the privacy, security and cognitive aspect should not be ignored. In this
work we presented the UAVs, their application, than we presented machine learning
and its application on UAVs, finally we spoke about the aspects of security and
privacy which are solvable.
References
1. Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification with deep convolutional

neural networks. In: Conference on neural information processing systems, p 10971105
2. Cavoukian A (2012) Privacy and drones: unmanned aerial vehicles information and privacy
commissioner of Ontario, Canada Ontario
3. Bou-Ammar H, Voos H, Ertel W (2010) Controller design for Quadrotor UAVs using rein-
forcement learning. In: Proceedings of IEEE international conference on control applications,
Japan, 8–10 September 2010
4. Carrio A, Sampedro C, Rodriguez-Ramos A, Campoy P (2017) A review of deep learning
methods and applications for unmanned aerial vehicles, Hindawi. J Sensors
5. Martinez C, Sampedro C, Chauhan A, Campoy P (2014) Towards autonomous detection
and tracking of electric towers for aerial power line inspection. In: Proceedings of the 2014
international conference on unmanned aircraft systems, ICUAS 2014, pp 284–295
6. Olivares-Mendez MA, Fu C, Ludivig P et al (2015) Towards an autonomous vision-based
unmanned aerial system against wildlife poachers. Sensors 15(12):31362–31391
7. Carrio A, Pestana J, Sanchez-Lopez J-L, et al (2016) Ubristes: UAV-based building rehabil-

itation with visible and thermal infrared remote sensing. In: Proceedings of the robot 2015:
second iberian robotics conference, Springer, pp 245–256
8. Li L, Fan Y, Huan X, Tian L (2016) Real-time UAV weed scout for selective weed control
by adaptive robust control and 12 Journal of sensors machine learning algorithm. In: Proceed-
ings of the 2016 ASABE annual international meeting. American Society of Agricultural and
Biological Engineers, p 1
9. Anwar Ma’sum M, Kholid Arrofi M, Jati G, Arifin F, Nanda Kurniawan M, Mursanto P, Jatmiko
W (2013) Simulation of intelligent unmanned aerial vehicle (UAV) for military surveillance,
ISBN: 978-979-1421-19-5, ICACSIS 2013
10. Mitchell TM (1997) Machine learning, vol. 45(37). McGraw Hill, Burr Ridge, Ill
11. El Naqa I, Murphy MJ (2015) Machine learning in radiation oncology: theory and appli-
cations. https://doi.org/10.1007/978-3-319-18305-3_1 What is Machine Learning? Springer,
Switzerland
12. Ifrah G (2001) The universal history of computing: from the abacus to the quantum computer.
Wiley, New York
13. Samuel AL (1959) Some studies in machine learning using the game of checkers. IBM J Res
Dev 3:210–29
14. Rosenblatt F (1958) The perceptron: a probabilistic model for information storage and
organization in the brain. Psychol Rev 65:386–408
15. Werbos PJ (1974) Beyond regression: new tools for prediction and analysis in the behavioral
sciences. PhD thesis, Harvard University
16. Quinlan JR (1986) Induction of decision trees. Mach Learn 1:81–106
17. Schapire RE (1999) A brief introduction to boosting. In: Proceedings of the 16th international
joint conference on artificial intelligence, vol 2. Morgan Kaufmann Publishers, Stockholm, pp
1401–6
18. Breiman L (2001) Random forests. Mach Learn 45:5–32
19. Rohan A, Rabah M, Kim S-H, et al (2019) Convolutional neural network-based real-time object
detection and tracking for parrot AR drone 2. IEEE Access
20. Christos Kyrkou GP, Theocharides T, Venieris SI, Bouganis C-S (2018) DroNet: efficient
convolutional neural network detector for real-time UAV applications design, automation and
test in Europe
21. Ren S, He K, Girshick R, Sun J (2017) Faster R-CNN: towards real-time object detection with
region proposal networks’. IEEE Trans Pattern Anal Mach Intell 39(6):1137–1149
22. Dai J, Li Y, He K, Sun J (2016) R-FCN: object detection via region-based fully convolutional
networks. In: Proceedings of NIPS, pp 379–387
23. Redmon J, Divvala S, Girshick R, Farhadi A (2016) You only look once: unified, real-time
object detection. In: Proceedings of IEEE conference computing visual pattern recognition
(CVPR), pp 779–788
24. Liu W, Anguelov D, Erhan D, Szegedy C, Reed S, Fu C-Y, Berg AC (2016) SSD: single shot
MultiBox detector. In: Proceedings of ECCV, pp 21–37
25. Jia Y, Shelhamer E, Donahue J, Karayev S, Long J, Girshick R, Guadarrama S, Darrell T (2014)
Caffe: convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093
26. Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, Erhan D, Vanhoucke V, Rabinovich
A (2015) Going deeper with convolutions. In: IEEE international conference on computer
vision and pattern recognition, pp 1–9
27. Castelluccio M, Poggi G, Sansone C, Verdoliva L (2015) Land use classification in remote
sensing images by convolutional neural networks cs.CV
28. Gonzalo-Martin C, Garcia-Pedrero A, Lillo-Saavedra M, Menasalvas E (2016) Deep learning
for superpixel-based classification of remote sensing images
29. Imanberdiyev N, Fu C, Kayacan E, Chen I-M (2016) Autonomous navigation of UAV by
using real-time model-based reinforcement learning. In: 2016 14th international conference on
control, automation, robotics and vision (ICARCV), pp 1–6
30. Sutton RS, Barto AG (1998) Reinforcement learning: an introduction. MIT Press, Cambridge,
MA
31. Kober J, Bagnell JA, Peters J (2013) Reinforcement learning in robotics: a survey. Int J Robot
Res 32(11):1238–1274
32. Volodymyr M, Koray K, David S, Andrei AR, Joel V (2015) Human-level control through deep
reinforcement learning. Nature 518:529–533
33. Rodriguez-Ramos A, Sampedro C, Bavle H, De La Puente P, Campoy PA (2019) Deep rein-
forcement learning strategy for UAV autonomous landing on a moving platform. J Intell Robotic
Syst 93:351–366
34. Wiering M, van Otterlo M (2012) Adaptation learning & optimization. In: Reinforcement
learning, vol 12. Springer
35. Christiano PF, Leike J, Brown TB, et al (2017) Deep reinforcement learning from human
preferences. In: 31st conference on neural information processing systems (NIPS 2017). Long
Beach, CA
36. Jordan MI, Mitchell TM (2015) Machine learning: trends, perspectives, and prospects Science.
Am Assoc Adv Sci 349:255–260
37. Yeon Choi S, Cha D (2019) Unmanned aerial vehicles using machine learning for autonomous
flight; state-of-the-art. Adv Robot 33(6):265–277. https://doi.org/10.1080/01691864.2019.158
6760
38. Barber D (2012) Bayesian reasoning and machine learning. Cambridge University Press,
Cambridge
39. Tai L, Liu M (2016) Deep-learning in mobile robotics-from perception to control systems: a
survey on why and why not. arXiv preprint arXiv:1612.07139
40. Allaire FCJ, Tarbouchi M, Labonté G, Fusina G (2008) FPGA implementation of genetic
algorithm for UAV real time planning. J Intell Robotic Syst
41. Klemas VV (2018) Coastal and environmental remote sensing from un-manned aerial vehicles:
an overview. J Coast Res 1260–1267. https://doi.org/10.2112/jcoastres-d-15-00005.1
42. Erdelj M, Natalizio E, Chowdhury KR, Akyildiz IF (2017) Help from the sky: leveraging UAVs
for disaster management. IEEE Pervasive Comput 16(1):24–32. https://doi.org/10.1109/mprv.
2017.11
43. Huang Y, Thomson SJ, Hoffmann WC, Lan Y, Fritz BK (2013) Development and prospect
of unmanned aerial vehicle technologies for agricultural production management. Int J Agric
Biol Eng 6(3):1–10. https://doi.org/10.3965/j.ijabe.20130603.001
44. Ham Y, Han KK, Lin JJ, Golparvar-Fard M (2016) Visual monitoring of civil infrastructure
systems via camera-equipped unmanned aerial vehicles (UAVs): a review of related works.
Visual Eng 4(1):1. https://doi.org/10.1186/s40327-015-0029-z
45. Fabre L, Touzet C (2016) Drones autonomes: peut-on embarquer la prise de décision? https://the
conversation.com/drones-autonomes-peut-on-embarquer-la-prise-de-decision-2-57758 (hal-
01355025)
46. Dufrene WR Jr, et al (2003) Application of artificial intelligence techniques in uninhabited
aerial vehicle flight. In: Digital avionics systems conference
47. Bortoff SA (2000) Path planning for UAVs. In: (Proceedings of) the 2000 American control
conference, vol 1. IEEE, pp 364–368
48. Koch W, Mancuso R, West R, Bestavros A (2018) Reinforcement learning for UAV attitude
control. arXiv:1804.04154v1,[CS.Ro]11
49. Kahn G, Villaflor A, Ding B et al (2018) Self-supervised deep reinforcement learning with
generalized computation graphs for robot navigation. In: 2018 IEEE international conference
on robotics and automation (ICRA). IEEE, pp 1–8
50. Lee D, Kim S, Suk J (2018) Formation flight of unmanned aerial vehicles using track guidance.
Aerosp Sci Technol 76:412–420
51. Loquercio A, Maqueda AI, delBlanco CR et al (2018) Dronet: learning to fly by driving. IEEE
Robot Automat Lett 3(2):1088–1095
52. Lei G, Dong M-z, Xu T et al (2011) Multi-agent path planning for unmanned aerial vehicle based
on threats analysis. In 2011 3rd international workshop on intelligent systems and applications
(ISA). IEEE, pp 1–4
53. Ross S, Melik-Barkhudarov N, Shankar KS et al (2013) Learning monocular reactive UAV

control in cluttered natural environments. In: 2013 IEEE international conference on robotics
and automation. IEEE, pp 1765–1772
54. Junell JL, VanKampen E-J, de Visser CC et al (2015) Reinforcement learning applied to a
quadrotor guidance law in autonomous flight. In: AIAA guidance, navigation, and control
conference, p 1990
55. Christos Kyrkou GP, Theocharides T (2019) Deep-learning-based aerial image classification
for emergency response applications using unmanned aerial vehicles. In: CVPR workshop on
computer vision for UAVS 2019
56. Alexandrov D, Pertseva E, Berman I, Pantiukhin I, Kapitonov A (2019) Analysis of machine
learning methods for wildfire security monitoring with an unmanned aerial vehicles. In: 2019
24th conference of open innovations association (FRUCT), pp 3–9. Moscow, Russia. https://
doi.org/10.23919/fruct.2019.8711917
57. Cherian A, Andersh J, Morellas V et al (2009) Autonomous altitude estimation of a UAV
using a single onboard camera. In: IEEE/RSJ international conference on intelligent robots
and systems. IROS 2009. IEEE, pp 3900–3905
58. Wu M, Xie W, ShiThird X et al (2018) Real-time drone detection using deep learning approach.
In: International conference, MLICOM 2018, Hangzhou, China, 6–8 July 2018
59. Suzuki T, Tsuchiya T, Suzuki S, Yamaba A (2016) Vegetation classification using a small UAV
based on superpixel segmentation and machine learning. J Remote Sens Soc Jpn
60. Floreano D, Wood RJ (2015) Science, technology and the future of small autonomous drones
nature. Nature Publishing Group, pp 521, 460
61. Gatteschi V, Lamberti F, Paravati G, Sanna A, Demartini C, Lisanti A, Venezia G (2015)
New frontiers of delivery services using drones: a prototype system exploiting a quadcopter
for autonomous drug shipments. In: Proceedings of IEEE 39th annual computer software and
applications conference, pp 2, 920–927
62. Sato S, Anezaki T (2017) Autonomous flight drone for infrastructure (transmission line) inspec-
tion (2). In: Proceedings of international conference on intelligent informatics and biomedical
sciences (ICIIBMS), pp 294–296
63. Froomkin A, Colangelo P (2015) Self-defense against robots and drones. Connecticut Lab Rev
48
64. Kreps S, Zenko M (2014) The next drones wars preparing for proliferation. Foreign Aff 93, 68
65. Dung ND, Rohacs J (2018) The drone-following models in smart cities. In: 2018 IEEE 59th
international scientific conference on power and electrical engineering of Riga Technical
University (RTUCON), pp 1–6
66. Agarwal S, Ogier Du Terrail J, FrÉdÉric J (2019) Recent advances in object detection in
the age of deep convolutional neural networks, computer vision and pattern recognition, p 20.
arXiv:1809.03193
67. Cruz H, Eckert M, Meneses J, Martınez J-F (2016) Efficient forest fire detection index for
application in unmanned aerial systems (UASs). Sensors 16(6):893. https://doi.org/10.3390/
s16060893
68. Ambroziak L, Gosiewski Z (2015) Two stage switching control for autonomous formation
flight of unmanned aerial vehicles. Aerosp Sci Technol 46:221–226
Fuzzy H∞ Delay-Independent
Stabilization of Depth Control
for Underwater Vehicle with Input
Constraint
Mohamed Nasri, Dounia Saifia, Salim Labiod, and Mohammed Chadli
Abstract This paper investigates the depth control of unmanned underwater vehicle
(UUV) in the presence of external disturbances, time delay, and actuator satura-
tion. Firstly, a Takagi–Sugeno model is elaborated based on the speed variation.
Then, using the polyquadratic representation with Lyapunov-Krasovskii function,
an H∞ criteria is applied to design a new LMI based stabilization to calculate PDC
controller gains taking into consideration the input limitation. Finally, simulation
results demonstrate the superiority of the proposed approach.
Keywords Takagi-Sugeno model · Delay-independent stability · H∞

performance · Actuator saturation · Lyapunov Krasovskii function
1 Introduction
Over last decade, unmanned underwater vehicle have take a growing attention.
Indeed, the high ability of UUV devices to replace humans tasks, allowing to
various sectors such as: military, scientific research …etc. [1, 2] to use it. Partic-
ularly in missions where human life is in danger. However, these works cannot be
achieved without implementation of precise depth controller for UUV. For this raison,
many studies have developed various depth control strategies such as: sliding mode
controller [3], fuzzy controller [4], adaptive controller [2].
M. Nasri (B) · D. Saifia · S. Labiod

Faculty of Science and Technology, LAJ, University of Jijel, Ouled Aissa, 18000 Jijel, Algeria
D. Saifia
S. Labiod
M. Chadli
IBISC Lab, University of Paris-Saclay, 40 rue Pelvoux, Evry 91020, France

https://doi.org/10.1007/978-981-15-6403-1_21
306 M. Nasri et al.
Among depth control strategies, the approaches based on Takagi-Sugeno models

[5], have been largely utilized [6, 7], either parallel distributed compensation (PDC)
or static output feedback (SOF) controllers. In any case, T-S-based approaches have
many advantages; especially have facilitated maneuvering in control by the imple-
mentation of dedicated techniques for linear systems. The depth control challenge
consists at stabilizing the UUV under deferent obstacle like: sea waves, water current,
time delay…etc. for this raison, the previous studies, have been focused on solving
these problems (for example, see [4, 6]. However, the input limitation, which is the
hydroplane angle, has never been taken into account. Indeed, actuator saturation is
an important factor to the performances degradation of closed-loop systems; in some
cases, it can drive the system to instability [8]. This influence motivate the scientists
to research a control approach that can guarantee the closed-loop performances in
the presence of input constraints [9, 10]. In this context, the saturation nonlinearity
can be modeled by sector nonlinearity, by dead zone representation, or by polytopic
representation [8]. Among control solutions, the nicknamed saturated controller [11],
gives better stabilization performances, even though, it can cause chattering.
Besides, the stabilization problem of T-S systems with time delay is very chal-
lenging. Indeed, time delay is often encountered in many practical systems [6, 12].
Generally, the studies for time delay systems can be classified into two categories, the
first one uses delay-independent criteria, and the second one uses delay-dependent
criteria [13]. For our case, we rely on the first criterion to deduce the stabilization
conditions of the closed-loop system.
Motivated by the above discussion, this paper investigates the depth control for
UUV system in the presence of external disturbances, time delay, and actuator satu-
ration. The nonlinear model of the UUV is first expressed as a Takagi-Sugeno model.
Using polyquadratic form with Lyapunov Krasovskii function and H∞ criterion, a
new LMI stabilization conditions are derived. Finally, simulation results are given to
demonstrate the effectiveness of the proposed approach.
The rest of this paper is organized as follows. Section 2 gives the system modeling.
Section 3 gives the T-S representation of the UUV system. Saturated controller for
delayed UUV system is developed in Sect. 4. In Sect. 5, simulation results are given.
A conclusion in Sect. 6 finishes the paper.
2 Mathematical Modeling of Under Actuated UUV System
The body fixed coordinate system and the earth coordinate system are depicted in
Fig. 1, where x, y, z are the Cartesian position of UUV, and φ, θ , ψ are the vehicle
altitude with respect to the earth fixed coordinate system. u, v, w are surge, sway,
heave velocities. p, q, r are the angular velocities (roll, pitch, yaw).
Fuzzy H∞ Delay-Independent Stabilization of Depth Control … 307
Fig. 1 The reference of the

UUV
Based on the works in [6], the mathematical model of the depth control system
can be simplified as follows
Heave equation of motion:

m ẇ − uq − x G q̇ − z G q 2 = Z q̇ q̇ + Z ẇ ẇ + Z uq uq+
(1)
Z uw uw + (W − B0 ) cos(θ ) + Z uu u 2 δs
Pitch equation:
I yy + m[x G (uq − ẇ) + z G wq] = Mq̇ q̇ + Mẇ ẇ + Muq uq

+Muw uw − (x G W − x B B0 ) cos(θ )− (2)
(z G W − z B B0 ) sin(θ ) + Muu u 2 δs
Pitch and depth:
ż = w cos(θ ) − u sin(θ )
(3)
θ̇ = q
where (x G , yG , z G ) and (x B , y B , z B ) are the centre of gravity and buoyancy.

I yy , Z {.} , M{.} are the constant parameters of UUV, B0 is the buoyancy of the
UUV; W is the weight of the UUV; u is the speed of UUV; δs is the hydroplane
angle.
Now, let us define:
ze = zd − z (4)
308 M. Nasri et al.
with z d is the desired depth value. By assuming that:
• (x G , yG , z G ) = (x B , y B , z B ) = 0
• The heave velocity w has little effect on the depth control
sin(θ ) = θ, cos(θ ) = 1
From the above equations and assumptions, we can give the mathematical model
of depth control system as follow:
⎧
⎨ ż e (t) = u(t)θ (t)
θ̇ (t) = θ (t) (5)
⎩
q̇(t) = a(t)q(t) + b(t)δs (t)
M u(t) Muu u 2 (t)

with: a(t) = I yyuq−Mq̇ , b(t) = I yy −Mq̇
T
Let x = z e θ q
Therefore,
ẋ(t) = A(t)x(t) + B(t)δs (t) (6)
where
⎡ ⎤ ⎡ ⎤
0 u(t) 0 0
A(t) = ⎣ 0 0 1 ⎦, B(t) = ⎣ 0 ⎦
0 0 a(t) b(t)
A. Dynamic model with time delay
Based on the time-delay model of Network Control Systems NCS in [14], the
mathematical model of depth control can be expressed as follows [6]:
⎧
⎨ ẋ(t) = Ãx(t) + Ãd x(t − τ ) + Bδs (t) + Bϕ ϕ(t)ẋ(t) = Ãx(t) + Ãd x(t − τ )
+ Bδs (t) + Bϕ ϕ(t)
⎩
x(t) = η(t), t ∈ [−τ, 0]
(7)
where Ã = β A(t) and Ãd = (1 − β)A(t), β is the steady coefficient of the depth
control system, ϕ(t) is the external perturbations, and η(t) is the initial state of the
system.
3 Takagi-Sugeno Model for the Depth Control System
In this section, a T-S model is developed based on the sector nonlinearity approach
[15]. In this context, unlike previous works [6], the UUV speed u is supposed time
varying.
The T-S fuzzy representation of a delayed continuous system is complies with the
following IF-THEN form:
Rulei : if ς1 (t) is about M1i and … and ςq (t) is about M1q

ẋ(t) = Ãi x(t) + Ãdi x(t − τ ) + Bδs (t) + Bϕ ϕ(t)
(8)
y(t) = C x(t)y(t) = C x(t)
In our case the nonlinearity in the depth control system is the UUV speed which
is time varying and bounded (u min ≤ u(t) ≤ u max ). Consequently, by applying the
sector nonlinearity method, the T-S fuzzy model of the depth control system is
expressed as follows:
⎧
⎪ 8
⎪
⎪ ẋ(t) = μ (ς (t)) Ãi x(t) + Ãdi x(t − τ )+
⎪
⎨
i
i=1
Bδs (t) + Bϕ ϕ(t) (9)
⎪
⎪
⎪
⎪ y(t) = C x(t)
⎩
ψ(t) = C1 x(t)
where y(t) is the output and ψ(t) is the controlled output. μi (ς (t)) are the
membership functions, which are calculated as follows:
Let us write,
V1min ≤ u(t) = V1 ≤ V1max

(10)
V2min ≤ u 2 (t) = V2 ≤ V2max
Using sector nonlinearity method, we get
V1 −V1min
F11 = V1max −V1min
, F12 = 1 − F11
V2 −V2min (11)
F21 = V2max −V2min
, F22 = 1 − F21
Consequently, the membership functions can be written as:
μ1 = F11 F21 , μ2 = F11 F22

(12)
μ3 = F12 F21 , μ4 = F12 F22
with
310 M. Nasri et al.
⎡ ⎤
0 βV1max 0
⎢ ⎥
Ã1 = ⎣ 0 0 1 ⎦;
Muq
0 0 β I yy −Mq̇ V1max
⎡ ⎤
0 βV1max 0
⎢ ⎥
Ã2 = ⎣ 0 0 1 ⎦;
Muq
0 0 β I yy −Mq̇ V1max
⎡ ⎤
0 βV1min 0
⎢ ⎥
Ã3 = ⎣ 0 0 1 ⎦;
Muq
0 0 β I yy −Mq̇ V1min
⎡ ⎤
0 βV1min 0
⎢ ⎥
Ã4 = ⎣ 0 0 1 ⎦;
Muq
0 0 β I yy −Mq̇ V1min
⎡ ⎤ ⎡ ⎤
0 0
⎢ ⎥ ⎢ ⎥
B1 = ⎣ 0 ⎦; B2 = ⎣ 0 ⎦;
Muu Muu
V
I yy −Mq̇ 2max
V
I yy −Mq̇ 2min
⎡ ⎤ ⎡ ⎤
0 0
⎢ ⎥ ⎢ ⎥
B3 = ⎣ 0 ⎦; B4 = ⎣ 0 ⎦;
Muu Muu
I yy −Mq̇
V2max I yy −Mq̇
V 2min
⎡ ⎤
0 (1 − β)V1max 0
⎢ ⎥
Ãd1 = ⎣ 0 0 1 ⎦;
Muq
0 0 (1 − β) I yy −Mq̇ V1max
⎡ ⎤
0 (1 − β)V1max 0
⎢ ⎥
Ãd2 = ⎣ 0 0 1 ⎦;
Muq
0 0 (1 − β) I yy −M V1max
⎡ ⎤
q̇
0 (1 − β)V1min 0
⎢ ⎥
Ãd3 = ⎣ 0 0 1 ⎦;
Muq
0 0 (1 − β) I yy −Mq̇ V1min
⎡ ⎤
0 (1 − β)V1min 0
⎢ ⎥
Ãd4 = ⎣ 0 0 1 ⎦.
Muq
0 0 (1 − β) I yy −M q̇
V 1min
4 Fuzzy H∞ Delay-Independent Stabilization of UUV

System Under Input Constraint
In this section, the goal is to design a PDC controller to guarantee the closed-loop
stability of the depth control system in the presence of time delay, actuator saturation,
and external disturbances.
A. Saturation modeling
In this work, the polytopic representation of saturation will be used.
Let us consider that the input saturation is σ (U ), where is a nonlinear function of
its input defined as
σ (U ) = sat(U ) = sat(δs ) (13)

Now, let us define the set ℵ H j as follows

j
ℵ H j = x(t) ∈ R n / h i x ≤ Ūi (14)
j
with, H j is a matrix m × n, h i is the ith row of the H j , n is the number of the state
variables, m is the number of the control inputs and Ūi is the saturation level of the
control input signal.
The polytopic representation of the saturation is described as follows:
⎧
⎪
2m
⎪
⎪ σ (u) = ηs E s U + Ē s ν
⎪
⎪
⎪
⎨ s=1
r
ν(t) = μ j (ς (t))H j x(t) (15)
⎪
⎪ j=1
⎪
⎪
⎪
⎪
2m
⎩ ηs = 1, 0 ≤ ηs ≤ 1
i=1
with, E s represents all components of E, Ē s = I − E s and E s ∈ {0, 1}

In this case, the PDC control law is given by

r
U (t) = μi (ς (t))K i x(t) (16)
i=1
Therefore, the closed-loop depth control system is given by

⎧
⎨ ẋ(t) = A T + B2T E T K T + Ē T HT x(t)
+ AdT x(t − τ ) + BϕT ϕ(t) (17)
⎩
ψ(t) = z(t) = C1T x(t)ψ(t) = z(t) = C1T x(t)
with
312 M. Nasri et al.

r
r
AT = μi (ς (t))Ai , BϕT = μi (ς (t))Bϕi ,
i=1 i=1
r
r
BT = μi (ς (t))Bi , C1T = μi (ς (t))C1i
i=1 i=1
B, H∞ performance
The H∞ criterion allows us to give the results attained. It is defined by
T T
ψ (τ )ψ(τ )dτ < γ
T 2
ϕ T (τ )ϕ(τ )dτ (18)
0 0
It is equivalent to the following optimization problem:

minδ
(19)
V̇ (t) + ψ T (t)ψ(t) − γ 2 ϕ T (t)ϕ(t) < 0
With V̇ (t) is the derivative of the Lyapunov-Krasovskii function (LKF), which is

chosen based on the delay-independent stability technique. The contribution here is
that the LKF is depending on the membership function, which gives less conservatism
stability conditions and more relaxation. In this context, the new LKF is described
as
r −1 t

V (t) = x (t)
T
μi Pi x(t) + x T (s)x(s)ds (20)
i=1 t−τ
with Pi = PiT > 0, = T > 0.

For a constant ρ > 0 and a symmetric positive definite matrix P, we define an
ellipsoid as

ε(P, ρ) = x ∈ n / x T P x ≤ ρ (21)
Lemma 1 Henrion
and Tarbouriech
[16] The ellipsoid ε(P, ρ) is included in the
polyhedron = x/ m iT x ≤ λi , i ∈ Im if and only if
−1
P
(m i ) T
m i ≤ λi2 (22)
ρ
Theorem 1 For all k = 1, . . . , r − 1; μ̇k (t) ≥ φk , the closed-loop system (17),

is asymptotically stable with an attenuation level of disturbances γ . if there exist
matrices Q i = Q iT > 0, S = S T > 0, matrices F j ∈ R m×n , Z j ∈ R m×n , satisfying
the following LMIs:
⎡ γ
min
r
−1
⎢ Ai Q i + Bi E s F j + Bi Ē s Z j + (∗) + S − φk (Q k − Q r )
⎢ k=1
⎢
⎢ Q i AdTi
⎢
⎣ BϕT
C1 Q i (23)
⎤
∗ ∗ ∗
−(1 − α)S ∗ ∗ ⎥⎥<0
0 −γ I ∗ ⎦
2
0 0 −I
i ∈ Ir , j ∈ Ir

Ū j
zi
ρ ≥0 (24)
∗ Qi

r
Q k − Q r ≥ 0, K j = F j Q −1
T , QT = μi (t)Q i , α ≥ τ̇ ,
with: i=1
j
z i is the i th element o f Z j
Proof Let us define the new LKF as
t
V (t) = x (t)PT x(t) +
T
x T (s)x(s)ds (25)
t−τ
−1

r
with PT = μi Pi
i=1
Applying the H∞ criterion, the closed-loop system is stable if
V̇ (t) + ψ T (t)ψ(t) − γ 2 ϕ T (t)ϕ(t) < 0 (26)
The derivative of LKF (25) is given by
V̇ (t) =ẋ T (t)PT x(t) + x T (t)PT ẋ(t) + x T (t) ṖT x(t)

+ x T (t)x(t) − (1 − τ̇ )x T (t − τ )x(t − τ ) (27)
Using (16) and (17), on can write

V̇ (t) =x T (t) A TT PT + E T K T + Ē T HT B T PT + (∗) + ṖT + x(t)
+ x T (t − τ )PT AdTT x(t) + x T (t)PT AdT x(t − τ )
− (1 − τ̇ )x T (t − τ )x(t − τ ) (28)
314 M. Nasri et al.
In compact form
V̇ (t) = x̄ T (t) Ā x̄(t) (29)

!
A TT PT + E T K T + Ē T HT B T PT + (∗) + + ṖT ∗
with Ā =
AdTT PT −(1 − α)
With the H∞ criterion, one gets
!T !
x̄(t) x̄(t)
V̇ (t) + ψ (t)ψ(t) − γ ϕ (t)ϕ(t) =
T 2 T
T <0 (30)
ϕ(t) ϕ(t)
where
⎡
A TT PT + E T K T + Ē T HT B T PT + (∗) + Π + Ṗ
⎢ AdTT PT
T = ⎢
⎣ BϕT PT
C
⎤
∗ ∗ ∗
−(1 − α) ∗ ∗ ⎥ ⎥
0 −γ I ∗ ⎦
0 0 −I
Pre-and post multiplying by = diag(Q T , Q T , I, I ) with the change of the

variables: Q T = PT−1 ,
F j = K j Q T , Z j = H j Q T , S = Q T Π Q T , we get
⎡ T =

_˙___
−1
⎢ A T Q T + BT E T FT + BT Ē T Z T + (∗) + S + Q T Q T Q T
⎢
⎢ T
Q T A dT
⎢
⎣ BϕT
C1 Q T (31)
⎤
∗ ∗ ∗
−(1 − β)S ∗ ∗ ⎥
⎥
∗ −γ I ∗ ⎦
2
0 0 −I
According to [17], on can write Q T Q −1

T Q T = − Q̇ T , this last can be approximated
by
r −1

Q̇ T ≥ φk (Q k − Q r ) (32)
k=1
Using this approximation with the inequality (31), the condition (23) is hold.
The inequality (24) can be obtained using Lemma 1 and Schur complement [9].
In this section, the results obtained previously have been applied to depth control
for UUV system. The desired depth has been chosen at −5, β = 0.2, the hydroplane
angle has been supposed to be limited between (−0.6 rad) and (0.6 rad), with system
parameter values as in [6]. Using LMI toolbox of Matlab, the resolution of LMI
optimization problem in Theorem 1 gives:
⎡ ⎤ ⎡ ⎤
1.663 −1.0037 1.4282 1.2313 −0.5539 0.6288
⎢ ⎥ ⎢ ⎥
Q1 = ⎢ ⎥ ⎢ ⎥
⎣ −1.0037 0.8293 −1.7818 ⎦, Q 2 = ⎣ −0.5539 0.3264 −0.5269 ⎦
1.4282 −1.7818 5.1649 0.6288 −0.5269 1.1283
⎡ ⎤ ⎡ ⎤
1.1796 −1.0190 0.9322 0.8378 −0.3129 0.0522
⎢ ⎥ ⎢ ⎥
Q3 = ⎢ ⎥ ⎢ ⎥
⎣ −1.0190 1.2685 −1.9412 ⎦, Q 4 = ⎣ −0.3129 0.2405 −0.1472 ⎦
0.9322 −1.9412 4.1856 0.0522 −0.1472 0.3496

F1 = F2 = F3 = F4 = −0.0607 0.0850 24.3951 ,
γ = 1.1348.
Using a disturbance model as

10 6
ϕ(t) = 0.5 sin t + 0.3 sin t .
π π
Fig. 2 Saturated input signal

316 M. Nasri et al.
Fig. 3 The depth signal
The obtained results have been depicted in Fig. 2. for control input, Fig. 3 for the
depth z(t), Fig. 4 for pitch angle, and Fig. 5 for angular velocity.
Fig. 4 The pitch angle

signal
Fig. 5 The angular velocity

signal
From the shown Figs. 2, 3, 4 and 5, one can see that the proposed approach ensures
the closed-loop stability for deferent time delay and input constraint, in the presence
of waves and sea current disturbances.
6 Conclusion
In this paper, the depth control of the underactuated UUV under the state-delay,
external disturbances, and actuator saturation has been investigated. First, a new
Takagi-Sugeno delayed model has been elaborated based on the NCS model and
speed variation. Then, using the LKF, polyquadratic form, and H∞ cretirion, a
new LMI stabilization conditions have been designed. Finally, simulation results
demonstrated the effectiveness of the proposed designed control approach.
References
1. Zheng Z, Ruan L, Zhu M (2019) Output-constrained tracking control of an underactuated

autonomous underwater vehicle with uncertainties. Ocean Eng 175:241–250
2. Tabataba’i-Nasab FS, Keymasi Khalaji A, Moosavian SAA (2019) Adaptive nonlinear control
of an autonomous underwater vehicle. Trans Inst Meas Control 41(11):3121–3131
3. Guzmán MJV, Lara JLR, Hernández JAR, Alazki H (2018) Robust control for the dynamics
of an unmanned surface vehicle under the perturbations (Marine Waves and Currents). In:
The 15th international conference on electrical engineering, computing science and automatic
control (CCE). 5–7 September 2018, Mexico City, Mexico, pp 1–6
4. Ma C, Qiao H, Kang E (2018) Mixed H∞ and passive depth control for autonomous underwater
vehicles with fuzzy memorized sampled-data controller. Int J Fuzzy Syst 20(2):621–629
5. Takagi T, Sugeno M (1985) Fuzzy identification of systems and its application to modeling
and control. IEEE Trans Syst Man and Cyber 1115(1):116–132
6. Zhang X, Xiao Y, Zhang G, Sun X, Zhao X (2016) Delay-independent stabilization of
depth control for unmanned underwater vehicle. In: 2016 IEEE international conference on
mechatronics and automation, 7–10 August 2016. Harbin, China, pp 2494–2499
7. Silvestre C, Pascoal A (2007) Depth control of the INFANTE AUV using gain-scheduled
reduced order output feedback. Control Eng Pract 15(7):883–895
8. Tarbouriech S, Garcia G, da Silva Jr JMG, Queinnec I (2011) Stability and stabilization of
linear systems with saturating actuators. Springer
9. Nasri M, Saifia D, Chadli M, Labiod S (2019) H∞ static output feedback control for electrical
power steering subject to actuator saturation via fuzzy Lyapunov functions. Trans Inst Meas
Control 41(12):3340–3351
10. Saifia D, Chadli M, Labiod S, Guerra TM (2012) Robust H∞ static output feedback stabilization
of TS fuzzy systems subject to actuator saturation. Int J Control Autom Syst 10(3):613–622
11. Nasri M, Saifia D, Chadli M, Labiod S (2019) H∞ fuzzy control for electrical power
steering subject to actuator saturation. In: Advanced control engineering methods in elec-
trical engineering systems, ICEECA 2017. Lecture Notes in electrical engineering, vol. 522,
Springer
12. Li H, Liu H, Gao H, Shi P (2011) Reliable fuzzy control for active suspension systems with
actuator delay and fault. IEEE Trans Fuzzy Syst 20(2):342–357
13. Ting CS (2008) A robust fuzzy control approach to stabilization of nonlinear time-delay systems
with saturating inputs. Int J Fuzzy Syst 10(1):50–60
318 M. Nasri et al.
14. Zhang L, Wu Y, Gong L, Chen H (2008) A novel research approach on network control
systems. In: 2008 international conference on internet computing in science and engineering,
28–29 January 2008, Harbin, China, pp 262–265
15. Morère Y (2001) Mise en oeuvre de lois de commande pour les modèles flous de type Takagi-
Sugeno. Ph.D. thesis, LAMIH, Université de Valenciennes et du Haut Cambrésis, France
16. Henrion D, Tarbouriech S (1999) LMI relaxations for robust stability of linear systems with
saturating. Automatica 35:1599–1604
17. Guelton K, Bouarar T, Manamanni N (2008) Fuzzy Lyapunov LMI based output feedback
stabilization of Takagi Sugeno systems using descriptor redundancy. In: IEEE International
Conference Fuzzy Systems, 1–6 June 2018, Hong Kong, pp 1212–1218
Petri Type 2 Fuzzy Neural Networks
(PT2FNN) for Identification and Control
of Dynamic Systems—A New Structure
and a Comparative Study
Youssouf Bibi, Mohamed Seghiri, Omar Bouhali, Abdelkarim Nemra,

and Tarek Bouktir
Abstract The techniques and theories used for identification and control of uncer-
tain dynamic systems based on the universal approximators are diverse, but can
mainly be divided into two categories, some are basic universal approximators as
fuzzy systems (FS), neural networks (NN) and wavelet neural networks (WNN).
And other hybrid approximators which are combinations between two basic approx-
imators as fuzzy neural wavelet networks (FWNN) and petri fuzzy systems (PFS).
These approximators can be used to estimate the dynamic of system in an aim to
construct a controller, as we can use it directly to estimate the control law using
specific learning algorithms. In this paper, a hybrid structure between type 2 fuzzy
logic and petri networks (PT2FNN) is proposed as a new approximator to alleviate
the problem of uncertainties with an optimal cost. By incorporating Petri layers,
the number of rules is optimized. Moreover, the time consumed is reduced using a
new inference type 2 method. The parameter update algorithms are derived based
on the gradient method. This hybrid structure is tested for the identification and the
control of uncertain dynamic plants. The simulation results showed that the proposed
structure performs better compared with other approximators.
Keywords Identification and control · Uncertain dynamic systems · Soft

computing · Basic approximators · Hybrid approximators · Update algorithms
Y. Bibi (B) · O. Bouhali

Laboratoire mecatronique LMT, Université de Jijel, BOX 98 Ouled Aissa, Jijel, Algerie
M. Seghiri · A. Nemra
Laboratoire véhicules autonomes et intelligents, école militaire polytechnique, Bordj El Bahri,
Algerie
T. Bouktir
Département de génie electrique, Université de Sétif, Sétif, Algerie

https://doi.org/10.1007/978-981-15-6403-1_22
320 Y. Bibi et al.
1 Introduction
Most dynamic systems are characterized by uncertainties in nonlinear terms of

model and parameters. These uncertainties cannot be considered by deterministic
models, and therefore, Traditional controllers based on such models cannot ensure the
required performances. To overcome this problem, the soft computing methodologies
(artificial intelligence) can be emerged as a powerful alternative.
In literature, numerous theories and structures are proposed for solving identifica-
tion and control problems, and improve the performances such as the precision, the
fast convergence and the robustness. Without defining a deterministic plant model,
these structures can be used as a universal approximator to estimate the dynamic of
system aiming to constructing a controller. Most of them offers a way of dealing
with modeling problems by implementing linguistic, nonformally expressed control
laws derived from expert knowledge [1–7].
Neuro-Fuzzy Networks (FNN), Neural Networks (NNs) and Wavelet Neural
Networks (WNN) are the basic universal approximators, which can approximate
continuous unknown functions to any desired accuracy, based on a specific learning
algorithms [8–10]. Numerous researches and comparative studies investigate the
capabilities of approximation for each approximator, according to fixed criteria as its
precision, learning algorithms, cost of implementation and the speed of convergence
[7, 11, 12].
In order to improve the performances of approximation, a hybrid structures are
proposed as new universal approximators which are combinations of basic approx-
imators; such as Fuzzy Wavelet Neural Networks (FWNN) [13] and Petri Fuzzy
Systems [14].
The well-known optimization learning algorithm used for tuning the parameters
of these approximators is gradient descent (GD) algorithm [15]. The convergence
speed may be slow when the search space is complex. To overcome these issues,
sliding mode control (SMC) theory-based algorithms are proposed for adjusting the
parameters of these approximators. These new SMC theory-based learning algo-
rithms ensure faster convergence and make the overall system more robust than the
traditional learning techniques in online tuning [16].
In [17] the authors developed a new universal approximator by the mixing of
Petri networks into type two fuzzy neural networks (T2FNN). By incorporating Petri
layers, the amount of rules is optimized. Moreover, a replacement inference type two
fuzzy system was developed to scale back the time consumed within the old inference
procedure. The Petri type two fuzzy neural networks (PT2FNN) approximator was
applied in the field of adaptive control [18, 19]. Where the adaptive laws of the
adjusted parameters are derived based on the Lyapunov function.
In this work we consider the application of the developed (T2FNN) to solve iden-
tification and control problems. Our goal is to evaluate the performance of this struc-
ture for both identification and control purposes. In order to achieve faster conver-
gence, the gradient descent learning algorithm has been proposed to tune the adjusted
parameters.
Petri Type 2 Fuzzy Neural Networks (PT2FNN) for Identification … 321
This paper is organized as follows; Sect. 2 presents the PT2FNN six-layer struc-
ture. In Sect. 3, the parameter update laws based on the gradient descent method are
derived. Then, we consider a plant models from the literature in order to compare the
performances of the suggested network with other approaches reported. Concluding
remarks and perspectives are given in Sect. 5.
2 Petri Type 2 Fuzzy Neural Networks (PT2FNN)
The integration of fuzzy type 2 theory and Petri networks is developed to alleviate
the problem of uncertainties in the design of controller. The PT2FNN uses a new
inference type 2 fuzzy system to improve the computational power of the neuro-
fuzzy type 2 system. Moreover, the number of rules is optimized by incorporating
Petri layers.
The proposed PT2FNN is constructed on the basis of a set of fuzzy rules. Each
rule includes an IF part and a consequent part. In the PT2FNN, a novel step between
these two parts is introduced using the petri layer that filters out the rules and passes
just those rules which are more important. The rules used have the following form:

j1
j2
jn
if x1 is F 1 and x2 is F 2 . . . and xn is F n then y = θk .k = 1, . . . , N

ji
where [x1 , x2 , . . . , xn ] is the vector of input variables, F i ⊂ U ji i =
1, 2, . . . , n and j = 1, 2, . . . , m and k = 1, 2, . . . , N are the antecedent
type 2
membership functions to describe the linguistic terms, θk = θk , θk are the lower
and the upper singleton consequent type 2 fuzzy sets, y is the output.
The proposed six-layered PT2FNN structure is depicted in Fig. 1, and the duties
of each layer are detailed as follows:
Layer 1 (Input Layer): Each node in this layer is a linear function where the number
of nodes is equal to the number of input signals.
Layer 2 (Membership Function Layer): Each node corresponds to one linguistic
term. In this stage the fuzzification operation is executed using type 2 Gaussian
ji
membership functions Fi defined as:
⎧ x +c
⎪
⎨μ
i −
F i (x i ) = a exp − 2
1 j
ji ji
σj

Fi = (1)
⎪
⎩ x +c̄
μ F j i (x i ) = exp − 21 iσ j j
i
Layer 3 (Petri Layer): This layer produces tokens and makes use of competition
equations as follows to select suitable fired nodes:
322 Y. Bibi et al.
Fig. 1 Architecture of petri type 2 fuzzy neural network
⎧
⎨ 1, μ
ji ≥ dt h
t ij = Fi (xi )
(2)
⎩ 0, μ
ji < dt h
Fi (xi )
where t ij is the transition and dt h is a variable threshold as a function of the tracking

error et , that is calculated by the following equation:
ka exp(−kb et )
dt h = (3)
1 + exp(−kb et )
where ka and kb are positive constants. The tracking error et is inversely proportional
to the threshold values dt h It means that if the tracking error et become large, therefore
many rules will be considered by the PT2FNN and vice versa.
Layer 4 (Rule Layer): This layer computes the implication of each rule using
AND (min) operation. The firing interval of the kth rule ϕ k can be computed using
the following equation:
⎧
⎪
⎪ f¯k = μ F k1 (x1 ) ∗ . . . ∗ μ F ik (xi )
⎪
⎨ ¯k
f = μ F̄1k (x1 ) ∗ . . . ∗ μ F k (xi )
i (4)
⎪
⎪
⎪
⎩ ϕ k = f k , f¯k , k = 1, . . . N
−
Layer 5 (Upper and lower output Layer): Each node in this layer is a consequent
node and functions as a linear model. Each rule node in layer 4 has a corresponding
consequent node θk in layer 5.
There are many such methods of defuzzification to calculate the output of the
whole network. The most commonly used one is the center-of-sets type-reducer:
N
ϕ k (xi )θk
Y (x) = k=1
N
= γl , γr (5)
k=1 ϕ k (x i)
where γl and γr can represent by:

N
k=1 f kθk
N
−
γl = N = ξ k θ k = ξ (x)θ T (6)
k=1 fk k=1
−
−
N
f¯k θ̄k
N
γr = k=1 = ξ̄ k θ̄k = ξ̄ (x)θ̄ T (7)
N
k=1 f¯k
k=1

where θ = θ 1 , θ 2 , . . . , θ N , θ̄ = θ̄1 , θ̄2 , . . . , θ̄ N , ξ = ξ 1 , ξ 2 , . . . , ξ N , ξ̄ =
1 2
ξ̄ , ξ̄ , . . . , ξ̄ N are the vectors of the basis functions with:
fk
− f̄
k
ξ = N
k
, ξ̄ = N
k
p,k = 1, 2, . . . , N (8)
p=1 fp
− p=1 f̄
Layer 6 (Modulation Layer): In this layer an adaptive ponderation between the

upper output γr and the lower output γl is performed using the proposed adaptive
factor α
The output of the PT2FNN Y x, θ, θ is given by:

Y x, θ− , θ̄ = α γl + (1 − α)γr (9)
The constructed PT2FNN structure has the following advantages: The PT2FNN
uses an adaptive modulation as inference method and type 2 membership functions.
The use of such a structure allows the PT2FNN to approximate complex functions
more effectively and with minimum cost because the number of rules is optimized.
324 Y. Bibi et al.
3 Petri Parameter Update Algorithms for Learning
The gradient-based learning algorithm is used to update the parameter of the consid-
ered peti type 2 fuzzy neural networks. The adaptive parameters are the centers cij
and the standard deviations σij of the Gaussian membership functions in the second
layer (memberships
layer) of the network, the parameters of the consequent part
θk = θk , θk and the adaptive modulation.
The value of the cost function E is defined as:
1 d 2
o
E= y − yi (10)
2 k=1 i
where, o is the number of output signals of the PT2FNN network (in our case, we have
one output), yid and yi are the desired and the current output values of the PT2FNN

network, respectively. The parameters of the consequent part θk = θ k , θ k and the
adaptive modulation α, the parameters of the Gaussian membership functions ci j and
σi j of the PT2FNN structure are adjusted according to the following formulas:
⎧ ∂E
⎪
⎪ θ k (t + 1) = θ k (t) − γθ ∂θ
⎪
⎪ k
⎪
⎪ θ (t + 1) = θ (t) − γ ∂E
⎪
⎪
k k θ ∂θ−
⎪
⎪ k
⎪
⎪ α(t + 1) = α(t) − γ ∂E
⎨ α ∂α
c− (t + 1) = c− (t) − γ−c ∂∂cE (11)
⎪
⎪ ij ij −
ij
⎪
⎪ (t + = (t) − γ ∂E
⎪
⎪ c̄i j 1) c̄i j c̄ ∂
⎪
⎪
c̄i j
⎪
⎪ σ (t + 1) = σ (t) − γ σ
∂E
⎪
⎪
ij ij − ∂σ i j
⎩ ∂E
σ i j (t + 1) = σ i j (t) − γσ̄ ∂σ ij
The following formulas represent the values of derivatives in the Eq. (11)
⎧ fk
⎪
⎪ ∂E ∂y
= ∂∂Ey ∂θ = − yid − yi α(t) N f p
⎪
⎪ ∂θ
⎪
⎪
k k
d p=1
f¯k
⎪
⎪ ∂E
= ∂ E ∂y
= − y − y (1 − α(t)) N
⎪
⎪ ∂θ k ∂ y ∂θ k i i ¯p
⎪
⎪ N p=1 f
⎨ ∂ E = ∂ E ∂ y = − y d − y k=1 f k θ k − k=1
⎪ N k
f θk
∂α ∂ y ∂α i i N k N k
k=1 f k=1 f (12)
⎪
⎪ ∂ E ∂y − ∂ E ∂ k
∂ E ∂ y ∂ f¯k
⎪
⎪
f
⎪ ∂∂cE =
⎪ , ∂ c̄i j = ∂ y ∂ f¯k ∂ c̄i j
⎪
⎪ − ∂ y ∂ f k ∂−c
⎪ ij
⎪
k − ij k
⎪
⎪ ∂ E ∂ y ∂ −f ∂ E k
∂ E ∂ y ∂ f¯k
⎪ ∂E
⎩ ∂σ = , ∂σ i j =
ij ∂ y ∂ f k ∂σ ij ∂ y ∂ f¯k ∂σ i j
k − k
The derivatives in (12) are determined using the following formulas:

⎧ d
⎪ ∂E
⎪ ∂ y = − yi − yi
⎪
⎪
⎪ ∂y
⎪
⎪
⎪ ∂ −f k
= α(t) θk
, ∂∂fy¯k = (1 − α(t)) θk
⎪
N N
⎪ f k
f¯k
⎪
⎨ k=1
−

k=1
∂ fk 2 xi− − c (12)
, ∂∂ c̄fi j = f¯ ( σi−2 i j )
¯k
⎪ ∂ c− = f k
⎪
ij 2 x c̄
⎪
⎪ − − σ i2j
⎪
⎪ i j

2
ij
⎪
⎪
⎪ −
⎪ ∂ f k x c
i− −
¯k (xi− c̄i j )2
⎪
⎩ ∂σ = f k σ 3 i j , ∂∂ σ̄fi j = f¯ σ 3
ij − ij ij
4 Simulation Studies
In order to validate the effectiveness and evaluate the performances of the proposed
approximator PT2FNN for both identification and control objectives, we chose plant
models from literature [7].
A. PT2FNN Identification Performance Studies
This section studies the process design of both Fuzzy Neural Networks (FNN) and
Petri Type 2 Fuzzy Neural Networks (PT2FNN). In this numerical example, we use
Fuzzy Neural Networks (FNN) and Petri Type 2 Fuzzy Neural Networks (PT2FNN)
to model the same nonlinear system described as:
y(k) = 0.72y(k − 1) + 0.025y(k − 2)u(k − 1) + 0.01 u 2 (k − 2) + 0.2u(k − 3)

(14)
The structure of the identification processes is shown in Fig. 1, where y(k) is

plant output, yn (k) is output of PT2FNN system. The excitation signal used for the
identification of plant (14) is defined as (Fig. 2):
⎧ πk
⎪
⎪ sin 25 k < 250
⎨
1 250 ≤ k < 500
u(k) = (15)
⎪ −1
⎪
⎩ πk πk 500 ≤ k < 750
0.3 sin πk
25
+ 0.1 sin 32
+ 0.6 sin 10
750 ≤ k < 1000
For each input y(k) and u(k),we define five Gaussian membership functions.
Where the initial parameters are chosen as:
c(0) = [c1 , c2 , c3 , c4 , c5 ] = [−0.5, −0.25, 0, 0.25, 0.5], σ(0) =
π
[σ1 , σ2 , σ3 , σ4 , σ5 ], σ1 = σ2 = σ3 = σ4 = σ5 = 12 , ka = 0.4, kb =
0.03, a = 0.87, γ = γθ , γθ , γα , γc , γc̄ , γσ− , γσ̄ = [60, 60, 35, 35, 35, 35, 35]

θ(0) = θ(0), θ(0) = [0.5, 0.5], α̂(0) = 0
326 Y. Bibi et al.
Fig. 2 Identification scheme using petri type 2 fuzzy neural network
Fig. 3 Résultat of identification using petri type 2 fuzzy neural network
The parameters of the PT2FNN networks are adjusted using the algorithm (11),
in order to ensure a fast convergence of yn (k). The Fig. 3 presents the evolution of
the plant output y(k) and the PT2FNN identifier yn (k),where solid line denotes the
output of the plant y(k), dashed line denotes the PT2FNN output yn (k)
The cost function E, the Mean Absolute Error (MAE), the Root Mean Squares
Error (RMSE) given by the Eqs. (10), (16) and (17) respectively, the number of
activated rules and the number of adjusted parameters are fixed as a performance
criteria, In order to evaluate the efficiency of PT2FNN compared with fuzzy neural
networks approximator.

k
i=1 y(k + 1) − yn (k + 1)
M AE = (16)
k

k
i=1 (y(k + 1) − yn (k + 1))2
RMSE = (17)
k
where K = 1000 is the number of steps and the step size h = 0.01. The values of these
performance criteria are depicted in Table 1. The first column of Table 1 contains the
performance results of the fuzzy neural networks approximator. The second column
Table 1 Comparative of the

The estimator FNN PT2FNN
obtained performances
E 0.0062 0.0028
MAE 0.0062 0.0028
RMSE 0.5341 0.1988
Number of activated rule 36 [3, 10]
Number of adjusted parameter 18 37
Fig. 4 Identification error obtained during learning using petri type 2 fuzzy neural network
contains the results of PT2FNN. For a better illustration of the results in Table 1 we
show a visual representation of the identification error e(k) in Fig. 4.
From the comparison results, it can be concluded that both Fuzzy Neural Networks
(FNN) and Petri Type 2 Fuzzy Neural Networks (PT2FNN) are successfully designed
and implemented for the identification of the plant (14) with varying degree of preci-
sion. Based on the simulation results, it is found that, Petri Type 2 Fuzzy Neural
Networks (PT2FNN) can get a very precise estimation, but the number of adjusted
parameters is more important compared with Fuzzy Neural Networks (FNN) which
causes significant cost of implementation. The number of rules involved by the
PT2FNN is less than FNN, therefore the computation time is optimized.
B. PT2FNN Control Performance Studies
In this section, the efficiency of the suggested PT2FNN is explored for control
purposes. The PT2FNN-based control system’s design is as shown in Fig. 5. Here,
y(k) is the output signal of the system, yd (k) is the reference signal e(k) and ė(k) are
the error and the change in error, respectively. Using these signals, the parameters
of the PT2FNN structure are updated in order to ensure the tracking of the reference
and the stability of the closed-loop. The suggested PT2FNN is used in this example
to control the dynamic system described by the following difference equation:
y(k − 1)y(k − 2)(y(k − 1) + 2.5)

y(k) = + u(k) (11)
1 + y(k − 1)2 + y(k − 2)2
328 Y. Bibi et al.
Fig. 5 Structure of PT2FNN based control system
The PT2FNN controller has 5 membership functions for each input and 45 param-
eters to be updated. The initial values of these PT2FNN parameters are generated
as:
c(0) = [c1 , c2 , c3 , c4 , c5 ] = [−2, −1, 0, 0.25, 0.5], σ(0) =
π
[σ1 , σ2 , σ3 , σ4 , σ5 ], σ1 = σ2 = σ3 = σ4 = σ5 = 12 , ka = 0.4, kb =
0.03, a = 0.87, γ = γθ , γθ , γα , γ−c , γc̄ , γσ , γσ = [60, 60, 35, 35, 35, 35, 35],

θ(0) = θ(0), θ(0) = [0.5, 0.5], α̂(0) = 0
The Fig. 6 shows the tracking of the reference in the presence of the external
disturbance.
Fig. 6 The reference

tracking using PT2FNN
5 Conclusion
In this paper, a novel adaptive approximator is proposed for the identification and
control of dynamic systems. The developed approximator combines the advantages
of the type 2 fuzzy logic and Petri networks. The structure’s parameter update rules
are obtained based on the gradient algorithm. For identification and control purposes,
several simulation experiments are performed. The plant models are extracted from
the literature to allow a direct comparison of performance with existed approaches.
In both identification and control schemes, It is seen that PT2FNN can converge
faster and is more adaptive to changes of the dynamic. The performance is much
better, leading to lower RMSE values, despite the smaller number of rules, which is
optimized by incorporating Petri layers. Future works on the topic includes the use
of sliding mode control (SMC) theory-based learning algorithms which can ensure
faster convergence and make the overall system more robust than the traditional
learning techniques in online tuning.
References
1. Narendra KS, Parthasarathy K(1990) Identification and control of dynamical systems using
neural networks. IEEE Trans Neural Netw 1:4–27
2. Polycarpou MM, Ioannou PA (1992) Modeling, identification and stable adaptive control of
continuous-time nonlinear dynamical systems using neural networks. In: Proceedings of the
American control conference, pp 36–40
3. Sanner R, Slotine JJ (1992) Gaussian networks for direct adaptive control. IEEE Trans Neural
Networks 3(6):837–864
4. Sadegh N (1993) A perceptron network for functional identification and control of nonlinear
systems. IEEE Trans Neural Netw 4(6):982–988
5. Ho DWC, Zhang P-A, Xu J (2001) Fuzzy wavelet networks for function learning. IEEE Trans
Fuzzy Syst 9(1)
6. Abiyev RH (2005) Controller based of fuzzy wavelet neural network for control of technological
processes. In: Proceedings of IEEE International CISMA, GiardiniNaxos, Italy
7. Abiyev RH, Kaynak O (2008) Fuzzy wavelet neural networks for identification and control of
dynamic plants—a novel structure and a comparative study. IEEE Trans Ind Electron 55(8)
8. Cervanten J, Yu W, Salazar S, Chairez Takagi–Sugeno I (2017) Dynamic neuro-fuzzy controller
of uncertain nonlinear systems. IEEE Trans Fuzzy Syst 25(6)
9. Orlovskyi I (2019) Transformation method of nonlinear mathematical models of the DC series
drive into the form of modified recurrent neural network. Can J Electr Comput Eng 42(2)
10. Fayez F, El-Sousy M, Ali Abuhasel K (2018) Adaptive nonlinear disturbance observer using
a double-loop self-organizing recurrent wavelet neural network for a two-axis motion control
system. IEEE Trans Ind Appl 54(1)
11. Basterretxea K, Manuel Tarela J, del Campo I, Bosque G (2007) an experimental study on
nonlinear function computation for neural/fuzzy hardware design. IEEE Trans Neural Netw 18
12. Rubio-Solis A, Melin P, Martinez-Hernandez U, Panoutsos G (2019) General type-2 radial
basis function neural network: a data-driven fuzzy model. IEEE Trans Fuzzy Syst 27(2)
13. Zhao J, Chih-Min L (2019) Wavelet-TSK-type fuzzy cerebellar model neural network for
uncertain nonlinear systems. IEEE Trans Fuzzy Syst 27(3)
14. Kiaei I, Lotfifard S (2019) Fault section identification in smart distribution systems using
multi-source data based on fuzzy petri nets. IEEE Trans Smart Grid
330 Y. Bibi et al.
15. Kayacan E, Kayacan E, Ahmadieh Khanesar M (2015) Fuzzy basis function universal
approximation and orthogonal least square learning. IEEE Trans Ind Electr 62(3):1716–1724
16. Kayacan E, Kayacan E, Ahmadieh Khanesar M (2015) Identification of nonlinear dynamic
systems using type-2 fuzzy neural networks—a novel learning algorithm and a comparative
study. IEEE Trans Ind Electron 62(4)
17. Bibi Y, Bouhali O, Bouktir T (2017) Petri type 2 fuzzy neural networks approximator for
adaptive control of uncertain non-linear systems. IET control theory and applications, pp 3130–
3136
18. Bibi Y, Bouhali O, Bouktir T (2019) Robust direct adaptive controller for a class of uncertain
nonlinear systems using petri type 2 fuzzy neural networks (PT2FNN) as a new approximator.
J Electr Syst 15(2):181–196
19. Bibi Y, Bouhali O, Bouktir T (2018) Intelligent petri type 2 fuzzy direct adaptive control
for uncertain nonlinear systems. In: The 3rd international conference on recent advances in
electrical systems
20. Chien Y-H, Wang W-Y (2012) Online hybrid intelligent tracking control for uncertain nonlinear
dynamical systems. In: Advanced mechatronic systems, Tokyo, Japan, 18–21 September 2012
21. Zhu R (2012) A new petri net model and stability analysis of fuzzy control system. In: Advanced
mechatronic systems, Tokyo, Japan, 18–21 September 2012
22. Benbrahim M, Essounbouli N, Hamzaoui A (2013) A betta adaptive type-2 fuzzy sliding mode
controller for SISO, nonlinear systems subject to actuator faults. Int J Autom Comput
23. Wang LX, Mendel JM (1992) Fuzzy basis function universal approximation and orthogonal
least square learning. IEEE Trans Neural Netw 3
24. Chen C-S, Lin W-S (2003) Adaptive wavelet sliding mode control of uncertain multivariable
nonlinear systems. In: Proceedings of the American control conference, Denver, Colorado, 4–6
June 2003
25. Hai-bo ZHOU, Hao YING, Ji-an DUAN (2011) Adaptive control using interval type-2 fuzzy
logic for uncertain nonlinear systems. Central South University Press and Springer, Berlin
Heidelberg
26. Kulkarni A, Kumar A (2015) Adaptive wavelet backstepping control for a class of MIMO
underactuated systems. J Autom Control Eng 3(4)
27. Sharma M, Verma A (2012) Wavelet neural network observer based adaptive tracking control
for a class of uncertain nonlinear delayed systems using reinforcement learning. Int J Intell
Syst Appl
Design Fractional Order PI Controller
with Decoupler for MIMO Process Using
Diffusive Representation
Sami Laifa, Badreddine Boudjehem, and Hamza Gasmi
Abstract This paper presents a new design method of decentralized fractional order
PI controller (FO-PI) with decoupler for the pilot plant binary distillation column
(multivariable process). The process is transformed into two independent SISO
systems by introducing a simple decoupler to reduce the interaction. The Optimal
parameters of the FO-PI controllers are tuning by minimizing performance index
criterion as objective function for each SISO system. The non standard transfer func-
tion of the fractional order controllers are performed by means of diffusive approach.
The simulation results show the superior performance obtained by decentralized
fractional PI controller (FO-PI) in comparison with classical control such as Non-
Dimensional Tuning (NDT) method and Simplified Internal Model Control (SIMC)
technique.
Keywords Decentralized fractional order PI (FO-PI) controller · MIMO process ·

Performance index · Decoupler · Diffusive representation
1 Introduction
Control of multivariate systems (multi-input and multi-output (MIMO)) is different

from controlling single-variable systems (single-input and single-output (SISO))
because of the interconnection between inputs and outputs, making control of these
systems difficult, this coupling can pose major difficulties of design and tuning PI/PID
S. Laifa (B) · H. Gasmi

Advanced Control Laboratory, Department of Electronics and Telecommunications, Université 8
Mai 1945 Guelma, BP 401, 24000 Guelma, Algeria
H. Gasmi
B. Boudjehem
Advanced Control Laboratory, Department of Electrotechnics and Automatic, Université 8 Mai
1945 Guelma, BP 401, 24000 Guelma, Algeria

https://doi.org/10.1007/978-981-15-6403-1_23
332 S. Laifa et al.
controller. Among the solutions proposed to reduce the interconnection between these
rings is the use of the decoupler [1–3]. The control design of multivariate systems
is generally divided into two types: centralized control and decentralized control
[4–8]. Decentralized control is the most used in industry, given the simplicity of
implementation and maintenance, as well as the performance that can be achieved
by focusing on the performance of individual loops.
Many methods are proposed for the design and tuning of decentralized classical
PI/PID controllers for MIMO systems:
Maghade and Patre [9] presented a Decentralized Conventional PI/PID Controllers
based on gain and phase margin specifications for two-input and two-output (TITO)
processes.
In [10], a decentralized Conventional PI (PID) controller is obtained using the
non–dimensional tuning method for TITO processes.
In [6], presents a decentralized PID controller design method for two input two
output (TITO) systems with time delay using characteristic ratio assignment (CRA).
In [11], a decentralized PI with decoupler for the distillation column is designed
using Non-Dimensional Tuning (NDT) method and Simplified Internal Model
Control (SIMC) technique.
Recently, fractional order calculus has been applied to many fields of engineering
such as control of linear and nonlinear systems, electrical circuits, signal processing
and communications, [12–16] etc. The fractional order proportional integral deriva-
tive was first introduced by Podlubny [17] and it is considered as the generalization
case of classical PID controllers.
The concept of fractional calcul has been applied successfully for the control of
(SISO) systems.
In [18], fractional order model are used to predict the process output using smith
predictor. Fractional PID are also used to enhance control system performance in
comparison with other method.
In [19], analytical design method based on using fractional order reference system
as reference model.
In [20], An Optimal Design of a Variable Coefficient Fractional Order PID (V-
FOPID) Controller by using PSO and AIS Optimization Algorithms.
In [21], the Fractional PID Controller Design Based on Minimizing Performance
indices.
This success has encouraged researchers to think of introducing this concept to
large multi-input and output systems (multivariable systems):
In [22], the application of robust Fractional Order Proportional-Integral (FO-
PI) autotuning control strategy is presented and applied to heterogeneous dynamic
systems using decentralized control.
In [23], A Fractional Order Proportional-Integral (FOPI) controller tuning method
by combining Biggest Log-modulus Tuning (BLT) and Internal Model Control (IMC)
method applied to Distillation Column.
In [24], Fractional order IMC based PID controller design using Novel Bat
optimization algorithm for TITO Process.
Design Fractional Order PI Controller with Decoupler … 333
Auto-tuning of FOPI Controllers design for TITO Processes with Experimental

Validation presented in [25].
This paper discusses the design of a decentralized fractional PI controller for
efficient control of the MIMO system (Pilot Plant Binary Distillation Column) [11,
26, 27]. The idea is to combine a decoupler with the fractional PI controller imple-
mented using the diffusive approach. The simple decoupler [3] was used to reduce
interactions between variables and decoupled subsystems were formed.
The mathematical formulation of the optimization problem are posed directly to
fractional parameters. The major objective of this paper is to show that the decentral-
ized fractional order (FO-PI) controller performance is better than the decentralized
classical (C-PI) controller such Non Dimensional Tuning (NDT) control technique
and Simplified Internal Model Control (SIMC) technique [11] for MIMO systems.
This paper is organized as follows: In Sect. 2, discuss Decoupler design, In Sect. 3
a mathematical notions on fractional order operators and fractional order controller
are given. Section 4 shows the basises of the proposed design method. The results
of the simulation are given in Sect. 5 and the conclusion in Sect. 6.
2 The Decoupler Design for a TITO Process
Consider the TITO (Two-Input-Two-Output) process is as follows:

g (s)e−τ 11s g12 (s)e−τ 12s
G(s) = 11 (1)
g21 (s)e−τ 21s g22 (s)e−τ 22s
Decoupler is designed to eliminate and minimize the interaction between the

variables of the given system G(s) in Eq. (1). Decoupler was suggested by Wang
et al. [2] for TITO systems.
The following Two cases are considered in the design of the decoupler matrix:
Case 1: The off-diagonal elements of G(s) have no RHP poles and diagonal elements
of G(s) have no right-half-plane (RHP) zeros. The decoupler has the following
transfer function matrix.

v1 (s) d12 (s)v2 (s)
D(s) = (2)
d21 (s)v1 (s) v2 (s)
where:

1 τ21 ≥ τ22
v1 (s) =
e(τ21 −τ22 )s τ21 < τ22

1 τ12 ≥ τ11
v2 (s) =
e(τ12 −τ11 )s τ12 < τ11
334 S. Laifa et al.
g12 (s) −(τ12 −τ11 )s

d12 (s) = − e
g11 (s)
g21 (s) −(τ21 −τ22 )s
d21 (s) = − e (3)
g22 (s)
Case 2: If there are no right-half-plane (RHP) poles in diagonal and no right-half-

plane (RHP) zeros in off- diagonal elements of G(s), The decoupler matrix is:

d11 (s)v3 (s) v3 (s)
D(s) = (4)
v4 (s) d22 (s)v4 (s)
where:

1 τ22 ≥ τ21
v3 (s) = (τ22 −τ21 )s
e τ22 < τ21

1 τ11 ≥ τ12
v4 (s) =
e(τ11 −τ12 )s τ11 < τ12
g22 (s) −(τ22 −τ21 )s
d11 (s) = − e
g21 (s)
g11 (s) −(τ11 −τ12 )s
d22 (s) = − e (5)
g12 (s)
The resulting of transfer function of G(s) ∗ D(s) is a diagonal system H (s) as

presented in Eq. (6).
H (s) = G(s) ∗ D(s) = diag(h 11 (s), h 22 (s)) (6)
where:

h 11 (s) 0
H (s) = G(s) ∗ D(s) =
0 h 11 (s)
The diagonal elements h 11 (s) and h 22 (s) calculated by using Eq. (6), they are
controlled by decentralized fractional order controllers Ci (s). where i = 1, 2.
3 Fractional Order Operators and Controllers
A. Fractional Order Operators
Fractional calculus is a generalization of integration and differentiation to non-integer

order operator aDαt where a and t denote the limits of the operation and a denotes
the fractional order is as follows [28]:
⎧ dα
⎪
⎪ α α>0
⎨ dt
α
aDt = t1 α=0 (7)
⎪
⎪ α
⎩ ∫(dτ) α < 0
a
Ou α ∈ R est l’ordre de l’opération.

There are various definitions of fractional differentiation-integrals available The
following definitions of fractional calculus are used widely in the area of control
system [29]. A commonly used definition of fractional differo-integral is the
Riemann-Liouville definition which is given by:

n t
α 1 d f (τ )
t0 Dt f (t) = dτ (8)
Γ (n − α) dt (t − τ )α−n+1
t0
For (n − 1 < α < n) and Γ (α) is Euler’s Gamma function

By Grunwald-Letnikov method the αth order differ-integration of a function f (t)
is defined as:
∞
1 j α
Dtα f (t) := lim α (−1) f (t − j h) (9)
h→0 h j
j=0
where:

α α! Γ (α + 1)
= = (10)
j j!(α − j)! j!Γ (α − j + 1)
Denotes the binomial co-efficient The Laplace transform of Grunwald-Letnikov

fractional differ-integration is:
∞
∫ e0−st Dtα f (t)dt = s α F(S) (11)
0
where F(s) is the normal Laplace transformation f (t).

336 S. Laifa et al.
B. The Diffusive Representation
Diffusive approach [30] is one of the methods used to implement fractional order
integration, the usefulness of diffusive representation are to transform problems of
a hereditary nature of fractional operators (non-rational) into nature non-hereditary
(rational).
The diffusive realization of the pseudo differential operator
H (transfer function
(non-rational)), with impulse response h, u → y = H dtd u is defined by the
dynamic input-output system:
⎧ ∂ψ(ξ,t)
⎪
⎪ = −ξ ψ(ξ, t) + u(t)
⎨ ∂t +∞
y(t) = ∫ μ(ξ )ψ(ξ, t)dξ (12)
⎪
⎪
⎩ 0
ψ(ξ, 0) = 0 , ξ > 0
The system (12) is the diffusive canonic realization of operator H.

The impulse response h(t) is expressed from h(t) by:
+∞
h(t) = e−ξ t μ(ξ )dξ (13)
0
The transfer function of the operator H is given by:
+∞
μ(ξ )
H ( p) = dξ (14)
p+ξ
0
The transfer function associated with the diffusive representation. In other words,
μ(ξ) the representation is obtained directly by the inverse Laplace transformation of
the impulse response h(t) La The diffusive representation is therefore a “counter-
sense” use of the Laplace transform:
Lξ Lt
μ(ξ ) → h(t) → H ( p), μ(ξ ) = L−1 (h(t)) (15)
For example, the diffusive symbol of the fractional integrator s of order α is

∂ 1
H oú H ( p) = a for 0 < a < 1 (16)
∂t p
The diffusive symbol is expressed as (see [31]):
sin(απ)
μ(ξ ) = ξ −α ξ >0
π
where α is the order of integration.

And the diffusive realization of the controller P I a with input u and output y may
expressed by [21]:
⎧
⎨ ∂t ψ(ξ, t) = −ξ ψ(ξ, t) + u(t)
+∞ (17)
⎩ y(t) = ∫ v(ξ )ψ(ξ, t)dξ
0
the transfer function of the fractional controller P I a by means of diffusive

representation are:
+∞ v(ξ )
C(s) = ∫ dξ (18)
−∞ p+ξ
where v the diffusive representation know as follows:

1 sin(απ) −α
v(ξ ) = K P δ(ξ ) + ξ (19)
Ti π
C. Fractional Order P I a Controller
The differential equation of the fractional order P I a (FO-PI or P I a ) is given by:

1
u(t) = K p e(t) + D −a e(t) (20)
Ti
where:
D → is the derivative operation, K P → the proportional gain
TI → is the integral time constant, α → is the integration order
and the transfer function of the fractional order PI (FO-PI) controller is given by:

1 KI
C(s)frational = K P 1 + = KP + = K P + K I s −a
TI s a sa
as:
K I = K P /TI (21)
If a = 1. we get classical PI controller

1 KI
C(s)classical = K P 1 + = KP + = K P + K I s −1 (22)
TI s s
338 S. Laifa et al.
4 Roposed Design Method
Figure 1 shows the diagram of the decentralized fractional P I a controller imple-

mented on a multivariable system (a Pilot Plant Binary Distillation Column) in the
presence of a decoupler.
The mathematical model is identified by Vinaya and Arasu for a Pilot Plant Binary
Distillation Column [27] is expressed by the following transfer matrix:

−0.13e−0.03s 0.18e−0.03s
G( p) = 1.14s+1 0.64s+1
−0.34e−1.22s 0.18e−0.03s (23)
1.23s+1 0.32s+1
The transfer function matrix of decoupler is:

1.3846(1.14s+1)
1
D(s) = 1.888(0.32s+1)e−1.19s
0.64s+1
(24)
1.23s+1
1
The resulting diagonal elements of H (s) = G(s)D(s) = diag(h 11 (s), h 22 (s))

are:
−0.13e−0.03s 0.3398(0.32s + 1)e−1.22s

h 11 (s) = + (25)
1.14s + 1 (0.64s + 1)(1.23s + 1)
0.18e−0.03s 0.47076(1.14s + 1)e−1.22s

h 22 (s) = + (26)
0.32s + 1 (1.23s + 1)(0.64s + 1)
The decoupled systems (h 11 (s) and h 22 (s)) is controlled by two fractional PI

controllers F O − P I1 and F O − P I2 (see in Fig. 2). The transfer functions of the
two fractional controllers are given by (21).
Fig. 1 Schematic diagram of the decentralized fractional P I a controller designed with decoupler
for a pilot plant batch distillation column
Fig. 2 Diagram illustrating the proposed design method
Thus the transfer function of the two fractional P I a1 and P I a2 controllers by

means of diffusive Approach are:
+∞ v1 (ξ )
F O − P I1 = ∫ dξ (27)
−∞ p+ξ
+∞ v2 (ξ )
F O − P I2 = ∫ dξ (28)
−∞ p+ξ
Our objectives are to setting the six optimal parameters (K p1 , K i1 , K p2 , K i2 α1 ,

α2 ) of the decentralized fractional PI controllers that minimizing the performance
index error defined by objective function that may be given as follows
∞
JIAE, ISE and ITAE K p1 , K i1 , K p2 , K i2 α1 , α2 = ∫ f (e(t))dt (29)
0
With f is a function of the error e(t) that based to the criteria (ISE, IAE and ITAE).
These three commonly used measures: Integral Squared Error (ISE), Integral
Absolute Error (IAE) and Integral Time weighted Absolute Error (ITAE), and are
defined as:
∞ ∞ ∞
I AE = ∫ |e(t)|, I T AE = ∫ t|e(t)|dt, I S E = ∫ e(t)2 dt (30)
0 0 0
The steps to determine the fractional PI controllers parameters can be summarized

as follows:
340 S. Laifa et al.
• Implement the feedback control system of controlled system with fractional PI

controller in Matlab/Simulink including diffusive realization of the fractional
operator through Simulink model.
• Select the appropriate function of Matlab optimization toolbox to minimize the
objective function J.
• The stability margin based Non-Dimensional Tuning method (NDT) [10] is used
for determine the initial controller parameters.
• Calculate the error (IAE, ISE and ITAE).
For more details of proposed design method (see in Fig. 2).
The clasical PI controller based on the NDT-PI and SIMC-PI methods is designed
using Non-Dimensional Tuning method and Simplified Internal Model Control
technique respectively [10–32], the methods can be summarized in the following
steps:
• decoupled subsystems (h 11 (s) and h 22 (s)) are reduced to first order plus dead
time (FOPDT) model by fitting of frequency response For each FOPDT reduced
subsystem.
K ii
Q iiFOPDT (s) = e−L ii s (31)
Tii s + 1
where : i = 1, 2; K ii = process gain; L ii = delay time; Tii = time constant
• The Classical PI controllers (C − P I1 and C − P I2 ) are given as follows
1 1
C − P I1 = K p1 + K i1 , C − P I2 = K p2 + K i2
s s
The Parameters of the PI controllers are tuned by NDT method as follows.
Tii 1
K pii K ii = + (32)
2L ii 14

TI ii 1 L ii L ii
= min 1 + ,9 (33)
Tii 7 Tii 14
• also for determining PI parameters by SIMC technique as follows

1 Tii
K pii = (34)
K ii Tcii + L ii
TI ii = min(Tii , 4(Tcii + L ii )) (35)
For more details of NDT and SIMC design methods (see in [10, 11]).
As an simulation studie, we consider the mathematical model identified by Vinaya

and Arasu for a Pilot Plant Binary Distillation Column [11, 26, 27] given by matrix
transfer function, see in Eq. (23). The results and performance of proposed fractional
PI controller compared with Conventional PI controller such NDT-PI [10] and SIMC-
PI [32], is discussed using MATLAB/Simulink.
The optimized parameters of the fractional PI controller (FO-PI) based on ISE
criterion were organized in Tables 1 and 2. Attached with the classical controllers
(NDT-PI and SIMC-PI). The results of the performance (IAE, ISE, ITAE, Overshoot
(%) and Settling time (sec)) were organized in Table 3.
The two Figs. 3 and 4, shows the Step response of the first and second outputs (Y1
and Y2) of the controlled system based on design method of decentralized fractional
and classical controllers.
Figures 5 and 6 shows the result for output response (first and second outputs
(Y1 and Y2)) of closed loop with different setpoint for a Distillation Column. With
fractional PI (FO-PI) and classical PI (NDT and SIMC) controllers. The photographic
view of the experimental setup [11] of a pilot plant binary distillation column (MIMO
process) is as shown in Fig. 7.
To demonstrate the efficiency of the proposed design method, we compare our
results with those obtained using deentralized classical PI controller NDT-PI and
Table 1 Tuning parameter for first output (Y1)

P I1 controller Parameter
K P1 KI1 α1
FO-PI1 1.9278 0.010479 0.13567
NDT-PI1 1.9473 1.19 –
SIMC-PI1 1.6071 1.1904 –
Table 2 Tuning parameter for second output (Y2)

P I2 controller Parameter
K P2 KI2 α2
FO-PI2 −0.9786 −0.01038 1.1451
NDT-PI2 −1.0665 −1.116 –
SIMC-PI2 −0.8202 −1.116 –
342 S. Laifa et al.
Table 3 Performance analysis results for FO-PI, NDT-PI, and SIMC-PI controllers
Performance Input(u)–Output(y) Method
FO-PI NDT-PI SIMC-PI
IAE u 1− y1 5.123 5.946 6.375
u 2− y2 3.55 3.951 4.431
ISE u 1− y1 4.561 4.797 4.945
u 2− y2 3.725 3.948 4.063
ITAE u 1− y1 15.06 23.22 28.55
u 2− y2 6.14 8.134 11.44
Overshoot % u 1− y1 6.2 26 27.42
u 2− y2 5.01 23.2 23.95
Settling time (s) u 1− y1 8.02 15.12 17.15
u 2− y2 7.45 8.11 8.54
1.5
1
Y1
0.5 FO-PI Controller

NDT-PI Controller
0 SIMC-PI Controller
-0.5
0 20 40 60 80 100
Time (s)
Fig. 3 Step responses for first output (Y1) of the controlled system with fractional PI (FO-PI) and
classical PI (NDT and SIMC) controllers
1.5
1
Y2
0.5 FO-PI Controller

NDT-PI Controller
0 SIMC-PI Controller
-0.5
0 20 40 60 80 100
Time (s)
Fig. 4 Step responses for second output (Y2) of the controlled system with fractional PI (FO-PI)
and classical PI (NDT and SIMC) controllers
100
80
Y1 and Y2 60
40
FO-PI Controller method of Output Y1
20 FO-PI Controller method of Output Y2
NDT-PI Controller of Output Y1
0 NDT-PI Controller of Output Y2
SIMC-PI Controller of Output Y1
-20 SIMC-PI Controller of Output Y2
-40
0 20 40 60 80 100
Time (s)
Fig. 5 Step responses for the first and second outputs (Y1 and Y2) of controlled system with
Setpoint-reference1 = 50 and Setpoint-reference2 = 70
100
80
60
Y1 and Y2
40
20 FO-PI Controller method of Output Y1

FO-PI Controller method of Output Y2
0
SIMC-PI Controller of Output Y1
-20 SIMC-PI Controller of Output Y2
-40
0 20 40 60 80 100
Time (s)
Fig. 6 Step responses for the first and second outputs (Y1 and Y2) of controlled system with
Setpoint reference1 = 64 and Setpoint-reference2 = 76
SIMC-PI) are given as follows: Figs. 3, 4 and Table 1 show clearly that the decen-
tralized fractional controllers give a good performance (settling time and Overshoot)
in comparison with classical one.
It is observed also that different performance indices (IAE, ISE and ITAE) as
objective functions, with fractional PI controller gives very good and superior results
When compared to the classical controller (C-PI), see Table 3. It is clear that error
indices and peak overshoot are less in proposed method.
Figures 5 and 6 shows the result for output response of closed loop with different
setpoint for controlled system (a Distillation Column). We can say that the use of
the decentralized fractional PI controller with decoupler provides good performance
and robustness.
344 S. Laifa et al.
Fig. 7 Laboratory setup of binary distillation column
6 Conclusion
In this paper, decentralized fractional PI controller with decoupler is designed and

implemented for a Pilot Plant Binary Distillation Column (multivariable process),
The problem of assigning the optimum possible fractional controllers parameters for
multivariable system using Matlab optimization toolbox is addressed based on mini-
mizing the performance indices criterion (IAE, ISE and ITAE). simulation results
show clearly that the time response performances (settling time and maximum over-
shoot) given by of the controlled system with fractional-order controller are give a
good performance and acceptable in comparison with the classical PI controller such
(Non-Dimensional Tuning (NDT) method and Simplified Internal Model Control
(SIMC) technique).We can say that the performance of the fractional PI controller
based on the diffusive Approach is clearly superior to the performance of classical
PI controllers.
References
1. Lee J, Kim D, Edgar T (2005) Static decouplers for control of multivariable processes. Am
Inst Chem Eng 51(10):2712–2720
2. Wang QG, Huang B, Guo X (2000) Auto-tuning of TITO decoupling controllers from step
tests. ISA Trans 39:407–418
3. Nordfeldt P, Hagglund T (2006) Decoupler and PID controller design of TITO systems. J
Process Control 16(9):923–936
4. Ram DV, Chidambaram M (2015) Simple method of designing centralized PI controllers for
multivariable 155 systems based on SSGM. ISA Trans 56:252–260
5. Majaaz V, Bhat S, Thirunavukkarasu I, Shanmuga Priya S (2015) Centralized controller tuning
for MIMO process with time delay. In: IEEE Xplorer, 4th international conference on renewable
energy research and applications, Palermo, Italy, pp 660–664
6. Hajare VD, Patre BM (2015) Decentralized PID controller for TITO systems using character-
istic ratio assignment with an experimental application. ISA Trans 59:385–397
7. Hu W, Cai WJ, Xiao G (2010) Decentralized control system design for MIMO processes with
integrators/differentiators. Ind Eng Chem Res 49(24):12521–12528
8. Labibi B, Marquez HJ, Chen T (2009) Decentralized robust PI controller design for an industrial
boiler. J Process Control 19(2):216–230
9. Maghade DK, Patre BM (2012) Decentralized PI/PID controllers based on gain and phase
margin specifications for TITO processes. ISA Trans 51(4):550–558. https://doi.org/10.1016/
j.conengprac.2005.06.006
10. Tavakoli S, Griffin I, Fleming PJ (2006) Tuning of decentralised PI (PID) controllers for
TITO processes. Control Eng Pract 14(9):1069–1080. https://doi.org/10.1016/j.conengprac.
2005.06.006
11. Santhosh Kumar PL, Selva Kumar S, Thirunavukkarasu I, Bhat VS (2018) Decentralized PI
controllers with decoupler for the distillation column. Int J Pure Appl Math 118(20):9–14
12. Vinagre BM, Feliú V (2002) In: Proceedings: 41st IEEE conference on decision and control,
Las Vegas, 9 December 2002
13. Podlubny I, Dorcak L, Kostial I (1997) On fractional derivatives, fractional-order dynamic
system and PID-controllers. In: Proceedings of the 36th conference on decision & control, vol
5, pp 4985–4990
14. Çelik V, Demir Y (2010) Effects on the chaotic system of fractional order PI α controller.
Nonlinear Dyn 59(1–2):143–159
15. Sabatier J, Agrawal OP, Tenreiro Machado JA (2007) Advances in fractional calculus:
theoretical developments and applications in physics and engineering. Springer, the Netherlands
16. Monje CA, Chen YQ, Vinagre BM, Xue DY, Feliu V (2010) Fractional-order systems and
controls: fundamentals and Applica-tions. Springer, London
17. Podlubny I (1999) Fractional order systems and PI λ Dμ controllers. IEEE Trans Autom Control
44:208–214
18. Boudjehem D, Sedraoui M, Boudjehem B (2013) A fractional model for robust fractional order
smith predictor. Nonlinear Dyn 73:1557–1563
19. Boudjehem B, Boudjehem D (2013) Fractional order controller design for desired response. J
Syst Control Eng 227:243–251
20. Aydogdu O, Korkmaz M (2019) Optimal design of a variable coefficient fractional order PID
controller by using heuristic optimization algorithms. Int J Adv Comput Sci Appl (IJACSA)
10(3)
21. Boudjehem B, Boudjehem D (2016) Fractional PID controller design based on minimizing
performance indices. IFAC-PapersOnLine 49(9):164–168
22. Cajo R, Muresan CI, Ionescu R, Keyser D (2018) Plaza-multivariable fractional order pi
autotuning method for heterogeneous dynamic systems. IFAC-PapersOnLine 51-4:865–870
23. Edet E, Katebi R (2018) On fractional-order PID controllers. In: 3rd IFAC conference on
advances in proportional integral-derivative control Ghent, Belgium, 9–11 May, IFACpaper-
sOnline 51-4:739–744
346 S. Laifa et al.
24. Lakshmanaprabu SK, Sabura Banu U, Hemavathy PR (2017) Fractional order IMC based
PID controller design using Novel Bat optimization algorithm for TITO process. Energy Proc
117:1125–1133
25. Gargi B, Somanath M, Chitralekha M (2018) Auto-tuning of FOPI controllers for TITO
processes with experimental validation. Int J Autom Comput. https://doi.org/10.1007/s11633-
018-1140-0
26. Bhat VS, Thirunavukkarasu I, Janani R (2017) Design and implementation of MSC based
multi-loop pid controller for pilot plant binary distillation column. In: International conference
on circuits power and computing technologies [ICCPCT]. IEEE
27. Bhat VS, Thirunavukkarasu I, Priya SS (2016) Design and implementation of decentralized pi
controller for pilot plant binary distillation column. Int J ChemTech Res 10(2):284
28. Podlubny I (1999) Fractional differential equations. Academic Press, California
29. Oldham KB, Spanier J (1974) The fractional calculus, theory and applications differentiation
and integration to arbitrary order. Elsevier. ISBN 0486450015
30. Montseny G (2004) Simple approach to approximation and dynamical realization of pseudo
differential time operators such as fractional ones. Int EEE Trans Circuits Syst II 51:613–618.
https://doi.org/10.1109/TCSII.2004.834544
31. Laudebat L, Bidan P, Montseny G (2004) Modeling and optimal identification of pseudo defer-
ential dynamics by means of diffusive representation part i: modeling. Int EEE Trans Circuits
Syst 51:1801–1813. https://doi.org/10.1109/TCSI.2004.834501
32. Skogestad S (2003) Simple analytic rules for model reduction and PID controller tuning. J
Process Control 13:291–309
Direct Sliding Mode Control of Transient
Power in Microgrid During Grid Failure
“Unintentional Islanding”
I. Ameur, N. Gazzam, and A. Benalia
Abstract An important feature of a Microgrid is the ability to disconnect or connect

to the main grid with a transparent transfer of power between the two modes. During
a sudden grid loss (due to defects on the main grids), the Microgrid passes to the
Unintentional Islanding Mode (UIM). In this case, a transient power in the AC bus
increase the voltage at the DC link and it can turn off the inverter thus reduces the
reliability of the Microgrid. This article proposes a new control strategy based on the
Sliding Mode Control (SMC) in order to regulate the DC voltage and improve the
performance of the Microgrid when the UIM occurs. Finally, the simulation results
in the MATLAB/ Simulink environment made it possible to check the robustness of
the proposed controllers.
Keywords Microgrid · Unintentional islanding · DC link voltage · Sliding mode

controller
1 Introduction
Today, the rising fuel prices and the environmental constraints (the global warming,
CO2 emissions) encourage various governments around the world to integrate
Renewable Energy Sources (RES) in the electrical networks. From this, the concept
of a Microgrid has emerged and started to develop especially in order to reduce
the cost and the losses of energy in the transport. A Microgrid comprises a set of
renewable energy systems (solar energy, wind energy …) or non-renewable energy
(diesel …), in the form of a distributed generation (DG), and a storage means (battery,
I. Ameur (B) · N. Gazzam · A. Benalia

Faculty of Technology, LACoSERE Lab, University of Laghouat, Laghouat, Algeria
N. Gazzam
A. Benalia

https://doi.org/10.1007/978-981-15-6403-1_24
348 I. Ameur et al.
DC/DC DC/AC
PV array converter Ipv_dc DC- Link inverter LCL Filter
Ipv
P1
+ +
Vpv Vdc Vac
- -
U_PV1 U_inv1
photovoltaic Voltage
Controller Controller
Secondary Control Vac_ref
S.T.S
Droop
photovoltaic Utility
P 1* Controller Grid
Controller Primary Control
Secondary Control
U_PV2 DC/AC
PV array DC- Link LCL Filter
inverter
Load
Ipv_dc
Ipv P2
+ +
Vpv Vdc Vac
- -
DC/DC U_inv2
bid.converter
Ibat
Voltage
+
Vbat
- Controller
Vac_ref
U_bat1 U_bat2
Droop
Battery
Battery P 2* Controller
Controller AC-Link
Primary Control
Secondary Control
Fig. 1 Typical structure of microgrid adopted
storage tank of H2…) associated with charges (small of consumers) [1]. Our Micro-
grid is composed of three distributed generation units, two photovoltaic systems and a
battery bank as an Energy Storage System (ESS). These three elements are connected
to the grid via a Static Transfer Switch (STS), as illustrated in Fig. 1. The reliability
of Microgrid is afected when one or several renewable energy sources are used, due
to their low and intermittent efficiency, since they are associated with climate change
(temperature, sunlight …), and therefore storage (ESS) is necessary to balance any
mismatch between load demand and the available power of DG sources. In the Micro-
grid, it is important to share the power to satisfy the load demand. The droop control is
used to improve the power sharing between the inverters of the distributed generators
that connect in parallel in a Microgrid. This sharing is made it without any physical
communication between DG units [2, 3]. There are many strategies in the control of
droop, according to the type of grid and line impedance. In a medium/high voltage
grid, the Q/V strategy is used, but in the case of low voltage grid, which is the case
here, the active power/voltage (P/V) droop strategy is used. On the other hand, if the
dominant line impedance is indicative, the active power/frequency (P/f) is used, and
Direct Sliding Mode Control of Transient Power in Microgrid … 349
respectively, if the dominant line impedance is resistive, the active power/voltage

(P/V) droop strategy is used [3–5].
In normal operation, the Microgrid can operate in two modes via a Static Transfer
Switch (STS). The grid-connected mode: where the Microgrid operates in conjunc-
tion with the main grid in order to feed the load or inject the excess energy produced
in the main grid. The stand-alone mode: the Microgrid operates without the main
grid, the droop control works to correct the deviation of the voltage and frequency
of the AC-link [2]. In the abnormal (unintentional) case, the Microgrid operates in
unintentional islanding mode (UIM), this mode occurs during a sudden grid loss and
the Supervisory Controller (SC) is unable to select quickly the power setpoints, due
to the lack of knowledge about the state of the grid because of the inherent delay of
the communication channels [6, 7].
According to IEEE 1547 standards, the anti-islanding controller that detects the
failure of the grid must not exceed 2s to inform the supervisory controller for opening
the STS, and set the appropriate power setpoints for all inverters. The Major problem
occurs before the activation of the anti-islanding controller when the transient power
may flow to the inverters that received a low power setpoint, which lead to increase
the DC bus voltage (instability), and the inverter may close to prevent damage [8, 9].
There are a number of publications discussed the seamless continuous transfer
of power in the Microgrid, some have addressed the problem of stability of the DC
link voltage in the case of unintentional islanding like Issa et al. [10] who proposed
a controller that prevent the shutdown of the inverter. However, this control strategy
has certain limitations: it is used only for two inverters and the results are validated
without load. The authors in [8] studied the current flow effect between several
inverters in the Microgrid, they proposed a regulator to control the discharge circuit
switch (resistor), and the limitation in this regulator is the dissipation of power and
the disturbance of the DC source.
This article proposes a controller based on the sliding mode technique for each
DC source incorporated with droop control for regular DC bus especially in the
unintentional islanding mode, thus, we improve the stability and the reliability of the
Microgrid.
The rest of the article embodies five sections. The next section discusses the
operation of droop control. Analyzing the phenomenon of power import into the
inverters is presented in Sect. 2. Accordingly, Sect. 3 explores the sources of energy
used in the microarray. Furthermore, the proposed controllers are disclosed in Sect. 4.
Hence, the simulation results that prove the performance of the proposed control
strategy are presented in Sect. 5.
2 Operation Modes of a Microgrid and Problem Statement
The energy management in the supervisory controller is charged to generate the

setpoint of each inverter in the Microgrid. During grid connected mode the supervi-
sory delivered different setpoint power to the inverters, in order to provide different
350 I. Ameur et al.
amount of energy to the Microgrid according to the load requirement, the voltage
and frequency of the AC bus are fixed by the utility grid. When the Microgrid is
operating in islanded mode, all inverters have a constant setpoint power. Before the
transition toward islanded mode, the supervisory controller operates all the inverter
in their power rating. The unintentional islanding mode occurs when the utility grid
is in failure situation, while the STS remains in grid-connected mode. The different
value of the set-point power operations the inverters may cause the importation of
power in the inverter.
For further clarification, we use the droop control method to explain the issue
of the unintentional islanding where each inverter has a droop control. The droop
control equations can be expressed as: [3].

ωr e f = ωn − K ω Pavg Pi − Pi∗ (1)

Vr e f _i = Vn − K V Q avg Q i − Q i∗ (2)
where ωn and Vn are the nominal frequency and nominal voltage respectively,
where K ω and K V is the frequency and voltage proportional drooping coefficient,
respectively.Pavg and Q avg are the average measured active and reactive powers, Pi∗
and Q i∗ are the active and reactive set points power. The K ω and K V droop slopes
are determined according to the power rating of the inverter and according to the
maximum allowable variations in output frequency and voltage [11].
Figure 2 show the droop control characteristics, we assumed that the two inverters
have the same power ratings (Pn1 = Pn2 ) which means they have the same drooping
gains (K ω1 = K ω2 = K ω ).
During grid-connected mode, the frequency of the Microgrid is fixed and main-
tained by the utility grid and equals to the nominal frequency ω M.g = ωgrid = ωn .
Fig. 2 Structure (P - ω) characteristics of two inverters in droop control

(P1+P2)>PL
Mode 1 Mode 2
Gridon Gridoff
= =
(P1+P2)<PL
Mode 3
Unitentional
islanding
=
Fig. 3 The different operation modes of the Microgrid
In this mode, the two inverters generate different power values P1(grid) and P2(grid)
because
∗ we assumed that the two inverters have different power set-points, such as
P1 = P2∗ .
When the microgrid transits to the unintentional islanding
mode,
the frequency of
the Microgrid deviates from its nominal value ω M.g = ωn and the inverters
generated different mounts of power P1(u.island) and P2(u.island) where the
P2(u.island) < 0 which means that the inverter 2 is importing power from the
inverter 1 (P1(u.island) > 0).
The conditions leading to the switching between the different operating modes of
the Microgrid can summarized in Fig. 3 [10].
The importing power occurs in the inverter that receives lower setpoint power
resulting high DC voltage to last, if this value exceeds the maximum voltage limit
allowable for the DC bus, the inverter will shut down by protection circuit element.
This process can affect the reliability of this Microgrid.
3 Dynamic Controllers Design
To overcome the problem of reliability of the Microgrid, the system needs the regu-
lation of the DC bus voltage. Two controllers can be used in this study. The Primary
Control designed to control the inverter and the Secondary Control designed to regu-
late the DC link voltage by controlling the DC-DC converter. The design and the
synthesis of the two controllers are given in the following subsections.
A. The Primary Control
The primary controller consists of two control loops: the inner loop represented by
the voltage control and the outer loop represented by droop controller.
352 I. Ameur et al.
The outer control loop

The inverter controllers received the signal from the supervisory controller (SC)
about the state of STS (gridon or gridoff). In case of grid-connected mode, the active
and reactive power of inverters (Pi , Q i ) with i = 1, 2 are regulated at their reference
values (Pi∗ , Q i∗ ) that are delivered by the SC. In case of standalone mode, the nominal
power values (Pn , Q n ) become the reference values, which ensure the regulation of
the frequency. When the unintentional islanding mode happens, the reference values
become the reference DC bus voltage in order to detect and prevent the circulating
of the transit power among the inverters as illustrated in Fig. 4.
The power error is the input of the droop controller, this controller consists of the
Eqs. (1, 2), the output of this controller is the reference voltage signal of AC link
Vac_r e f .
The inner control loop
In order to ensure the stabilization of the output voltage of the inverter (Vac ) around
its desired value (Vac_r e f ) that is obtained by the outer loop, a Proportional-Integral
(PI) controller with PWM techniques is used, as well which is given, by

Iac_r e f = K P Vac_r e f − Vac + K I Vac_r e f − Vac dt (3)
Fig. 4 The unintentional islanding controller in the outer control loop

Fig. 5 The inner control loop
where K P and K I are the parameters of PI controller, obtained by using the PSO
algorithm.
It should be noted that we can use the Feedforward control instead of the integral
action in the PI controller (Fig. 5).
B. The Secondary Control (Proposed Controller)
This section discusses the proposed Dc-bus voltage controller. In this paper, the regu-
lation of the DC-bus voltage is accomplished by the photovoltaic system (structure1
in Fig. 1) and the storage system—battery (structure2 in Fig. 1).
The control of photovoltaic system
The PV boost controller has two functions: extract the maximum power available
from the PV array through MPPT algorithm, and regulate the DC link voltage by
reducing the injection of the current from the DC source (PV) when the Vdc increased
especially if the inverter imports the power (unintentional islanding case) and cannot
handle with the Vdc .
The Second Order Sliding Mode (SOSM) approach can used in order to track
the reference of DC bus voltage (Vdc_r e f ) and improve the transient response with a
fast stabilization of the system against parametric variation and large range of load
variation. In order to determine the control law, let us first recall the dynamic model
of the system PV-boost converter which is highly nonlinear [12]:
⎧
⎪
⎪ ẋ1 = λ1 x3 − λ1 x2 (1 − u)
⎨
ẋ2 = λ2 x1 (1 − u) − λ3 x2
(4)
⎪
⎪ ẋ = λ4 I pv − λ4 x1
⎩ 3
y = s(x, t)
where λ1 = L1 , λ2 = C1dc , λ3 = RC1 dc , λ4 = C1pv , x1 = i L , x2 = Vdc , x3 = V pv

t
and x = x1 x2 x3 R3 is the state vector and, uR is the bounded input control,
s(x, t) is the smooth output- feedback function . The control objective is making
output function s = 0.
354 I. Ameur et al.
Supposing that the entire variable can be measured, the sliding mode surface is
combined with the integration of the error is chosen as:

e = x2 − x2−r e f
(5)
s(x, t) = e + k e dt
where k is a constant.
we choose to apply the second order sliding mode control.
The control law depends on the switching function as follows:
u = u1 + u2 (6)
With

u̇1 = −λ1 sign(s)
(7)
u2 = −λ2 sγ
where λ1 > 0, λ2 > 0 and 0 < γ ≤ 1/2.

Where λ1 and λ2 are positive real constant parameters that can be selected prop-
erly and aθ = |a|θ sign(a), ∀a R, θ > 0 [Lagh15]. In order to ensure that the
real sliding order is a maximum, we choose γ = 1/2 and can reach the second order
sliding [79-thes sahraoui].
The reader may refer to [13–18] for a detailed proof (Fig. 6).
Fig. 6 The proposed control structure of the PV-boost converter

The control of battery system

The bidirectional DC-DC converter can work in the two direction as a boost in one
direction (discharging mode operation), or as buck in the other direction (charging
mode operation), depending on the battery current orientation.
The main objective to control the battery is to regulate the DC-bus voltage, dual-
loop feedback control strategy can be used (outer and inner loops). The first can
realized by the PI control to generate the reference of the battery current which is
given by Eqs. (8, 9) and illustrate in Fig. 7.
maintain the stability of the DC bus voltage which can be expressed as:

Idc−r e f = K p .(Vdc−r e f − Vdc ) + ki . Vdc−r e f − Vdc dt (8)
I Bat−r e f = 1/Vbat (Idc−r e f .Vdc + Iinv .Vdc − I pv .V pv ) (9)
and the latter loop can be achieved through a sliding mode controller (SMC) to
control the charging/discharging current value.
In order to design the Sliding Mode controller, let us first recall the dynamic model
of the battery-bidirectional converter that can be defined as [20]:
For buck mode:
⎧
⎨ ẋ1 = λ1 Vbat − λ1 x2 u bat1
ẋ = −λ2 x1 u bat1 + λ3 x2 (10)
⎩ 2
y = s(x, t)
For boost mode:

⎧
⎨ ẋ1 = λ1 Vbat − λ1 x2 (1 − u bat2 )
ẋ = λ2 x1 (1 − u bat2 ) − λ3 x2 (11)
⎩ 2
y = s(x, t)
where λ1 = L1 , λ2 = C1dc , λ3 = RC1 dc , λ4 = C1bat , x1 = Ibat , x2 = Vdc , and

t
x = x1 x2 R2 is the state vector and, (u bat1 , u bat2 )R is the bounded input
control, s(x, t) is the smooth output- feedback function. The control objective is
making output function s = 0.
Supposing that the entire variable can be measured, the sliding mode surface is
chosen as:
s(x, t) = x1 − x1−r e f (12)
where x1−r e f is the reference of input current.

The control law depends on the switching function as follows:
356 I. Ameur et al.
Fig. 7 The proposed control structure of the battery converter
• For the charging battery process “buck mode”: the control law corresponding this
mode is:
1
u bat1 = (1 − sign(s)) (13)
2
• For the discharging battery process “boost mode”: the control law corresponding
this mode is:
1
u bat2 = (1 + sign(s)) (14)
2
For a more detailed, the reader may refer to [19, 20].
The proposed control of the Microgrid, presented in Fig. 2, is validated through

simulations for a typical converter. We care about the steady state of the modes. The
performance and robustness of the sliding mode controller will be checked through
unintentional islanding scenario.
The Fig. 8 illustrates the continuous state evolution of the DC Microgrid, when
the power of the inverter 300 W to the grid, while the inverter 2 is exporting 600 W.
At t = 3 s, the grid was lost and the transient power phenomena is occurred, which is
corresponding to the circulating of the power from inverter 2 to the inverter1, and this
results the increase of the DC bus voltage. The primary controller of the inverter 1 and
2 responses to prevent the circulating of the power. While the secondary controller
is interested to the DC voltage regulation, the PV-1 controller involve by reducing
the PV generated power. On the other hand, the DC voltage of the inverter 2 increase
also and need to regulate. Since the SOC of the battery is low, the battery controller
intervenes to regulate the DC voltage by adjusting the output power of the inverter
because the islanding mode has not been detected yet. At t = 3.25 s, the DC voltage
was regulated and the disturbance is eliminated.
5 Conclusion
In this paper, a new control scheme is developed for controlling a Microgrid in case
of transient power between paralleled inverters during unintentional islanding mode.
Different order of Sliding Mode Control strategies are used in the system, which is
advantageous to regulate the DC link with excellent dynamic responses compared
to the classical control method. The simulation results confirm the effectiveness and
the robustness of the suggested controllers under different modes operation of the
Microgrid with the seamless transfer of power between them.
As a future work, a complete study of the voltage and frequency restoration in
Islanded Microgrid will be investigated, to ensures the power sharing and the perfect
control of frequency restoration.
358 I. Ameur et al.
a. Power inverters d. PV2 power
b. DC bus voltage e. Battery power
c. PV 1 power
f. PCC voltage
g. Grid current
Fig. 8 State evolution of the DC Microgrid when Pinv2 > Pinv1 under different mode operation
References
1. Gao M, Chen M, Jin C, Guerrero J, Qian Z (2012) Analysis, design, and experimental evaluation
of power calculation in digital droop-controlled parallel microgrid inverters. J Zhejiang Univ
Sci C (Computers & Electronics). https://doi.org/10.1631/jzus.c1200236
2. Issa W, Abusara M, Sharkh S (2014) Control of transient power during unintentional islanding
of microgrids. IEEE Trans Power Electron 30(8):4573–4584
3. Abusara MA, Guerrero JM, Sharkh SM (2014) Line-interactive UPS for microgrids. IEEE
Trans Ind Electron 61(3):1292–1300
4. Laaksonen H, Saari P, Komulainen R (2005) Voltage and frequency control of inverter based
weak LV network microgrid. In: Int Conf Future Power Syst. IEEE, p 6
5. Hossain MA, Azim MI, Mahmud MA, Pota HR (2015) Active power control in an islanded
microgrid using dc link voltage status. In: IEEE Innovation Smart Grid Technology-Asia (ISGT
ASIA), pp 1–6
6. He J, Lu C, Jin X, Li P (2008) Analysis of time delay effects on wide area damping control.
In: IEEE Asia Pacific conference on circuits system, APCCAS 2008. IEEE, pp 758–761
7. Liu F, Gao H, Qiu J, Yin S, Fan J, Chai T (2014) Networked multirate output feedback control
for setpoints compensation and its application to rougher flotation process. IEEE Trans Ind
Electron 61(1):460–468
8. Hossain M, Pota H, Haruni A, Hossain MJ (2014) DC-link voltage regulation of inverters to
enhance microgrid stability during network contingencies. IEEE Trans Ind Electron 61(1):460–
468
9. IEEE Standard for interconnecting distributed resources with electric power systems, IEEE
Standard 1547–2003, New York, Jun 2003
10. Issa W, Abusara M, Sharkh S (2015) Control of transient power during unintentional islanding
of microgrids. IEEE Trans Power Electron. https://doi.org/10.1109/tpel.2014.2359792
11. Mohamed YR, Zeineldin HH, Salama M, Seethapathy R (2012) Seamless formation and robust
control of distributed generation microgrids via direct voltage control and optimized dynamic
power sharing. IEEE Trans Power Electron 27:1283–1294
12. Mahmood H, Michaelson D, Jin J (2012) Control strategy for a standalone PV/battery hybrid
system. In: 38th annual conference on IEEE industrial electronics society—IECON 2012, pp
3412–3418
13. Mu C, Sun C, Qian C, Zhang R (2013) Super-twisting sliding mode control based on Lyapunov
analysis for the cursing flight of hypersonic vehicles. In: 10th IEEE international conference
on control and automation (ICCA), Hangzhou, China, 12–14 June 2013
14. Koo B, Yoo Y, Won S (2012) Super-twisting algorithm-based sliding mode controller for a
refrigeration system. In: 12th international conference on control, automation and systems, In
ICC, 17–21 October 2012, Jeju Island, Korea
15. Utkin V (2014) Mechanical energy-based Lyapunov function design for twisting and super-
twisting sliding mode control. IMA J Math Control Inf, 1 of 13 https://doi.org/10.1093/ima
mci/dnu010
16. Chalanga A, Kamal S, Fridman M, Bandyopadhyay AM (2016) Implementation of super-
twisting control: super-twisting and higher order sliding mode observer based approaches.
IEEE Trans Ind Electron. https://doi.org/10.1109/tie.2016.2523913
17. Derafa L, Benallegue A, Fridman L (2011) Super twisting control algorithm for the attitude
tracking of a four rotors UAV. J Franklin Inst. https://doi.org/10.1016/j.jfranklin.2011.10.011
18. Sahraoui H, Drid S, Alaoui L, Ouriagli M, Bussy P (2014) Robust control of the boost converter
applied in photovoltaic systems using second order sliding mode. In: 15th international confer-
ence on sciences and techniques of automatic control & computer engineering—STA’2014,
Hammamet, Tunisia, 21–23 December 2014. IEEE 978-1-4799-5907-5/14/$31.00
19. Ameur K, Hadjaissa H, Cheknane A, Essounbouli N (2017) DC-bus voltage control based on
power flow management using direct sliding mode control for standalone photovoltaic system.
Electric Power Comp Syst. https://doi.org/10.1080/15325008
20. Tsang KM, Chan WL (2013) Model based rapid maximum power point tracking for
photovoltaic. Energy Convers Manage 70:83–89
New Virtual Synchronous Generator
Control Technique of Distributed
Generator Unit to Improve Transient
Response of the Microgrid
Yacine Daili and Abdelghani Harrag
Abstract To allow a high penetration of the renewable energy into the grid,
Distributed Generator (DG) based Virtual Synchronous Generator (VSG) control
technique has been proposed. DG units are often connected to low and medium
line voltage, where the impedance line is not purely inductive. The active and reac-
tive powers are nonlinear functions, which makes the control of active and reactive
power more complicated. Thus, dramatic oscillations of the delivered power and
frequency of the DG based VSG may occurred in dynamic steady state. The oscil-
lation of frequency and power could be suppressed by increasing the VSG damping
coefficient. However, the response time of the system is increased. The main contri-
bution of this work is to propose a new VSG control technique to improve the
dynamic performances of the conventional one. With the proposed VSG technique
the damping coefficient is auto-adjustaed to eliminate the power and frequency oscil-
lation without scarifying the response time of the system. The effectiveness of the
proposed VSG control strategy is verified and compared with basic VSG control
technique by simulation, the results demonstrate the superior performances of the
proposed VSG control strategy.
Keywords Virtual synchronous generator (VSG) · Micro-grid · Power electronics

converter · Renewable energy sources · Distributed generator · Grid stability
Y. Daili (B) · A. Harrag

Mechatronics Laboratory (LMETR) - E1764200, Optics and Precision Mechanics Institute,
University of Ferhat Abbas – Sétif 1, 19000 Sétif, Algeria
A. Harrag

https://doi.org/10.1007/978-981-15-6403-1_25
362 Y. Daili and A. Harrag
1 Introduction
With rapid growth penetration of renewable energy sources, such as wind turbines,
fuel cell and photovoltaic into the grid, the conventional Synchronous Generators
(SGs) are more and more substituted by renewable energy sources usually interfaced
to the main grid through power electronic inverters. The absence of inertia with
inverters leads to decrease the total inertia of the power system. Therefore, high
penetration of the renewable energy into the grid can cause a stability problem of the
grid [1–5].
The droop-based control method and Virtual Synchronous Generator (VSG)
control method are widely employed in Distributed Generator (DG) units [6–8].
These control techniques can ensure both grid connected and island modes oper-
ating of the system. With a droop-based control method, the injected active and
reactive powers into the grid are performed by imitating static operation character-
istics of SGs. The problem associated with DG based this controller is the absence
of the inertia [9]. To solve this issue, virtual synchronous generator known also as
synchronverters was introduced by [8, 10]. The idea of the VSG is to control the DG
based inverter to emulate the essential dynamic behavior of synchronous generators,
the inertia characteristic emulated by the VSG controller participates in providing
the necessary frequency support for the grid.
The authors of [3] prove that the VSG control technique and droop control
strategy have the same dynamic if an appropriate low-pass filter is added to the
droop controller.
DG units are often coupled to the grid via a low and medium voltage line, where the
line impedance is not dominantly inductive and the power angle is not enough small.
The active and reactive powers are nonlinear functions and strongly coupled [11]. As
results, using the conventional VSG control technique could introduce an oscillatory
mode of the system, which causes low frequency oscillations of the inverter output
power and frequency. These oscillations deteriorate the dynamic performance of
VSG, and it may even result in unnecessary trips of DG units, because inverters have
very limited over-current capability compared to real synchronous generator [1].
Consequently, the parameters design the VSG controller to fulfil the requirements of
stability, oscillation damping and dynamic performances of the system is challenging
task [11–14].
The main contribution of this paper is to propose a new VSG control method
to improve the transient response of the DG system. With the proposed technique,
the damping coefficient of VSG controller is auto-adjustated to eliminate the power
and frequency oscillations without scarifying the response time of the system. The
effectiveness of the proposed VSG control strategy is tested and compared with basic
VSG controller by simulation.
The paper is organised as follows: basic operation principal of VSG controller is
discussed in Sect. 2. The principal of the proposed VSG controller is presented in
Sect. 3. The simulation results with proposed and conventional VSG controller are
discussed in Sect. 3 to verify the theoretical analysis. Finally, this paper is summarized
in Sect. 4.
New Virtual Synchronous Generator Control Technique … 363
2 Principle and Modelling of a VSG Control Strategy
A. Principal of a VSG control technique
The control circuit diagram of the VSG is illustrated in Fig. 1. To simplify studies, the
DC source is used in place of distributed generation (DG). The three phases voltage
source inverter (VSI) is connected to the grid at the point of common coupling (PCC)
through an LC filter, where L F and C F are its elements and line impedance Z L , which
is included the resistance and inductance line respectively RL and L L .
Where vm and vg represents the terminal voltage of a VSG and the grid voltage
respectively.
In this schema, the delivered active power Pout and reactive power Qout to the grid
are computed from vg and ig in αβ reference frame as follows:

Pout = Vgd Igd + Vgq Igq
(1)
Q out = Vgβ Igα − Vgα Igβ
PCC
LF RL LL
VSI
CF
PWM
abc
Capacitor voltage
controller
Active power controller Pout

Active & Reactive power calculation
(Eq 1)
Vg
Vg PLL
reference
synthese
Voltage
(Eq 5)
Reactive power controller
Fig. 1 Control block diagram of VSG

The VSG mimics the swing equation and damping characteristic of a synchronous
generator. Thus, the power equation can be expressed as [2, 15]:

Pin −Pout
ωg
= D ωg − ωm + J dω
dt
m
dθg (2)
dθm
dt
= ωm and dt
= ωg
where θ g and θ m represent respectively the grid voltage phase and VSG virtual phase,
J is the rotational inertia, D is the damping coefficient, Pin is the input power (as same
the prime mover power in a conventional SG), Pout is the electrical output power,
ωg is the grid angular velocity and ωm is the virtual rotor angular velocity. The grid
angular velocity ωg is detected at the PCC by the phased looked loop PLL system,
and then the virtual rotor speed ωm is computed by resolving the Eq. (2). By passing
ωm through an integrator, the virtual mechanical phase angle θ m is produced.
The input power Pin is generated by frequency droop regulator according to
Eq. (3), which simulates the governor of the conventional SG [16]:

Pin = Pr e f + K p ω0 − ωg (3)
Pref is the reference value of active power, ω0 is the rated angular velocity, K p is the
active droop coefficient.
The reactive power controller emulates the conventional excitation circuit of a
real SG. The control loop includes a reactive power droop mechanism Eq. (4) and a
voltage regulator (5) [17]:

Q in = Q r e f + K q V0 − Vg
(4)
Vm = KSi (Q in − Q out ) + V0
where V g is the amplitude value of the grid voltages, V 0 is the no-load voltage, K q
is the reactive droop coefficient, K i is the integral coefficient of the voltage regulator
and Qref is the reference value of reactive power.
According to (3) and (4) the active and reactive power delivered by the DG unit
change with respect to the PCC frequency and the amplitude of its voltage
The amplitude of electromotive force V m and the virtual mechanical phase angle
θ m information are used for synthesizing the references voltages (vma , vmb ,vmc ), which
are the inputs of voltages controllers. The voltage references can be written as:
⎧ √
⎨ vma = √2Vm sinθ m
v = 2V sin θ − 2π
(5)
⎩ mb √ m m 3
vmc = 2Vm sin θm + 2π
3
Finally, the voltages controller is introduced to track the voltage references. The
outputs of the voltage controller are fed into the PWM modulator to the drive signals
of the inverter.
It worth noting that, the coefficient K p and K q in (3) and (4) are selected based
on the local grid standard [11, 18].
B. Modelling of a VSG control strategy
Considering the equivalent model of the system illustrated in Fig. 1 where tow nodes
V m and V g separated by a line impedance Z L . The delivered active and reactive
powers to the grid are expressed as [2]:
⎧
⎨ Pout = R L Vm2 −R L Vm Vg cos δ+X L Vm Vg sin δ
R 2L +X 2L
X L Vm2 −X L Vm Vg cos δ−R L Vm Vg sin δ (6)
⎩ Q out =
R 2L +X 2L
where δ is the load angle has the following expression:

δ = θg − θm = ∫ ωg − ωm dt (7)
For a large synchronous generator coupled to the grid through a high voltage line,
the line impedance is purely inductive. Thus, the active power is controlled by the
angle and the reactive power by amplitude output voltage V m .
However, for a small and medium DG systems coupled to a low or medium voltage
line, the impedance line is not dominantly inductive, and as a consequence, the active
and reactive powers are nonlinear functions. Additionally, they are not independent,
changes of V m and affect both Pout and Qout variations. This behaviour makes the
control of active and reactive more complicated.
Considering small perturbations of the active and reactive powers denoted by
ΔPout , ΔQout resulting from small variations of the input variables δ and ΔV m . By
linearizing (6) around an equilibrium point, the following functions are obtained:
⎧

⎨ Pout = P0 + R2L Vm02 Vm + −Q0 + X2L Vm02 δ = K P−V Vm + K P−δ δ

2
Vt0 RL +XL
RL +X
L m
⎩ Q out =
2
Q R V
0
Vt0
+ R2 +X2 Vm + P0 − R2 +X2 δ = K P−Vm Vm + K P−δ δ
X L V m0 L m0
L L L L
(8)
P0 and Q0 are the active and reactive powers equilibrium point respectively.
The small signal of the active and reactive controllers (2)–(4) could be obtained
as:
Pin − Pout
= D ωg − ωm + J s ωm (9)
ωg0
Pin = Pr e f − K p ωg (10)
Q in = Q r e f − K q Vg (11)
Ki
Vm = ( Q in − Q out ) (12)
s
From (8) to (10), the transfer functions of active and reactive powers have the
following expressions:
K P−Vm K P−δ J ωg0 s 2

Pout (s) = Vm (s) + Pr e f (s)
K P−δ J ωg0 s 2 + 1 K P−δ J ωg0 s 2 + 1
K P−δ J ωg0 s 2 K P−δ J D K p s 2
+ ω g (s) − δ(s) (13)
K P−δ J ωg0 s 2 + 1 K P−δ J ωg0 s 2 + 1
Ki Ki Kq K Q−δ s
Q out (s) = Q r e f − Vg (s) + δ(s) (14)
s + Ki s + Ki s + Ki
The frequency and voltage amplitude deviations of the micro-grid could be a result
of connection/disconnection of other DG systems or a transit of the micro-grid load.
The parameters determining the dynamic behavior of the active and reactive
powers loop are the rotational inertia J, the damping coefficient D and the integral
coefficient of the voltage regulator K i respectively.
3 Proposed VSG Control Technique
As demonstrated in the previous section, both active and reactive powers loops are
strongly coupled, which makes the the parameters design of active and reactive loops
to fulfil the requirements of stability, oscillation damping and dynamic performances
of the system is not straightforward task.
Assuming that, the coupling effect on the active power control loop is neglected,
one can observe that the transfer function from reference power to the output power
(Eq. 13) that the overshoot and the response time for a fixed inertia J depend only
on the damping coefficient D. So that, a small damping coefficient leads to large
oscillations of the DG output power.
The output power and frequency oscillations could be considerably reduced by
choosing a large value of damping coefficient. However, the system takes a long time
to reach the set point.
In order to improve the response time without causing large oscillations of the
output power, a new technique is proposed in this paper. It consists in adjusting auto-
matically the VSG damping coefficient. Figure 2 summarizes how this coefficient is
adjusted. This figure shows the reference power in pointed line with power measured
in solid line. From this figure, we can distinguish two phases, a first one is when
the system is divergent from the set point and the second phase is the divergent of
the system from the set point. To improve a system response time and weaken the
overshoot and oscillations of output power and frequency, this coefficient is selected
Measured power
Reference power
Small D Big D Small D Big D Small D
Fig. 2 Rules of damping coefficient selection
small when the system is convergent to the set point, while in the divergence phase
the coefficient is chosen big to damp rapidly the kinetic energy stored in the virtual
energy and reduce oscillations of the output power.
In order to detect the divergence and the convergence of the system from the set
point, both power error and derivative of the measured power are used. So that, the
system is convergent in two cases. The first one is when the active power derivative
is positive and the error is negative. Secondly, when the derivative of the measured
active power is negative and the error is positive.
While, it is divergent in case of positive error and positive derivative or when the
error is negative and also the derivative is negative.
The synoptic schema of the proposed VSG control technique is illustrated in
Fig. 3, the additional damping coefficient D́ is added or removed by switching S
Fig. 3 Control diagram of the proposed VSG technique

according to either the system is divergent or convergent from the set point. The
switch S is controlled by two signals, the derivative of the measured power and the
power error. The hysteresis comparator are used instead of a simple comparator to
make the system insensitive to the noise of measurement signals.
To verify the improvement of the proposed VSG compared with a conventional one,
the distributed generation system shown in Fig. 4 with a proposed and conventional
VSG are simulated in Matlab/Simulink environment under several power scenario
are tested. The system under study is presented in Fig. 4. The connected load is
powered by the DG system and the grid, the DG is coupled to the PCC trough
resistive-inductive line impedance to get more realistic case. The system is running
in grid connected mode. The essential parameters of the system are listed in Table 1.
Throughout the simulation validation, the comparative study between the conven-
tional and proposed VSG techniques is performed. Figure 5 shows the dynamic
response of the system illustrated in Fig. 4 controlled with proposed and conventional
VSG techniques.
This figure presents the behavior of the DG system with both proposed and conven-
tional VSG techniques under step up and steep down of the reference active power.
The active power increases from 2000 to 7000 W in the right column and decreases
from 7000 to 2000 W in the lift column, in both cases the reactive power reference is
set to 0 VAR, and it is remained unchanged. As can be observed from this figure by
applying the conventional algorithm, the injected active power at the PCC presents
large oscillations around its reference with an overshoot about 33.7% and the setting
time is equal to 0.347 s.
By using the proposed algorithm, the active power reacts with excellent dynamic.
Comparing with the conventional technique, the overshoot and setting time are
reduced by almost half. So that, the proposed algorithm detects successfully the
Vg
+ RL LL
VSI
-
Grid
PCC
Load : P=8kW,
Q=4 kvar
Fig. 4 The simulated micro-grid platform

Table 1 System parameters

Parameters Values
Nominal utility line-line voltage (rms) 220 V
Nominal utility frequency 50 Hz
DC link voltage 450 V
Inverter switching frequency 4 kHz
Inverter output filter inductance 5 mH
Inverter output filter capacitor 80 μF
Inverter rated power 10 kW
Line resistance 0.641
Line inductance 0.32 mH

Virtual inertia 0.5 kg m2
Damping coefficient 10
Active droop coefficient 7957
Reactive droop coefficient 1250
divergence/convergence of the active power to set point in all cases (see Fig. 5b), and
it is switch on/off the additional coefficient D’ to improve a system response time
and weaken the overshoot.
By observing waveforms of the inverter output frequency by applying both VSG
control approaches (Fig. 5c), one can note that the two techniques present the same
overshoot in dynamic steady state. However, the system with the proposed controller
is less oscillatory, the system reaches the steady state rapidly with lower setting time.
Consequently of the oscillations of the output power with the conventional VSG
technique, the inverter output current is oscillatory and presents a considerable over-
shoot (see Fig. 5d). It worth to note that currents oscillations may result in unnecessary
trips of the DG unite, because the inverter have very limited over-current capability
compared to the conventional synchronous generator. Thanks to the proposed tech-
nique, the overshoot and low frequency oscillations of the output current are reduced
and system reaches the steady state with lower setting time (Fig. 5e).
5 Conclusion
This work interests on the development of a new VSG control strategy applied to DG
unit to improve transient response of the conventional VSG, the proposed control
schema consists in auto-adjusting the damping coefficient of VSG controller to elim-
inate the power and frequency oscillation without scarifying the response time of the
system, So that, the dumping coefficient is selected small when the system is conver-
gent to the active power set point, while in the divergence phase the coefficient is
chosen big to reduce oscillations of the output power.
8000
(a) Reference power 9000
7000 (a)
Conventional VSG 8000 1685 W
Tr-Pop=0.184 S
6000 Proposed VSG 870 W
7000
Active Power [W]
5000 Tr-Conv=0.347 S
6000
4000 Reference power
5000
Tr-Pop=0.192 S
Conventional VSG
3000
4000 Proposed VSG
2000 Tr-Conv=0.349 S
866 W 3000
1000
1680 W 2000
0
5.95 6 6.05 6.1 6.15 6.2 6.25 6.3 6.35 6.4 6.4 1000
Time [Sec] 8.5 8.55 8.6 8.65 8.7 8.75 8.8 8.85 8.9 8.95 9
Time [Sec]
(b) (b)
1 1
0.8 0.8
Control Signal
Control Signal
0.6 0.6
0.4 0.4
0.2 0.2
0 0
5.95 6 6.05 6.1 6.15 6.2 6.25 6.3 6.35 6.4 6.4 8.5 8.55 8.6 8.65 8.7 8.75 8.8 8.85 8.9 8.95 9
Time [Sec] Time [Sec]
50.06 (c) Conventional VSG

50.02
(c) Conventional VSG 50.05 Proposed VSG
50.01 Proposed VSG
50.04
Frequency [Hz]
50
Frequency [Hz]
50.03
49.99
50.02
49.98
50.01
49.97 50
49.96 49.99
49.95 49.98
8.5 8.55 8.6 8.65 8.7 8.75 8.8 8.85 8.9 8.95 9
49.94
5.95 6 6.05 6.1 6.15 6.2 6.25 6.3 6.35 6.4 6.4 Time [Sec]
Time [Sec]
20 20
(d)
15 (d) 15
10 10
Current [A]
5 5
Current [A]
0 0
-5 -5
-10 -10
-15 -15
-20 -20
5.95 6 6.05 6.1 6.15 6.2 6.25 6.3 6.35 6.4 6.45 8.5 8.55 8.6 8.65 8.7 8.75 8.8 8.85 8.9 8.95 9
20 20
(e)
(e)
15 15
10 10
Current [A]
5 5
Current [A]
0 0
-5 -5
-10 -10
-15 -15
-20 -20
5.95 6 6.05 6.1 6.15 6.2 6.25 6.3 6.35 6.4 6.45 8.5 8.55 8.6 8.65 8.7 8.75 8.8 8.85 8.9 8.95 9
Fig. 5 Dynamic performance comparison between proposed and conventional VSG under step
up (right column) and step down (lift column) of the reference active power. a DG output active
power with both techniques, b switching signal of the proposed algorithm, c DG output frequency
with both techniques, d DG output current with conventional technique, e DG output currents with
proposed technique
The proposed VSG has been implemented and compared to the basic VSG scheme
using Matlab/Simulink environment. Simulation results show the superior perfor-
mances of the proposed VSG control strategy, Thus, based on the proposed new
VSG controller, the overshoot and setting time are reduced by almost half for tested
reference active power steps.
All these results demonstrate the effectiveness of the proposed VSG control
technique.
Acknowledgment The Algerian Ministry of Higher Education and Scientific Research via the
DGRSDT supported this research (PRFU Project code: A01L07UN190120180005).
References
1. Liu J, Miura Y, Ise T (2019) Fixed-parameter damping methods of virtual synchronous

generator control using state feedback. IEEE Access 7:99177–99190
2. Daili Y, Harrag A (2019) New model of multi-parallel distributed generator units based on
virtual synchronous generator control strategy. Energy Ecol Environ, 1–11
3. Liu J, Miura Y, Ise T (2016) Comparison of dynamic characteristics between virtual
synchronous generator and droop control in inverter-based distributed generators. IEEE Trans
Power Electr 31:3600–3611
4. Zhong Q-C (2016) Virtual synchronous machines: a unified interface for grid integration. IEEE
Power Electron Mag 3:18–27
5. Huang L, Xin H, Wang Z (2019) Damping low-frequency oscillations through VSC-HVDC
stations operated as virtual synchronous machines. IEEE Trans Power Electron 34(6):5803–
5818
6. Monica P, Kowsalya M (2016) Control strategies of parallel operated inverters in renewable
energy application: a review. Renew Sustain Energy Rev 65:885–901
7. Shuai Z, Hu Y, Peng Y, Tu C, Shen ZJ (2017) Dynamic stability analysis of synchronverter-
dominated microgrid based on bifurcation theory. IEEE Trans Ind Electron 64:7467–7477
8. Zhong Q-C, Weiss G (2011) Synchronverters: inverters that mimic synchronous generators.
IEEE Trans Ind Electron 58:1259–1267
9. Zhong Q-C, Konstantopoulos GC, Ren B, Krstic M (2018) Improved synchronverters with
bounded frequency and voltage for smart grid integration. IEEE Trans Smart Grid 9:786–796
10. Beck H-P, Hesse R (2007) Virtual synchronous machine. In: Proceedings of 9th international
conference on electrical power quality and utilization, pp 1–6
11. Li B, Zhou L, Yu X, Zheng C, Liu J (2017) Improved power decoupling control strategy based
on virtual synchronous generator. IET Power Electron 10:462–470
12. Wu T, Liu Z, Liu JJ, Wang S, You ZY (2016) A unified virtual power decoupling method for
droop-controlled parallel inverters in microgrids. IEEE Trans Power Electron 31(8):5587–5603
13. Bin L, Lin Z, Xirui Y, Chen Z, Jinhong L, Bao X (2016) New control scheme of power
decoupling based on virtual synchronous generator. In: Proceedings of IEEE power and energy
conference at Illinois (PECI), pp 1–8
14. Li M, Wang Y, Xu N et al (2016) A power decoupling control strategy for droop controlled
inverters and virtual synchronous generators. In: Power electronics motion control conference.
IEEE, pp 1713–1719
15. Yibin H, Xiangwu Y, Dongxue L, Xinxin L, Bo Z (2017) Experimental study of micro sources
inverter based on virtual synchronous generator. In:Procedings of IEEE conference on electrical
and electronic engineering (ICEEE), pp 2433–2438
16. Meng J, Shi X, Wang Y, Fu C (2014) A virtual synchronous generator control strategy for
distributed generation. In: Proceedings of IEEE China international conference on electricity
distribution, pp 495–498
17. D’Arco S, Suul JA, Fosso OB (2015) Small-signal modeling and parametric sensitivity of a
virtual synchronous machine in islanded operation. Int J Electr Power Energy Syst 72:3–15
18. Dong S, Chen YC (2017) Adjusting synchronverter dynamic response speed via damping
correction loop. IEEE T Energy Conver 32:608–619
Design and Analysis of Robust Nonlinear
Synergetic Controller for a PMDC Motor
Driven Wire-Feeder System (WFS)
Noureddine Hamouda, Badreddine Babes, and Amar Boutaghane
Abstract In this paper a robust non-linear synergetic control (NSC) algorithm

was developed for the speed control of a permanent magnet DC motor (PMDC)
driven wire-feeder systems (WFSs) of gas metal arc welding (GMAW) process. The
proposed control approach allows the PMDC motor to track desired trajectories by the
output voltage of a full bridge converter with respect to the large torque disturbances,
unexpected changes in rotor inertia and makes the wire-feeder unit (WFU) has fast
and stable starting process as well as excellent dynamic characteristics. The effective-
ness of the proposed control scheme is illustrated by numerical simulations within
the MATLAB/SIMULINK environment software. The simulation results clearly
demonstrate the significant improvement rendered by the proposed controller in the
wire-feeder system’s reference tracking performance, torque disturbance rejection
capability, and transient recovery time.
Keywords Component · GMAW process · Wire-feeder system (WFS) ·

Non-linear synergetic control (NSC) algorithm · Simulation results
1 Introduction
Nowadays, joining metals is a fundamental aspect of modern industrialized oper-

ations such as the ship building, automotive, and construction industries. It can
be accomplished by different arc welding techniques. Among the various types
of welding the typical GMAW is the most frequently employed and economically
important welding process for joining metals. It is preferred for its flexibility, rapidity
and can be utilized for both manual and automatic modes of welding for wide range
of ferrous and non-ferrous metal pieces [1].
N. Hamouda · B. Babes (B) · A. Boutaghane

Research Center in Industrial Technologies (CRTI), Algiers, Algeria
N. Hamouda

https://doi.org/10.1007/978-981-15-6403-1_26
374 N. Hamouda et al.
Consistency and high quality welding procedures are the key issues to maintain
and increase the overall product quality. During GMAW, the electrode wire is melted
and liquid droplets are formed at the tip of the electrode. When detaching from the
electrode, the droplets transfer both mass and heat into the weld pool. In order to
achieve quality welds, the transfer process must be controlled.
One of the strategies to control the quality of the weld is to maintain the set values
of welding current and arc length to achieve the preferred values of heat and mass
transfer to the work-piece. The control variables selected are, the wire feed rate and
the open circuit voltage, which are utilized to control the current and arc length of the
GMAW process [2]. Large swing in wire feed rate results in large increased stress in
PMDC motor and in welding current during welding process. This changes in wire
feed rate causes the arc breaking, affecting the arc stability which ceases the welding
operation.
Therefore, wire-feeder system (WFS) is an important subsystem of typical auto-
matic GMAW process. It should not only prevent the fracture and vibration of wire
from occurring, but also guarantee the high speed of wire feed to meet the need of
high rate of automatic welding production. Therefore, the control of wire feed rate
for the WFS is the key technologies of the wire-feeder units (WFUs), which has a
strong impact on the welding quality.
The available WFUs are designed for constant wire feed rate and feature a large
inertia and static friction due to the reduction gearbox and the eccentricities in the
wire roller mechanism, and also due to the wire spool and the important frictions
of the wire feed path [3]. Thus, this mechanical dynamic is very slow as compared
to the arc welding melting process [4]. Many solutions have been introduced in the
literature to improve the wire feed speed responses of the WFUs either by developing
new mechanisms with various types of permanent magnet DC (PMDC) motors or
by designing robust wire feeder controllers [4, 1].
The accurate design of wire feeder controller is essential to provide productivity,
wide range welding capability, and comfort level to the welder user. However, the
design of the accurate wire feeder controller for PMDC motor of the WFUs using
the traditional PID method require accurate modeling of PMDC motor that considers
the non-linear dynamics. It is possible to use a traditional PID regulator to control
the wire feed speed of the WFUs [5]. However, traditional PID controllers do not
yield reasonable performance over a wide range of working conditions because of
the fixed gains utilized. Thus, the PID parameters need to be automatically adjusted
by a fuzzy set [6].
Accordingly, many non-linear control techniques have been introduced to solve
the wire feeder control problem of WFUs, and hopefully realize certain performance
objectives such as variable structure sliding mode control. Paul [7] used a sliding
mode to control the wire feeding rate to achieve a desired constant welding current
easily and quickly. A similar method with variable structure sliding mode control has
been applied by Ngo et al. [8] to ensure robustness in the GMAW process. Another
application of sliding mode control function was also elaborated by Paul [9], who
developed an accurate speed control of wire feeding using back-EMF as feedback
signal. However, their studies typically needed large time-consuming computations.
Design and Analysis of Robust Nonlinear Synergetic Controller … 375
Although the non-linear synergetic control (NSC) is recently proposed, numerous

applications of these techniques are successfully reported in various domains of
engineering. These controllers are successfully applied to a number of domestic
and industrial manufacturing. The nonlinear power system stabilizers [10], power
converters for pulse current charging [11], DC-DC boost converters [12] and control
of chaotic oscillation in power systems [13] are such of these applications.
In this paper, the new way to controlling the wire feed speed of WFUs with non-
linear synergetic control (NSC) strategy has been investigated. More specifically,
the model of PMDC motor is simplified which makes the NSC algorithm simple
to design. Moreover, the NSC controller performance in presence of parametric
uncertainties and disturbance is discussed.
2 Dynamics of GMAW Process
In the GMAW process, the welding inverter controls the open circuit voltage V oc (V )
between the contact tip tube and the work-piece. In addition, the wire feed servo
motor rotates a set of pinch rollers, which force the wire into the torch head and
through the contact tube whereupon the wire is consumed by the GMAW process
as illustrated in Fig. 1. The wire feed servo motor is in itself a feedback controlled
system which is capable of delivering wire to the weld process at a controlled wire
feed rate; increasing or decreasing the wire feed speed V f (m/s) on the wire feeder
servo motor increases or decreases the welding current I w (A) as well as the metal
transfer mode [1]. In majority of cases, the value of V f is reserved constant at desired
value [14].
Fig. 1 Physical Pinch roller

representation of WFS for Vf
GMAW process
Guides Liner
Rs L1 I
W
Wire spool
Contact tube
ls
Voc CT
larc Varc
Workpiece
For the purposes of this derivation, the wire feed rate is considered to be the input.
The arc dynamics form the plant and the arc current I arc (A) is taken to be the output.
CT = larc + ls = constant (1)
where CT (m) is the contact tube height, larc (m) is the arc length, and ls (m) is the
stick-out length.
The dynamic equation for the electrical circuit of the GMAW process, is
dIW
Voc = L1 + Rs IW + Varc (2)
dt
where V oc (V ) represents the open circuit voltage of arc welder power supply, I w (A)
is the instantaneous welding current, Rs (Ω) is the Thevenin resistance of arc welder
power supply plus cabling resistance and L 1 (mH) is the inductance of arc welder
power supply.
The dynamic equation of arc voltage V arc (V ) is expressed as
Varc = Earc larc + Rarc IW + Vp (3)
where E arc (V /m) is the electric field intensity of the arc plasma, V p (V ) is the output
voltage of power source, and Rarc (Ω) is the arc resistance.
The arc resistance can be written as a function of the arc length
ρ ρ
Rarc = ls = [CT − larc ] (4)
A A
where ρ(Ω.m) is the resistivity of the weld wire and A(m2 ) is the cross-sectional area
of the weld wire.
The dynamic equation of arc length larc , is
dlarc
= Vm − Vf (5)
dt
where V m (m/s) represents the wire melting rate may be expressed as
Vm = km IW (6)
where k m indicates the coefficient of wire melting rate.

The dynamic equation of the power source V p (V ), is
Vp = (Ru + k1 IW )k0 (7)
where Ru (V ) is the control input of the power source, k 0 is gain of power source and
k 1 is feedback gain. Taking the Laplace transform of above equations and connecting
k1 IW (s)
Vf (s)
Voc(s)
_ _
+ + Vp(s) 1 1
Ru(s) k0 Km larc(s)
+ + L1s+Rs + s
Rarc
Earc
Fig. 2 Block diagram of a GMAW for a constant wire feed-rate system
them according to their physical relationship, one can obtain the configuration of the
control unit as shown in the block diagram of Fig. 2.
3 Dynamics of PMDC Motor Driven WFSs
Advances in permanent magnet equipments and attractive characteristics such as

light weight, low cost, low speed, fast response etc., enlarged the exploitation of
permanent magnet DC (PMDC) motor particularly for wire-feeding units (WFUs) in
GMAW process [4]. The graphic illustration of power circuit of the wire-feed servo
motor control is depicted in Fig. 3. The PMDC motor control variables are DC input
voltage V a (V ) and load torque T L (N.m). In WFUs, T L depends on diameter d(cm) of
electrode wire and its material composition. For a particular application it could be
considered as constant. The PMDC motor output variables are the wire feed speed
V f (m/s), the angular displacement of the motor shaft θ (rad) and the armature current
ia (A).
D1 D3 S1 S3
PMDC motor
Ra La
Vac Va A B
C
D2 D4 Ea
S2 S4
Fig. 3 Power circuit for the wire-feed servo motor control

The PMDC motor’s dynamic equations can be developed based on the Kirchhoff’s
voltage law around the armature circuit and the Newton’s moment law using Eqs. (8)–
(10)
dia dθ
Va = La + Ra ia + k2 (8)
dt dt
d 2θ dθ
J +f = k3 ia (9)
dt dt
dθ
Vf = k4 (10)
dt
where, V a (V ) is the armature voltage, ia (A) is the armature current, Ra (Ω) is the arma-
ture resistance, L a (mH) is the armature inductance, θ (rad) is the angular displace-
ment of the motor shaft, V f (m/s) is the wire feed rate, J(kg·m2 ) is the moment of
inertia of the motor and mechanical load converted to the motor shaft, f (Nm·s) is
the coefficient of viscosity of the motor and mechanical load converted to the motor
shaft, k 2 , k 3 and k 4 are constants.
Substitution of (8) into (9) and (10) yields a state space model of the PMDC motor
dynamics
ẋ = Ax + Bu (11)
where
⎡ ⎤ ⎡ ⎤
− RLaa − k4kL2 a 0 1
⎢ k4 k3 ⎥ ⎢ La
⎥
x = ia Vf θ , A = ⎣ J − kJ4 f 0 ⎦, B = ⎣ 0 ⎦ (12)
1
0 k4
0 0
Based on the state space model of PMDC motor of formula (12), manifold and
NSC algorithm can be designed for wire feed speed controller [15].
3.1 Torque Disturbance Model
The wire feed servo motor turns a set of spring-loaded pinch rollers which pull wire
from a large spool, through a series of guides, and down to the torch head (See Fig. 1).
The resulting torque load on the PMDC motor shaft has two principal components.
There is a large constant load due to the stiffness of the wire and the friction of
the guides, rollers, and bearings. In addition, mechanical imperfections in the pinch
roller assembly cause an eccentricity that gives rise to an approximately sinusoidal
term [3].
The total torque load on the PMDC motor shaft is modeled as [3]
TG (t) = A + B sin(2π fD t) (13)
This torque at the gearbox is a product of the torque load on the PMDC motor,
the torque efficiency of the gearbox, and the gear reduction ratio. Reflect the gearbox
torque back through the reduction gears to obtain the torque load on the PMDC motor
shaft [3]
TL (t) = (0.66)−1 (56.5)−1 · TG (t) = 0.0231TG (t) (14)
The frequency of the sinusoidal component is proportional to the wire feed rate
1 1revolution 1min
fD = Vf = 3.26 × 10−3 Vf (15)
π 1.626 in 60 s
where 1.626 inches (4.13 cm) is the diameter of each pinch roller and the wire feed
rate V f is given in inches per minute (IPM).
Extensive studies [3] demonstrate that the typical values for constants A and B in
Eq. (13) are 141 and 15.0 (oz.in) respectively. Reasonable wire feed rates lie in the
200–400 (IPM) range, resulting in disturbance frequencies of 0.65–1.3 (Hz). These
facts yield a typical torque load at the PMDC motor shaft of
TL (t) ≈ 3.27 + 0.35 sin(2π fD t) (oz.in) (16)
where
0.65 Hz ≤ fD ≤ 1.3 Hz
The scaling factor k 4 converts motor angular velocity ω(rad/s) into wire feed rate
V f (IPM)
1 π 1.626 in 1 revolution 60 s IPM.s

k4 = = 0.7447 (17)
65.5 1revolution 2π rad 1 min rad
4 Wire Feed Servo Controller Design
Generally, the wire feed servo controller is designed to realize accurate and robust
tracking of the preferred speed from no load to full load conditions. The level of
uncertainty on the PMDC motor shaft could be huge. There is a perception that
NSC algorithm concept could be capable to achieve accurate and stable tracking in
finite time under such conditions. Hence, a simple NSC control law is employed in
this paper to achieve the design objectives as closely as possible. The design proce-
dure based on NSC algorithm follows the procedure design of aggregated controller
methods [15].
Given that Eq. (12) is in the structure described by (18)
ẋ = f (x, u, t) (18)
where x is the state vector, u is the control input vector, and t is time. Our objective now
is to calculate the control signal u such that the closed-loop system is exponentially
stable at the operating point in finite time. To this end, we define the macro function
σ = σ (x, t). The macro function σ (x, t) contains all the information of PMDC motor
control index. Since the control parameters of the PMDC motor consist of wire feed
rate, angular displacement and, armature current, consider the macro function of the
NSC algorithm as

σ (x, t) = λ1 Vf − Vf∗ + λ2 θ − θ ∗ + λ3 ia − ia∗ (19)
where, λ1 , λ2 and λ3 for weight coefficient, θ * reference value of angular displace-

ment, Vf∗ is the reference value of wire feed speed, ia∗ is the reference value of
armature current.
The control signal u will force the PMDC motor to operate on the manifold
σ (x, t) = 0.
When the PMDC motor is in the manifold σ (x, t) = 0, the stability after the
introducing of control parameter can be ensured at the same time, and it will converge
to the manifold balance. The preferred dynamic evolution of the manifold is defined
by
τ σ̇ (x, t) + σ (x, t) = 0 (20)
where τ is a positive constant which imposes the speed of convergence of macro

function to the manifold σ (x, t) = 0.
The chain rule of differentiation yields
d σ (x, t)
σ̇ (x, t) = · ẋ (21)
dx
Substitution of (18) and (20) into (21) gives
dσ
τ f (x, u, t) + σ (x, t) = 0 (22)
dx
Resolving (22) for u provides the NSC control law as
u = g(x, σ (x, t), τ, t) (23)

Combined system state space model with synergetic control theory, it can obtain
the NSC control algorithm in Eq. (24).

k2 λ2 La λ1 fLa λ1 k3 La
u= − + Vf + Ra − ia . . .
k4 λ3 k4 λ3 k4 J λ3 J
La

− λ1 θf − θf∗ + λ2 Vf − Vf∗ + λ3 ia − ia∗ (24)
τ λ3
To prove stability, we chose the following Lyapunov function candidate
1
V (t) = σ (x, t)2 ≥ 0 (25)
2
After differentiation, this leads to
1
V̇ (t) = σ (x, t) σ̇ (x, t) = − σ (x, t)2 (26)
τ
Because (25) is satisfied and τ > 0, stability is therefore guaranteed for we now
have
V̇ (t) ≤ 0 (27)
By substituting the selected control parameter λ1 , λ2 , λ3 and τ into (24), with

the Lyapunov stability theory, one can obtain that speed of controlled PMDC wire
feeding system is asymptotic stability in the equilibrium state. Figure 4 shows block
diagram of PMDC motor control WFS with NSC algorithm.
TL
Converter PMDC
Motor
u Vf*
Vf - Vf* θf*
NSC θf - θf* ia*
Algorithm ia - ia*
Fig. 4 Block diagram of PMDC motor control WFS with NSC algorithm
In this section, a PMDC motor simulation model is carried out by using

MATLAB/SIMULINK software in order to verify the possibility and capability of
the controlled PMDC wire feeding system using the proposed NSC control law. The
system parameters are set as shown in Table 1 when process simulating.
The tuning parameters required for the proposed controller are λ1 = 0.01, λ2 = 0.
1, λ3 = 100 and τ = 0.000001.
A set of simulation tests is considered, which are detailed in Table 2.
5.1 Tracking Performance Without Load Torque Disturbance
Figure 5 shows the simulation responses for the step change of wire feed speed (0–
280 IPM) without load torque disturbance. It is observed that excellent trajectory
tracking responses are obtained by using the proposed NSC control law due to the
proper selection of the control parameters.
Table 1 Specifications of the PMDC motor used in simulation work

Specifications of the PMDC motor
Major parameters Parameter value
Armature resistance Ra 0.25
Armature inductance L a 0.5 mH

Rotary inertia J 1 × 10−4 kg m2
Viscous damping coefficient f 1.29 × 10−3
Back EMF constant k 2 N.m.s/rad
Electromagnetic torque constant k 3 0.0573 V.s/rad
0.0573 N.m/A
DC input voltage V a 24 V
Table 2 Simulation tests

Test Details
Test 1 Phase of start-up: a step change in V f (0 − 280 IPM)
Load torque disturbance: none
Load torque disturbance: TL = 3.27 + 0.35 sin(2π fD t)

J 0 ≤ t ≤ 0.05 s
Rotor inertia variations: J =
5 ∗ J t > 0.05 s
Test 4 Steady-state response of the proposed NSC, SMC and PI controllers
300
250
Wire feed rate (IPM)

282
200
()
280
150
278
100
0.066 0.068 0.07 0.072 0.074 0.076
50
0
0 0.02 0.04 0.06 0.08 0.1
Time (s)
80
Armature current (A)
60
40
20
0
0 0.02 0.04 0.06 0.08 0.1
Time (s)
8
Angular displacement (rad)
0
0 0.02 0.04 0.06 0.08 0.1
Time (s)
Fig. 5 Wire feed speed, armature current and angular displacement tracking responses of PMDC
motor in test 1
5.2 Tracking Performance Under Load Torque Disturbance
In order to demonstrate the effectiveness of the proposed NSC controller in the

presence of external disturbances, a torque disturbance is included in this test. The
torque disturbance applied to the PMDC motor is chosen as in Eq. (16). The wire feed
speed, armature current and angular displacement of PMDC motor in association
with the desired trajectories under load torque disturbance are shown in Fig. 6,
respectively. It is obvious that the ripples of the wire feed speed are small under the
action of the proposed NSC controller, this validating the better disturbance rejection
ability of the proposed control method.
300
290
200
280
270
100
260
0.1 0.12 0.14 0.16 0.18
0
0 0.02 0.04 0.06 0.08 0.1
Time (s)
80
60
40
20
0
0 0.02 0.04 0.06 0.08 0.1
Time (s)
8
0
0 0.02 0.04 0.06 0.08 0.1
Time (s)
motor in test 2
5.3 Recovery Performance Under Rotor Inertia Variation
Considering that the sudden rotor inertia variation may be involved in the running
automatic GMAW process, we can assume that the rotor inertia of PMDC motor
increases to 5*J after t ≥ 0.05 s. The responses obtained are shown in Fig. 7, which
300
281
200
280
279
100 278
277
0.04 0.045 0.05 0.055 0.06
0
0 0.02 0.04 0.06 0.08 0.1
Time (s)
80
60
40
20
0
0 0.02 0.04 0.06 0.08 0.1
Time (s)
8
0
0 0.02 0.04 0.06 0.08 0.1
Time (s)
motor in test 3
350
300

250
200
PI
150 SMC
100 NSC
50
0
0 0.05 0.1 0.15 0.2
Time (s)
Fig. 8 Wire feed speed tracking responses of PMDC motor in test 4
demonstrates the high tracking performance and the strong robustness of the proposed
NSC control strategy against the model uncertainties.
5.4 Steady-State Response of the Proposed NSC, SMC and PI

Controllers
To better show the advantage of the considered technique, two other types of wire
feed speed controllers are also considered in the simulations for the purpose of
comparison, which are sliding mode control (SMC) of Paul [9] and conventional PI
control of Chaouch et al. [5]. The comparative simulations are performed with the
same conditions given above without torque disturbance. It can be noticed from the
obtained results that the control performance of the system using the proposed NSC
algorithm was better than that using the SMC and conventional PI controllers with
regard to, not only the settling time, but also the overshoot, rising time, and steady
error (Fig. 8).
6 Conclusion
In this paper, a key solution to the problem of wire feed rate regulation in GMAW
process has been introduced using a robust non-linear synergetic control theory.
Based on the desired manifold and the Lyapunov function method, a digital NSC
control law has been developed to ensure the rapidity and accuracy of tracking control
of WFUs. The proposed NSC algorithm assumes no knowledge of the inertia of the
PMDC motor and is thus unconditionally robust with respect to this mechanical
parametric uncertainty. The major advantage of the proposed NSC control law is its
robustness with respect to torque disturbances, inertia uncertainty and change of the
operating point. Numerical simulations show that global asymptotic tracking can be
obtained by the proposed control scheme. In the future work, the authors intend to
carry out the experimental part to complete and finalize the project.
References
1. Naidu DS, Ozcelik S, Moore KL (2003) Modeling sensing and control of gas metal arc welding.
Elsevier
2. Kahla S, Boutaghane A, Abdallah L, Dehimi S, Hamouda N, Babes B, Amraoui R (2017)
Grey wolf optimization of fractional PID controller in gas metal arc welding process. In:
Proceedings of 5th international conference on control engineering and information technology
CEIT, Sousse, Tunisia
3. Greene BW (1988) Arc current control of a robotic welding system: modeling and control
system design. University of Illinois, Champaign
4. Jiluan P (2003) Arc welding control. Woodhead Publishing, CRC Press, Cambridge
5. Chaouch S, Hasni M, Boutaghane A, Babes B, Mezaache M, Slimane S, Djenaihi M (2018)
DC-motor control using arduino-uno board for wire-feed system. In: Proceedings of the IEEE,
3rd CISTEM’18—Algiers, Algeria, 29–31 October 2018
6. Truong DQ, Ahn KK (2009) Force control for hydraulic load simulator using self-tuning grey
predictor–Fuzzy PID. Mechatronics 19:233–246
7. Paul AK (2014) Experimental design approach to explore suitability of PI and SMC concepts
for power electronic product development. Int J Power Electron 6(1):42–65
8. Ngo MD, Duy VH, Phuong NT, Kim HK, Kim SB (2007) Development of digital gas metal
arc welding system. J Mater Process Technol 189:384–391
9. Paul AK (2019) Robust PMDC motor control for accurate wire feeding in GMAW using back
EMF. IEEE Trans Industr Electron. https://doi.org/10.1109/TIE.2019.2896131
10. Jiang Z (2009) Design of a nonlinear power system stabilizer using synergetic control theory.
Electric Power Syst Res 79:855–862
11. Jiang Z, Dougal RA (2004) Synergetic control of power converters for pulse current charging of
advanced batteries from a fuel cell power source. IEEE Trans Power Electron 19(4):1140–1150
12. Santi E et al (2003) Synergetic control for DC-DC boost converter: implementation options.
IEEE Trans Ind Appl 39(6):1803–1813. https://doi.org/10.1109/TIA.2003.818967
13. Ni J et al (2014) Variable speed synergetic control for chaotic oscillation in power system.
Nonlinear Dyn 78(1):681–690
14. Wu GD, Richardson RW (1989) The dynamic response of self-regulation of the welding arc. In:
Proceedings of recent trends in welding science & technology, Tennessee, USA, pp 929–933
15. Kolesnikov A, Veselov G (2000) Modern applied control theory: synergetic approach in control
theory. TSURE Press, Moscow-Taganrog
Simulation Study of the Dual Star
Permanent Magnet Synchronous
Machine Using Different Modeling
Approaches
Elyazid Amirouche, Khaldi Lyes, Ghedamsi Kaci, and Aouzellag Djamal
Abstract Four mathematical models of a dual star permanent magnet synchronous

machine (DSPMSM) are presented, one in the natural reference frame, and three
others in a synchronously rotating d-q reference frame. In order to simplify the
machine analysis, a proper transformation is needed to transform the time varying
quantities of the machine into constant variables, many transformations are reported
in the literature. In this paper, three transformations are selected and applied to
a DSPMSM. The models are implemented and simulated using Matlab/Simulink
environment, the differences between the models are discussed. Finally, the limits of
these transformed models are highlighted.
Keywords Dual star permanent magnet synchronous machine · Machine

modeling · Different modeling methods · Limits of simplified d-q models
1 Introduction
Nowadays, power electronic converters are used to supply alternative current

machines, consequently they are decoupled from the grid, so the number of the
machine’s phases is not limited to three anymore, making possible the exploitation
of multiphase machines in domains requiring some specific advantages, such as lower
E. Amirouche (B) · K. Lyes · G. Kaci · A. Djamal

Laboratoire de Maitrise des Energies Renouvelables, Faculté de Technologie, Université de
Bejaia, 06000 Béjaïa, Algerie
K. Lyes
G. Kaci
A. Djamal

https://doi.org/10.1007/978-981-15-6403-1_27
390 E. Amirouche et al.
torque repulsion [1] and fault tolerance [2]. The voltage source inverter fed double
star machine is one of the widely discussed topics in the literature [3, 4].
Because the machine’s variables are not constant in time in steady state, the
use of an adequate transformation in order to obtain constant variables presents
many advantages, constant variables are easier to analyze, and the rotor position
dependency of inductances that characterizes salient pole machines is eliminated,
simplifying the model as well as the control of the machine.
Many modeling technics are proposed in the literature, the double winding
approach represents the machine with two pairs of d-q windings with a mutual
coupling between each other [5], so this model is an extension of the conventional
three phase machine model. In order to eliminate this coupling, a second consecutive
transformation is proposed by [6], called decoupled d-q model, which simplifies the
model analysis and implementation. An extended park transformation is reported
by [7], it proposes a model where the stator circuits are completely decoupled by
applying only one transformation.
This paper focuses on the double star permanent magnet synchronous machine
modeling, many modeling approaches will be presented, simulated and compared,
and the limitation of these models will be presented at the end of this article.
2 Double Star Electrical Machines
Multiple star machines are a particular case of multiphase machines, where the phases
are divided into multiple three phase sets, with isolated neutral points.
In a double star machine, the angle between the sets can have any value from 0°
to 60°, the study of [1] shows that the torque characteristics of a double star with a
displacement angle of 30° between the sets is better than any other value, the torque
ripple is greatly reduced, also the predominant frequency has been shifted 12 times
the supply frequency [1].
The use of multiphase machines instead of conventional three phase machines
presents multiple advantages, in addition to the torque ripple reduction, as stated
above, the machine’s power is divided among more inverter legs, so the power elec-
tronics rating is reduced, also, the reliability of the machine is increased in case of
loss of one or more machine phases [8].
Because of the fault tolerant behaviour of multiphase machines, they are inten-
sively used in applications requiring continuous service even with partial power, for
example, in ship propulsion systems, aerospace applications, electrical vehicles, and
in renewable energy conversion systems [8].
Simulation Study of the Dual Star Permanent Magnet Synchronous … 391
3 Natural Model (ABC) of DSPMSM
Before doing any transformation, the abc model of the machine must be developed,
Fig. 1 shows the coils configuration of a dual star synchronous motor.
In order to simplify the modeling of the machine, the following assumptions
are made: all the sets are identical, and the windings are sinusoidally distributed
around the air gap, mutual leakage inductances, saturation and eddy current are not
considered.
3.1 Electric Model
Starting from the natural reference frame, the electric equation of the system can be
written:
dϕabc
vabc = r i abc + (1)
dt
where
t
vabc = va1 vb1 vc1 va2 vb2 vc2 (2)
t
i abc = i a1 i b1 i c1 i a2 i b2 i c2 (3)
t
ϕabc = ϕa1 ϕb1 ϕc1 ϕa2 ϕb2 ϕc2 (4)
⎡ ⎤
rs · · · 0
⎢ ⎥
r = ⎣ ... . . . ... ⎦ (5)
0 · · · rs
Fig. 1 Dual star C1 C2 d

synchronous machine
A1
A2
q
B2
B1
ϕabc is the resultant air gap flux, r a 6 × 6 matrix with rs the armature resistance.
The resultant air gap flux can be written according to the inductance matrix and
permanent magnet flux vector as follow:

[L 1 ] [M12 ]
ϕabc = i abc + ϕ P Mabc (6)
[M12 ]t [L 2 ]
where [L i ] the inductance matrix of the same three phase set windings, [M12 ] the
mutual inductance matrix between the different sets, and ϕ P Mabc the flux produced
by the permanent magnets.
⎡ ⎤
cos(θ )
⎢ cosθ − 2π ⎥
⎢ 3 ⎥
⎢ ⎥
⎢ cos θ + 2π ⎥
ϕ P Mabc = ψP M ⎢ 3 ⎥ (7)
⎢ cos(θ − α) ⎥
⎢ ⎥
⎣ cos θ − 3 − α ⎦
2π

cos θ + 2π
3
−α
With α the electric displacement angle between the sets, and Ψ pm the magnitude
of the permanent magnet flux.
3.2 Inductance Matrix
Considering the rotor position dependency of the inductances, the self and mutual
inductances are generally expressed by Fourier expansions [8]. Taking into account
higher order harmonics, the self inductance of winding i is written as follow:
∞

L i (θ ) = L s0 + L S2n cos 2n(θ + θi ) + ϕin (8)
n=1
where L s0 is a constant average value, L s2n is the coefficient of inductance harmonics

(2nd, 4th,…), θi is the electric displacement angle of the ith winding magnetic axis
from the reference axis (chosen to be the magnetic axis of coil A1 as shown in Fig. 1),
while ϕin is the displacement angle of the corresponding harmonic component.
The constant term L s0 can be expressed as the sum of the stator leakage inductance
L Sl and the stator magnetizing inductance m [7].
L s0 = L sl + m (9)
Similarly, the mutual inductance between the phases can be expressed as:
∞

Mi j (θ ) = mcos θi − θ j + Ms2n cos n(2θ + θi + θ j + ϕin ) (10)
n=1
Considering only the first inductance harmonic (n = 1), the self and mutual
inductance equations become:
L i (θ ) = L s0 + L s2 cos(2θi ) (11)

Mi j (θ ) = mcos θi − θ j + Ms2 cos 2θ + θi + θ j (12)
Comparing to the general equation given by [7], the coefficients L s2 and Ms2 are
equal.
The full inductance matrix of the windings is given by (13).
⎡ ⎤
L a1 Ma1 b1 Ma1 c1 M a1 a2 Ma1 b2 Ma1 c2
⎢M Mb1 c2 ⎥
⎢ a1 b1 L b1 Mb1 c1 Ma2 b1 Mb1 b2 ⎥
⎢ ⎥
⎢M Mb1 c1 L c1 Ma2 c1 Mc1 b2 Mc1 c2 ⎥
L s (θ ) = ⎢ a1 c1 ⎥ (13)
⎢ M a1 a2 Ma2 b1 Ma2 c1 L a2 Ma2 b2 Ma2 c2 ⎥
⎢ ⎥
⎣ Ma1 b2 Mb1 b2 Mc1 b2 Ma2 b2 L b2 Mb2 c2 ⎦
Ma1 c2 Mb1 c2 Mc1 c2 Ma2 c2 Mb2 c2 L c2
The four matrices presented earlier in (6) can be recognised in this matrix.
Note that every single term of this matrix is θ dependent, as a consequence, the
inverse of this matrix must be calculated at every time step, making this model very
complex, time consuming, and requires a high processing power.
3.3 Electromagnetic Torque
The electromagnetic torque is derived starting from the electromagnetic power

divided by the angular mechanical speed.
Pe = ea1 i a1 + eb1 i b1 + ec1 i c1 + ea2 i a2 + eb2 i b2 + ec2 i c2 (14)
With ei the induced electromotive force of winding i, expressed by:
dϕabc
[e] = − (15)
dt
Or
ei = vi − ri i (16)
Then the electromagnetic torque is given by (17):
Pe
te = (17)
ω
4 Double d-q Winding Model
In order to simplify the analysis of the dual star machine behaviour, the park trans-
formation is applied to the abc system, leading to a couple of d-q winding with a
mutual coupling between theme [9].
The standard park transformation matrix for a three phase system is defined as:
⎡ ⎤
cos(θ ) cos θ − 2π cos θ + 2π
2⎢ 3
2π ⎥
3
T (θ ) = ⎣ −sin(θ ) −sin θ − 3 −sin θ + 3 ⎦
2π
(18)
3 √1 √1 √1
2 2 2
Starting from this, the transformation matrix T1 for a dual star system can be
deduced [9].

T (θ ) 03,3
T1 = (19)
03,3 T (θ − α)
With 03,3 a 3 × 3 null matrix.

The transformation is given by (20).
X dq = T1 X abc (20)
where X can be any phase variable vector (current, voltage or flux).

Applying (20) to the stator flux (21) will result in (22)
ϕabc = L s i abc + ϕ P Mabc (21)

ϕdq = T1 L s T1t i dq + ϕ P Mdq (22)
With
√ √ t
ϕ P Mdq = 2
6
ψ pm 00 2
6
ψ pm 00 (23)
After several lines of trigonometric simplifications, the transformed inductance

matrix can be obtained.
⎡ ⎤
L d1 0 0 Md1 0 0
⎢ 0 0 ⎥
⎢ L q1 0 0 Mq1 ⎥
⎢ ⎥
⎢ 0 0 L 01 0 0 0 ⎥
L T1 = T1 L s T1 = ⎢
t
⎥ (24)
⎢ Md2 0 0 L d2 0 0 ⎥
⎢ ⎥
⎣ 0 Mq2 0 0 L q2 0 ⎦
0 0 0 0 0 L 02
In this matrix, we can see four square 3 × 3 sub-matrices, related to two couples of
orthogonal windings dq1 and dq2 , with a mutual coupling between theme expressed
by Md and Mq .
m 3
L d1 = L d2 = L s0 + + L s2 (25)
2 2
m 3
L q1 = L q2 = L s0 + − L s2 (26)
2 2
L 01 = L 02 = L s0 − m (27)
3 3
Md1 = Md2 = m + L s2 (28)
2 2
3 3
Mq1 = Mq2 = m − L s2 (29)
2 2
As can be seen, the resultant matrix inductance is not diagonal, adding some
complexity to the model simulation and analysis. Also, Eq. 23 shows that the magnet
flux is present in both sub-machines, so the total developed torque is the sum of the
torques developed by each sub-machine. Despite this complexity, it can be used in a
current control scheme such proposed by [10] in which a decoupled voltage equations
are introduced to improve the PI controllers’ performance, but these equations induce
more errors since the current derivatives are ignored.
4.1 The Current Model in the Coupled d-q Reference Frame
The machine electric model can now be obtained by applying the transformation
matrix to the electric Eq. (1) and combining the equations aforementioned, leading
to the electric equation in the dq reference frame (30).
dϕdq
vdq = r i dq + + ωjϕdq (30)
dt
The current model of the machine (31) is obtained by developing (30).

di dq dϕ P Mdq
= L −1
T1 vdq − ωjϕ P Mdq − − r + ωj L i
T1 dq (31)
dt dt
where
⎡ ⎤
0 −1 0 0 0 0
⎢1 0⎥
⎢ 0 0 0 0 ⎥
dT1−1 ⎢ ⎥
1 ⎢0 0 0 0 0 0⎥
j= T1 =⎢ ⎥ (32)
ω dt ⎢0 0 0 0 −1 0⎥
⎢ ⎥
⎣0 0 0 1 0 0⎦
0 0 0 0 0 0
ϕdq = L T1 i dq + ϕ P Mdq (33)
Starting from the instantaneous output power, the electromagnetic power as well as
torque can be derived.
Pout = vd1 i d1 + vq1 i q1 + vd2 i d2 + vq2 i q2 (34)
By replacing the voltage by it’s equation given by (30), ignoring the resistive voltage
drop, and dividing by the angular speed, the electromagnetic torque is obtained:

te = P ϕd1 i q1 − ϕq1 i d1 + ϕd2 i q2 − ϕq2 i d2 (35)
5 Decoupled D-Q Model
The matrix of the further transformation required to eliminate the coupling between
the two d-q windings, is given by T2 , leading to a final decoupled D1 -Q1 and D2 -
Q2 windings, note the introduction of a coefficient √12 to get an invariant power
[6].
⎡ ⎤
1 0 0 1 0 0
⎢ 0 0⎥
⎢ 1 0 0 1 ⎥
1 ⎢
⎢ 0 0 1 0 0 0⎥
⎥
T2 = √ ⎢ ⎥ (36)
2⎢
⎢
0 1 0 0 −1 0⎥
⎥
⎣ −1 0 0 1 0 0⎦
0 0 0 0 0 1
The application of T2 to (24) yields

⎡ ⎤
L D1 0 0 0 0 0
⎢ 0 ⎥
⎢ L Q1 0 0 0 0 ⎥
⎢ ⎥
⎢ 0 0 l0∗ 0 0 0 ⎥
L T2 = T2 L T1 T2 = ⎢
t
⎥ (37)
⎢ 0 0 0 L D2 0 0 ⎥
⎢ ⎥
⎣ 0 0 0 0 L Q2 0 ⎦
0 0 0 0 0 l0∗∗
where
L D1 = L d1 + Md1 (38)
L Q 1 = L q1 + Mq1 (39)
L D2 = L q 1 − M q 1 (40)
L Q 2 = L d1 − Md1 (41)
L 01
L ∗0 = L ∗∗
0 = (42)
2
The final transformation matrix can be found by combining the two transformation
matrices.

T (θ ) π T (θ − α)π
1
T3 = T2 T1 = √ (43)
2 T θ −α+ 2 T θ −α− 2
In this system, the fundamental wave and (12 k ± 1)th harmonics (k = 1, 2, 3,…) in
the original quantities are mapped into D1 -Q1 pair quantities, resulting as constant
terms, and 12 kth harmonics, while the (6(2 k + 1) ± 1)th harmonics (k = 0, 1, 2,…)
are mapped into D2 -Q2 pair quantities, resulting as (6(2k + 1))th harmonics [6].
5.1 The Current Model in the Decoupled D-Q Reference

Frame
Same as in Sect. 4.1, applying the transformation matrix T3 to (1), yields the current
model of the machine in the decoupled d-q reference frame (44).

di D Q dϕ P M D Q
= L −1
T2 v D Q − ωjϕ P M DQ
− − r + ωj L T2
i D Q (44)
dt dt
where
⎡ ⎤
0 −1 0 0 0 0
⎢1 0⎥
⎢ 0 0 0 0 ⎥
dT3−1 ⎢ ⎥
1 ⎢0 0 0 0 0 0⎥
J= T3 =⎢ ⎥ (45)
ω dt ⎢0 0 0 0 −1 0⎥
⎢ ⎥
⎣0 0 0 1 0 0⎦
0 0 0 0 0 0
ϕ P M D Q can be easily deduced by applying the transformation matrix T3 to ϕ P Mabc

√ t
ϕ P M D Q = T3 ϕ P Mabc = 3ψ pm 0 0 0 0 0 (46)
The electromagnetic torque is derived starting from the instantaneous power divided
by the mechanical angular speed , yielding an equation similar to the one derived
for the coupled d-q model.

te = P ϕ D 1 i Q 1 − ϕ Q 1 i D 1 + ϕ D 2 i Q 2 − ϕ Q 2 i D 2 (47)
6 Extended 0dq Model
Another transformation is proposed by [7] where a unique transformation matrix is

used to fully decouple the inductance matrix.
Starting from the following three phase park transformation matrix T (θ ), the full
transformation matrix T p (θ ) for the dual star system can be constructed.
⎡ ⎤
√1 cos(θ ) −sin(θ )
2⎢ 2 ⎥
T (θ ) = √1
⎣ cos θ − 2π −sin θ − 2π3 ⎦ (48)
3 2 3
√1
2
cos θ + 2π 3
−sin θ + 2π3

1 T (θ ) T (θ )
T p (θ ) = √ (49)
2 T (θ − α) −T (θ − α)
Note the use of the same coefficient as in T2 to get an invariant power.

The application of this transformation to the inductance matrix L s (θ ) will result in
the following matrix:
⎡ ⎤
L n0 0 0 0 0 0
⎢ 0 0 ⎥
⎢ L nd 0 0 0 ⎥
⎢ ⎥
⎢ 0 0 L nq 0 0 0 ⎥
L T = Tp L s Tp = ⎢
t
⎥ (50)
⎢ 0 0 0 L a0 0 0 ⎥
⎢ ⎥
⎣ 0 0 0 0 L ad 0 ⎦
0 0 0 0 0 L aq
where
L n 0 = L a0 = L ad = L aq = L s0 − m (51)
L n d = L s0 + 3L s2 + 2m (52)
L nq = L s0 − 3L s2 + 2m (53)
This transformation splits the machine into two decoupled systems, one called
normal system, and the other is called anti system, in the normal system, the current
support each other, creating a normal rotational field in the air gap, and in the anti
system, the currents oppose each other leaving only leakage fluxes in the stator [7].
6.1 The Current Model in the Extended 0dq Reference Frame
As done in Sect. 4.1, the current model of the machine in the extended 0dq reference
frame is given by:

di 0dq dϕ P M0dq
= L −1
T v0dq − ωjϕ P M0dq − − (r + ωj L )i
T 0dq (54)
dt dt
With
⎡ ⎤
00 0 00 0
⎢ ⎥
⎢ 0 0 −1 0 0 0 ⎥
t−1
dT p ⎢ ⎥
1 ⎢0 1 0 0 0 0 ⎥
j= T pt =⎢ ⎥ (55)
ω dt ⎢0 0 0 0 0 0 ⎥
⎢ ⎥
⎣ 0 0 0 0 0 −1 ⎦
00 0 01 0
√ t
ϕ P M0dq = T p ϕ P Mabc = 0 3Ψ pm 0 0 0 0 (56)
6.2 Electromagnetic torque
As done in Sect. 4.2, the electromagnetic torque is given by (57).

te = P ϕnd i nq − ϕnq i nd + ϕad i aq − ϕaq i ad (57)
All the three models have been simulated using Matlab/Simulink environment, the
machine is simulated as a generator driven with constant speed, and feeding a resistive
load. The simulation results are presented and discussed below. The parameters of
the considered machine are shown in Table 1.
The Figs. 2, 3 and 4 present the currents in the transformed reference frames,
the decoupled and extended models are similar, they describe the machine in the
same way, but they are different from the double winding model, which describes
the machine differently.
In the coupled reference frame, the d components (i d1 and i d2 ) are collapsed
together, and have a non null value, also the q (i q1 and i q2 ) and homopolar components,
Table 1 Machine parameters

Description Value
[11]
Stator resistance 0.53
Average self inductance L s0 18.93 mH
Second harmonic coefficient L s2 −3.03 mH
Mutual inductance m 10.53 mH
Permanent magnet flux 1.8328 Wb
Number of pole pairs 8
Rotational speed 350 rpm
Fig. 2 Currents in the

coupled reference frame

decoupled reference frame

extended reference frame
but in the decoupled and extended models, only two components are active, i D1
and i Q 1 for the decoupled model, and i n d and i nq for the extended model, all the
other component are null, meaning that they do not actively affect the machine
comportment.
After applying the invers transformation, the currents appear identical and match
the current of the natural frame model, confirming the validity of the simulated
models. Because the currents are similar, the Fig. 5 is used to show the currents of
the transformed models, and the Fig. 6 represents the current of the natural frame
model (Fig. 7).
The torque is shown in Fig. 8, a zoom into the plot shows an error of about 0.0063%
of the transformed models compared to the natural frame model in the steady state,
this error in negligible and do not influence the comportment of the models.
Fig. 5 The currents of the

transformed models
Fig. 6 The currents of the

natural frame model
Fig. 7 Electromagnetic
torque of the four models
Fig. 8 Electromagnetic
torque of the double winding
model
In the double winding model, each sub machine generates half of the total machine
torque, like shown in Fig. 8, and the total developed torque is the same as for the
other models.
8 Electrical Machine d-q Model Limitations
The transformed model of an electrical machine simplifies not only the analysis of the
machine behaviour, but also the control, for example, in a predictive current control
strategy [12], since it deals with constant parameters, but it presents some limits and
inaccuracies.
Such a model ignores the higher order harmonics in voltage and flux, assuming
pure sinusoidal waveforms, but in faulty conditions, harmonics are present, and the
pure sinusoidal variation is removed, making the model mathematically incorrect
[13].
Also, the simplified model ignores the effect of saturation, magnetic circuit geom-
etry, and assume that all the coils are identical, also, mutual inductances are supposed
to be equal [14]. All these assumptions give inaccuracies to the model.
9 Conclusion
To simplify the analysis and control of double star machine, an adequate mathemat-
ical model must be constructed, for this purpose, many models are reported in the
literature.
The natural frame model gives more details about the machine since it includes
all the voltage harmonics, and can easily take into account saturation effect and
other phenomena, but it’s very complex to solve and requires a high computational
power. Simplified models have been proposed, the double winding model is very
simple to compute, and allows a separate control of the two windings, but the mutual
coupling between the two winding couples decrease the dynamic performance of
the control scheme. The use of decoupled equations in order to minimize the effect
of such coupling would induce more errors because the derivatives of the currents
can not be considered. Also, in this model, all the current components are active, so
the analysis of the machine behaviour becomes more complex comparing to other
simplified models.
To solve the problems encountered with the model aforementioned, a second
transformation is proposed to fully decouple the windings, this new approach is
called decoupled d-q approach, it’s comparable to another model called extended
d-q model, these two models require less processing power. Used in a PI controller
based current control strategy give better results, since the performance of the PI
controllers is higher when the controlled variable is decoupled from other variables
in the system. However, these models don’t permit a separate control of the two
winding sets. The major drawback of the decoupled d-q model is the use of two
consecutive transformations.
Despite of their differences, they describe the machine behaviour in a rotating
reference frame, eliminating the dependency of machine inductances on the rotor
position, the main difference between all these models is the complexity and the
difficulty of implementation.
References
1. Nelson RH, Krause PC (1974) Induction machine analysis for arbitrary displacement between
multiple winding sets. IEEE Trans Power Appar Syst PAS-93(3):841–848
2. Levi E, Bojoi R, Profumo F, Toliyat HA, Williamson S (2007) Multiphase induction motor
drives—a technology status review. IET Electr Power Appl 1(4):489–516
3. Star D, Magnet P, Machine S, Order S, Mode S (2018) Direct torque control using second order
sliding mode of a double star permanent, vol. 80
4. Zaimeddine R, Berkouk EM (2012) A novel DTC scheme of double-star induction motors
using three-level voltage source inverter
5. Barrero F, Arahal MR, Gregor R, Toral S, Durán MJ (2009) A proof of concept study of
predictive current control for VSI-driven asymmetrical dual three-phase AC machines. IEEE
Trans Ind Electron 56(6):1937–1954
6. Kallio S, Andriollo M, Tortella A, Karttunen J (2013) Decoupled d-q model of double-star
interior-permanent-magnet synchronous machines. IEEE Trans Ind Electron 60(6):2486–2494
7. Knudsen H (1995) Extended Park’s transformation for 2 × 3-phase synchronous machine and
converter phasor model with representation of AC harmonics. IEEE Trans Energy Convers
10(1):126–132
8. Cardinale D (2016) Double star Pm machine : analysis and simulations. Politecnico di Milano
9. Fuchs EF, Rosenberg LT (1974) Analysis of an alternator with two displaced stator windings.
IEEE Trans Power Appar Syst PAS-93(6), 1776–1786
10. Karttunen S, Kallio S, Peltoniemi P, Silventoinen P, Pyrhonen O (2012) Dual three-phase perma-
nent magnet synchronous machine supplied by two independent voltage source inverters. In:
International symposium on power electronics power electronics, electrical drives, automation
and motion, pp 741–747
11. Karttunen J, Peltoniemi P, Kallio S, Silventoinen P, Pyrhönen O (2014) Determination of
the inductance parameters for the decoupled d-q model of double-star permanent-magnet
synchronous machines. IET Electr Power Appl 8(2):39–49
12. Zhang Y, Gao S, Xu W (2016) An improved model predictive current control of perma-
nent magnet synchronous motor drives. In: Conference proceedings—IEEE applied power
electronics conference and exposition - APEC, 2016, vol. 2016, pp 2868–2874
13. Mohammed OA, Liu S, Liu Z (2004) Phase-variable model of PM synchronous machines for
integrated motor drives. IEE Proc Sci Meas Technol 151(6):423–429
14. Mohammed OA, Liu S, Liu Z (2005) Physical modeling of PM synchronous motors for
integrated coupling with machine drives. IEEE Trans Magn 41(5):1628–1631
Modeling and Experimental
Identification of Salient-Pole
Synchronous Machine
Khalil-Errahmane Sari and Bilal Sari
Abstract This paper is a complete procedure to determine a valid mathematical

model for the salient-pole synchronous machine. It represents: (i.) a development of
a suitable model for the synchronous machine formed as a lumped-parameter equiva-
lent circuit, and (ii.) an experimental identification of the different machine’s param-
eters. Taking into account the presence of damper windings, an electrical equivalent
circuit of the machine is found in the dq-stator reference frame. In order to accurately
measure the machine’s parameters: X d , X d , X d , X q , Td , and Td , etc., experimental
identification tests including: the open-circuit test, the sustained short-circuit test, the
slip test, and the sudden three-phase short-circuit test were performed. The reduction
factor of the field winding is obtained directly from the short-circuit test and without
any knowledge of design information of the synchronous machine. The accuracy of
the developed model is validated by comparing model simulation results to practical
measurements.
Keywords Salient-pole synchronous machine · Lumped-parameter equivalent

circuit · Experimental identification tests · Sudden three-phase short-circuit test
1 Introduction
Synchronous machines are very important electro-mechanical energy-conversion

devices. They are widely used in both: electricity production as generator units, and
in motor applications. In fact, most power plants rely on synchronous machines to
provide power to the grid. For that, modeling of the machine is considered essential
for power systems simulation and analysis, and control purposes [1, 2].
The appropriate representation of a synchronous machine is subjected to the phys-
ical construction of its rotor, the availability of data, and computational considera-
tions [3]. The author in [3] also stated that: “A complete synchronous machine model
K.-E. Sari (B) · B. Sari

DAC HR Laboratory, Department of Electrical Engineering, Faculty of Technology, University
Ferhat Abbas Sétif 1, Sétif 19000, Algeria

https://doi.org/10.1007/978-981-15-6403-1_28
408 K.-E. Sari and B. Sari
consists of the combination of a model structure and a set of parameter values.”

This definition can be broken into two interconnected distinct parts: (i) the construc-
tion of the model structure, which is the basic form for machine representation,
and (ii) the evaluation of the model parameters. The model structure can be formed
as a lumped-parameter equivalent circuit, differential equation representation, etc.,
whereas, the model parameters can be obtained from manufacturers’ data, experi-
mental test results, or based on other analytical techniques [3]. Synchronous machines
are usually acknowledged to be accurately modeled by a lumped-parameter equiva-
lent circuit representing the d-axis and the q-axis [1], where the stator and the rotor
windings are included without using the same magnetic circuit. In this work, we
consider a more general method in which the same magnetic circuit is used. This
is done by referring rotor quantities to the stator reference frame. An extra work,
however, is in the determination of the reduction factors [2]. To study the machine’s
behavior during transient and steady-state conditions, accurate knowledge of the
machine’s equivalent circuit parameters is required [4]. IEC standard 34-4-1 [5]
and IEEE standard 115 [6] include the laboratory tests procedures for the practical
determination of these parameters.
The present work is carried out in five main sections. After the introduction given in
Sect. 1, the rest is as follows: Sect. 2 covers the modeling of the synchronous machine.
Starting from the classical machine equations, a mathematical model is developed
in the stator reference frame and an electrical equivalent circuit of the machine is
found. Three reduction factors: K f , K D and K Q are used in reducing rotor quantities
to the stator. Section 3 describes the experimental determination of the machine’s
parameters. The laboratory procedures of the open-circuit test, the sustained short-
circuit test, the slip test, and the sudden three-phase short-circuit test are explained
and analyzed, whereby, the values of the required parameters are extracted. Section 4
sums the relationships between the developed model parameters and the parameters
extracted experimentally. Based on these relationships, it is found that K D and K Q
do not affect the terminal behavior of the machine, and therefore, finding their exact
values is outside the scope of this paper. The value of K f , on the other hand, is of
great importance and is found directly from the short-circuit test. Section 5 represents
the validation of the deduced model under MATLAB and Simulink. A comparison
between practical measurements and the corresponding simulation results is done.
This work finally ends with some concluding remarks.
2 Synchronous Machine Modeling
The schematic of a three-phase synchronous machine conventional model structure

is illustrated in Fig. 1. The machine consists of two essential elements: armature
windings and excitation, or field winding. The magnetic axis of the machine is
defined as the direct- or d-axis, and an orthogonal quadratic- or q-axis is located 90
electrical degrees ahead of the d-axis [7]. The machine is modeled with one damper
winding for the d-axis and one damper winding for the q-axis [2, 7].
Modeling and Experimental Identification of Salient-Pole … 409
Fig. 1 Synchronous
machine conventional
structure, with one damper
winding for the d- axis and
one damper winding for the
q-axis
2.1 Mathematical Formulation and dq-Model
The model is to be formed as a set of equations that completely describes the electrical
behavior of the synchronous machine. Before proceeding to develop a suitable model
for the machine, some assumptions are made [7, 8]:
• Magnetic saturation and hysteresis are neglected.
• Stator currents are assumed to set up magneto-motive force sinusoidally
distributed in space around the air-gap. Therefore, the effect of space harmonics
is neglected.
Applying Maxwell’s equation to the configuration shown in Fig. 1, the phase-
voltage equations in the natural-reference frame are simply [7, 8]:
dψa
= i a Rs + va
dt
dψb
= i b Rs + vb
dt
dψc
= i c Rs + vc
dt
dψ f
− = if Rf − vf
dt
dψ D
− = i D RD
dt
dψ Q
− = i Q RQ , (1)
dt
where:
• vabc , iabc : The stator voltages and the stator currents.
• vf , if : The field voltage and the field current.

• iD , iQ : The dampers currents.
• ψ abc : The stator total fluxes.
• ψ f : The field total flux.
• ψ D , ψ Q : The dampers total fluxes.
The above equations completely describe the electrical behavior of the
synchronous machine. These equations, however, contain parameters which vary
with θ, the rotor position, which varies with time [8].
Further development of the machine’s model can be done using Park’s transfor-
mation to mathematically transform the three-phase time-varying stator quantities
into time-invariant d- and q-axes quantities under steady-state conditions [1]. The
rotor windings: f , D, and Q are already aligned along with the d- and q-axes. That is,
only stator equations that need to be transformed. The transformation from abc- to
dq-variables can be done by applying the transformation matrix given by (2) [7, 8]:

2 cos(−θ ) cos(−θ + 23 π ) cos(−θ − 23 π )
P(θ ) = . (2)
3 sin(−θ ) sin(−θ + 23 π ) sin(−θ − 23 π )
Therefore:
vdq = P(θ )vabc , i dq = P(θ )i abc and, ψdq = P(θ )ψabc .
The analysis of the machine’s model in terms of dq-variables is considerably

simpler for the following reasons [8]:
• For balanced steady-state operations, the stator quantities have constant values.
• For balanced conditions, the zero sequence quantities disappear.
• The parameters associated with d- and q-axes may be directly measured from
terminal tests.
The inverse transformation is [7, 8]:
3
[P(θ )]−1 = [P(θ )]T . (3)
2
Applying the dq-transformation, the following equations in terms of transformed
components of voltages, currents, and flux linkages result [7, 8]:
dψd
= i d Rs + vd + ωr ψq
dt
dψq
= i q Rs + vq − ωr ψd
dt
r
dψ f
− = i rf R rf − vrf
dt
dψ Dr
− = i rD R rD
dt
dψ Qr
− = i rQ R rQ , (4)
dt
with:
ψd = −L d i d + M f i rf + M D i rD
ψq = −L q i q + M Q i rQ
3 3
ψ rf = (L rf l + L f m )i rf − M f i d + M f D i rD
2 2
3 3
ψ Dr = (L rDl + L Dm )i D − M D i d + M f D i rf
r
2 2
3
ψ Qr = (L rQl + L Qm )i Q − M Q i q ,
r
(5)
2
and:
dθ
= ωr . (6)
dt
2.2 dq-Model in the Stator Reference Frame
To simulate the machine’s model presented by (4) (5), the different mutual induc-
tances: M f , M D , M Q , M fD , etc., need to be evaluated. By reducing the rotor variables
to the stator, the number of inductances reduces, and the electrical equivalent circuit
of the machine is found [2, 7].
Rewriting the machine’s model equations referring all variables to the stator, the
following result [7]:
dψd
= i d Rs + vd + ωr ψq
dt
dψq
= i q Rs + vq − ωr ψd
dt
dψ f
− = if Rf − vf
dt
dψ D
− = i D RD
dt
dψ Q
− = i Q RQ . (7)
dt
The flux-current relations are now [7]:
ψd = −L sl i d + L dm (−i d + i D + i f )
ψq = −L sl i q + L qm (−i q + i Q )
ψ f = L f l i f + L dm (−i d + i D + i f )
ψ D = L Dl i D + L dm (−i d + i D + i f )
ψ Q = L Ql i Q + L qm (−i q + i Q ), (8)
where:
• L dm , L qm : The direct and the transverse stator main leakage inductances.
• L sl : The stator leakage inductance.
• L fl : The field leakage inductance, referred to the stator.
• L Dl , L Ql : The dampers leakage inductances, referred to the stator.
Comparing (5) to (8), the following definitions of current reduction factors are
considered valid [7]:
i f = i rf K f , i D = i rD K D and, i Q = i rQ K Q , (9)
with:
Mf
Kf = : The reduction factor of the field winding
L dm
MD
KD = : The reduction factor of the d - axis damper winding
L dm
MQ
KQ = : The reduction factor of the q - axis damper winding.
L qm
Using the reduction factors, and from (5), the following is obtained [7]:
2 L dm
ψ rf = ψ f = L f l i f + L dm (−i d + i D + i f )
3 Mf
2 L dm
ψ Dr = ψ D = L Dl i D + L dm (−i d + i D + i f )
3 MD
2 L qm
ψ Qr = ψ Q = L Ql i Q + L qm (−i q + i Q ). (10)
3 MQ
From (10), it follows that:
2
L f l = L rf l
3K 2f
2
L Dl = L rDl
3K D2
Fig. 2 Electrical equivalent circuit, with rotor quantities referred to the stator
2
L Ql = L rQl . (11)
3K Q2
Reducing the rotor circuit resistances and the field voltage to the stator is done by
power equivalence [7], this yields:
2
R f = R rf
3K 2f
2
R D = R rD
3K D2
2
R Q = R rQ , (12)
3K Q2
and:
2
v f = vrf . (13)
3K f
Using the above equations, the electrical equivalent circuit of the machine with
rotor quantities referred to the stator is deduced [7, 2] (See Fig. 2).
3 Experimental Determination of the Machine’s

Equivalent Circuit Parameters
In order to obtain the synchronous machine’s electrical equivalent circuit parameters,

different experimental identification tests need to be performed [4]. A LabVolt salient-
pole synchronous machine of 1.5 KVA with damper windings is used in this work.
Figure 3 shows the experimental test bench.
Fig. 3 Experimental test bench
3.1 Open-Circuit Test
The open-circuit test is performed on the synchronous machine at no-load to obtain

the no-load saturation curve, or the open-circuit characteristic [4]. The synchronous
machine being tested is firstly driven, using a prime mover, at its rated speed with
no excitation being applied. The field winding current is then supplied and gradually
increased while measuring the terminal voltage until reaching the saturation condition
[4]. Figure 4 shows the obtained I-V curve.
Fig. 4 Open-circuit
characteristic
3.2 Sustained Short-Circuit Test
The objective of sustained short-circuit test is to obtain the short-circuit charac-

teristic of the machine. This last is used with the previously obtained open-circuit
characteristic curve to determine the synchronous reactance X d [4].
During the test, the machine is rotating at its rated speed with its stator termi-
nals permanently short-circuited. The excitation current is supplied and gradually
increased while measuring the short-circuit current flowing in armature windings.
The short-circuit current is plotted versus the field winding current to obtain the short-
circuit characteristic curve (See Fig. 5). While performing this test, it is important to
keep track of the stator current, it might exceed the rated value by only a small toler-
ance to prevent stator windings damage. The d-axis unsaturated impedance Z d(unsat)
can be obtained as a quotient of the voltage on the open-circuit characteristic curve
at a field current point, and the short-circuit armature current on the short-circuit
characteristic curve corresponding to the same field current point. That is [4]:
VL−L
Z d(unsat) = √ . (14)
3Isc
The d-axis unsaturated synchronous reactance is therefore:

X d(unsat) = 2
Z d(unsat) − Rs2 . (15)
where:
• Rs : The stator resistance, and it is measured directly.
Fig. 5 Short-circuit
characteristic
3.3 Slip Test
The aim of the slip test is to obtain the saliency ratio. Using this ratio and the value
of X d obtained previously, X q can be determined [4].
During the test, the synchronous machine is driven at a speed slightly different
from the rated speed of about 1% to achieve a very small slip. With no excitation, a
balanced three-phase voltage of about 25% of the rated value is applied across the
machine’s terminals. Figure 6 shows a recording of the armature current and voltage.
Clearly, both quantities oscillate, having maximum and minimum values.
It is remarkable also that when the armature voltage is of maximum amplitude,
the armature current is minimum value. This is, in fact, the time when the d-axis is in
line with the armature mmf, and thereby the offered air-gap reactance represents the
d-axis synchronous reactance. Similarly, when the voltage is of minimum amplitude
and the current is maximum value, the q-axis is aligned with the armature mmf, and
the presented air-gap reactance is the q-axis synchronous reactance [4].
Using the obtained current and voltage recordings, the following calculations can
be made:
Vt (L−L)max
X d(sli p) = √
3Imin
Vt (L−L)min
X q(sli p) = √ . (16)
3Imax
The saliency ration is:

Xq X q(si p)
= . (17)
Xd sli p X d(sli p)
Finally, X q is obtained:
Fig. 6 Real current and

voltage waveforms from slip
test

Xq
Xq = Xd . (18)
Xd sli p
3.4 Sudden Three-Phase Short-Circuit Test
Many transient and subtransient parameters can be extracted from a suitable oscillo-
gram of a short-circuit current recorded during the sudden three-phase short-circuit
test [5, 6]. The interest here is to calculate the d-axis transient and subtransient
synchronous reactances, X d and X d respectively. The values of the q-axis transient
and subtransient synchronous reactances are beyond the scope of this paper.
The synchronous machine is firstly driven at its rated speed, and is excited so that
its terminal voltage is equal to the rated value at no-load condition. A sudden three-
phase short-circuit is then applied to the machine’s terminals. The short-circuit arma-
ture currents are recorded [5, 6]. Figure 7 shows the short-circuit current waveform
of phase a. The current waveform consists of: an ac component and a dc component.
The ac component can be resolved into three different components: subtransient
I , transient I and steady-state I s components [4]. Also, it is clear that the current
envelope obeys an exponential decay. The time constant controlling the rapid decay
during the subtransient period is called the d-axis subtransient time constant Td ,
whereas, the time constant controlling the slower decay during the transient period
is called the d-axis transient time constant Td [5, 6].
The rms amplitude of the total ac component of the short-circuit current in one
phase can be expressed as [4]:
Fig. 7 Phase-a real

short-circuit current
Fig. 8 Typical polynomial

curve fitting of a three-phase
short-circuit current envelope
−t −t
Isc (t) = I − I e Td + I − Is e Td + Is (19)
or,
−t −t
1 1 Td
1 1 Td
1
Isc (t) = E 0 − e + − e + . (20)
X d X d X d Xd Xd
By extrapolating the current envelope back to the zero time, or the time of the
short-circuit occurrence (See Fig. 8), the transient and subtransient time constants
and currents, Td , Td , I and I respectively, can be evaluated [4, 5]. Finally, the
values of X d and X d result in [4]:
E0
X d =
I
E0
X d = . (21)
I
where:
• E 0 : The rms line-to-neutral open-circuit prefault terminal voltage.
3.5 Results from Experimental Identification Tests
The previously discussed experimental identification tests were realized for several
tries. For each try, data were recorded, and the mean value was taken. This is done in
order to accurately identify the different machine’s parameters. Table 1 summarizes
the results from experimental identification tests, as well as, the nominal values of
the synchronous machine used in this work.
Table 1 Parmeters of the Labvolt Ems 8557-1 Salient-Pole Synchronous Machine

Measured parameters

X d () X q () X ’d () X d () Rs ()
75.443 46.556 10.309 8.5298 2.2

Rrf () T ’d (s) T d (s) T ’d0 (s)a
127 0.0776 0.0147 0.235
Nominal values
Power S n Voltage V n Current I n fn tr/min
1.5KVA 220 V 4A 50 Hz 1500
a Measured having the synchronous machine under a short-circuit then a sudden open-circuit is
made [2]
4 Mathematical Relationships Between Model Parameters

and Measured Parameters
The machine’s model presented by (7) (8) requires the determination of leakage
inductances. The parameters of Table 1, however, are in terms of reactances and
time constants. For that, mathematical relationships between these parameters and
leakage inductances need to be found.
According to IEC standard [5], these relationships can be written as follows:
⎧
⎪
⎪ X = Lωr
⎪
⎪
⎪
⎪ X d = X sl + X dm
⎪
⎪
⎪
⎪ X q = X sl + X qm
⎪
⎪
⎪
⎪ X = X sl
X X
+ X dmdm+Xf lf l
⎪
⎪ d
⎪
⎪
⎪
⎪ X d = X sl +
X Dl X f l
⎪
⎪ X Dl +X f l
⎪
⎪ X X
⎨ X q = X sl
⎪ + X qmqm+XQlQl

Td = ωr1R f X f l + XXdmdm+X X sl (22)
⎪
⎪ sl
⎪
⎪
= ωr1R f X dm + X f l
⎪
⎪ Td0

⎪
⎪
⎪
⎪ Td = ωr 1R D X Dl + X f lf +X
X l X sl
⎪
⎪ sl
⎪
⎪
⎪
⎪ T
= 1
X +
X f l X dm
⎪
⎪ d0 ωr R D Dl X f l +X dm
⎪
⎪
⎪ T = 1 X Ql + X qm X sl
⎪
⎪
⎪ q ωr R Q X qm +X sl
⎩ T = 1 X + X
⎪
q0 ωr R Q qm Ql
Based on the above expressions, and after some mathematical manipulation,

relationships between model parameters and measured parameters are deduced [2,
9]:
⎧
⎪
⎪ X = Td0
ωr R f X d − X d
⎪
⎪
dm
⎪
⎪
⎪ X sl = X d − X dm
⎪
⎪
⎪
⎪ X qm = X q − X sl
⎪
⎪
⎪
⎪
⎪ X f l = Td0 ωrR − X dm
⎪
⎪
⎪ X Dl = X f l (+X
⎪
⎨
X X d −X sl )

fl sl −X d
X qm ( X q −X sl ) (23)
⎪ X Ql = X qm +X sl −X
⎪
⎪
⎪
q
⎪
⎪ R D = ωr 1T X Dl + X f lf +X
X l X dm
⎪
⎪ d0
⎪
⎪ dm
⎪
⎪ R = 1
X + X
⎪
⎪
Q ωr Tq0 qm Ql
⎪
⎪
⎪
⎪ T
= T X d
⎪
⎪
d0 d X d
⎩ T = T X q
q0 q X q
Using the relationships of (23), the model parameters can now be computed.
Notice that the value of X dm requires the determination of Rf , the rotor resistance
referred to the stator. This implies the determination of the field winding reduction
factor K f .
4.1 Determination of the Reduction Factors
The reduction factors used to reduce rotor quantities to the stator may be calculated
through analytical or numerical methods, and may also be measured [2, 7].
(1) The dampers reduction factors: The dampers’ parameters of (23) are computed
without using the dampers reduction factors K D and K Q . In fact, K D and K Q are
essentially needed to find the dampers currents I D and I Q [2]. Practically, these
currents are not measured, and since it is possible to simulate the machine’s
electrical behavior without using these factors, finding their exact values is
outside the scope of this work.
(2) The field winding reduction factor: The value of the field winding reduction
factor K f can be extracted directly from the short-circuit test [2]. This is done
by simply applying:
If
Kf = √ . (24)
3Isc
The obtained values of K f at different values of field winding currents are

illustrated in Fig. 9. By taking the mean value, K f is finally found [2].
Fig. 9 Reduction factor of

the field winding, K f
5 Validation of the Developed Model
Having the synchronous machine’s parameters of Table 1, the mathematical relation-

ships of (23), and the value of the field winding reduction factor K f , the simulation
of the machine’s electrical behavior is now possible.
The system of equations given by (7) (8) can be written as:
di f
vd = −Rs i d + L sl + L qm ωr i q − L qm ωr i Q + L dm
dt
di d di D
− (L sl + L dm ) + L dm
dt dt
vq = −Rs i q − (L sl + L dm )ωr i d − L dm ωr i f + L dm ωr i D
di q di Q
− L sl + L qm + L dm
dt dt
di d di D
v f = −R f i f + L f l + L dm ωr i d − L dm + L dm
dt dt
di D di d di f
0 = −R D i D + (L Dl + L dm ) + L dm − L dm
dt dt dt
di Q di q
0 = −R Q i Q + L Ql + L qm − L qm . (25)
dt dt
In matrix form:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
vd id id
⎢v ⎥ ⎢i ⎥ ⎢i ⎥
⎢ q ⎥ ⎢ q ⎥ d⎢ q ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢vf ⎥ = Z ⎢ i f ⎥ + L ⎢ i f ⎥, (26)
⎢ ⎥ ⎢ ⎥ dt ⎢ ⎥
⎣0 ⎦ ⎣ iD ⎦ ⎣ iD ⎦
0 iQ iQ
with:
⎡ ⎤
−Rs L sl + L qm ωr 0 0 −L qm ωr
⎢ −(L + L )ω −Rs L dm ωr L dm ωr ⎥
⎢ sl dm r 0 ⎥
⎢ ⎥
Z =⎢ 0 0 Rf 0 0 ⎥
⎢ ⎥
⎣ 0 0 0 RD 0 ⎦
0 0 0 0 RQ
⎡ ⎤
−(L sl + L dm ) 0 L dm L dm 0
⎢ − L sl + L qm ⎥
⎢ 0 0 0 L qm ⎥
⎢ ⎥
L=⎢ −L dm 0 L f l + L dm L dm 0 ⎥, (27)
⎢ ⎥
⎣ −L dm 0 L dm L Dl + L dm 0 ⎦
0 −L qm 0 0 L Ql + L dm
in state-space representation:
Ẋ = AX + BU
Y = C X, (28)
with:
T
A = −L −1 Z , B = L −1 , X = id iq i f i D i Q and,
T
U = vd vq v f 0 0 .
Validation of the synchronous machine’s model given by (25) is done using

MATLAB and Simulink software. To check the fitness of the mathematical model
developed and the accuracy of the identified parameters, a sudden three-phase short-
circuit test is performed under Simulink. With the help of the Simscape library, a
three-phase short-circuit is applied at 55% rated terminal voltage. The armature
currents, as well as, the field winding current obtained by experimental tests and
those obtained by simulation are plotted on the same graphs (please refer to Figs. 10
and 11). Clearly, the resulting currents have approximately the same waveforms. At
the moment of the sudden three-phase short-circuit, and in the subtransient-state, it
is noticeable that the real currents decrease more rapidly than the model currents.
This is because the measured subtransient time constant Td is defined long. In the
transient-state, the real and model currents tend to converge. However, in the steady-
state, a very good agreement between real and model currents is achieved. This is
also the case during the prefault period.
The synchronous machine’s model developed closely reflects the actual machine’s
electrical behavior. Hence, successful validation of the model is achieved. The pres-
ence of some discrepancies is mainly due to the neglected effects of saturation,
hysteresis, and harmonics of the main field mmf, as well as measurements and
rounding errors.
Fig. 10 Real and model armature short-circuit currents

Fig. 11 Real and model field winding short-circuit current
6 Conclusion
A complete procedure to determine a valid model for the salient-pole synchronous

machine is presented in this work. Taking into account the presence of damper
windings, an electrical equivalent circuit of the machine is developed in the dq-stator
reference frame. Experimental identification tests are performed to measure the exact
constants of the machine’s electrical equivalent circuit, as well as, the field winding
reduction factor. Validation is done using MATLAB and Simulink software where a
sudden three-phase short-circuit test is performed on the simulation model developed.
The machine states before and after the sudden short-circuit are presented. The
simulation current graphs show a very good agreement to those measured practically.
The small difference is essentially due to modeling assumptions and measurement
errors.
References
1. Zhang Y (2018) Advanced synchronous machine modeling. Diss, University of Kentucky

2. Barakat A, Tnani S, Champenois G, Mouni E (2010) Analysis of synchronous machine modeling
for simulation and industrial applications. Simul Model Pract Theory 18(9):1382–1396
3. IEEE Guide (2003) Synchronous generator modeling practices and applications in power system
stability analyses. In: IEEE Std 1110–2002 (Revision of IEEE Std 1110-1991), pp 1–80, 12 Nov
2003. https://doi.org/10.1109/IEEESTD.2003.94408
4. Ghanim DFD (2012) Experimental determination of equivalent circuit parameters for a laborator
salient-pole synchronous generator. Diss. Memorial University of Newfoundland
5. IEC 60034-4-1 (2018) Rotating electrical machines - Part 4-1: methods for determining
electrically excited synchronous machine quantities from tests
6. IEEE Guide (2010) Test Procedures for Synchronous Machines Part I—Acceptance and Perfor-
mance Testing Part II—Test Procedures and Parameter Determination for Dynamic Analysis.
In: IEEE Std 115-2009 (Revision of IEEE Std 115-1995), pp 1–219, 7 May 2010. https://doi.
org/10.1109/IEEESTD.2010.5464495
7. Boldea I (2018) Synchronous generators. CRC Press
8. Kundur P (1994) Power system stability and control. McGraw-hill, New York
9. Kirtley Jr, J (2013) 6.685 Electric Machines. Fall 2013. Massachusetts Institute of Technology:
MIT OpenCourseWare, https://ocw.mit.edu. License: Creative Commons BY-NC-SA
Power Quality Improvement of PWM
Rectifier-Inverter System Using Model
Predictive Control for an AC Electric
Drive Application
Abdelkarim Ammar
Abstract The AC–DC–AC conversion system is extensively used in a variety of

applications, such as power system, industry applications, and transportation systems.
The back-to-back converter with adjustable DC-link provides a bi-directional power
flow between the load and the grid. The grid-side-converter (GSC) can control inde-
pendently active and reactive power. Moreover, it functions as an active power filter
to compensate harmonics and provide a sinusoidal grid current. The Finite state
model predictive control (FS-MPC) has been widely used in power electronics and
electrical drives field due to its promising results. This technique incorporates the
power converter model in the control design and does not need any modulation
unit. In this work, the predictive model control will be applied to both back-to-back
converter stages. Predictive power control and predictive torque control are applied
respectively on the rectifier and the inverter which fed the motor. MPC evaluates
the control variables in a cost-function to generate optimal switching states in a
sampling period. This optimization reflects on power quality and harmonics elimi-
nation. The effectiveness of the presented algorithms is investigated by simulation
using MATLAB/Simulink software.
Keywords Back-to-back converter · PWM-rectifier · Induction motor · Direct

power control · Model predictive control
1 Introduction
The AC-DC-AC conversion system in electrical drives application uses a full-diode

bridge for the grid side rectifying where the power transfer cannot be reversible. In
addition, the diode-bridge causes high grid current distortion and thereby it reduces
A. Ammar (B)
Signals and Systems Laboratory (LSS), Institute of Electrical and Electronic Engineering,
University of M’hamed BOUGARA of Boumerdes, Boumerdes, Algeria
LGEB Laboratory, Electrical Engineering Department, Biskra University, Biskra, Algeria

https://doi.org/10.1007/978-981-15-6403-1_29
428 A. Ammar
the power factor and the quality of the energy [1]. Recently, and due to breakthrough
advancement in the power devices developments, the application of three-phase pulse
width modulation (PWM) rectifier has been risen quickly [2]. The back-to-back
converter is consisting of a PWM-rectifier (AC-DC) and an inverter a (DC-AC)
with an intermediate DC-link. This converter is widely used in many applications
such as electrical drives and renewable energy conversion systems. It offers plenty
of advantages, such as: a bi-directional energy flow ability, a controllable DC-link
voltage and reactive power and four-quadrant operation [3].
Several strategies have been proposed either for the rectifier power control or for
the inverter. They can be classified into two groups, indirect methods with cascade
current control like Field/Voltage Oriented Control (i.e. FOC and VOC) [4, 5] or
direct methods like the Direct Power/ Torque Control (i.e. DTC and DPC) [6, 7],
which directly controls the input active and reactive power. Despite that the Direct
power/torque control methods show high performance over the indirect methods
such as the fast dynamic and the eliminating of the current loops, they present some
drawbacks. The high power/torque ripples and variable switching frequency were the
two most notable disadvantages [8]. Therefore, another method has been appeared
and attracted wide attention in the area of power electronics which is known by the
Model Predictive Control (MPC) [9].
Nowadays, MPC has become very popular research topic. It can be classified into
two main categories, named by the continuous model predictive control and finite-
state model predictive control (FS-MPC) [10]. The continuous MPC like the classical
generalized predictive control (GPC) can demonstrate good performance. However, it
has high order of complexity and needs a modulator in control design [11]. Contrari-
wise, the finite-state MPC, known as finite-control-set in other reference, (FCS-
MPC), eliminates the use of modulation block, it incorporates the converter model in
the control design and respects its discrete nature [9]. The converter switching states
are considered in order to minimize a predefined cost function which consists of the
errors between references and predicted measured control variables.
Due to its high flexibility, MPC has been well applied in power electronics and
drives field where MPC can suit the different power converters structures. Among
them, there are predictive control strategies for active front-end-rectifier like predic-
tive power control (PPC) [12] and predictive current control (PCC) [9]. Others for
inverters that feed motors like the predictive torque control (PTC) [13]. In particular,
PPC and PTC are featured by stationary reference frame unlike PCC. Moreover,
they keep a similar structure of traditional DPC or DTC, where the switching table is
replaced by the online optimization procedure [14]. In finite-state MPC, the control
variables such as active and reactive power in case of PPC or torque and flux in case
of PTC are predicted for the finite number of possible switching states of a power
converter. The optimal selection of switching states which is obtained by actuating
the cost function can reduce power and torque ripples. Then, the selected state is
directly applied to the converter in the next sampling time [15].
This paper presents a model predictive control strategy applied to the back-to-
back fed induction motor. This topology has been proposed for the sake of power
quality improvement and performance enhancement for the induction motor and the
Power Quality Improvement of PWM Rectifier-Inverter System … 429
grid-tied converter. A comparative study has been done between the MPC controlled
rectifier-inverter system and other traditional structures that don’t employ grid-side
control or employ a traditional DPC strategy. The simulation results have obtained
using been MATLAB/Simulink in order to examine the effectiveness of the presented
algorithm.
2 System Modeling and Configuration
The main circuit’s topology of two-level back-to-back PWM converter is shown in

Fig. 1. it can by divided to a rectifier with RL-filter in the grid side, DC-link and an
inverter which drive the induction motor in the motor side.
2.1 Active-Front-End Rectifier Model
The three PWM rectifier is a full controlled bridge consisting of six power transistors
(IGBTs). The grid delivers three-phase voltages vg using the L-filter inductances and
resistances.
Transforming the three-phase mathematical model into two-phase stationary αβ
frame, the model of PWM rectifier can be expressed as:
di g
eg = Ri g + L g + vr ec (1)
dt
where eg , ig and vrec , present, grid voltage and current vectors and rectifier voltage
vector.
with
2
eg = ega + egb e j(2π/3) + egc e j(4π/3) (2)
3
2
ig = ega + egb e j(2π/3) + egc e j(4π/3) (3)
3
Fig. 1 Circuit’s topology of

the PWM Rectifier-Inverter
system driving and induction
motor
430 A. Ammar
2
vr ec = Vdc Sa + Sb e j(2π/3) + Sc e j(4π/3) (4)
3
V dc : is the DC-link voltage.
The rectifier control bases on the logic values S i , where:
S i = 1, Ti is ON and Ti is OFF.
S i = 0, Ti is OFF and Ti is ON.
with: i = a, b, c.
The DC-link dynamics of the back-to-back converter can be modeled as follows
Switching states:
d Vdc (t) 1 1
= Idc (t) = Ig (t) − Im (t) (5)
dt C C
where I g and I m represent the DC-link components of the grid and load side (machine)
currents.
There are eight possible positions from the combinations of switching states. Six
are active vectors (V1, V2 … V6) and two are zero vectors (V0, V7).
2.2 Voltage Source Inverter Model
In this work, a two-level voltage source inverter (VSI) fed the controlled IM. The
voltage vector is generated by the following equation:
2
vinv = Vdc Sa + Sb e j(2π/3) + Sc e j(4π/3) (6)
3
2.3 Induction Machine Model
The dynamic model of the induction machine can be employed in the stationary
frame (i.e. stator-fixed reference frame) (α, β) in complex matrix from by:
1
d ī s ī
= A s + σ L s v̄s (7)
dt ψ̄s ψ̄s 1
where:

− σ1τs + 1
σ τr
+ jω 1
σ Ls
1
τr
− jω
A=
−Rs 0
M2
τs = LRss ; τr = LRrr ; σ = 1 − L s Lsrr
v̄s is the stator voltage vector;
ψ̄s is the stator flux vector.
ī s is the current.
Rs and Rr are the stator and rotor resistances.
L s , L r and M sr are stator, rotor and mutual stator-rotor inductance respectively.
ω is the electrical velocity
The electromagnetic torque expressed in terms of stator current and flux is given
by:

Te = p · Im ψ̄s · ī s (8)
p is the number of pole pairs.
3 Conventional Direct Power and Torque Contral
The direct torque control uses a switching table to select an appropriate voltage
vector. This selection is related directly to the variation of the stator flux and the
torque and the position of the flux vector. The relationship between the stator voltage
and the stator flux change can be established as:
ψs = Vs Tz (9)
From (9), it can be deduced that the flux can be varied by the application of stator
voltage during a sampling period T z . Then, the torque of the induction motor can be
expressed in terms of stator and rotor flux vectors, the rotor flux vector is supposed
invariant.
Msr
|Te | = p |ψs ||ψr | sin(δ) (10)
σ Ls Lr
where:
δ load angle between the stator and rotor flux vectors.
According to (10), the torque can be controlled by adjusting the load angle δ if the
stator and rotor flux amplitudes are maintained constant. The hysteresis comparators
are used for flux and torque control.
The twelve sector switching divides the circular flux locus into 12 regions instead
of 6. This table makes all the six states used per sector to solves the ambiguity in
torque control. Therefore, it is necessary to define small and large torque variations
which require to divide the hysteresis band of the torque into four parts instead of
three. Then, all six voltage vectors can be used.
432 A. Ammar
Table 1 12-sectors switching

S1 S12
table
Flux Increase V1, V2 and V6 V1, V2 and V6
Decrease V3, V4 and V5 V3, V4 and V5
Torque Increase V2, V3 and V4 V1, V2 and V3
Decrease V5, V6 and V1 V4, V5 and V6
Table 1 shows the voltage vectors of the first and the 12th sectors, the rest sectors
can be deduced.
The control scheme of basic direct torque control strategy is shown in Fig. 1.
In similar way, the basic principle of direct power control for PWM rectifiers is
similar to direct torque control (DTC) in motor drives. It selects the voltage vector
using a switching table, according to the grid voltage position [8]. The used regulators
are hysteresis comparators for instantaneous errors of active and reactive power [16]
(Fig. 2).
In DPC control design, six sectors switching table has been chosen (Table 2).
k indicates to the sector number.
Fig. 2 Block diagram of basic direct torque control algorithm
Table 2 6-sectors switching

Sector
table for DPC strategy
Active power Increase V k+3 , V k-1
Decrease V k+1 , V k
Reactive power Increase V k+3 , V k+1
Decrease V k-1 , V k
4 Model Predictive Control Strategies
4.1 Grid Side Converter (GSC) Control
To predict currents into the next step the Euler forward discretization equation is
used:
dx x(k + 1) − x(k)
≈ (11)
dt Tz
The predicted input current is given by:

Rg Tz Tz
ī g (k + 1) = 1 − · ī s (k) + + ī g (k)
Lg Ls

+ vg (k) − vr ec (k) (12)
The predicted instantaneous active and reactive power can be expressed by

equations in term of the input voltage and current vectors as:

P(k + 1) = Re eg (k + 1).i g (k + 1) (13)

Q(k + 1) = Im eg (k + 1).i g (k + 1) (14)
For a high sampling frequency with respect to the grid fundamental frequency, it
can be assumed that eg (k + 1) ≈ eg (k).
The control variable for the rectifier are the input active and reactive powers.
In order to select the optimal voltage vector, the cost function that summarizes the
desired behavior of the rectifier is given as follows:

g1 = P ∗ − P(k + 1) + λ1 Q ∗ − Q(k + 1) (15)
λ1 is the weighting factor between active power and reactive power, P* and Q*
are the active and reactive power references.
The reference value of the reactive power Q* is usually set to zero to obtaining
unity power factor operation.
4.2 Machine Side Converter (MSC) Control
FS-PTC algorithm includes three main steps, the estimation of the stator flux and
the torque is the first step. The second step is the prediction of the next-instant of
434 A. Ammar
current ī s (k + 1), flux ψ̄s (k + 1) and torque Te (k + 1). The cost function design is
done finally [17].
From the IM model, the stator current and the stator flux can be described as
follows:

d ψ̄ s
= v̄s − Rs ī s (16)
dt

1 d ī s 1
ī s = − Lσ · − kr · − jω ψ̄r − v̄s (17)
Rσ dt Tr
where:
k r =M sr /L r, Rσ = Rs + kr2 · Rr , L σ = σ.L s .
After discretization with the sampling time T z , the stator flux and current
prediction can be obtained as:

ψ̄ s (k + 1) = ψ̄ s (k) + Tz v̄s (k) − Rs · Tz ī s (k) (18)

Tz Tz
ī s (k + 1) = 1 − · ī s (k) + ·
Tσ Tσ

1 1
kr · − jω(k) ψ̄r (k) + v̄s (k) (19)
Rσ Tr
with
Tσ = σ L s /Rσ
By the predictions of the stator flux and current, the electromagnetic torque can
be predicted as following:

Te (k + 1) = p · Im ψ̄s (k + 1) · ī s (k + 1) (20)
The cost function in the MPC strategy compares the predicted and reference values
[10]. The classical cost function for the PTC method is:

g2 = Te∗ − Te (k + 1) + λψ̄s∗ − ψ̄s (k + 1) (21)
Te∗ is the reference torque and T̂e (k+1) is the predicted torque for a given switching
state, ψ̄s∗ is the reference stator flux and ψ̄s (k + 1) is the predicted stator flux.
λ is the weighting factor which defines a trade-off between the torque and flux
tracking [18].
Figure 3 illustarate the global control scheme of the model predictive power/torque
control for the back-to-back converter that feed in induction motor.
Fig. 3 Diagram of predictive power/torque control for an induction motor drive using back to back
converter
This section presents the simulation results of the proposed model predictive control
for a back-to-back converter that feed an induction motor. The following figures
displayed the performance analysis of the grid-side-converter and the motor-side-
converter as a comparative study. The predictive power/torque control (PPC-PTC) is
compared to another two topologies. A conventional topology (DTC) that consists
of an uncontrolled rectifier (full diode bridge) in the grid-side and a traditional DTC-
controlled inverter in the motor-side. The second topology (DPC-DTC) consisted of
DPC-controlled rectifier in the grid side and DTC controlled inverter in the motor-
side. The figures are specified by ‘a’ for DTC, ‘b’ for DPC-DTC and ‘c’ for PPC-PTC
strategies.
The grid-side-converter (GSC) control perfromance analysis is shwon in Fig. 4.
The figure presents from top to buttom, the active power, reactive power, grid currents
and the DC-link voltage. Generally, the active power depends to the load demand. In
436 A. Ammar
4000
P[W]
P[W]
P[W]
2000
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

4000
Q[VAR]
Q[VAR]
Q[VAR]
2000
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

5
ig[A]
ig[A]
ig[A]
-5
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
540
Vdc [V]
Vdc [V]
Vdc [V]
520
500
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Times [s] Times [s] Times [s]
Fig. 4 Grid-side-converter performance: Active power “P”, Reactive power “Q”, Grid current “ig ”,
DC-link Voltage “V dc ”
Fig. 4a, the case of the uncontrolled rectifier, it can be seen that active power shows
high ripples, while the reactive power and DC-link voltage cannot be controlled. In
Fig. 4b, the direct power control has been applied in GSC, in this case, the active
power provides reduced ripples. In addition, the control of reactive power and DC-
voltage has been achieved. the reactive power has been set to zero in order to have a
unity power factor. The control of DC-link rises the system stability during the speed
variation, moreover, this topology is featured by a bi-directionality which gives the
ability to control the motor in four quadrants. The predictive power control (PPC)
in Fig. 4c, provide a higher performance compared to the two previous topologies.
The smooth powers waveform and the minimized voltage overshoot/undershoot is
clearly observed.
Next, Figs. 5 and 6 present the stator current and its FFT analysis, it can be seen that
the traditional topology with diode rectifier in Figs. 5a and 6a shows highly distorted
input current, where the total harmonics distortion (THD) is >31% and this can be
4 4
4
2 2
2
iga[A]
iga[A]
iga[A]
0 0
0
-2 -2
-2
-4 -4
-4
1.6 1.61 1.62 1.63 1.64 1.65 1.66 1.6 1.61 1.62 1.63 1.64 1.65 1.66
1.6 1.61 1.62 1.63 1.64 1.65 1.66
Fig. 5 ZOOM of grid current waveform iga
100 100
100
90
90 90
Mag (% of Fundamental)
80
80 80
70
70 70
60
60 60
50
50 50
40
40 40
30
30 30
20
20 20
10
10 10
0
0 50 100 150 200 250 300 350 400 450 500 0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz) Frequency (Hz) Frequency (Hz)
Fig. 6 Spectrum of total harmonics distortion (THD) for the grid current
Te[rN.m] W[rpm]
Te[rN.m] W[rpm]
Te[rN.m] W[rpm]
100 100 100
0 0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
10 10
10
5 5
5
0
0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
10
10 10
isa[A]
isa[A]
isa[A]
0
0 0
-10
-10 -10
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
ψs [Wb]
ψs [Wb]
ψs [Wb]
1 1
1
0.5 0.5
0.5
0 0
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Fig. 7 Motor-side-converter performance: Rotor speed “W ”, Torque “T e ”, stator current “isa ”,

Flux Linkage “ψ s ”
easily observed for, current waveform. Furthermore, the THD in cases of DPC-DTC
and PPC-PTC strategies is 6.28% and 0.82% respectively. These results show the
importance of the PWM-rectifier over an uncontrolled rectifier in one hand and the
merit of the MPC application in the other hand. the optimum voltage selection of PPC
strategy reflect on the quality of signals specially the current. Then, Fig. 7 displayed
the motor side converter control performance. The figure shows from top to bottom,
rotor speed, torque, stator phase current and stator flux magnitude. The machine
behavior of DTC and DPC-DTC in Fig. 7a, b is the same since both topologies use
the traditional direct torque control in the motor side. The high toque and flux ripples
can be clearly observed However, in Fig. 7c, PPC-PTC has a reduced torque and flux
and smoother current waveform due to the optimized MPC.
6 Conclusion
This paper presents a model predictive strategy for PWM-rectifier-inverter that feed
induction motor drive. MPC has been employed for grid and motor-sides. The
proposed algorithms have been compared to conventional techniques that use a look-
up switching table. The simulation results have been obtained under different condi-
tions in order to investigate the performances such as steady state, speed and load
variation.
The use of PWM-rectifier offers different advantages, in particular, current quality
improvement and unity power factors. Moreover, the application of MPC can enhance
the control performance and reduce power/torque ripples and current distortion.
Generally, the finite-state model predictive control techniques are presented as an
alternative solution in electrical drives and power electronics field. They preserve
numerous good proprieties of conventional drives and offer more advantages. Never-
theless, they still a challenging subject for researches concerning performance
enhancement.
438 A. Ammar
Appendix
The different parameters for simulation in SI units are:

Grid-side-converter:
Line voltage vg = 220 V

Line resistance R = 0.4
Line inductance Lg = 0.015 H
Frequency f = 50 Hz
DC-link voltage Vdc = 500 V
DC-link capacitor C = 2040 μF
Motor-side-converter:
Motor’s power P = 1.1 kW

Motor’s rated current I = 2.5 A
Rated stator flux ψs = 0.95 Wb
Rated Torque Te = 6 N·m
Rated speed ωr = 1450 rpm
Stator resistance Rs = 6.75
Rotor resistance Rr = 6.21
Stator inductance Ls = 0.5192 H
Rotor inductance Lr = 0.5192 H
Mutual inductance Msr = 0.4957 H
Number of pole pairs p=2
Friction coefficient fr = 0.002 N·m·s
The moment of inertia J = 0.01240 kg·m2
References
1. Lai Y-S, Lin C-K, Chuang F-P, Yu J (2017) Model-free predictive current control for three-phase
AC/DC converters. IET Electr Power Appl 11(5):729–739
2. Mehreganfar M, Saeedinia MH, Davari SA, Garcia C, Rodriguez J (2019) Sensorless Predictive
Control of AFE Rectifier With Robust Adaptive Inductance Estimation. IEEE Trans Ind Inf
15(6):3420–3431
3. Wang W, Yin H, Guan L (2009) Study on back-to-back PWM converter based on direct power
control for induction motor drive. In: Proceedings - 12th International conference electrical
machine system ICEMS 2009, no. 50777022, pp 1–4
4. Zhang Y, Zhang Q, Li Z, Zhang Y (2013) Comparative study of model predictive current
control and voltage oriented control for PWM rectifiers. In: 2013 International conference on
electrical machines and systems (ICEMS), pp 2207–2212
5. Wang F, Zhang Z, Mei X, Rodríguez J, Kennel R (2018) Advanced control strategies of induc-
tion machine: field oriented control, direct torque control and model predictive control. Energies
11(1)
6. Idris NRN, Yatim AHM (2004) Direct torque control of induction machines with constant
switching frequency and reduced torque ripple. IEEE Trans Ind Electron 51(4):758–767
7. Zhang Z et al (2015) Predictive control with novel virtual-flux estimation for back-to-back
power converters. IEEE Trans Ind Electron 62(5):2823–2834
8. Zhang Y, Li Z, Zhang Y, Xie W, Piao Z, Hu C (2013) Performance improvement of direct power
control of PWM rectifier with simple calculation. IEEE Trans Power Electron 28(7):3428–3437
9. Rodriguez J et al (2007) Predictive current control of a voltage source inverter. IEEE Trans Ind
Electron 54(1):495–503
10. Vazquez S, Rodriguez J, Rivera M, Franquelo LG, Norambuena M (2017) Model predic-
tive control for power converters and drives: advances and trends. IEEE Trans Ind Electron
64(2):935–947
11. Alkorta P, Barambones O, Cortajarena JA, Zubizarrreta A (2014) Efficient multivariable gener-
alized predictive control for Sensorless induction motor drives. IEEE Trans Ind Electron
61(9):5126–5134
12. Liang J, Wang H, Yan Z (2019) Grid voltage sensorless model-based predictive power control
of PWM rectifiers based on sliding mode virtual flux observer. IEEE Access 7:24007–24016
13. Ammar A, Talbi B, Ameid T, Azzoug Y, Kerrache A (2019) Predictive direct torque control
with reduced ripples for induction motor drive based on T-S fuzzy speed controller. Asian J
Control, p asjc. 2148
14. Zhang Y, Xia B, Yang H (2015) Model predictive torque control of induction motor drives with
reduced torque ripple. IET Electr Power Appl 9(9):595–604
15. Yin Z, Li W, Zhang Y, Tang R (2019) Model predictive torque control based on a disturbance
observer for induction motor drives. In: 2019 14th IEEE conference on industrial electronics
and applications (ICIEA), 2019, pp 1871–1875
16. Aissa O, Moulahoum S, Colak I, Kabache N, Babes B (2016) Improved performance and power
quality of direct torque control of asynchronous motor by using intelligent controllers. Electr
Power Component Syst 44(4):343–358
17. Ammar A, Kheldoun A, Metidji B, Talbi B, Ameid T, Azzoug Y (2018) An experimental
assessment of direct torque control and model predictive control methods for induction machine
drive. In: 2018 International conference on electrical sciences and technologies in Maghreb
(CISTEM), 2018, no. October, pp 1–6
18. Habibullah M, Lu DD-C (2015) A speed-sensorless FS-PTC of induction motors using extended
kalman filters. IEEE Trans Ind Electron 62(11):6765–6778
Novel Smart Air Quality Monitoring
System Based on UAV Quadrotor
Mehdi Zareb, Benaoumer Bakhti, Yasser Bouzid, Hamza Kadourbenkada,

Kamel Bouzgou, and Wahid Nouibat
Abstract In tomorrow’s society called Smart community, the air quality and pollu-
tants will take the attention of researchers and leaders to develop intelligent systems
to better protect tomorrow’s citizens from the risks of Smart Risk that can take
place in Smart City. Also, Unmanned Aerial Vehicles (UAVs) due to their flexibility,
mobility in 3d and ability to carry air low-cost sensing units have become a cheap
alternative to monitor pollution values in a given area. However, enough still needs
to be learned as these types of studies are expected to increase in the future years. In
this paper, we propose an electronic system based on low-cost sensors designed for
air quality monitoring. This system is embedded in a quadrotor type UAV in which,
the mobility of the quadrotor is controlled with an autopilot. Experiment tests are
made on the campus of Mascara University, where the obtained results have proved
the effectiveness of the proposed UAV-based air quality monitoring system.
Keywords UAV · Quadrotor · Air quality monitoring · AQM · AQI · Path

planning · Smart risk · Smart city · IoT
M. Zareb (B)
Electronic Department, USTO-MB, University of Mascara, Mascara, Algeria
B. Bakhti
Physics Department, University of Osnabrueck, University of Mascara, Osnabrueck, Germany
Y. Bouzid
CSCS Laboratory, Ecole Militaire Polytechnique, Bordi El Bahri, Algeria
H. Kadourbenkada
Electronic Department, University of Mascara, Mascara, Algeria
K. Bouzgou · W. Nouibat
Electronic Department, LEPESA Laboratory, USTO-MB University, Oran, Algeria
W. Nouibat

https://doi.org/10.1007/978-981-15-6403-1_30
442 M. Zareb et al.
1 Introduction
For many families in the word, bad air quality is causing more and more health issues
and dangerous conditions for all kinds of communities. This is due to the pollutants
in the air. The pollutant arising from artificial sources contaminating are the most
relevant with regard to human/animal health. It can be grouped into first, stationary
sources including as example power stations, where coal, natural gas, or petroleum
are burned. Second, mobile sources including gasoline and diesel-fueled vehicles.
However, air pollutants bear similarities based on which are categorized into gaseous
pollutants cased directly by fossil fuel combustion as Carbon Monoxide CO, Nitrogen
Oxide NO, ozone O3 , nitrogen dioxide NO2 , Volatile Organic compounds VOCs,
Sulfur Dioxide SO2 , etc. Heavy metals include of mercury Hg, lead Pb, cadmium
Cd and silver Ag, etc. Can be transported by air and their most common sources
are an industrial and mining waste, vehicle emissions, fertilizers, and acid batteries,
etc. Other important categories of pollutants are the Particulate matter PM. Waste
incineration facilities, dust, fires, vehicles are a major source of PM. The particle
size ranges between 2.5 μm (PM2.5) and 10 μm (PM10).
The literature has well-established the link between long-term exposure to air
pollution and rate of morbidity and mortality in different country as mentioned in
[1–3, 4]. As an example, in [5] they estimate 1.1 million annual premature deaths
from PM2.5 in India between 2012 and 2018. The relationship between the pollution
levels around the days and death for 22 million adults aged 65 and older based on
death records from 2000 to 2012 in the USA is investigated in [6, 7]. In fact, short-term
exposure to Air Pollutants can cause pulmonary and cardiovascular diseases neurode-
generation, neurological disorders, etc. as studied in [8–10]. Also, the different parts
of the respiratory system are affected by PM depends upon the size of the particle as
presented in [11]. Also, it may affect to people with allergic diseases [12].
The first idea in this topic was to use vision UAV base guide solution by analyzing
imagery interpretation in the investigation of environment pollution by Zang et al.
[13] and Yang et al. [14]. Where, they demonstrate the effectiveness of applying UAVs
than the use of satellite or manned aerial remote sensing, which is too expensive and
take time.
Then, Gas Sensing System (GSS) based on the low-cost sensor is developed, as an
example, in [15] a GSS system was developed to investigate the Gas leakage. Another
example of ecological monitoring with the prime price is proposed by Danilov et al.
[16]. Where a GSS was developed and mounted on the Small DJI hexacopter UAV to
build 3D models of the pollution of the atmosphere. This approach is carried out in two
steps, first, one the trajectory of flight is presented on the spiral of Archimedes, then
the trajectory of the flight on the second stage is proposed on a helicoidal trajectory.
The application of the UAV-based GSS system for Smart Cities was investigated: first
for multiple air pollutants by Gu and Jia [17] and by Alvear et al. [18, 19]. Where the
experiments tests have verified the feasibility and precision of the system. Where they
found that the on-board sensor devices did not affect the UAV’s power consumption
and flight time. However, the UAV during flight operations produced an important
Novel Smart Air Quality Monitoring System … 443
noise into the sensors and the data acquisition modules, which resulted in the variation
of sensor readings as the case of NO2 sensor for example. To overcome this problem,
they propose as a solution, the use of a separate power source for the UAV and the on-
board devices. Secondly, UAV-based GSS system for a specific pollutant as in [20]
where they trend relating PM concentration with elevation and traffic conditions near
the specific urban area. Also, in where Volatile organic compounds VOCs in the air
was analyzed. The design methodology problems for the UAV-based AQM system
can be summarized: First, the location where the sensors should be mounted. Second,
the influence of sensor orientation on the captured values. Third, the influences of
incomplete measurement. Also, noise caused by the UAVs propellers during the
flight. In this paper, we propose a UAV-based air quality monitoring system AQM.
Where design methodology that we follow consider the AQM problems already cited.
First, based on results of the work [21], for example, the AQM system is positioned
between the central tile positions of quadrotor UAV (we make the choice on this type
of UAV for this study). A data acquisition module based on SD card is added to AQM
electronic system to overcome the influences of incomplete measurement. Also, a
separate power source for the UAV and the onboard AQM devices are used to reduce
the noise caused by the UAVs propellers during the flight. Finally, the trajectories of
flight are presented on the Boustrophedon, and helicoidal.
The rest of the paper is organized as follow: Section two presentation of the
developed AQM UAV-based system, section three results and discussion. Finally,
conclusion.
2 Presentation of the Developed AQM UAV-based System
In general, drones are equipped with low-cost micro sensors that have the advantage
of capturing the spatial and temporal variability of air pollutants. In addition, drones
provide spatial resolution and vertical profiling near the surface of air pollution due to
their mobility in both horizontal and vertical dimensions. In this report, we design an
AQM electronic system assembled on a quadrotor UAV rotor as shown in Fig. 1. The
objective of the system is to inspect and monitor air quality in 3D. This system has
been tested in the University of Mascara, whose results have proven the effectiveness
of using the drone as an AQM system.
The proposed system is a two-tier system:
• The electronic air quality measurement system responsible for data acquisition
and analysis
• Unmanned aerial vehicles (UAVs) responsible for flying the system in a three-
dimensional space
444 M. Zareb et al.
Fig. 1 Photo of the realized UAV-based AQM system taken during experiments in Mascara
university
2.1 The Electronic Air Quality Measurement System
The AQM System is composed of different elements ranging from the microcontroller
to a variety of sensors and an external power supply, etc. Where the sensor elements
are tested to test their reliability. The Arduino Mega 2560, a microcontroller board
based on Atmega2560, is chosen for program processing because it contains enough
digital and analog I/O pins.
The MQ-7 sensor is chosen to detect the concentration of carbon monoxide CO.
Just as the MQ-7 MQ131 is a semi-conductor gas sensor, the sensitive material of
this gas sensor is O3 . To detect temperature and humidity the DHT22 is used as a
low-cost basic digital temperature and humidity sensor. It uses a capacitive humidity
sensor and a thermistor to measure ambient air. For the particle mater, Shenyei
PPD42NS is chosen, it’s a low-cost PM sensor that operates according to the light
scattering method, it continuously detects airborne particles. It is possible to obtain
a pulse output corresponding to the concentration per unit volume of particles, using
an original detection method based on a light scattering principle similar to that of
the particle counter, this sensor is very sensitive to very fine particles. The rest of
the hardware used includes Arduino SD card shield and a real-time clock for data
storage. An LCD module for data display. An external power supply. By also adding
other elements such as assembly box, wires, connection card, etc. Figure 2 presents
the electrical schematic of the AQM measurement system that shows the connection
of each sensor to the Arduino Mega. Thus, it represents the photo of the AQM
electronic system produced. In Fig. 3 the flowchart of the program that manages the
operation of the AQM system is shown, where the sampling time chosen to take the
concentration of O3, CO and PM equal to 60 s, whose measured data are displayed
and recorded at the same time in an SD card. For the software part of the project, we
have programmed a Final Sketch (in C Arduino program) that groups all the other
programs of each sensor with modification and alteration taking into account the
different properties of each element (response time, calibration, etc.).
However, for implementation, each sensor is tested and verified on its own.
Starting with the MQ-7 sensor where the sensor’s test process was performed by
lighting a burning paper and exposing the sensor to smoke, then MQ-131 and Shenyei
PPD42NS sensors. Figure 4a shows the test result obtained from the MQ-7 sensor
for the concentration of CO. The concentration of O3 obtained from the MQ-131
sensor is presented in Fig. 4b. In addition, the PM2.5 and PM10 concentrations are
presented in Fig. 4c, d.
The calibration of the gas sensors, MQ-7, MQ-131 is an important procedure,
where we have followed the method represented by Kularatna et al. [22].
Fig. 2 Electrical schematic of the AQM measurement system

446 M. Zareb et al.
Fig. 3 The flowchart of the AQM program
35 60
CO
30
55 O3
25
50
20
PPM
/M3
45
15
∝G
40
10
5 35
0 30
0 50 100 150 200 250 300 350 400 0 100 200 300 400 500 600 700 800 900 1000
TEMPS [SEC] TEMPS [SEC]
(a) CO evaluation test result (b) O3 evaluation test result
4
X 10
X 10 6 4.5
9 PM2.5
PM10
8 4
7 3.5
6 3
µM3
2.5
/M3
5
2
∝G
4
3 1.5
2 1
1 0.5
0
0 0 100 200 300 400 500 600 700
0 100 200 300 400 500 600 700
TEMPS [SEC] TEMPS [SEC]
(c) PM10 evaluation test result (d) PM25 evaluation test result
Fig. 4 Pollutants sensor reliability tests
2.2 UAV System
An overview of the UAV system is presented. The Quadrotor or Quadcopter which

is defined as a UAV with four rotors [23] (Fig. 1). Where, according to [24] this type
of UAV has the largest use, for this reason, we have made the choice to use it. The
UAV is equipped with an autopilot that contains the planned trajectory (as proposed
by [25]). Figure 9 presents the schematic block of the Quadrotor UAV autopilot.
Where the first bock contains the autonomous trajectory generation (path planer)
for UAV-AQM. We will focus on the Bous-trophedon trajectory (as Figs. 5 and 6
present) that follows the Lawnmower search and the Archimedes Spiral trajectory
that follows the Spiral Search as Fig. 7. Second, in the schematic block of the UAV
(Fig. 9) a PID controllers are used to guide and maintain the UAV Quadrotor to a
specific location (x, y and z) with a three desired orientation φ, θ and ψ that represent
respectively the roll, pitch and yaw angles of the UAV. The desired states (xd yd z d
and ψ) are calculated by the path planer block to follow the proposed trajectory. The
Fig. 5 Boustrophedon
trajectory with a circular foot
print
Fig. 6 Boustrophedon
trajectory with a square foot
print
Fig. 7 Archimedes spiral

448 M. Zareb et al.
1.35
CO 28
O3
1.3
26
1.25
PPM
24
G / µM 3
1.2 22
1.15 20
1.1 18
0 100 200 300 400 500 600 700 800 900 0 100 200 300 400 500 600 700 800 900
Time [sec] Time [sec]
(a) CO concentration test result (b) O3 concentration test result
250 36
PM2.5 Trajectory 2 Temperature
200 34
∝ g/m3
degree
32
150
30
100
28
50 26
500 1000 1500 2000 0 100 200 300 400 500 600 700 800 900
Time [sec] Time [sec]
(c) PM10 concentration test result (d) Temperature test result
Fig. 8 Experiment tests of the concentration of the pollutants
U PID
+ e(t) Quadrotor
Trajectory PID Controller System
x, y, z, ϕ, θ and ψ
Fig. 9 UAV Autopilot schematic block
PID control low is presented in Eq. 1 (Fig. 6).
t
de(t)
UPID = K p .e(t) + K i e(τ ).dτ + K d . (1)
dt
0
where Kp , Ki , and Kd present respectively the proportional, integral and derivative

gains. Also, e(t) is the error calculated between desired and measured quadrotor
displacements, that can be resumed respectively as follows:
e1 (t) = xd (t) − x(t) (2)
e2 (t) = yd (t) − y(t) (3)

e3 (t) = z d (t) − z(t) (4)
e4 (t) = ψd (t) − ψ(t) (5)
According to their official website, the University of Mas-cara has 16875 under-
graduate and graduate students, 293 Ph.D. students and about 1746 staff members
(professors and agents). It is a large number of people who enter and exit through the
main door daily, we saw this as an opportunity to test our system to determine the air
quality to which all these people are exposed daily. That is why we have organized
a series of tests at different locations on campus to measure pollutant concentration
levels, which helps us determine the overall air quality in the region and we have
conducted these tests at times when we suspected the presence of air pollutants. All
these tests will be used to validate the system we propose.
3.1 First Test
For the first test, we chose a trajectory that most students follow when they enter
university, as shown in Fig. 10, point A is the starting point and point B is the ending
point, where the UAV flights at an altitude equal to 15 meters. The measurements
obtained from the first test are presented from Fig. 8a, d. First, the AQM electronic
system has successfully detected all required pollutant levels. We notice the presence
of a peak of ozone (a secondary pollutant), high temperature and humidity contribute
to its formation, carbon monoxide CO was low because this pollutant intervenes in
the event of a gas leak or a form of combustion now.
Second, the AQI (Air Quality Index) of the area inspected for the first test can be
calculated by using pollutant concentration data with the Eq. 3:
(P Mobs − P Mmin ) ∗ (Imax − Imin )

Ip = + Imin (3)
(P Mmax − P Mmin )
where, PM obs is the observed average concentration μg/m3 , PM min is the maximum
concentration of AQI color category that contains PM obs , PM min is the minimum
concentration of AQI color category that contains PM obs , I max is the maximum AQI
value for color category that corresponds to PM obs and Imin is the minimum AQI
value for color category that corresponds to PM obs (These values can be taken from
the Table 1). The AQI value tells us how clean or unhealthy your air is. Table 1
presents the AQI six levels of health concern. Let’s calculate the AQI around the
450 M. Zareb et al.
Fig. 10 Mapping of UAV air quality monitoring at the main entrance of the University of Mascara
Table 1 The AQI six levels of health concern

Air quality index Levels of health PM10
When the AQI is in this range Concern air quality conditions Breakpoints (μg/m3 ) When the
are: Break-point is in this range:
0 to 50 Good 0–50
51 to 100 Moderate 51–150
101 to 150 Unhealthy for sensitive groups 151–350
151 to 200 Unhealthy 351–420
201 to 300 Very unhealthy 421–500
301 to 500 Hazardous 500–600
main entrance of the campus corresponding to PM10 concentration of 200 μg/m3

taken from Fig. 8c where:
(200 − 500) ∗ (200 − 100)

I P M10 = + 100
(350 − 150)
I P M10 = 125
This AQI value of the fine particles (PM) means that the air quality conditions are
unhealthy of sensitive groups around the main entrance of the campus we thought it
was the high presence of motor vehicles and their emission rate that caused these bad
levels of p.m. At the end of the trajectory, most levels have decreased, which means
that both indoor places and classes have better air quality, so we advise students not
to stay too long at the main entrance. Figure 10 presents the mapping of the air quality
index for the first test, where the orange circle means that the air quality is unhealthy
for sensitive groups, close to bad (the main door of the University of Mascara). The
green circle means that the air quality is good (zone B) in Fig. 10.
3.2 Second Test
For the second test, we followed a similar trajectory, but this one is horizontal to the
other. As shown in Fig. 11, point A is the starting point and point B is the endpoint.
Figure 12 presents the PM concentration during the second test UAV trajectory,
where it reaches 65 μg/m3 with an AQI value equals to 85 at the starting point (A)
(Fig. 11) which corresponds to moderate AQI category. At the end of the trajectory
(point b), the PM concentration reaches 42 g = m3 with an AQI value equals to
39 which corresponds to good air quality. Also, in Fig. 11 yellow and green circles
present respectively the moderate and good air quality zone. Finally, Fig. 13 presents
the trajectory made by the UAV quadrotor during the AQM tests at the points A and
Fig. 11 Mapping of UAV air quality monitoring of the second test

452 M. Zareb et al.
Fig. 12 PM10 concentration 80

PM
result of the second test
70
60
50
40
30
0 500 1000 1500 2000 2500 3000
TIME [SEC]
Fig. 13 3D trajectory made

by the UAV quadrotor during
25
the AQM at the zone A and B
20
Altitude [meters]
15
10
0
15
10 15
5 10
0 5
−5 0
−10 −10 −5
−15 −15
B. Where the autopilot follows an Archimedes Spiral path, in which In the UAV
moves to several altitudes 5 meters after 15 meters, etc. to properly inspect the area.
4 Conclusion
The UAV equipped with low-cost micro-sensors that have the advantage of capturing
the spatial and temporal variability of air pollutants has enabled spatial resolution
and vertical profiling near the surface of air pollution due to their mobility in hori-
zontal and vertical dimensions. In this paper, the pro-posed AQM electronic system
assembled on a quadrotor UAV have properly inspected and monitored the 3D air
quality at Mascara University. Unlike the AQM systems already existing in the liter-
ature, it measures several pollutants such as PM 10 and 2.5, bad ozone O3 , carbon
monoxide at the same time. According to the tests made, the air quality is moderate,
close to bad at the main entrance to the campus of Mascara University indicated by
an orange zone. While the green zone (good air quality) is located in the center of
the university. So, the results obtained proved the effectiveness of using the UAV as
an AQM system. In future work, we propose to carry out wireless communication
with a base station to do air monitoring online. Also, Adaptive and automatic moni-
toring techniques that ensure better battery consumption of the UAV, with sufficient
accuracy of air quality index will be investigated.
References
1. Khajavi A, Khalili D, Azizi F, Hadaegh F (2019) Impact of temperature and air pollution on
cardiovascular disease and death in iran: a 15-year follow-up of tehran lipid and glucose study.
Sci Total Environ 661:243–250. http://www.sciencedirect.com/science/article/pii/S00489697
19302013
2. Chen R, Yin P, Meng X, Wang L, Liu C, Niu Y, Liu Y, Liu J, Qi J, You J, Kan H, Zhou M (2019)
Associations between coarse particulate matter air pollution and cause-specific mortality: a
nationwide analysis in 272 chinese cities. Environ Health Perspect 127(1):017008
3. Lelieveld J, Barlas C, Giannadaki D, Pozzer A (2013) Model calculated global, regional and
megacity premature mortality due to air pollution. Atmosp Chem Phys 13(14):7023–7037.
https://www.atmos-chem-phys.net/13/7023/2013/
4. Achilleos S, Al-Ozairi E, Alahmad B, Garshick E, Neophytou AM, Bouhamra W, Yassin
MF, Koutrakis P (2019) Acute effects of air pollution on mortality: a 17-year analysis in
kuwait. Environ Int 126:476–483. http://www.sciencedirect.com/science/article/pii/S01604
12018323419
5. David LM, Ravishankara AR, Kodros JK, Pierce JR, Venkataraman C, Sadavarte P, Prema-
ture mortality due to pm 2.5 over india: effect of atmospheric transport and anthro-
pogenic emissions. GeoHealth 3(1):2–10. https://agupubs.onlinelibrary.wiley.com/doi/abs/10.
1029/2018GH000169
6. Jhun I, Kim J, Cho B, Gold DR, Schwartz J, Coull BA, Zanobetti A, Rice MB, Mittleman
MA, Garshick E, Vokonas P, Bind M-A, Wilker E H, Dominici F, Suh H, Koutrakis P (2019)
Synthesis of harvard epa center studies on traffic-related particulate pollution and cardiovas-
cular outcomes in the greater boston area. J Air Waste Manag Assoc, pMID: 30888266. https://
doi.org/10.1080/10962247.2019.1596994
7. DeFlorio-Barker S, Crooks J, Reyes J, Rappold AG (2019) Car-diopulmonary effects of fine
particulate matter exposure among older adults, during wildfire and non-wildfire periods, in
the united states 2008–2010. Environ Health Perspect 127(3):037006
8. Ghosh R, Hertz-Picciotto I (2014) Chapter 24—the epidemiology of air pollution and childhood
lung diseases. In: The Lung (Second Edition), second ed., R. Harding and K. E. Pinkerton (Eds.)
Boston: Academic Press, 2014, pp. 423–437. http://www.sciencedirect.com/science/article/pii/
B9780127999418000249
9. Lim CC, Hayes RB, Ahn J, Shao Y, Silverman DT, Jones RR, Thurston GD, Mediterranean diet
and the association between air pollution and cardiovascular disease mortality risk. Circulation
10. Gore and RW, Deshpande DS (2017) An approach for classification of health risks based on
air quality levels. In: 2017 1st International Conference on intelligent systems and information
management (ICISIM), Oct 2017, pp 58–61
11. Morabet RE (2018) Effects of outdoor air pollution on human health. In: Reference module in
earth systems and environmental sciences. Elsevier. http://www.sciencedirect.com/science/art
icle/pii/B9780124095489110127
12. Stewart GA, Peden DB, Thompson PJ, Ludwig M (2012) 5 - allergens and air pollutants.
In: Holgate ST, Church MK, Broide DH, Martinez FD (eds.) Allergy (FourthEdition), fourth
454 M. Zareb et al.
edition. Edinburgh: W.B. Saunders, pp 73–128. http://www.sciencedirect.com/science/article/

pii/B9780723436584000093
13. Zang W, Lin J, Wang Y, Tao H (2012) Investigating small-scale water pollution with uav remote
sensing technology. World Autom Congress 2012:1–4
14. Yang Y, Hu Z, Bian K, Song L (2019) Imgsensingnet: Uav vision guided aerial-ground air
quality sensing system. In: Proceedings of the IEEE international conference on computer
communications (INFO-COM). IEEE, pp 1–9
15. Rossi M, Brunelli D, Adami A, Lorenzelli L, Menna F, Re-mondino F (2014) Gas-drone:
portable gas sensing system on uavs for gas leakage localization. IEEE Sensors 2014
Proceedings, pp 1431–1434
16. Danilov A, Smirnov UD, Pashkevich MA (2015) The system of the ecological monitoring of
environment which is based on the usage of uav. Russian J Ecol 46:14–19
17. Gu Q, Jia C (2019) A consumer uav-based air quality monitoring system for smart cities. In:
2019 IEEE international conference on consumer electronics (ICCE), Jan 2019, pp 1–6
18. Alvear O, Calafate CT, Hernández E, Cano J-C, Manzoni P (2015) Mobile pollution data
sensing using uavs. In: Proceedings of the 13th international conference on advances in mobile
computing and multimedia. ACM, pp 393–397
19. Alvear O, Zamora W, Calafate C, Cano J-C, Manzoni P (2016) An architecture offering mobile
pollution sensing with high spatial resolution. J Sens 2016
20. Mayuga GP, Favila C, Oppus C, Macatulad E, Lim LH (2018) Airborne particulate matter
monitoring using uavs for smart cities and urban areas. In: TENCON 2018–2018 IEEE Region
10 conference. IEEE, pp 1398–1402
21. Villa TF, Salimi F, Morton K, Morawska L, Gonzalez LF (2016) Development and validation
of a uav based system for air pollution measurements. Sensors
22. Kularatna N, Sudantha B (2008) An environmental air pollution monitor-ing system based on
the ieee 1451 standard for low cost requirements. IEEE Sens J 8(4):415–422
23. Zareb M, Nouibat W, Bestaoui Y, Ayad R, Bouzid Y (2019) Evolutionary autopilot design
approach for uav quadrotor by using ga. Iranian J Sci Technol Trans Electr Eng, https://doi.
org/10.1007/s40998-019-00214-6
24. Liew CF, DeLatte D, Takeishi N, Yairi T (2017) Recent developments in aerial robotics: a
survey and prototypes overview
25. Bouzid Y, Bestaoui Y, Siguerdidjane H, Zareb M (2018) Quadrotor guidance-control for flight
like nonholonomic vehicles. In: 2018 International conference on unmanned aircraft systems
(ICUAS), June 2018, pp 980–988
GA Based Adaptive Fuzzy Sliding Mode
for 3 DOF Spatial Arm Manipulator
Trajectory Tracking
Rabie Belloumi and Noureddine Slimane
Abstract In this paper a trajectory tracking control law for an arm manipulator
of three degrees of freedom is developed using an adaptive fuzzy sliding mode
control (AFSMC) with a proportional integral derivative (PID) sliding surface. The
introduction of this controller in the control loop permits that the joints track the
desired trajectory despite external disturbances. The developed control laws enable
us to note that the trajectories tracking performances of the manipulator are directly
related to the choice of the used adjustable parameter values. For that a genetic
algorithm optimization is used for the determination of the PID surface parameters
and the normalization gains of the fuzzy controllers. This approach allows us not only
to force the behavior system to a desired state very fast, but makes also the torques
functions smooth without chattering. The obtained results show an improvement of
the trajectory tracking with this approach and implied that this strategy is feasible,
effective and robust with regard to disturbances.
Keywords Genetic algorithm optimization · PID sliding surface · Fuzzy sliding

mode controller · 3 DOF manipulator · Adaptive fuzzy control
1 Introduction
The sliding mode control has shown its application simplicity and efficiency through
the reported theoretical studies; the advantage that gets such a command and
which makes it so important, is the robustness towards the model uncertainties and
disturbances.
R. Belloumi
Faculty of Technology, Department of Electronics, Batna 2 University, Batna, Algeria
N. Slimane (B)
Advanced Electronics Laboratory, Faculty of Technology, Electronics Department, Batna 2
University, Batna, Algeria

https://doi.org/10.1007/978-981-15-6403-1_31
456 R. Belloumi and N. Slimane
Several control approaches are proposed in the literature concerning the manip-
ulators with two degrees of freedom but limited work is carried out in the case of
manipulators to three degrees of freedom. In [1] a fuzzy adaptive control is applied for
a 2 DOF manipulator. Ngo et al. [2] propose a robust adaptive neural-fuzzy network
tracking control for robot manipulator. In [3] a PID fuzzy controller, considering
uncertainties of the model parameters, is proposed for the trajectory tracking. In [4]
a control by fuzzy sliding mode is proposed for MIMO systems by taking account
of the disturbances. Choi and Kim [5] use a fuzzy sliding mode controller for robust
tracking of robotic manipulators. Chen et al. [6] propose a fuzzy sliding mode control
of robot manipulators using the reaching law. In [7] fuzzy sliding mode controller
and particle swarm optimization are used for tracking the position of 2 DOF arm
manipulator, Shao and Ma [8] introduce a new method of backstepping sliding mode
control for tracking control of multi-joint robot manipulators. In the present paper,
genetic algorithm is proposed to optimize the gain parameters of sliding surface and
fuzzy controllers.
Because of the need for treating several problems varieties, the optimization algo-
rithms received an increased attention of the engineers treating these types of prob-
lems, which could not be solved with conventional techniques. The latter is inspired
by the natural evolution of the human beings and the mechanisms of evolution:
crossing, mutation, selections ….etc. The genetic algorithms try to optimize an objec-
tive function (fitness) from the evaluation of a population. The best individuals are
chosen to constitute the appropriate parents giving birth to better descents children,
among who will be fired the acceptable solutions of the handled problem [9–11].
In this work a genetic algorithm (GA) is proposed to optimize off line the param-
eters of the proposed PID sliding surface and the parameters of the fuzzy controllers
introduced in the control scheme. Automatically obtaining of these various param-
eters enables us to apply effectively on line this obtained adaptive fuzzy sliding
mode control. Simulation works are undertaken for trajectory tracking of the artic-
ular variables of an arm manipulator with three degrees of freedom, under these
two optimization algorithms by taking account of the external disturbances. The
obtained results are used to check the tracking performances in both cases and a
comparison is generated. This paper is structured as follows. In Sect. 2, the dynamic
equation of the manipulator is presented. In Sect. 3, an adaptive fuzzy sliding mode
controller with a PID surface and a supervisory fuzzy control is proposed using the
Lyapunov approach to guarantee the system stability. To illustrate the performance of
the proposed approach, simulation works are undertaken in Sect. 4. The conclusions
are given in Sect. 5.
2 Design of Manipulator Model
To define manipulator kinematics and dynamics and for the development of control
laws to be able to control it suitably, it is necessary to describe the various mathemat-
ical relations and to give a precise mathematical model, which makes it possible to
GA Based Adaptive Fuzzy Sliding Mode for 3 DOF Spatial Arm … 457
define the movements of this last in space. In general the manipulator model depends
on the considered application: geometrical, kinematic or dynamic. The manipulator
dynamic is defined by (1).
M(θ )θ̈ + C(θ, θ̇ )θ̇ + G(θ ) = U + P (1)
where θ, θ̇ and θ̈ are the vectors n × 1 of the articular variables and their derivative,
M(θ ) is the inertia matrix nxn of the manipulator, C(θ, θ̇ ) represent the Coriolis and
centrifugal torques vector n × 1, G(θ ) is the gravity vector nx1, P is the vector nx1
of external noise and U is the vector nx1 of torques and/or forces.
θ̈ = M(θ )−1 (U + P − C(θ, θ̇ ) − G(θ )) (2)
Posing x 1 = θ , x2 = ẋ1 et y = x 1

−1
ẋ1 = x2 (3)
ẋ2 = M(θ ) U + P − C(θ, θ̇ )θ̇ − G(θ )
⎧
⎨ ẋ1 = x2
ẋ2 = ga U + f a + b (4)
⎩
y = x1

where: f a = −M(θ )−1 C(θ, θ̇ )θ̇ + G(θ ) , ga = M(θ )−1 et b = M(θ )−1 P
With |b| < α, α represents the maximum value of the disturbance (Fig. 1).
Fig. 1 Manipulator structure

3 Adaptive Sliding Mode Control
The control by sliding mode rests on the principle of bringing in the first time the
dynamics of the system to reach a preset surface, called sliding surface, once arrived at
this surface it slips towards the equilibrium point. The first phase is called convergence
mode and the second it is the slip mode. The equation of surface is of the form:

s(t) = k p e(t) + ki e(t)dt + kdde(t)/dt (5)
where k p , k i and k d are gains with positive value.

The surface derivative is then:
ṡ(t) = k p ė + ki e + kd ë (6)
ṡ(t) = k p ė + ki e + kd (θ̈d − θ̈ ) (7)
ṡ(t) = k p ė + ki e + kd (−ga U − f a − b + θ̈d ) (8)
The resolution of the equation ṡ(t) = 0 without disturbance enables us to have

the following equivalent control:

u eq = (kd ga )−1 kd (θ̈d − f a ) + (k p ė + ki e) (9)
This equivalent control cannot ensure the performances wanted in the presence of
disturbances and/or of parametric uncertainties, then it is necessary to add a control
signal, which will eliminate the disturbances effect, by ensuring the stability and the
robustness of the system. The choice is:
u = u eq + u c = u eq + k f u f (10)
where k f is the standardization factor of the output variable of dimensions 3 × 3 and

uf is the output vector of the FSMC of dimensions 3 × 1
k f = f (e, ė) (11)
u f = [F S MC(s1 , ṡ1 ), F S MC(s2 , ṡ2 ), F S MC(s3 , ṡ3 )] (12)
The design of the proposed controller requires two inputs and one output, which
are respectively s, ṡ and the control signal uf , with the choice to use the same linguistic
variables for the three fuzzy units. [(s, μ(s)), (ṡ, μ(ṡ)), (uf , μ(uf ))] (Fig. 2).
The used linguistic variables are: NB: negative big, NM: negative middle, NS:
negative small, Z: zero, PS: positive small, PM: positive middle, PB: positive big.
Fig. 2 Diagram of adaptive fuzzy sliding mode (AFSMC) with surface PID
Fig. 3 Linguistic variables of s and uf
The range of input s and output uf of the controller is standardized and respectively
presented on Fig. 3.
The rules are arranged in the following table, for each entry of the controller there
are seven output fuzzy wholes, which gives 49 fuzzy rules (Table 1).
The used inference mechanism is of Mamdani type and the minimum function
of intersection is employed for the fuzzy implication. For the defuzzification, the
gravity center method is taken to calculate the values of the output vector uf . The
following control surface is obtained (Fig. 4).
The design of the supervisory controller requires two inputs and one output, which
are e, de/dt and the gain signal k f . The used linguistic variables for inputs are: NB:
negative big, NM: negative middle, NS: negative small, Z: zero, PS: positive small,
PM: positive middle, PB: positive big. For the output: VVS: very very small, VS:
very small, S: small, M: medium, B:big, VB: very big, VVB: very very big (Fig. 5).
The rules are arranged in the following table (Table 2).
The obtained surface is given below (Fig. 6).
Table 1 Fuzzy rules

uf S
PB PM PS Z NS NM NB
ds/dt PB NB NB NB Z Z Z Z
PM NB NB NB Z Z Z PS
PS NB NB NM Z Z PS PM
Z NB NM NS Z PS PM PB
NS NM NS Z Z PM PB PB
NM NS Z Z Z PB PB PB
NB Z Z Z Z PB PB PB
Fig. 4 Control surface of uf
Fig. 5 Linguistic variables of e and k f

Table 2 Fuzzy rules supervisory

kf S
NB NM NS Z PS PM PB
ds/dt NB M S VS VVS VS S M
NM B M S VS S M B
NS VB B M S M B VB
Z VVB VB B M B VB VVB
PS VB B M S M B VB
PM V M S VS S M B
PB M S VS VVS VS S M
Fig. 6 Fuzzy gain surface
The sliding surface written in Eq. (5) is selected with a Lyapunov function
V = ½ s2 .
with the derivative:

V̇ = s ṡ = s k p ė + ki e + kd ë (13)

V̇ = s k p ė + ki e + kd (θ̈d − ga u − f a − b) (14)
Substituent u and ueq by their values in (14) we obtain:

V̇ = s −kd ga k f F S MC(s, ṡ) − kd b(t) (15)
With |b| ≤ α, α represents the terminals ends of uncertainties and disturbances.

V̇ ≤ −skd ga k f + α (16)
To ensure the convergence of the state towards the sliding surface: V̇ ≤ 0


−skd ga k f + α ≤ 0 (17)
It is necessary that k f ≥ -g−

α α, thus the closed system is asymptotically stable,
1
and the error will converge towards zero.
Throughout this work one will simulate the worked out algorithms and the proposed
controllers on a model of arm manipulator to 3 DOF by using the matlab software.
The parameters of the used manipulator are in the Table 3 with the following
considerations: the distance between the gravity center of the second and third link
is r = 0.25; and stands for viscous friction of joints is c = 0.15.
The wished trajectory for the trajectory tracking of the three articulations is:

1 + sin(π(t + 3/4)) i f t ≥ 0
θd (t) =
0 if t < 0
The disturbance is taken as follows:
P(t) = 5 + 5sin(20(t − 1)) + 6sin(10(t − 0.5)) + 5u(t − 1) + 5u(t − 0.5)
The simulation results obtained by the use of these two optimization algorithms,
concerning the trajectories tracking of the three articulations, are illustrated in Fig. 7.
The trajectories tracking errors for the three joints are represented in Fig. 8.
Figure 9 represents the various torques delivered by the actuators of the three
articulations.
The simulation results are depicted in Figs. 7, 8, 9. Figure 7 shows the robot
manipulator angular positions Θ 1 , Θ 2 , Θ 3 and their desired references Θ 1d , Θ 2d ,
Θ 3d , respectively. Figure 8 represents the evolution of the tracking errors. Figure 9
depicts the generated torques of the three joints. The proposed control law shows its
improvement in tracking accuracy with generated minimal torque, without chattering.
Table 3 Manipulator
Joint 1 Joint2 Joint 3
Characteristics
Mass 3 kg 2 kg 1 kg
length 0.5 m 0.5 m 0.3 m
inertia 0.052 kg.m2 0.052 kg.m2 0.052 kg.m2
Fig. 7 Trajectory tracking

for the three joints
Fig. 8 Trajectory tracking

error for the three joints
Fig. 9 The generated

torques of the three joints
5 Conclusion
The genetic algorithm is used to obtain the parameters of the sliding surface and the
normalization gains of the fuzzy controllers, which are injected in real-time into the
fuzzy sliding mode controller suggested for the manipulator. This control enabled us
to reduce the chattering phenomenon existing in the traditional sliding mode control
and to minimize the torque values generated by the actuators. The presence of strong
disturbance forces the system to react negatively, but the suggested control with
the fuzzy supervisory makes it possible to preserve the trajectory tracking of the
various articulations with no increase in the torques generated by the actuators. The
normalization gains are not fixed any more, they vary and adapt automatically to the
fluctuations of the disturbances.
References
1. Guo Y, Woo PY (2003) Adaptive fuzzy sliding mode control for robotic manipulators. In:
Proceedings of the 42nd IEEE conference on decision and control. Hawaii
2. Ngo T, Wang Y, Mai TL, Nguyen MH, Robust JC (2012) Adaptive neural-fuzzy network
tracking control for robot manipulator. Int J Comput Commun Control 7: 341–352
3. Li W, Chang XG, Wahl FM, Farrell J (2001) Tracking control of a manipulator under uncertainty
by fuzzy PID controller. Fuzzy Sets Syst 122: 125–137
4. Aloui S, El Hajjaji A, Chaari A, Koubaa Y (2011) Improved fuzzy sliding mode control for a
class of MIMO nonlinear uncertain and perturbed systems. Appl Soft Comput 11:820–826
5. Choi S-B, Kim J-S (1997) A fuzzy sliding mode controller for robust tracking of robotic
manipulator. Mechatronics 7:199–216
6. Chen Z, Meng W, Wang H (2008) Fuzzy reaching law sliding mode control of robot manip-
ulators. In: Proceedings of 2008 Pacific-Asia workshop on computational intelligence and
industrial application
7. Soltanpour MR, Khooban MH (2013) A particle swarm optimization approach for fuzzy sliding
mode control for tracking the robot manipulator. Nonlinear Dyn 74(1–2): 467–478
8. Shao K, Ma Q (2014) Global fuzzy sliding mode controlfor multi-joint robot manipulators
based on backstepping. Found Intell Syst 277: 995–1004
9. Schmitt LM (2004) Theory of genetic algorithms II: models for genetic operators over the
string-tensor representation of populations and convergence to global optima for arbitrary
fitness function under scaling. Theoretical Comput Sci 310(1–3):181–231
10. McCall J (2005) Genetic algorithms for modelling and optimisation. J Comput Appl Math
184(1):205–222
11. Konak A, Coit DW, Smith AE (2006) Multi-objective optimization using genetic algorithms.
Reliab Eng Syst Saf 91(9):992–1007
Robust Adaptive Fuzzy Approach
with Unknown Control Gain Direction
and External Disturbance
Ouassila Bourebia and Nassira Zerari
Abstract This paper studies a fuzzy indirect adaptive approach to control uncertain
nonlinear system with unknown control gain sign (direction), in presence of external
disturbance. In the design procedure the unknown nonlinear functions are approxi-
mated by fuzzy logic system in view of their universal approximation properties. We
proposed using projection algorithm to avoid singularity problem and introducing
the Nussbaum Function to solve control direction problem. Then a robust adaptive
control is employed to build the overall controller architecture and deal with approx-
imation errors and disturbance. Furthermore, stability of the closed-loop system is
guaranteed using Lyapunov theory. Finally, simulated studies have demonstrated the
effectiveness of the proposed approach.
Keywords Fuzzy function · Adaptive control · NL system · Nussbaum function
1 Introduction
Fuzzy control has been successfully applied to many commercial products and indus-
trial systems, where controllers using fuzzy logic are generally considered applicable
to models that are mathematically complex, non-linear, where human experiences
are available to provide qualitative knowledge [1]. Therefore, fuzzy controllers are
supposed to work in situations where there is a great deal of uncertainty or unknown
variation in the parameters and structures of the system. Generally, the basic purpose
of adaptive control is to maintain the logical performance of a system in the presence
of these uncertainties [2, 3]. Fuzzy adaptive control provides a tool for using fuzzy
O. Bourebia (B)
Automatic and Robotics Laboratory, Faculty of Sciences and Technology, Department of
Electronics, Constantine1 University, Constantine, Algeria
N. Zerari
Signal Processing Laboratory SP-Lab, Faculty of Sciences and Technology, Department of
Electronics, Constantine1 University, Constantine, Algeria

https://doi.org/10.1007/978-981-15-6403-1_32
466 O. Bourebia and N. Zerari
information in a systematic and effective way. Several adaptive fuzzy control schemes
have been developed for uncertain nonlinear systems in [4]. In [5] the stability of
the underlying control systems has been investigated using a Lyapunov approach. A
common assumption in the schemes proposed in [4, 5] is that the sign of the control
gain is known apriori. However, in the general case, this assumption is by no means
realistic as pointed out in [6]. When there is no a priori knowledge about the control
gain sign, this problem has been solved by incorporating the Nussbaum-type function
[7, 8].
Inspired by [8, 9], we propose in this paper a robust adaptive fuzzy controller
for uncertain nonlinear systems to achieve the control objective. In the proposed
controller, a robustifying control term is added to the basic fuzzy controller to deal
with approximation errors. The Nussbaum gain function is introduced to solve the
problem of unknown control gain sign and the projection algorithm is applied to cope
with the problem of singularity. The stability of the closed-loop system is studied
using Lyapunov method.
In Sect. 2, we present basic ideas, in a constructive manner, of how to construct
indirect adaptive fuzzy controllers based on the fuzzy systems, and how to use
the projection algorithm to meet the control objectives. In Sects. 3 adaptive fuzzy
controllers are used to control the inverted pendulum to track a trajectory. Section 4
concludes the paper.
2 Construction of Fuzzy Adaptive Approach
(A) Problem formulation:
Consider a class of nonlinear systems modeled by the following differential equation:
x (n) = f (x) + g(x)u + d(t)

(1)
y = x1
T
where x n = [x1 , x2 , . . . , xn ]T = x ẋ · · · x (n−1) ∈ Rn is the state vector of the
system which is assumed to be available for measurement, uR, and yR are the
input and output of the system, respectively. d(t) is the external disturbance which
is assumed bounded. f (x) and g(x) are uncertain continuous functions.
Assumption 1 ([10]) The sign of the control gain g(x) is also unknown. And there
exist two unknown positive constants ḡ and g such us 0 < g ≤ |g(x)| ≤ ḡ.
We can rewrite (1) in the state space as follows:
ẋ = Ax + b( f (x) + g(x)u) + d (2)

Robust Adaptive Fuzzy Approach with Unknown Control … 467
⎡ ⎤ ⎡ ⎤
0 1 0 0 ··· 0
⎢0 0 1 0⎥ 0 ⎢0⎥
⎢ ⎥ ⎢ ⎥
⎢ .. ⎥ b = ⎢ .. ⎥
With A = ⎢ ... ..
.
..
..⎥⎥..
. ⎢.⎥
⎢ ⎢ ⎥
⎣0 0 ··· 0 1 ⎦ ⎣0⎦
0 0 ··· 0 0 1
The control objective is to design an adaptive fuzzy controller u(t) for system
(1) such that the system output y(t) follows a desired trajectory xd (t), while all the
signals of the closed-loop system are bounded.
Throughout this paper we make the following assumption:
Assumption 2 The desired trajectory xd = x d (t) =

(n−1)
T
xd (t), ẋd (t), · · · , xd (t) and its derivative are known and bounded.
Let define the tracking errors as
T
e = xd − x, ẋ d − ẋ · · · x d(n−1) − x (n−1)
T
= e ė · · · e(n−1) (3)
Then, from (2) we get

ė = Ae + b xd(n) − x (n)

= Ae + b − f (x) − g(x)u + xd(n) − d (4)
While the feedback control is given by:
1

u∗ = − f (x) + xd(n) (t) − k T e (5)
g(x)

Where k T = k0 , k1 , . . . , k(n−1) ∈ R n is the real number vector. The coefficients
k j , j = 0, . . . , n are selected so that the polynomial: S (n) +k(n−1) S (n−1) +· · ·+k0 = 0
are in the open left half plane (Hurwitz), [5] which implies that the tracking error
will converge to zero, i.e.limt→∞ |e(t)| = 0.
The control law (5) of system (1) can be easily implemented if f (x) and g(x) are
well known, the functions are generally uncertain, the goal is then to approach them
by fuzzy logic system.
(B) Development of the control law:
We propose to use fuzzy systems in this approach

1. Fuzzy logic system [6]
Lets a fuzzy system of q inputs v1……vq. The defuzzification by center of gravity
gives the output:
m q
l=1 j=1 μ Alj v j wl
y= m q (6)
l=1 j=1 μ Alj v j
where: m is the number of fuzzy ruler, Alj is the jth fuzzy set that corresponds to
the fuzzy ruler, μ Alj is the membership function of the fuzzy set Alj , and wl are the
centers of the lth fuzzy set of the fuzzy system.
The output is rewritten by:
q

m μ Alj v j wl
j=1
y= m q wl (7)
l=1 l=1 j=1 μ Alj v j
T
Such that: w = w1 w2 w3 · · · wm and
T
ξ = ξ1 ξ2 ξ3 · · · ξm
with
q q T
j=1 μ A1j v j j=1 μ A j v j
m
x= m q ··· m q (8)
l=1 j=1 μ Alj v j l=1 j=1 μ Alj v j
Exponent (7) by: y = w T ξ

C. Modeling unknowns:
f (x) and g(x) are unknown nonlinear functions. We can approximate using fuzzy
systems by w Tf ξ f (x) and wgT ξg (x) repectively, given w∗f and wg∗ constant and optimal
vectors such as:

w∗f = arg minwf {supx f (x) − wTf x f (x)} (9)

w∗g = arg minwg {supx g(x) − wgT xg (x)} (10)
Parameters w∗f and w∗f are unknown constants. From the above analysis, we can
write:

f (x) = fˆ x, w∗f + Φ f (11)

g(x) = ĝ x, w∗g + Φg (12)

f (x) − fˆ x, w∗f = w̃ Tf ξ f (x) + Φ f (13)

g(x) − ĝ x, w∗g = w̃gT ξg (x) + Φg (14)
With
w̃ f = w f − w∗f (15)
w̃g = wg − wg∗ (16)
Assumption 3 The fuzzy approximation

errors Φ f and Φg are bounded for all
x ∈ Ωx and Φ f ≤ Φ̄ f and Φg ≤ Φ̄g , where Φ̄ f and Φ̄g are unknown positive
constants. This assumption is reasonable, since we assume that fuzzy systems used
for approximating unknown functions have the universal approximation property.
Remark 1 Development that follows does not require knowledge of the sign of
the control gain, because the Nussbaum function technique will be used to tackle
this problem. The Nussbaum function technique was originally proposed in [7]. A
function is called a Nussbaum-type function if it has the following properties [7, 11]
s
(1) limt→∞ sup(1/s) N (ζ)dζ = ∞ (17)
0
s
(2) limt→−∞ inf(1/s) N (ζ)dζ = −∞ (18)
0
Here, an even function ζ2 cos(ζ) satisfying the above equalities will be employed
to handle unknown signs of control gain functions in this paper.

Lemma 1 ([12]) Let V (t) and ζ(t)are smooth functions on the interval [0, t f .
Suppose that V (t) > 0, ∀t ∈ [0, t f and N (ζ) is an even smooth Nussbaum-type
function. If the following inequality holds:
t
V (t) ≤ c0 + (g(τ )N (ζ(τ )) + 1)ζ̇ (τ )dτ (19)
0
where c0 and c1 are positive constants, and g(.)Is atime-varying parameter that
takes values in the unknown closed intervals I = l, l¯ with 0 ∈ / I , then V (t), ζ(t),
t
and g(τ )N (ζ)ζ̇ dτ are bounded on [0, t f ) with t f = ∞.
0
Proof of Lemma 1 To proof this lemma, see [12]

Controller design and stability analysis based on the above approximations and
the control law given by (5), we propose the following adaptive control law:
u(t) = u a (t) + u s (t) (20)

The control law (20) is a summation of an adaptive control term, u a (t), which
attempts to approximate the control law (5) by using the estimates of the system’s
nonlinearities, and additional control term u s (t), which is introduced to compensate
for reconstruction errors. The adaptive control term:
wgT ξg
ua = −wTf ξ f + xd(n) − ke − αwgT ξg N (ζ )e T Pb (21)
wgT ξg +
Where is a small positive constant, α positive constant. This additional control

term u s is called supervisory control, is given by:
u s = u ns N (ζ) (22)
Where
T
e P B
u ns = Φ̂ f¯ + Φ̂g u a − α N (ζ)e T Pb + ū (23)
eT P B

ū = T −wTf x f + xd(n) − ke − αwgT xg N (ζ)e T Pb (24)
wg xg +
N (ζ) = ζ2 cos(ζ) (25)

ζ̇ = e T Pbu ns + α e T Pb2 (26)
With the following parameter adaption laws with

ẇ f (t) = Pr ojw f Γ f e T Pbξ f (27)

ẇg (t) = Pr ojwg Γ g e T Pbξ g u a − α N (ζ )e T Pb (28)
Where Γ f Rr f ×r f and Γ g Rrg ×rg are diagonal matrices positives, the matrix P the
solution of Lyapunov equation AmT P + P Am = − Q for a given arbitrary n×n matrix
Q = Q T > 0. We can check, using the definition of Proj for each I.w fi ≤ w∗fi ≤ w̄ fi
And w gi ≤ wg∗i ≤ w̄gi , it means that:

wTf Γ −1
f Pr ojw f
Γ f e T
Pbξ f − e T
Pbξ f ≤0 (29)

wTg Γ −1 T T
g Pr ojwg Γ g e Pbξ g u a − α N (ζ )e Pb − e T Pbξg u a − α N (ζ )e T Pb ≤0
(30)

Φ̂˙ f¯ = γ f e T Pb (31)

Φ̂˙ g = γg e T Pbu a − α N (ζ )e T Pb (32)
Where γ f > 0, γg > 0.

D. Adaptive law derivation:
From (1) and (13)–(14) and (20), we get

ė = Ak e + b[−w̃ Tf ξ f (x) − Φ f − w̃gT ξg (x) u a − α N (ζ)e T Pb

− Φg u a − α N (ζ)e T Pb − g(x)u s + ū − αwgT xg N (ζ)e T Pb − d] (33)
Where Φ f¯ = Φ f + d
⎡ ⎤
0 1 0 ··· 0
⎢ 0 0 1 ⎥
0 0
⎢ ⎥
⎢ .. .. .. ⎥
.. ..
Ak = ⎢ . . . ⎥ . .
⎢ ⎥
⎣ 0 0 ··· 0 1 ⎦
−kn −kn−1 · · · · · · −k1
Choose the Lyapunov function candidate as:

1 T T −1 T −1 1 2 1 2
V = e Pe + w̃ f Γ f w̃ f + w̃g g w̃g + Φ̃ + Φ̃g (34)
2 γ f¯ γg
where Φ̃ f¯ = Φ f¯ − Φ̂ f¯ , Φ̃g = Φg − Φ̂g , and Φ̃˙ f¯ = −Φ̂˙ f¯ , Φ̃˙ g = −Φ̂˙ g , w̃i = wi − ŵi ,
w̃˙ i = −ẇi , i = f, g. The time derivative of V is given by
1 T
V̇ = ė Pe + e T P ė − w̃ Tf Γ f−1 ŵ˙ f − w̃gT Γg−1 ŵ˙ g
2
Φ̃ f¯ Φ̂˙ f¯ − Φ̃g Φ̂˙ g
1 1
−
γ f¯ γg
1 T T
≤ − e Qe + −w̃ f ξ f (x) − Φ f¯ − w̃gT ξg (x) u a − α N (ζ )e T Pb
2
− Φg u a − α N (ζ )e T Pb − g(x)u s + ū − αwgT xg N (ζ )e T Pb]
Φ̃ ¯ Φ̂˙ ¯ − Φ̃g Φ̂˙ g

1 1
− w̃ Tf Γ f−1 ŵ˙ f − w̃gT Γg−1 ŵ˙ g − (35)
γ f¯ f f γg
By substituting (23)–(26), and with the law adaptive (29)–(30), V̇ can be written
as follows:
1
V̇ ≤ − e T Qe + (1 + g(x)N (ζ ))ζ̇ (36)
2
We can obtain the following inequality

t
V (t) ≤ V (0) + (g N (ζ(τ )) + 1)ζ̇ (τ )dτ (37)
0
t Using Lemma 1, we can conclude from (36) the bounded of V (t), ζ(t) et
0 (g N (ζ(τ )) + 1) ζ̇ (τ )dτ with [0, t f .
According to [13], since no finite-time escape phenomenon may happen, then
t f = ∞.
Therefore, e, w̃i i = f, g, Φ̃ f¯ , Φ̃g , x(t) and u(t) are bounded. Moreover, by
invoking Barbalat’s lemma [14], we can conclude the asymptotic convergence of
e(t).
3 Application to Inverted Pendulum Tracking Control
In this section, we use our adaptive fuzzy controller to control the inverted pendulum
system to tack a sin wave trajectory.
The inverted pendulum system defined here is shown in Fig. 1, which is formed
from a cart, a pendulum and a rail for defining the position of the cart. The pendulum
is hinged in the center of the top surface of the cart and can rotate around the pivot
in the same vertical plane with the rail. The cart can move right or left on the rail
freely. It is given that no friction exists in the system between the cart and the rail or
between the cart and the pendulum [7, 8].
Let x1 = θ and x2 = θ̇ , the dynamic equation of the inverted pendulum system
can be expressed as [8],
Fig. 1 The inverted

pendulum system
⎧
⎪
⎪ θ̈ = f θ, θ̇ + g(θ )u + d(t)
⎪
⎨ y=θ
gsin(θ)−mlaθ θ̇ 2 −sin(2θ )/2 (38)
⎪
⎪ f θ, θ̇ =
⎪
⎩
4l/3−mlacos 2 (θ)
g(θ ) = − 4l/3−mlacos
a cos(θ)
2 (θ)
Where θ is the angular position of the pendulum, g = 9.8m/s 2 is the acceleration

due to gravity, M, is the mass of cart, m is the mass of pole, l is the half length of pole,
and u isthe applied force (control). We chose M = 1kg, m = 0.1 kg, and l = 0.5m,
= 1/m + M, d(t) = 0.1 sin(2 ∗ t).
It is easy to verify that g(θ) takes positive values and far from zero for all
π π
− ≤θ ≤
2 2

As the denominator of f θ, θ̇ is positive for any − π2 ≤ θ ≤ π2 and assume that

θ̇ is limited, and we have limited f θ, θ̇ .
We chose the reference signal θd (t) to be a sinusoid with amplitude of π/12 and
a frequency of 1 Hz. In our simulation we assume that the function f and g are
unknown, we used two fuzzy systems to approximate f and g with the same fuzzy
sets, shown in Fig. 2.
We choose 4 fuzzy sets for θ, θ̇ respectively which gives us 16 fuzzy rules possible:
Rule1 : if θ is NB and θ̇ is NB, Then y = w1

Rule2 : if θ is NB and θ̇ is NS, Then y = w2
Fig. 2 Fuzzy sets for θ and θ̇

..
.
Rule16 : if θ is PB and θ̇ is PB, Then y = w16
With wi , i = 1, . . . , 16 are adaptation parameters, the limits w f and wg respec-

tively w̄ fi = 100, w fi = −100, w̄gi = 1 and w gi = 0.1. all the initial values of the
components of w f are put to 0, and wg from to 0.9.
We choose d f = dg = 5, and all the elements of Γ f and Γg respectively 1000,
20, N (ζ ) = ζ 2 cos(ζ) 1 = 0.75, 2 = 159 and 3 = 0.25
We simply choose the vector k = 1 2 , to place the poles of Am at −1, and

1.5 0.5
Q = 0.25I 2 we solve Lyapunov equation to obtain P = with Eigen
0.5 0.5
√
values P = 1 ∓ 0.5 for the different initial values used θ0 = 0.5, θ0 = −0.5 .
Figure 3 shows the simulation results of the first adaptive fuzzy controller for
initial conditions θ0 = 0.5, we see that the initial parts of control are apparently
improved after incorporating these fuzzy rules. The result shows that the angle θ (t)
followed the desired trajectory after 3 s, and we notice that the system is stabilized.
The same remark for Fig. 4, where starting with, θ 0 = − 0.5 rad, which shows
the effectiveness of this approach. Figure 5 depicts the control input signal, the error
tends to 0, shown in Fig. 6.
In order to show the effectiveness of the proposed method a comparison with [15]
is considered. It is clear that the results obtained by the proposed approach are better
to that in the [15] in terms of performance since the cost of command signal in [15]
is more energetic [50–250](N).
variation de l'angle x1 pour x0=0.5

0.5
xd
x1
0.4
0.3
0.2
rad
0.1
-0.1
-0.2
-0.3
0 5 10 15 20 25 30 35
Temps (s)
Fig. 3 Response of the variable x1: θ 0 = 0.5

Variation de l'angle x1 pour x0= - 0.5

0.3
xd
x1
0.2
0.1
0
(rad)
-0.1
-0.2
-0.3
-0.4
-0.5
0 5 10 15 20 25 30 35
temps (s)
Fig. 4 Response of the variable x1: θ 0 = − 0.5
La commande u
30
25
20
15
10
Force (N)
-5
-10
-15
-20
0 5 10 15 20 25 30
temps (s)
Fig. 5 Control input signal
4 Conclusion
In this work we have proposed an approach to solve control problem of nonlinear

system with unknown control direction, to track a sin wave trajectory. We used
fuzzy indirect adaptive control to build a controller that guarantees the stability and
robustness of the system, whose approach based on the Lyapunov stability to build
the adaptation law, and fuzzy logic to approximate unknown functions to a robust
and stable controller. Compared with previous approaches, the proposed approach
obtains more compact control structure. The results are satisfactory.
Variation de l'erreur pour x1=0.5 rad et - 0.5 rad

0.5
erreur pour x1=0.5 rad
erreur pour x1= - 0.5 rad
0.4
0.3
0.2
0.1
Erreur (rad)
-0.1
-0.2
-0.3
-0.4
-0.5
0 5 10 15 20 25 30
Temps (s)
Fig. 6 Tracking error
References
1. Zadeh LA (1967) Fuzzy algorithms. Inf Control 12(2): 94–102

2. Becerikli Y, Chik BK (2006) Fuzzy control of inverted pendulum and concept of stability using
Java application. Mathematical and computer Modeling
3. Wang HQ, Chen B, Liu XP, Liu KF, Lin C (2013) Robust adaptive fuzzy tracking control
for pure-feedback stochastic nonlinear systems with input constraints. IEEE Trans Cybern
43(6):2093–2104
4. Chang YC (2001) Adaptive fuzzy-based tracking control for nonlinear SISO systems via VSS
and H∞ approaches. IEEE Trans Fuzzy Syst 9:278–292
5. Boulkroune A, Tadjine M, M’saad M, Farza M (2009) Adaptive fuzzy controller for non-affine
systems with zero dynamics. Int J Syst Sci 40(4): 367–382
6. Nussbaum RD (1983) Some remarks on the conjecture in parameter adaptive control. Syst
Control Lett 3:243–246
7. Liu Y-J, Li Y-X (2010) Adaptive fuzzy output-feedback control of uncertain SISO nonlinear
systems. Nonlinear Dyn 61:749–761
8. Boulkroune A, M’saad M (2010) On the design of observer-based fuzzy adaptive controller
for nonlinear systems with unknown control gain sign. Fuzzy Sets Syst 201 (2012): 71–85
9. Nassira Z, Mohamed C, Najib E (2015) Stable indirect adaptive HONN control for a class of
non affine SISO nonlinear systems. In: 3rd Inernational conference on control, engineering &
information technology (CEIT), pp 1–6. Tlemcen, Algeria
10. Liu L, Huang J (2008) Global robust output regulation of lower triangular systems with
unknown control direction. Automatica 44:1278–1284
11. Ye X, Jiang J (1998) Adaptive nonlinear design without a priori knowledge of control direction.
IEEE Trans Aut Control 43:1617–1621
12. Wang LX (1996) Stable adaptive fuzzy controllers with application to inverted pendulum
tracking. IEEE Trans Syst Man Cybern part b Cyber 26(5)
13. Boulkroune A (2016) A fuzzy adaptive control approach fornonlinear systems with unknown
controlgain sign. Neurocomputing
14. Slotine JJE, Li W (1991) Applied nonlinear control. Prentice-Hall, USA
15. luo L, luo F, Xu Y (2014) Direct adaptive fuzzy control of nonlinear systems with unknown
control direction. Inf Technol J 13(4): 697–702
Adaptive Neural-Network Control
Design for Uncertain CSTR System
with Unknown Input Dead-Zone
and Output Constraint
Zerari Nassira and Chemachema Mohamed
Abstract In this paper, the control problem for Continuously Stirred Tank Reactor
(CSTR) systems with input nonlinearity and unknown disturbances is addressed.
To ensure constraints satisfaction on the input and the output of CSTR we employ
a system transformation technique to transform the original constrained model of
CSTR into an equivalent unconstrained model, whose stability is sufficient to achieve
the tracking control of the original CSTR system with a priori prescribed perfor-
mance. In the proposed control methods, NNs are employed to approximate the
unknown dynamics of CSTR and additional adaptive compensators are introduced
to cope with NNs approximation errors and external disturbance; a robust control term
is introduced to overcome the effects of the unknown input dead-zone. The proposed
adaptive neural tracking controllers are designed with only one adaptive parameter
by using the second Lyapunov stability method. In comparison with the traditional
back-stepping based techniques, usually used in CSTR control, the structures of the
proposed controllers are much simpler with few design parameters since the causes
for the problem of complexity growing are completely eliminated. Simulation results
are presented to show the effectiveness of the proposed controllers.
Keywords Neural networks (NNs) · Adaptive control · Continuously stirred tank

reactor (CSTR) · Input nonlinearity · Prescribed performance
1 Introduction
The continuously stirred tank reactor (CSTR) is a typical nonlinear, and unstable
dynamic system that is applied in chemical industries. Control of the CSTR is used
to verify that a new control method has a strong ability to address nonlinear and
Z. Nassira · C. Mohamed (B)

Signal Processing Laboratory, Department Electronics, Faculty of Sciences of Technology,
Constantine 1 University, Constantine 2500, Algeria
Z. Nassira

https://doi.org/10.1007/978-981-15-6403-1_33
478 Z. Nassira and C. Mohamed
instability problems. The control design of CSTR systems have been widely studied
in the literature, by using control techniques ranging from classical approaches to
advanced ones. In [1, 2] robust PI stabilizing controllers are designed. State feed-
back controllers with state estimator and dynamic output feedback controllers are
designed in [3, 4] respectively. Standard nonlinear approaches are applied on CSTR
using feedback liearization and based on high gain and variable structure observers
[5–7]. More elaborated controllers are developed to deal with uncertainties in CSTR
systems such as fuzzy PID controller in [8] and fault tolerant control approach in [9]
to deal with CSTR uncertainties and faults occurrence, respectively. In [10] fractional
order PID based on firefly algorithm for optimal parameters settings is used in the
control of CSTR systems. The merits of the above mentioned approximation-based
control are that the adaptive laws of adjustable weights of neural networks or fuzzy
logic systems can be obtained on the basic of Lyapunov theorem, which guarantees
the stability of closed-loop system. Despite these efforts, the existing results [5–10]
were focused on uncertain CSTR systems without considering the input nonlinearity.
Practical engineering often requires the proposed control schemes to satisfy
certain quality of the performance indices such as overshoot, convergence rate, and
steady-state error. To address this problem, numerous existing works, including the
Barrier Lyapunov Function (BLF), demonstrate that it can guarantee the constraints
satisfaction. In [11, 12], a robust adaptive controller is designed for CSTR where
the barrier Lyapunov function is employed in the synthesis to prevent the velocity
constraint violation. On the other hand, an error transformation method, which was
originally presented in the prescribed performance control (PPC), has been used to
deal with constraints [13, 14]. Despite these efforts, the existing results [12–14] are
focused on uncertain CSTR systems without considering any input nonlinearities.
It is well known that input nonlinearities exist in real world. CSTR systems are
not an exception for which, in this paper, dead-zone nonlinearity is considered.
The existence of input dead-zone may lead to the instability of the closed-loop
system and hence makes the stabilization problem more difficult. Valuable works
on adaptive fuzzy/neural based controller design for CSTR dynamics preceded by
unknown dead-zone are proposed [15–17]. In [15] an adaptive NNs controller is
designed based on “n” NNs for the online learning of the CSTR dynamics. In [16, 17],
the non-affine pure-feedback discrete time nonlinear systems with non-symmetric
dead-zone inputs are transformed into a n-step ahead predictors where adaptive NNs
compensative terms are used to compensate for the dead-zone parameters. However,
from a point of view of practical applications, the methods developed in [15–17] are
computationally expensive because of the NNs and fuzzy approximators are used
at every step to online approximate the unknown dynamics, and hence render the
control laws computationally heavy and the stability analysis more complicated.
Motivated by the preceding remarks, in this paper, we are inspired to propose a
robust adaptive control approach for CSTR systems with input nonlinearity, output
constraints and unknown disturbance. To ensure constraints satisfaction with a priori
prescribed performance, we employ a system transformation technique to trans-
form the original constrained system into an equivalent unconstrained model whose
stability is sufficient to achieve the tracking control of the original system, and
guaranteeing the prescribed performance. Input dead-zone nonlinearity along with
Adaptive Neural-Network Control Design for Uncertain CSTR … 479
unknown external disturbance are considered in this tracking control design. In order
to compensate for the effects of dead-zone input and external disturbance, we intro-
duced an adaptive robust control structure. In the proposed control designs, NNs are
used to approximate the unknown nonlinearity based on their universal approxima-
tion properties. The adaptive NNs tracking controllers are designed with only one
adaptive parameter by using the common Lyapunov stability theorem. It is shown that
the proposed adaptive control methods can guarantee that all the signals in the closed-
loop system are bounded, and the tracking errors converge to a small neighborhood
of the origin. The main contributions of this paper are summarized as follows.
(a) The proposed approach is much simple since it avoids the well known explo-
sion of complexity problem of the back-stepping based methods caused by the
repeated differentiations of virtual controllers. This simplicity is explained by
simpler derivation of the adaptive laws without virtual controllers and hence
simpler stability analysis.
(b) The number of parameters of the proposed adaptive laws is highly reduced due to
the use of only one NNs, however, the traditional back-stepping based methods,
as in [11, 15–17], require tow NNs units with tow virtual controllers and their
derivatives leading to more complicated control laws with high computation
burden.
(c) The stability analysis of the closed loop and the derivation of the adaptive
learning laws of the proposed controllers are highly simplified.
(d) The simplified proposed work permits to cope with both of input nonlinearities
and output constraints along with unknown disturbances. In the contrary, the
back-propagation based approaches with recursive design procedures deal only
with input nonlinearities of CSTR in [15–17] and only output constraints of
CSTR in [11, 12].
(e) Furthermore, the proposed controllers guarantee the stability of the closed loop
of CSTR systems with a priori prescribed performance imposed by the designer
as shown theoretically and in a simulation example which is not the case for
most of the approaches in the literature.
The rest of the paper is organized as follows. The dynamics of the CSTR is
described in Sect. 2. In Sect. 3, we design the controls for the unknown input dead-
zone, while in Sect. 4, simulation results are presented to show the effectiveness of
the approach. Finally, conclusions are drawn in Sect. 5.
2 Problem Statement
2.1 Model Description
The dynamic of the process of CSTR considered in this paper can be written in the
following form [18, 19]:
⎧
⎪
⎪ ẋ1 = −2x1 + Da (1 − x1 )ex p[x2 /(1 + (x2 /ϕ))] + x1
⎨
ẋn = −x2 (1 − δ) + B Da (1 − x1 )ex p[x2 /(1 + (x2 /ϕ))]
(1)
⎪
⎪ δu(v) + d(t)
⎩
y = x1
where x1 and x2 denote the dimensionless reactant concentration and mixture temper-
ature, respectively; u ψ is the dimensionless coolant flow rate; Da , ϕ, B ψ and δ
ψ denote the Damokhler number, the activated energy, heat of reaction, and heat
transfer coefficient, respectively.
Then, we can rewrite (1) as
⎧
⎨ ẋ1 = f 1 (x1 , x2 ) + x1
ẋ = f 2 (x1 , x2 ) + δu(v) + d(t) (2)
⎩ 2
y = x1
with
f 1 (x̄2 ) = −x1 + Da (1 − x1 )ex p[x2 /(1 + (x2 /ϕ))] − x2

f 2 (x̄2 ) = −x2 (1 − δ) + B Da (1 − x1 )ex p[x2 /(1 + (x2 /ϕ))] (3)
where d(t) is the unknown external disturbance term, which is bounded; f i (., .) is the
unknown smooth function. The input nonlinearity u(v) which includes dead-zone.
To facilitate the control design, the following standard Assumptions and Lemmas
are introduced for system (2).
Assumption 1 For the system (2), there exists an unknown positive constant d̄ such
that |d(t)| < d̄.
Lemma 1 [7]: For ∀ς > 0 and ∀q ∈ R, the following inequality holds

ς
0 ≤ |q| − q tanh ≤ κς (3)
q
where κ = 0.2785
Lemma 2 (Young’s inequality [20]) For ∀(x, y) ∈ R 2 , the following inequality

holds:
εp p 1
xy ≤ |x| + q |y|q , (4)
p pε
where ε > 0, p > 1, q > 1, and ( p − 1)(q − 1) = 1.

In this paper, the control objective is twofold. Design a control input u(t) for the
CSTR system in (1) such that:
P1. The system output y(t) follows the desired output signal yd (t), while all the
signals of the closed-loop system are bounded.
P2. Both prescribed transient and steady state performance of the tracking error
e(t) = x1 (t) − yd (t) are achieved.
2.2 Dead-Zone Model and Its Properties
The control input with a dead-zone, is described by [15] as follows:

⎧
⎨ m r (v(t) − br ) v(t) ≥ br
u(v) = 0 br ≤ v(t) ≤ bl (5)
⎩
m l (v(t) − bl ) v(t) ≤ bl
where m r and m l are the general slope of the dead zone characteristic. It is assumed
that m = m r = m l ,br and bl denote the right and left dead zone breakpoints,
respectively.
Assumption 2 ([15]) The strictly positive constants m, dr and dl are unknown, but
their signs are known (m > 0, dr ≥ 0 and dl ≥ 0 respectively).
Assumption 3 ([15]) The dead zone parameters are bounded by known constants

m min ,m max , drmin , drmax , dlmin , dlmax such that m ∈ m min , m max , dr ∈
[drmin , drmax ] and dl ∈ [dlmin , dlmax ].
As a result, the output of the dead zone can be reformulated as
u(v) = mv(t) + h(v(t)) (6)
where h(v(t)) can be calculated from (5) and (6) as follows

⎧
⎨ −m r br v(t) ≥ br
h(v(t)) = −m(t)v(t) br ≤ v(t) ≤ bl (7)
⎩
−m l bl v(t) ≤ bl
From Assumption 2 and 3 we can conclude that h(v(t)) is bounded and satis-
fies
|h(v)| ≤ ρ, where ρ isthe upper-bound, which can be chosen as ρ =
max m max drmax , −m min dlmin .
2.3 Neural Networks
Neural networks (NNs) has been frequently used as function approximators, in this
study, a continuous function f (Z ) : R + → R will be approximated by neural
networks (NN) with the output given by:

fˆ Z /Ŵ = Ŵ T ∅(Z ) (8)
where Z ∈ Z ∈ R m is the input to NN, Ŵ ∈ RN is called the weight vector, and

N is the number of network nodes, ∅(Z ) = [∅1 (Z ), . . . , ∅ N (Z )]T is the so-called
vector of activation function chosen as the hyperbolic tangent function in our study

∅i (Z ) = 1 − e−Z / 1 + e−Z , i = 1, . . . , N (9)
According to the approximation property of the NNs in [21], any continuous

function over a compact set Ωz , there exists an integer N and optimal synaptic weight
vector W ∈ R N such that:

W ∗ = argmin W supz∈Ωz fˆ Z /W ∗ − f (Z ) (10)
In other terms, if the nodes number N is sufficiently large and the nodes number
terms are appropriately selected, then there exists a weight vector W such that
W T ∅(Z ) can approximate
f (Z ) to any
degree of accuracy, in a given compact set. In
ˆ
general, supz∈Ωz f (Z /W ) − f (Z ) becomes smaller as N increases. More precisely,
we may substitute, with no loss of generality, the unknown function by a linear in the
weights neural network structure, plus a modeling error term ∀Z ∈ z , obtaining:
f (Z ) = W ∗T ∅(Z ) + ε(Z ) (11)
where W ∗ is the optimal weights vector which minimizes the approximation error
ε(Z ). Owing to the approximation capabilities of the linear in the weights neural
network, there exists an unknown constant ε̄ > 0 such that:
|ε(Z)| ≤ ε̄, ∀Z ∈ z (12)
From the universal approximation results for neural networks stated in [22], it is
known that the constant ε̄ can be made arbitrarily small by increasing the NN nodes
number N.
3 Design of Adaptive Neural Prescribed Performance

Control with Input Nonlinearities
3.1 Prescribed Performance Function
To achieve the second objective P2, as defined in [13], a positive decreasing smooth
function μ(t) : R + → R + with lim μ(t) = μ∞ > 0 will be used as the prescribed
t→∞
performance function (PPF). As stated in [13, 14], the tracking error e(t) = x1 (t) −
yd (t) is satisfied using the following predefined inequality:
−δμ(t) < e(t) < δ̄μ(t), ∀t > 0 (13)
where δ and δ̄ > 0 are constants selected by the designer.

In this study, the performance function μ(t) is set as [13]:
μ(t) = (μ0 − μ∞ )e−κt + μ∞ (14)
where μ0 > μ∞ and κ > 0 are design parameters.

In (13) and (14), we know that δ̄μ0 defines the upper bound of the maximum
overshoot and −δμ0 defines the lower bound of the undershoot, the decreasing rate
κ introduces a lower bound on the convergence speed, and μ∞ denotes the allowable
steady-state tracking error [13]. Hence, the transient and steady state performances
can be designed a priori by tuning the parameters δ̄, δ, κ, μ0 and μ∞ .
To solve the control problem with the prescribed performance (13), and according
to [13], an output error transform will be introduced by transforming condition (13)
into an equivalent unconstrained one. In this study, we define a smooth, strictly
increasing function T (ξ1 ) of the transformed error ξ1 ∈ R, such that:
(1) − δ < T(ξ1 ) < δ̄, ∀ξ1 ∈ L∞
(2) lim T (ξ1 ) = δ̄ lim T (ξ1 ) = δ

ξ1 →+∞ ξ1 →−∞
From the properties of T(ξ1 ), condition (13) leads to:
e(t) = μ(t)T (ξ1 ) (15)
Since T (ξ1 ) is strictly monotonic increasing and the fact that μ(t) ≥ μ∞ > 0
holds according to (14), the inverse function of T (ξ1 ) exists and can be deduced as

−1 e(t)
ξ1 = T (16)
μ(t)
Note that PPF (14), T (ξ1 ) and the associated parameters δ̄, δ, κ, μ0 and μ∞
are all a priori designed. For any initial condition e(0), if parameters δ̄, μ0 and
δ.are selected such that −δμ(0) < e(0) < δ̄μ(0) and ξ1 can be controlled to be
bounded (i.e., ξ1 ∈ L ∞ , ∀t > 0), then −δ < T (ξ1 ) < δ̄ holds; thus, the condition
−δμ(t) < e(t) < δ̄μ(t) is guaranteed. Consequently, the tracking control problem
of system (1) is now transformed to stabilize the transformed system (16).
T −1 (.) denotes the inverse function of T (.). A typical choice of T (ξ1 ) can be
given by
δ̄eξ1 − δe−ξ1
T (ξ1 (t)) = (17)
eξ1 + e−ξ1
Then, from (16), the transformed error ξ1 is derived as

−1 e(t) 1 ρ(t) + δ
ξ1 = T = ln (18)
μ(t) 2 δ̄ + ρ(t)
where ρ(t) = e(t)/μ(t)

To stabilize the error system ξ1 and thus to achieve the guaranteed performance
of error e, we further deduce

∂ T −1 (.) 1 1 1 ė eμ̇
ξ̇1 = ρ̇(t) = − −
∂ρ(t) 2 ρ(t) + δ ρ(t) − δ̄ μ μ2

eμ̇
= ϕ x2 + f 1 (x1 , x2 ) − ẏd − (19)
μ

where ϕ = 1
− ρ(t)−
1
1
ρ(t)+δδ̄
can be calculated based on e(t) and μ(t) and fulfills
2μ

0 < ϕ < ϕ M = δ + δ̄ /μ∞ δ̄δ .
Moreover, one may obtain that

eμ̇
ξ̈1 = ϕ̇ x2 + f 1 (x1 , x2 ) − ẏd −
μ

∂ f 1 (x̄2 ) ∂ f 1 (x̄2 ) ėμ̇ eμ̈μ eμ̇2
+ ϕ ẋ2 + ẋ1 + ẋ2 − ÿd − − 2 + 2
∂ x1 ∂ x2 μ μ μ

eμ̇
= ϕ̇ x2 + f 1 (x1 , x2 ) − ẏd −
μ

∂ f 1 (x̄2 ) ∂ f 1 (x̄2 ) ėμ̇ eμ̈μ eμ̇2
−ϕ ẋ1 + ẋ2 − − 2 + 2
∂ x1 ∂ x2 μ μ μ
+ ϕ[ f 2 (x1 , x2 ) + δu(v) + d(t) − ÿd ] (20)
Define the filtered error as

T
es = [ 1] ξ1 , ξ̇1 (21)
where > 0 is a positive constant such that the tracking error ξ1 is bounded as long
as es is bounded.
Consequently, we have

eμ̇
ės = ( ϕ + ϕ̇) x2 + f 1 (x1 , x2 ) − ẏd −
μ
+ ϕ[ f 2 (x1 , x2 ) + δu(v) + d(t) − ÿd ]

∂ f 1 (x̄2 ) ∂ f 1 (x̄2 ) ėμ̇ eμ̈μ eμ̇2
−ϕ ẋ1 + ẋ2 − − 2 + 2
∂ x1 ∂ x2 μ μ μ
= ϕ[F(x1 , x2 , ẏd , ÿd , e, ϕ) + δu(v) + d(t) − ÿd ] (22)

where F(x1 , x2 , ẏd , e, ϕ) = ( + ϕ̇/ϕ) x2 + f 1 (x1 , x2 ) − ẏd − eμμ̇ −

∂ f 1 (x̄2 ) ∂ f 1 (x̄2 ) ėμ̇ eμ̈μ eμ̇2
∂ x1
ẋ 1 + ∂ x2
ẋ 2 − μ
− μ 2 + μ 2 + f 2 (x1 , x2 ) denotes the unknown
nonlinear function, which are approximated by the NNs (11).
F(x1 , x2 , ẏd , e, ϕ) = W ∗T ∅(Z ) + ε(Z ), (23)
where Z = [x1 , x2 , ẏd , e, ϕ]T
3.2 Stability Analysis
Consequently, taking (23) into (22) yields

ės = ϕ W1∗T ∅(Z ) + ε1 (Z ) + δu(v) + d(t) − ÿd (24)
Substituting (25) into (27) yields

ės = ϕ W1∗T ∅(Z ) + δmv(t) − ÿd + δh(v) + D1 (t) (25)
where D1 (t) = ε1 (Z ) + d(t).

It follows from Assumptions 1 that there exists an unknown upper bound τ ∗ such
that: |D1 | ≤ τ1∗ .
Furthermore, the adaptive NNs controller is chosen to be:

1 α1 es2 τ̂12
v(t) = − es + ÿd − Ŵ1 ∅(z) −
T
δm ϕ es τ̂1 tanh(es /) +
+ vc (t) (26)
where vc (t) is a compensative control term for h(v) and it is chosen as
vc (t) = −κ ∗ sign(es ) (27)
where κ ∗ is a control gain satisfying κ ∗ ≥ ρ/m min .

with the corresponding adaptive control laws
.
Ŵ = ϕ1 es ∅(Z ) + σ1 Ŵ1 (28)
1
.
τ̂ = ϕγ1 −|es | + σ2 τ̂1 (29)
1
where Ŵ1 and τ̂1 are the estimate values of the unknown parameters W1∗ and τ1∗ ,
respectively, and Γ1 ,γ1 , σ1 , σ2 , α1 are positives constants.
The stability of the closed-loop system will be proved via Lyapunov analysis.
Theorem 1 Consider the CSTR system (1) subject to the input dead-zone (6). Based
on Assumptions 1–4, consisting the controller (26), the compensative control term
(27), the adaptive laws (28)–(29), while the design parameters are chosen appro-
priately, we can conclude that all the signals in closed loop of the CSTR system are
bounded, and the tracking errors will converge to a small compact.
Proof Define the following Lyapunov function candidate:
1 2 1 T −1 1 2
V1 = es + W̃1 Γ1 W̃1 + τ̃ (30)
2 2 2γ1 1
where τ̃1 = τ1∗ − τ̂1 and W̃1 = W1∗ − Ŵ1 are the parameter estimation errors. The
derivative of (30) is given by:
. .
1
V̇1 = es ės − W̃1T Γ1−1 Ŵ − τ̃1 τ̂
1 γ1 1
∗T
= es ϕ W1 ∅(Z ) + δmv(t) − ÿd + δh(v) + D1 (t)
. .
1
− W̃1T Γ1−1 Ŵ − τ̃1 τ̂ (31)
1 γ1 1
According to the control law (26) and parameter adaptation laws (28)–(29), one has
V̇1 = −α1 es2 + ϕδmes {vc + h(v)/m} + ϕes D1 (t)

.
ϕes2 τ̂12 −1 1 .
− + W̃1 ϕes ∅(Z ) − Γ1 Ŵ − τ̃1 τ̂
T
es τ̂1 tanh(es /) + 1 γ 1
ϕes τ̂1
2 2
≤ −α1 es2 − + ϕ|es |τ1∗ + ϕ|es |τ̃1
es τ̂1 tanh(es /) +
− σ2 ϕ τ̃1 τ̂1 − σ1 ϕ W̃1T Ŵ1 (32)
Using Young’s inequality and invoking Lemma 2, we have

2
ϕ M σ1

∗ 2
−σ1 ϕ W̃1T Ŵ1
≤ W1 − W̃1 (33)
2
ϕ M σ2 ∗2
−ϕσ2 τ̃1 τ̂1 ≤ τ1 − τ̃12 (34)
2
In addition, using the following inequality, 0 ≤ xtanh(x/a) ≤ |x|, ∀x ∈ R, a > 0,
and by substituting (33) and (34) into (32), we obtain:
ϕ M ε|es |τ̂1 ϕ M σ1
2 ϕ M σ2 2
V̇1 ≤ −α1 es2 + − W̃1 − τ̃
|es |τ̂1 + 2 2 1
ϕ M σ1
W ∗ 2 + ϕ M σ2 τ ∗2
+ 1 (35)
2 2 1
Then, using the inequality 0 < x

x+y
< 1, ∀x ≥ 0, y > 0, it yields
ϕ M σ1
2 ϕ M σ2 2
V̇1 ≤ −α1 es2 − W̃1 − τ̃
2 2 1

ϕ M σ1 ∗ 2 ϕ M σ2 ∗2
+ W1 + τ + ϕM
2 2 1
≤ −θ1 V1 + β1 (36)

ϕ M σ1
where θ1 = min 2α1 , λ ,ϕ σ γ
max (Γ1−1 ) M 2 1
ϕ M σ1
W ∗ 2 + ϕ M σ2 τ ∗2 + ϕ M
β1 = 1
2 2 1
Moreover, by multiplying eθ1 t and integrating on both sides of (36) over [0, t], it
follows that:

β1 β1 −θ1 t β1
V1 (t) ≤ + V1 |t=0 − e ≤ V1 |t=0 + (38)
θ1 θ1 θ1
Combining (30) with (38), we can conclude that:

|es | ≤ 2( V1 |t=0 + θ1 /β1 )

W̃1 ≤ 2( V1 |t=0 + θ1 /β1 )/λmin Γ1−1

|τ̃1 | ≤ γ1 2( V1 |t=0 + θ1 /β1 ) (39)
Consequently, the parameter estimate errors W̃1 , τ̃1 and tracking error es are
bounded from (39), and remain in the compact sets Ωτ̃1 , ΩW̃1 and Ωes defined by:
√
Ωes = es ∈ R||es | ≤ Ω

−1
ΩW̃1 = W̃1 ∈ R W̃1 ≤ Ω/λmin Γ1

Ωτ̃1 = τ̃1 ∈ R||τ̃1 | ≤ γ1 Ω
where Ω = 2( V1 |t=0 + θ1 /β1 ) with θ1 and β1 defined as (33). The proof is completed.
4 Simulations
In order to show the effectiveness of the proposed method, simulation study is

described in this section. Consider the CSTR system excited by boundary distur-
bance d(t, x) = cos(t)cos(x2 ). The parameters of CSTR model are chosen as
Da = 0.072, δ = 0.3, ϕ = 20 and B = 8.The objective is to apply the proposed
adaptive NN controller such that the system output y tracks the desired trajectory
yd = 0.5 + 0.1sin(t).
In the simulation study, one NN system in the form of (8) is used for each case to
generate the unknown nonlinearities of the nonlinear CSTR system. The NNs input
vector is Z = [x1 , x2 , ẏd , e, ϕ]T , with N = 2 network nodes, and the activation
function ∅(.) is chosen as the hyperbolic tangent function in our study. The design
parameters used in this simulation are chosen by trial and error as follows:Γ1 =
20, Γ2 = 35, σ1 = σ3 = 0.1, σ2 = σ4 = 20, γ1 = γ2 = 10, = 0.1, α1 =
1.2,α2 = 2.8, κ ∗ = 4.5, λ = 10. The initial conditions are chosen as x1 (0) = 0.6,
x2 (0) = 0.5, τ̂1 (0) = τ̂2 (0) = 0, Ŵ1 (0) = Ŵ2 (0) = 0. The dynamic responses of
the CSTR system are simulated for two cases. In the first case, the CSTR dynamics
are simulated with the following input dead zone parameters: mmin = 0.6, mmax =
0.7, br = 2.09, bl = −1.5. The transient and steady state tracking error bounds are
prescribed through the performance function μ(t) = (μ0 − μ∞ )e−κt + μ∞ with the
parameters μ0 = 1.5, μ∞ = 0.05, κ = 0.09, δ = δ̄ = 1.
The trajectories of CSTR output along with the reference signals when applying
the proposed controllers are shown in Fig. 1. The boundedness of the control signals
v(t) and N (v(t)) is illustrated in Fig. 2. From the simulation results, we can observe
that the proposed methods are able to control CSTR system with effective perfor-
mance in the presence of dead-zone input nonlinearities and preserving smooth
trajectories for all the states and the control input signals. Indeed, the tracking errors
are guaranteed to be inside the imposed bounds (14) induced by the prescribed
performance function (13) as shown in Figs. 2.
0.7
y yd
0.6
0.5
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30
Time (sec)
Fig. 1 Tracking performance: desired trajectory yd , output y
20
v(t) u(v(t))
10
-10
-20
-30
-40
0 5 10 15 20 25 30
Time (sec)
Fig. 2 Trajectory of v(t) and N (v(t))
To further show the effectiveness of the proposed controller, a comparison is

conducted with traditional back-stepping CSTR control with input nonlinearities and
without considering the output constraints [15]. The obtained results when applying
the controller defined in [15] with input dead zone and without output constraints are
shown in Fig. 3. It is very clear that our approach outperform the approach proposed
in [15] since the obtained response in Fig. 3 shows that the prescribed performance
(13) is guaranteed.
5 Conclusions
In this paper, adaptive NN controller with dynamic robust control terms was proposed
for CSTR systems subjected to input dead-zone nonlinearity, output constraint and
external disturbance. Compared to the traditional back-stepping control, it was
proved that the developed controller exhibited more simplified control algorithms
0.1
Proposed method Back-stepping method PPF bound
0.08
0.06
0.04
0.02
0
-0.02
-0.04
-0.06
0 5 10 15 20 25 30
Time (sec)
Fig. 3 Tracking error of different controllers
with reduced design parameters thanks to the elimination of the causes of explosion
dimensionality problem. Then, the proposed adaptive neural tracking controller
was designed with only one adaptive parameter by using the second Lyapunov
stability method. Adaptive robust control term was used to deal with neural networks
approximation errors and external disturbance term. It was proved that the proposed
method was able to effectively control CSTR systems with guaranteed, a priori,
prescribed performances. Simulation results showed the effectiveness of the proposed
method.
References
1. Alvarez-Ramirez J, Femat R (1999) Robust PI stabilization of a class of chemical reactors.

Syst Control Lett 38(4–5):219–225
2. Pérez M, Albertos P (2004) Self-oscillating and chaotic behaviour of a PI-controlled CSTR
with control valve saturation. J Process Control 14(1):51–59
3. Viel F, Jadot F, Bastin G (1997) Robust feedback stabilization of chemical reactors. IEEE Trans
Automatic Control 42(4):473–481
4. Antonelli R, Astolfi A (2003) Continuous stirred tank reactors: easy to stabilise? Automatica
39(10):1817–1827
5. Biagiola SI, Figueroa JL (2004) A high gain nonlinear observer: application to the control of
an unstable nonlinear process. Comput Chem Eng 28(9):1881–1898
6. Jana AK, Samanta AN, Ganguly S (2005) Globally linearized control on diabatic continuous
stirred tank reactor: a case study. ISA Trans 44(3):423–444
7. Daaou B, Mansouri A, Bouhamida M, Chenafa M (2012) Development of linearizing feedback
control with a variable structure observer for continuous stirred tank reactors. Chin J Chem
Eng 20(3):567–571
8. So GB, Jin GG (2018) Fuzzy-based nonlinear PID controller and its application to CSTR.
Korean J Chem Eng 35(4):819–825
9. Wang ZY, Wang GX (2017) Temperature fault-tolerant control system of CSTR with coil
and jacket heat exchanger based on dual control and fault diagnosis. J Central South Univ
24(3):655–664
10. Alamdar Ravari M, Yaghoobi M (2019) Optimum design of fractional order pid controller
using chaotic firefly algorithms for a control CSTR system. Asian J Control
11. Li DJ, Li DP (2015) Adaptive controller design-based neural networks for output constraint
continuous stirred tank reactor. Neurocomputing 153:159–163
12. Li DJ (2015) Adaptive neural network control for a two continuously stirred tank reactor with
output constraints. Neurocomputing 167:451–458
13. Bechlioulis CP, Rovithakis GA (2008) Robust adaptive control of feedback linearizable MIMO
nonlinear systems with prescribed performance. IEEE Trans Autom Control 53(9):2090–2099
14. Bechlioulis CP, Rovithakis GA(2009) Adaptive control with guaranteed transient and steady
state tracking error bounds for strict feedback systems. Automatica, 45(2): 532–538
15. Li D (2014) Adaptive neural network control for a class of continuous stirred tank reactor
systems. Sci Chin Inf Sci 57(10):1–8
16. Li S, Gong M, Liu Y (2016) Neural network-based adaptive control for a class of chemical
reactor systems with non-symmetric dead-zone. Neurocomputing 174:597–604
17. Li DJ, Tang L (2014) Adaptive control for a class of chemical reactor systems in discrete-time
form. Neural Comput Appl 24(7–8):1807–1814
18. Zerari N, Chemachema, M, Essounbouli N (2018) adaptive neural-network output feedback
control design for uncertain cstr system with input saturation. In 2018 (CISTEM), pp 1–6.
IEEE
19. Henson MA, Seborg DE (1990) Input-output linearization of general nonlinear processes.
AIChE J 36(11):1753–1757
20. Polycarpou MM, Ioannou PA (1993) A robust adaptive nonlinear control design. In: 1993
American control conference, pp 1365–1369. IEEE
21. Park J, Sandberg IW (1991) Universal approximation using radial-basis-function networks.
Neural Comput 3(2):246–257
22. White DA, Sofge DA (eds) (1992) Handbook of intelligent control: neural, fuzzy, and adaptative
approaches. Van Nostrand Reinhold Company
Adaptive Fuzzy Fault-Tolerant Control
Using Nussbaum Gain for a Class
of SISO Nonlinear Systems
with Unknown Directions
Abdelhamid Bounemeur, Mohamed chemachema, Abdelmalek Zahaf,
and Sofiane Bououden
Abstract This paper treatsan indirect adaptive fuzzy fault-tolerant control using
fuzzy systems for a class of uncertain SISO systems with unknown control gain
sign and actuator faults. The uncertain nonlinearities of the systems and the actuator
faults are approximated by fuzzy systems that have been proven to be universal
approximators and the Nussbaum-type function is used to deal with the unknown
control gain sign. The proposed control scheme completely overcomes the singularity
problem that occurs in the indirect adaptive feedback linearizing control. Projection
in the estimate parameters is not required and the stability analysis of the closed-loop
system is performed using Lyapunov approach. Simulation results are provided to
verify the effectiveness of the proposed design.
Keywords Adaptive fuzzy control · Nonlinear systems · Lyapunov stability
1 Introduction
Universal approximators have been successfully applied to adaptive control problems

such as fuzzy systems [1, 2] and neural networks [3–5] because they need no accurate
mathematical models of the system under control. Fuzzy systems have been proven
to be able to approximate continuous nonlinear functions [6]. Based on this property
A. Bounemeur (B) · M. chemachema · A. Zahaf

Laboratory of Automatic, Robotic and Control Systems, Faculty of Engineering,
Department of Electronics, University of Constantine 1, Constantine, Algeria
M. chemachema
A. Zahaf
S. Bououden
Faculty of Sciences and Technology, University Abbes Laghrour, Khenchela, Algeria

https://doi.org/10.1007/978-981-15-6403-1_34
494 A. Bounemeur et al.
many researchers have presented adaptive control architecture for uncertain nonlinear
systems [2, 7–9]. Conceptually, there are two distinct approaches that have been
formulated in the design of a fuzzy adaptive control system: direct and indirect
schemes. In the direct approach which consists to approximate the ideal control law
by a fuzzy system [7, 8]. However, in the indirect approach the nonlinear dynamics
of the system are approximated by fuzzy systems to develop a control law based on
these systems [7, 9, 10].
However, all the previous results need the assumption that the input gain, which is
the function of the system states in general, is away from zero and its sign is known
a priori. The sign, called control direction, represents motion direction of the system
under any control, and knowledge of this sign makes adaptive control design much
easier. In the adaptive control literature, the unknown control direction is mainly
solved by using the Nussbaum-type function [11–14], and it has been successfully
applied for robust control design [12, 13], and adaptive control design for nonlinear
systems [15–18]. An alternative method, the so-called correction vector approach
[19], has been applied to first-order nonlinear systems [20]. Authors in [21], have
investigated an indirect adaptive fuzzy fault-tolerant scheme for a class of nonlinear
systems with both actuator and sensor faults. A combination of fuzzy systems and
backstepping technique allowed the online estimation of all adaptive parameters and
ensured the boundedness of all signals involved in the closed-loop system.
Inspired by [20, 21], we propose in this paper an indirect adaptive fuzzy controller
for a class of SISO nonlinear systems to achieve tracking of a desired output. New
learning algorithms are proposed in the presented controller which permits superior
control performance compared to the same class of controllers. In the proposed
controller, a robustifying control term is added to the basic fuzzy controller to deal
with approximation errors. The Nussbaum gain function is introduced to solve the
problem of unknown control gain sign and the regularized inverse is employed to
solve problem of singularity. The stability of the closed-loop system is studied using
Lyapunov method.
The outline of the paper is as follows. Section 2 presents a brief description of
the fuzzy system. In Sect. 3, a new control law and adaptive algorithms are proposed
and stability analysis is given. Simulation examples are illustrated in Sect. 4. The
conclusion is finally given in Sect. 5.
2 Description of Fuzzy Systems
It is shown that fuzzy systems are capable of approximating any real continuous
function on a compact set with arbitrary precision given by [6].
Sugeno and the employee have proposed a class of fuzzy system that allows to
represent knowledge that is expressed in analytical form, describing the internal
structure of system.
This class of fuzzy system is called Fuzzy systems Takagi-Sugeno (TS).
Adaptive Fuzzy Fault-Tolerant Control Using Nussbaum … 495
Note by x = [x1 , . . . , xn ]T is the input of fuzzy system and its output. For each
j
xi is associated m i fuzzy sets Fi in X i , as for xi ∈ X i there is at least one degree
of membership μ F j (xi ) = 0 where i = 1, 2, . . . , n et j = 1, 2, . . . , m i . The basic
i n
rules of the fuzzy system has N = i=1 m i fuzzy rules of the form:
k k
Rk : i f x1 is F 1 and . . . and xn is F n T hen y = f k (x), k = 1, . . . , N
k
where F i ∈ Fi1 , . . . , Fim i et f k (x) is a numerical function on the output space in
general, f k (x) is a polynomial function depending on variable inputs, but it can also
be an arbitrary function so that it can properly describe the behavior of the studied
system if f k (x) is a function:

n
f k (x) = a0k + aik xi (1)
i=1
Then it’s the first order Tackagi-Sugeno (TS1). If against.

f k (x) is a polynomial of zero order.
f k (x) = a k (2)
We have the Tackagi-Sugeno zero order (TS0).

In this work we will consider a fuzzy zero order (TS0).
Each rule has a numerical conclusion, the total output of the fuzzy system is
obtained by calculating a weighted average, and in this manner the time consumed
by the procedure of défuzzification is avoided. Then the output of fuzzy system is
given by following relationship:
N
k=1 μk (x) f k (x)
y(x) = N (3)
k=1 μk (x)
n k 1 mi
with μk (x) = μ
i=1 k, F
i ∈ Fi , . . . , Fi .
Fi
Which represents the degree of confidence or activation rule Rk .

We can simplify the output of fuzzy system as follows:
N
μk (x)a k
y(x) = k=1
N
(4)
k=1 μk (x)
By introducing the concept of fuzzy basis functions [1], the output of fuzzy system
TS0 can be written as:
y(x) = w T (x)θ (5)
with

• θ = a 1 . . . a N : Vector of parameters of the conclusion of rules fuzzy part.
• w(x) = [w1 (x) . . . w N (x)]T , where
μk (x)
w N (x) = N , k = 1, . . . , N (6)
j=1 μ j (x)
3 Controller Design and Stability Analysis
1. Problem Formulation
Consider the class of nonlinear systems that can be described by the following
deferential equations
⎧
⎪
⎪ ẋ1 = x2
⎪
⎪
⎪
⎨ ẋ2 = x3
.. (7)
⎪ .
⎪
⎪
⎪
⎪ ẋ = f (x1 , x2 , . . . xn ) + g(x1 , x2 , . . . xn )u
⎩ n
y = x1
Or equivalently

y (n) = f (x) + g(x)u
(8)
y = x1
where x = [x1 , . . . , xn ]T ∈ Rn , is the vector of the system; u ∈ R is the scalar

control input; y ∈ R is the scalar system output; f (x) and g(x) are unknown smooth
nonlinear functions.
2. Actuator Faults Model
The following Table 1 describe the studied faults.

where t f i denotes the time instant of failure
of the
ith sensor/actuator and
bidenotes
its accuracy coefficient such that b −b̄0i , b̄0i , where b̄0i > 0. k k̄i , 1 , where
k̄i > 0 denotes the minimum sensorand actuator effectiveness,
in which bi and k are
slowly varying respectively within −b̄0i , b̄0i and k̄i , 1 .
Table 1 Actuator faults

Actuators Faults kinds Conditions Faults names
u(t) u(t) + b b(t) = 0, bias (Lock in place)

b t f i = 0
u(t) + b(t) |b(t)| = λt, dri f t
0<λ1
f or all t ≥ t f i
u(t) + b(t) |bi (t)| < b̄0i , Loss o f accuracy
ḃi (t) → 0
k(t)u(t) 0 < k̄i ≤ k (t) ≤ 1 Loss o f e f f ectiveness
The faulty actuator will be written as
u(t) = k(t)u(t) + b(t)
If we introduce the actuator faults, Eq. (8) will take the following form

y (n) = f (x) + f a (x, u) + g(x)u
(9)
y = x1
with f a (x, u) = g((k − 1)u + b(t)).

In respect the dynamic of the system (9), the following assumptions will be made
Assumption 1 The order n of the system is know.
Assumption 2 The state vector is available for measurement.
Assumption 3 The control gain g(x) and its signare unknown, g ≤ |g(x)| ≤ ḡ with
g and ḡ are positive constants.
The objective is to design an adaptive fuzzy controller for system (9) such that
the system output y(t) follows a desired trajectory yd (t) while all signals in the
closed-loop system remain bounded.
Regarding the development of the control law, the following assumption should
also be made:
Assumption 4 The desired trajectory yd (t) and its time derivatives yd(i) , i = 1, . . . , n
are smooth and bounded.
Remark 1 In this paper, the nonlinear functions f (x), f a (x) and g(x) are unknown,
in this case two fuzzy systems of the form (5) are used, then we can write
⎧
⎪
⎨ fˆ x, θ f = w Tf (x)θ f

ĝ x, θg = wgT (x)θg (10)
⎩ fˆ x, θ = w T (x)θ
⎪
a fa fa fa
with w Tf , wgT , w Tf a are the vectors of fuzzy basic functions supposed properly fixed i
n prior by the user, θ f , θ f a and θg are vectors of the fitted parameters. The functions
f (x), f a (x) and g(x) can be expressed in terms of fuzzy approximations in the
following manner:
⎧
⎪
⎪
⎨ f (x) = fˆ x, θ ∗f + ε f (x)

g(x) = ĝ x, θg∗ + εg (x) (11)
⎪
⎪
⎩ f a (x) = fâ x, θ ∗ + ε f a (x)
fa
with ε f (x), f a (x) and εg (x) represent the fuzzy approximation errors, θ ∗f , θ ∗f a and θg∗
are respectively of the optimum parameters of θ f , θ f a and θg , the values of parameters
θ f , θ f a and θg respectively minimizing the approximation errors ε f (x) and εg (x).
These optimal parameters satisfy:

θ ∗f ˆ
= arg min sup f (x) − f x, θ f (12)
θf x

θg∗ = arg min supg(x) − ĝ x, θg (13)
θg x

θ ∗f a ˆ
= arg min sup f a (x) − f a x, θ f a (14)
θfa x
Note that the optimal parameters θ ∗f , θ ∗f a and θg∗ are unknown constants artificial
introduced only to the theoretical study ofthe stability of the control algorithm. Infact,
the knowledge of their values is not necessary for implementation of adaptive control
law. From the above analysis, we can write:

f (x) = fˆ x, θ ∗f + ε f (x) (15)

f a (x) = fâ x, θ ∗f a + ε f a (x) (16)

g(x) = ĝ x, θg∗ + εg (x) (17)
f (x) − fˆ(x, θ f ) = w Tf (x)

θ f + ε f (x) (18)
f a (x) − fâ (x, θ f a ) = w Tf a (x)

θ f a + ε f a (x) (19)
g(x) − ĝ(x, θg ) = wgT (x)

θg + εg (x) (20)
with

θ f = θ ∗f − θ f ,
θ f a = θ ∗f a − θ f a and
θ f = θg∗ − θg , are the parameter estimation
errors.
Assumption 5 The fuzzy approximation

errors ε f (x),
ε f a(x) and εg (x) are bounded
for all xx as ε f (x) ≤ ε f , ε f a (x) ≤ ε f a and εg (x) ≤ ε g , where ε f , ε f a and
ε g are unknown positive constants.
This assumption is reasonable, since we assume that fuzzy systems used for
approximating unknown functions have the universal approximator property.
Remark 2 Development that follows does not require knowledge of the sign of the
gain control, because the Nussbaum function technique will be used to tackle this
problem. The Nussbaum function technique was originally proposed in [11], and has
been used in controller design in solving the difficulty of unknown control [5].
A function is called a Nussbaum-type function if it has the following properties

[11, 13]:
⎛ ⎞
v
1
lim sup ⎝ N (τ )dτ ⎠ = +∞ (21)
v→+∞ v
0
⎛ ⎞
v
1
lim in f ⎝ N (τ )dτ ⎠ = −∞ (22)
v→+∞ v
0

For example, the continuous functions τ → eτ cos π2 τ and τ → τ 2 cos(τ )
2
are Nussbaum-type functions. In this paper, the even Nussbaum function τ 2 cos(τ )
is used.

Lemma1 ([14]) V (.) and τ (.) be smooth functions defined on 0, t f with V (t) ≥
0, ∀t ∈ 0, t f ; N (.) be an even smooth Nussbaum-type function. If the following
t
inequality holds for, ∀t ∈ 0, t f : V (t) ≤ c0 + 0 (g(ζ )N (τ (ζ )) + c1 )τ̇ (ζ )dζ where
g(t) is a piecewise continuous
time function which takes values in the unknown
closed interval I = g, ḡ with 0 ∈ / I ; c1 is any positive number and c0 repre-
t
sents some suitable
constant, then V (.), τ (.) and 0 (g(ζ )N (τ (ζ )) + c1 )τ̇ (ζ )dζ are
bounded on 0, t f .
The feedback control law and parameter adaptive laws are given by
u = u c + ur (23)
The control law (21) is the sum of two terms, the first u c is the adaptive control
term, and the second u r is the robustifying term
ĝ(x, θ)
uc = − fˆ(x, θ) − fâ (x, θ ) + y (n) + k T e + α ĝ(x, θ )N (τ )G (24)
ε0 + ĝ 2 (x, θ ) d
where ε0 is small positive constant, α positive constant and G = e T P B.

The robustifying term is given by
u r = u r b N (τ ) (25)
where
T
e P B ϕ − δ 2
ur b = (26)
eT P B
N (τ ) = τ 2 cos(τ ) (27)

ϕ = ε f + ε f a + ε g |u c − α N (τ )G| + |ū| (28)

ε0
ū = − fˆ x, θg − fâ (x, θ ) + yd(n) + k T e + α ĝ(x, θ )N (τ )G (29)
ε0 + ĝ x, θg
2
with the following parameter adaption laws

·
θ f = −γ f w f (x)e T P B (30)
·
θ f a = −γ f w f a (x)e T P B (31)
·
θg = −γg wg (x)e T P B(u c − α N (τ )G) (32)

ε̂˙ f = η f e T P B (33)

ε̂˙ f a = η f e T P B (34)

ε̂˙ g = ηg e T P B |u c − α N (τ )G| (35)
2
τ̇ = e T P Bu r b + αe T P B (36)
δ̇ = −σ0 δ (37)
where γ f > 0, γg > 0, η f > 0, ηg > 0, σ0 > 0 and δ(0) > 0.
Theorem 1 Consider the system (9) with the sign of the control gain assumed
unknown, we assume that the Assumptions 1–5 Are satisfied. The control law defined
by (23) and (24) with adaptions law (30–37) ensures the following properties:
• The tracking error and its derivatives converge to zero, e(i) (t) → 0 when for
i = 0, 1, . . . n − 1.
• The output of the system, and its derivatives up to the order (n −1) and the control
signal are bounded: .
Proof
e = yd − y (38)
e(n) = yd(n) − y (n) = yd(n) − f (x) − f a (x) − g(x)u (39)
using (21)
e(n) = yd(n) − f (x) − f a (x) − g(x)u c − g(x)u r (40)
by adding and subtracting
fˆ(x, θ ), fâ (x, θ ) and ĝ(x, θ )u c , (41)
becomes:

e(n) = yd(n) − f (x) − fˆ(x, θ )

− f a (x) − fâ (x, θ ) − g(x) − ĝ(x, θ ) u c
− ĝ(x, θ )u c − fˆ(x, θ ) − fâ (x, θ )

− g(x)u r (42)
using (17) and (18), (38) can be simplified to:
e(n) = yd(n) − w Tf (x)

θ f − ε f (x) − w Tf a (x)
θ f a − ε f a (x)

− wg (x)θg u c − εg (x)u c − ĝ(x, θ )u c
T
− fˆ(x, θ ) − fâ (x, θ ) − g(x)u r (43)

using the adaptive control law (22), (39) can be simplified to:
(n)
e(n) = yd − w Tf (x) θ f − ε f (x) − w Tf a (x) θ f a − ε f a (x) − wgT (x)
θg u c − εg (x)u c

ĝ(x, θ ) ˆ(x, θ ) − fâ (x, θ ) + y (n) + k T e + α ĝ(x, θ)N (τ )G
− ĝ(x, θ ) − f d
ε0 + ĝ 2 (x, θ )
− f (x, θ) − fâ (x, θ ) − g(x)u r
ˆ (44)
by adding and subtracting k T e and α ĝ(x, θ )N (τ )G then
e(n) = yd(n) − w Tf (x)

θ f − ε f (x) − w Tf a (x)
θ f a − ε f a (x)
ĝ 2 (x, θ ) fˆ(x, θ ) ĝ 2 (x, θ ) fâ (x, θ )
− wgT (x)
θg u c − εg (x)u c + +
ε0 + ĝ 2 (x, θ ) ε0 + ĝ 2 (x, θ )
ĝ 2 (x, θ )yd(n) ĝ 2 (x, θ )k T e
− − − fˆ(x, θ ) − fâ (x, θ ) − fâ (x, θ )
ε0 + ĝ 2 (x, θ ) ε0 + ĝ 2 (x, θ )
− g(x)u r + k T e − k T e + α ĝ(x, θ )N (τ )G − α ĝ(x, θ )N (τ )G (45)
using (27), one can rewrite (42) as
e(n) = −w Tf (x)
θ f − ε f (x) − w Tf a (x)
θ f a − ε f a (x) − wgT (x)
θg u c
− εg (x)u c − g(x)u r − k T e + ū − α ĝ(x, θ )N (τ )G (46)
by adding and subtracting αg(x)N (τ )G
e(n) = −w Tf (x)
θ f − ε f (x) − w Tf a (x)
θg (u c − α N (τ )G)
− εg (x)(u c − α N (τ )G) − g(x)u r − k T e + ū − αg(x)N (τ )G (47)
Then the dynamics of the error can be written as follows:

ė(t) = Ae(t) + B −w Tf (x) θ f − ε f (x) − w Tf a (x)

− εg (x)(u c − α N (τ )G) − g(x)u r + ū − αg(x)N (τ )G (48)
⎡ ⎤
0 1 0 0 ... 0 0 ⎡ ⎤
⎢ ⎥ 0
⎢ 0 0 1 0 ... 0 0 ⎥ ⎢ .. ⎥
⎢ .. .. .. .. . . .. .. ⎥ ⎢ ⎥
A=⎢ . . . . . . . ⎥B = ⎢ . ⎥
⎢ ⎥ ⎣0⎦
⎣ 0 0 0 0 ... 0 1 ⎦
1
−kn −kn−1 ··· ··· ... · · · −k1
Until (|s I − A|) = s (n) + k1 s (n−1) + · · · + kn is stable (A stable), we know that

there exists a symmetric positive definite matrix P(n, n) that satisfies the Lyapunov
equation:
A T P + P A = −Q (49)
where Q is a symmetric positive definite matrix of arbitrary dimensions (n × n).

Whether V , is the Lyapunov function, then
1 T 1 T 1 T 1 T 1 2 1 2
V = e Pe +
θf
θf +
θ f a
θfa +
θg
θg +
εf +
ε
2 2γ f 2γ f 2γg 2η f 2η f f a
1 2 1 2
+
ε + δ (50)
2ηg g 2σ0
where
⎧
⎨ εf = εf − εf

εfa = εfa − εfa (51)
⎩
εg = ε g − ε f
The time derivative of V is given by
1 T 1 T 1 T˙ 1 T ˙ 1 T˙
V̇ = ė Pe + e P ė + θ θ̃ f +
θ θ̃ f a + θ θ̃g
2 2 γf f 2γ f f a γg g
1 ˙ 1 1 ˙ 1
−
ε f ε̂ f − ε f a ε̂˙ f a − εg ε̂g + δ̇δ (52)
ηf ηf ηg σ0
Substituting (41), and using the equalities θ̃˙ f = −θ̇ f , θ̃˙ f a = −θ̇ f a and θ̃˙g = −θ̇g
it follows that
1 T T
V̇ = e A P + P A e + e T P B −w Tf (x) θ f − ε f (x) − w Tf a (x) θfa
2
− ε f a (x) − wgT (x) θg (u c − α N (τ )G) − εg (x)(u c − α N (τ )G) − g(x)u r + ū
1 T 1 T 1 T 1 ˙
− αg(x)N (τ )G] − θ f θ̇ f − θ f a θ̇ f a − θg θ̇g − ε f ε̂ f
γf γf γg ηf
1 1 ˙ 1
− ε f a ε̂˙ f a −
εg ε̂g + δ̇δ (53)
ηf ηg σ0
From (42) one can obtain

1
V̇ = − e T Qe + e T P B −w Tf (x) θ f − ε f (x) − w Tf a (x)
2
1 T 1 T
− εg (x)(u c − α N (τ )G) − g(x)u r + ū − αg(x)N (τ )G −
θ θ̇ f −
θ θ̇ f a
γf f γf fa
1 T 1 1 1 ˙ 1
−
θ θ̇g − ε f ε̂˙ f −
ε f a ε̂˙ f a −
εg ε̂g + δ̇δ (54)
γg g ηf ηf ηg σ0
1
V̇ = − e T Qe − αe T P Bg(x)N (τ )G + V̇1 + V̇2 (55)
2
where

1 ·
V̇1 = −
θ Tf θ f +w f (x)e T P B
γf

1 ·

− θfa T
θ f a +w f a (x)e T P B
γf

1 ·

θgT θg +e T P Bwg (x)(u c − α N (τ )G) (56)
γg
Using (26) and (27), (49) can be simplified to
V̇1 = 0 (57)
V̇2 = −e T P Bg(x)u r − e T P Bε f (x) − e T P Bε f a (x)

− e T P Bεg (x)(u c − α N (τ )G) + e T P B ū
1 ˙ 1 1 ˙ 1
−
ε f ε̂ f − ε f a ε̂˙ f a − εg ε̂g + δ̇δ (58)
ηf ηf ηg σ0
using Assumption 5, V̇2 can be upper bounded by:

V̇2 ≤ −e T P Bg(x)u r + e T P B ū + e T P B ε f
1
+ e T P B ε f a + e T P Bu c (u c − α N (τ )G)ε g − ε f − ε f ε̂˙ f

ηf
1 1 1
ε f a − ε f a ε̂˙ f a − ε g − ε g ε̂˙ g + δ̇δ

− (59)
ηf ηg σ0
with (28), (29) and (31), one can rewrite V̇2

V̇2 ≤ −e T P Bg(x)u r + e T P B ū + ε f e T P B + ε f a e T P B

+ ε g e T P B(u c − α N (τ )G) − δ 2

(60)
using (24), (53) can be simplified to

V̇2 ≤ −e T P Bg(x)u r + e T P B ϕ − δ 2 (61)
by adding and subtracting e T P Bu r b

V̇2 ≤ −e T P Bg(x)u r + e T P B ϕ − δ 2 + e T P Bu r b − e T P Bu r b (62)
using (22) one obtains
V̇2 ≤ −e T P Bg(x)u r + e T P Bu r b (63)
Substituting the results given by (50) and (56), V̇ can be bounded by
1
V̇ ≤ − e T Qe − αe T P Bg(x)N (τ )G − e T P Bg(x)u r + e T P Bu r b (64)
2
from (21), one can obtain
1
V̇ ≤ − e T Qe − αe T P Bg(x)N (τ )G − e T P Bg(x)u r b N (τ ) + e T P Bu r b (65)
2
Thus, V̇ can be rewritten as
1 2
V̇ ≤ − e T Qe − g(x)N (τ ) α e T P B + e T P Bu r b + e T P Bu r b (66)
2
2
by adding and subtracting αe T P B
1 2
V̇ ≤ − e T Qe − g(x)N (τ ) α e T P B + e T P Bu r b
2
2 2
+ e T P Bu r b + α e T P B − α e T P B (67)
Substituting (30)
1 2
V̇ ≤ − e T Qe − α e T P B − g(x)N (τ )(τ̇ ) + τ̇ (68)
2
where α > 0.
We can obtain the following inequality
t
V (t) − V (0) ≤ −(g(v)N (v) + 1)τ̇ (v)dv (69)
o
Equation (63) can be simplified to

t
V (t) ≤ V (0) + −(g(v)N (v) + 1)τ̇ (v)dv (70)
o
t Using Lemma 1, we can conclude from (64) the boundedness of V (t), N (t), and
o −(g(v)N (v) + 1) τ̇ (v)dv for t 0, t f .
According to (Liu06 and Liu08), since no finite-time escape phenomenon may
happen, then t f → ∞.
θ f (t), e(t),
Therefore,

θg (t), ε f (t), ε g (t), δ(t), x(t) and u(t) are bounded. As

an intermediate result e(t) is square integrable and ė(t) is bounded. Moreover, by
invoking Barbalat’s lemma, we can conclude the asymptotic convergence of e(t).
In this section, to demonstrate the effectiveness of the indirect adaptive fuzzy control
algorithms, we consider the tracking control of the second-order system. The dynamic
equations of such system are given by [22–24].
⎧
⎨ ẋ1 = x2
ẋ2 = f (x) + f a (x, u) + g(x)u (71)
⎩
y = x1
where

f (x) = 1.5 1 − x12 x2 − x1 (72)
g(x) = 1 + x12 + x22 + x1 x2 (73)
f a (x) = 1.5 ∗ r ect(1 − u ∗ t) − t + u (74)
The control objective is to force the system output y(t) = x1 (t) to track the desired
trajectory yd (t) = sin(t).
The actuator faults are applied from the t ≥ 0 until the end of the simulation.
Within this simulation, two fuzzy systems in the form of (5) are used to approxi-
mate the unknown smooth functions f (x) and g(x). The input variables of the used
fuzzy systems are x1 and x1 , and we added the control input u as an input variable
of the used fuzzy systems for approximating f a (x). For each input variable we have
define five Gaussian membership functions with center −1, −0.5, 0, 0.5 and 1, and a
variance equal to 4. The initial conditions are x(0) = [0.5, 0]T , and the initial values
of the adjustable parameters θ f (0) and θg (0) are set equal to zero.
p = [13 0.25; 0.25 0.52], Q = diag(10, 10), k = [20; 10], ε f (0) = 0, εg (0) =
0, τ (0) = 0, δ(0) = 1, σ0 = 5, ε0 = 0.1, γ f = γg = 20, η f = ηg = 0.001.
Simulation results for the case of an unknown positive control direc-
tion (sgn(g(x)) = 1) are shown in Figs. 1 and 2. To show the controller
behavior in the case of a negative control direction, we consider the case
g(x) = − 1 + x12 + x22 + x1 x2 . Using the same simulation design parameters,
simulation results are shown in Figs. 3 and 4. We can see that actual trajectories
converge to the desired ones and that the control input signal is bounded.
1.5
X1 X2
1
0.5
-0.5
-1
-1.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)
Fig. 1 System’s response (g(x) > 0): actual (blue lines); desired (red lines)
-1
0 2 4 6 8 10 12 14 16 18 20
Time (s)
Fig. 2 Control input signal (g(x) > 0)

1.5
X1 X2
1
0.5
-0.5
-1
-1.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)
Fig. 3 System’s response (g(x) < 0): actual (blue lines); desired (red lines)
-2
-4
-6
-8
-10
0 2 4 6 8 10 12 14 16 18 20
Time (s)
Fig. 4 Control input signal (g(x) < 0)
5 Conclusion
In this paper, an indirect adaptive fuzzy control for class of unknown (SISO) nonlinear
systems with unknown control sign is developed. The scheme consists of an adap-
tive fuzzy controller with a robuste control term used to compensate for approxima-
tion errors. New adaptive parameters update law are used, besides the Nussbaum-
type function, is used to overcome the problem of unknown control gain sign. The
proposed adaptive schemes allow initialization to zero of all adjustable parameters
of the fuzzy systems, guarantee the boudedness of all signals in the closed-loop, and
the tracking error has been shown to be asymptotically stable. Simulation results
have verified the effectiveness of the proposed adaptive control approach.
References
1. WangL X (1994) Adaptive fuzzy systems and control. Prentice-Hall, EnglewoodCliffs,

NewJersey
2. Passino KM, Yurkovich S (1998) Fuzzy control. Addison-Wesley Longman Inc.
3. Chemachema M, Belarbi K (2011) Direct adaptive neural network controller for a class of
nonlinear systems based on fuzzy estimator of the control error. Int J Syst Sci 42(7):1165–1173
4. Chemachema M (2012) Output feedback direct adaptive neural network control for uncertain
SISO nonlinear systems using a fuzzy estimator of the control error. Neural Netw 36:25–34
5. Ge SS, Hong F, Lee TS (2004) Adaptive neural control of nonlinear time-delay systems with
unknown virtual control coefficients. IEEE Trans Syst Man Cybern Part B 34(1):499–516
6. Wang LX, Mendel JM (1992) Fuzzy basis functions, universal approximation and orthogonal
least-squares learning. IEEE Trans Neural Netw 3(5):807–814
7. Essounbouli N, Hamzaoui A (2006) Direct and indirect robust adaptive fuzzy controller for a
class of nonlinear systems. Int J Control Autom Syst 4(2):146–154
8. Labiod S, Guerra TM (2007) Directe adaptive fuzzy control for a class of MIMO nonlinear
systems. Int J Syst Sci 38(8):665–675
9. Boukezoula R, Galichet S, Foulloy L (1998) Apprentissage de lois de commande floues pour
des systèmes non linéaires (synthèse directe et indirecte). Actesdes rencontresfrancophones
sur lalogiquefloueetsesapplications, Rennes, France, pp 19–27
10. Labiod S, Boucherit MS (2006) Indirect fuzzy adaptive control of a class of SISO nonlinear
systems. Arab J Sci Eng 31(1B):61–74
11. Nussbaum RD (1983) Some remarks on the conjecture in parameter adaptive control. Syst
Control Lett 3:243–246
12. Liu L, Huang J (2006) Global robust stabilization of cascade-connected systems with dynamic
uncertainties without knowing the control direction. IEEE Trans Autom Control 51:1693–1699
13. Liu L, Huang J (2008) Global robust output regulation of lower triangular systems with
unknown control direction. Automatica 44:1278–1284
14. Ye X, Jiang J (1998) Adaptive nonlinear design without a priori knowledge of control direction.
IEEE Trans Autom Control 43:1617–1621
15. Zhang TP, Ge SS (2007) Adaptive neural control of MIMO nonlinear state time-varying delay
systems with unknown dead-zones and gain signs. Automatica 43:1021–1033
16. Liu YJ, Wang ZF (2009) Adaptive fuzzy controller design of nonlinear systems with unknown
gain sign. Nonlinear Dyn 58:678–695
17. Boulkroune A, Tadjine M, M’Saad M, Farza M (2010) Fuzzy adaptive controller for MIMO
nonlinear systems with known and unknown control direction. Fuzzy Sets Syst 161:797–820
18. Li X, Wang D, Yao Y, Lan W, Peng Z, Sun G (2010) Adaptive fuzzy control for a class
of nonlinear time-delay system with unknown control direction. In: Proceedings of the 29th
Chinese control conference, pp 2593–2597
19. Lozano R, Collado J, Mondie S (1990) Model reference adaptive control without a priori
knowledge of the high frequency gain. IEEE Trans Autom Control 35:71–78
20. Labiod S, Bouberyakh H, Guerra TM (2011) Indirect adaptive fuzzy control for a class of
uncertain nonlinear systems with unknown control direction. Int J Fuzzy Syst Appl 1(4):1–17
21. Bounemeur H, Chemachema M, Essounbouli N (2018) Indirect adaptive fuzzy fault-tolerant
tracking control for MIMO nonlinear systems with actuator and sensor failures. ISA Trans.
https://doi.org/10.1016/j.isatra.2018.04.014
22. Abdelhamid B, Mouhamed C, Najib E (2017) Optimal indirect robust adaptive fuzzy control
using PSO for MIMO nonlinear systems. In: International conference on electrical engineering
and control applications. Springer, Cham, pp 208–224
23. Bounemeur A, Chemachema M, Essounbouli N (2014) Robust indirect adaptive fuzzy control
using Nussbaum gain for a class of SISO nonlinear systems with unknown directions. In: 2014
15th international conference on sciences and techniques of automatic control and computer
engineering (STA). IEEE, pp. 748–754
24. Bounemeur A, Chemachema M, Essounbouli N (2014) New approach of robust direct adaptive
control of a class of SISO nonlinear systems. In: 2014 15th international conference on sciences
and techniques of automatic control and computer engineering (STA). IEEE, pp 725–730
Renewable Energy (RE)
Optimal Integration of Renewable
Distributed Generation Using the Whale
Optimization Algorithm
for Techno-Economic Analysis
Samir Settoul, Rachid Chenni, Mohamed Zellagui, and Hassan Nouri
Abstract A new method for optimal integration of renewable energy sources based
on photovoltaic solar panels and wind turbines in the distribution network is presented
with the objective of reducing the Active Power Loss (APL) index, to improve the
Total Voltage Variation (TVV) index and the Total Operating Cost (TOC) index.
The objectives are achieved by optimal integration of Distributed Generation (DG)
based solar photovoltaic (PV) and wind turbine (WT) renewable sources using a
novel optimization algorithm, namely the Whale Optimization Algorithm (WOA)
for determining the optimal DG location and sizing subject to the constraints such
as power conservation, distribution line constraints, DG capacity limits, and DG
penetration limit. The proposed algorithm is evaluated on standard IEEE 12, 33,
and 69 bus distribution networks. A numerical simulation including comparative
studies is presented to demonstrate the performance and applicability of the Whale
Optimization Algorithm (WOA). The validity of the proposed WOA technique is
demonstrated by comparing the obtained results with those reported in literature
using other optimization algorithms.
Keywords Renewable energy · Optimal integration · Power distribution systems ·

Active power loss index · Voltage variation index · Total operating cost index ·
Whale optimization algorithm
S. Settoul (B) · R. Chenni

Department of Electrotechnic, Mentouri University of Constantine 1, Constantine, Algeria
M. Zellagui
Department of Electrical Engineering, École de Technologie Supérieure, Montréal, Canada
H. Nouri
Department of Engineering Design and Mathematics, University of West of England, Bristol, UK

https://doi.org/10.1007/978-981-15-6403-1_35
514 S. Settoul et al.
1 Introduction
The rapid development of Distributed Generation (DG) in different forms and capac-
ities is transforming the conventional planning of power distribution networks.
Despite the benefits offered by renewable DG technologies, several economic and
technical challenges can result from the inappropriate integration of DG in existing
power distribution networks. Therefore, the optimal planning of DG is of paramount
importance to ensure that the performance of the distribution network can meet the
expected power quality, voltage stability, power loss reduction, reliability and prof-
itability [1]. Decentralized power generation from renewable energy sources (RES)
is a long-term solution that addresses the present environmental threats because of its
widespread availability, sustainability, nonpolluting generation and eco-friendliness.
The most widely used renewable distributed generation types are WT and PV
systems, of which their generation is intermittent. The rapid development of DG tech-
nologies in various forms and capacities is significantly reshaping the conventional
planning of power distribution networks [2].
DGs are a reliable and economic solution for supplying electricity to customers
and are normally directly connected to the distribution network or on the customer
site. It is therefore necessary to allocate DGs optimally (size, placement and type)
to obtain commercial, technical, environmental and regulatory advantages of power
systems [3].
There are numerous algorithms to address the optimal DG planning problem.
Metaheuristic algorithms are often used as they offer more flexibility, particularly
for objective function planning problems without the pursuit of a globally opti-
mized solution. For example, the Particle Swarm Optimization (PSO) algorithm for
optimal placement of WT and PV sources based DGs for power loss minimization
and voltage stability improvement [4], using PSO and nonlinear programming for the
optimal multiples DG planning on the power loss sensitivity [5]. Applied Bat Algo-
rithm (BA) for optimal DG based planning to maximize the Voltage Stability Index
(VSI) and also to minimize the total active power losses [6]. Analytical approach
based Sensitivity Analysis Technique (SAT) for optimal allocation of DG and shunt
capacitor [7]. Applied Backtracking Search Optimization Algorithm (BSOA) for
Optimal WT and PV sources based DGs for power loss minimization [8]. Flower
Pollination Algorithm (FPA) to determine the optimal DG-unit’s size and location
in order to minimize the total system real power loss and improve the bus voltage
[9]. Krill Herd Algorithm (KHA) for the solution of optimal placement and sizing
of DG in order to minimize the line losses considering various technical constraints
[10]. Applied Ant Lion Optimization Algorithm (ALOA) for optimal integration
of PV and WT sources and applied in various systems [11]. Optimal multiples DGs
placement by multi-objective opposition based Chaotic Differential Evolution (CDE)
algorithm for techno-economic analysis [12]. Improved Chaos Stochastic Fractal
Search (CSFS) algorithm for optimal determination of location, size, and quantity
of DGs in complex distribution systems [13]. A newly effective Stochastic Fractal
Search Algorithm (SFSA) for optimal DGs integration into the distribution systems
Optimal Integration of Renewable Distributed Generation … 515
based on three objectives involving active power loss, voltage deviation, and VSI
[14]. Simulated Annealing (SA) algorithm for a novel approach to identify optimal
access point and sizing of multiple DGs in a small, medium and large scale radial
distribution systems [15]. Loss Sensitivity Factor (LSF) that determines the most
sensitive bus for DG installation and Bacterial Foraging Optimization Algorithm
(BFOA) for minimization of power loss, voltage deviation and TOC of the distri-
bution system [16]. Teaching-Learning Based Optimization (TLBO) and Grey Wolf
Optimizer (GWO) algorithm for multi-objective placement of renewable DGs with
objective to reduce losses and improve reliability [17]. Symbiotic Organism Search
(SOS) algorithm for Optimal number, location, and size determination of DGs in
order to minimize the power loss [18] and finally the Spider Monkey Optimiza-
tion (SMO) algorithm for optimal allocation of renewable DG to the demand side
management, using distribution system stability indicator [19].
This paper presents a new approach to optimal integration of renewable energy
source based DG in electrical power systems using the WOA algorithm for techno-
economic analysis. In this study, optimal integration of renewable DG based solar
photovoltaic and wind turbine source was installed in three different standard distri-
bution systems using the Whale Optimization Algorithm (WOA). The optimal inte-
gration has been selected to minimize the total active power losses, total voltage
variation and annual losses cost value using the proposed Whale Optimization Algo-
rithm. The validity of the proposed WOA algorithm is demonstrated by comparing
the obtained results with those reported in literature, which used other optimization
techniques. Analysis of the obtained results shows that the proposed algorithm is able
to achieve better minimization of the active power loss index and the total operating
cost index when compared with existing optimization techniques.
2 Problem Formulation
A. Objective function
The proposed objective function (OF) is used to reduce the Active Power Loss
(APL) index, to improve the Total Voltage Variation (TVV ) index and the Total
Operating Cost (TOC) index. It presents a multi-objective function for the distri-
bution system. The DG locations and their sizing are optimally obtained by solving
the following objective function:

N
bus N bus
O F = min α1 .A P L i j + α2 .T V Vi + α3 .AT OCi j (1)
i=1 j=2
where, α 1 , α 2 and α 3 are weighting factors. The sum of the absolute values of the
weights assigned to all impacts should add up to one as shown in the following
equation:

3
αm = |α1 | + |α2 | + |α3 | = 1 (2)
m=1
In this paper, α 1 is taken as 0.5 while α 2 and α 3 are taken as 0.25. The value of
these factors is based on the practical, technical and economic indicators. The three
indices are presented in Eqs. (4), (6) and (8) as described in the following:
A f ter DG
PLoss
A P L = Be f or e DG
(3)
PLoss
and,

N
bus N bus Pi2j + Q i2j
PLoss = Ri j (4)
i=1 j=2
Vi2
where, PLoss is the total active loss on the distribution system. Pij and Qij are active
and reactive powers that flow through branches from bus i to bus j. Nbus is the number
of buses, Rij is the resistance of the distribution line and V i is the voltage at bus i.
T V V A f ter DG
T V V = (5)
T V VBe f or e DG
and,

N bus
TVV = |1 − Vi | (6)
i=2
The Total Operating Cost (TOC) is another DGs installation merits in the distri-
bution networks [20]. The operational cost is composed of two components. The
first component is related to the active power supplied from the substation, which
can be reduced by minimizing the total real power loss of the system. The second
is the cost of active power supplied by those DGs units that are integrated into the
network. This cost could be reduced by minimizing the amount of real power drawn
from the DGs. The Total Operating Cost (TOC) index is calculated by:
T OC A f ter DG
T OC = (7)
T OC Be f or e DG
The TOC before and after the integration of the DG is calculated by the following
expressions:
Be f or e DG
T OC Be f or e DG = K 2 × PLoss (8)

A f ter DG Be f or e DG
T OC A f ter DG = K 1 × PLoss + K 2 × PLoss (9)
where, K 1 is the cost coefficient of real power supplied by the substation and DG
and K 2 represents the maintenance and installation costs of DGs units. K 1 and K 2
are taken to be 4 and 5 $/kW [20].
B. Power conservation constraint
Active and reactive power injected into each bus should equate the total load bus
power. The power flow equations are defined as equality constraints in the optimal
DG problem. The mathematical representation is given by [12–18]:
PG + PDG = PD + PLoss (10)
Q G + Q DG = Q D + Q Loss (11)
where PG , and PDG , PD and PLoss are active power of generator, DG, load and total
active losses respectively. Similarly, QG , QDG , QD and QLoss are respectively the
reactive power of the generator, DG, load and total losses respectively.
Bus voltage limits: The magnitude of voltage at each bus must be limited by the
following equation:
Vmin ≤ |Vi | ≤ Vmax (12)
where, V min and V max are the minimum and maximum specified voltage at each bus,
respectively.
Voltage drop limit: The limits of the voltage drop (V ) can be considered as the
flowing equation:
|V1 − Vi | ≤ Vmax (13)
where, V 1 is the voltage at the generating station is equal to 1.0 p.u. and V max is
the maximum voltage drop at each branch.
Line capacity constraint: Power flow through any distribution feeder must comply
with the thermal capacity of the line as given by the following equation.

Si j ≤ |Smax | (14)
where, S ij represents the apparent power in the distribution line between bus i and
S max denotes its corresponding maximum power.
C. DG constraints
DG capacity limits: The used DG units must have the allowable size in their following
range:
min ≤ P ≤ P max
PDG (15)
DG DG
Q min max
DG ≤ Q DG ≤ Q DG (16)
where, Pmin and Pmax are the minimum and maximum active power output limits of
the DG respectively; Qmin and Qmax are the minimum and maximum reactive power
output limits of the DG respectively. In the case of PV, the DG source only delivers
active power. While, in the case of WT, the DG source delivers active and reactive
power.
Power Factor of DG: The PF DG injected by the DG unit at each bus must be
limited between the minimum and maximum value as represented by the following
equation:
min ≤ P F ≤ P F max
P FDG (17)
DG DG
where,
PDG
P FDG = (18)
2
PDG + Q 2DG
Position of DG: The DG should not be positioned at bus 1 as it represents the

substation or slack bus. The position of DG is varied between the second bus and all
the buses in the distribution power system:
2 ≤ DG Position ≤ Nbus (19)
Number of DG: The number of DGs installed in the distribution power system is
limited by the maximum DGs (N DG.max ) as governed by the following equation:
N DG ≤ N DG.max (20)
Location of DG: This constraint presented one DG installed in the distribution

power system at one location or bus:
n DG,i /Location ≤ 1 (21)

Table 1 Limits values of constraints

Parameters Cases studies
12-bus 33-bus 69-bus
Bus number 12 33 69
PDG (kW) Min 10 10 10
Max 3000 3000 3000
QDG (kVar) Min 10 10 10
Max 2000 2000 2000
V i (p.u.) Min 0.95 0.95 0.95
Max 1.05 1.05 1.05
V max (%) 5 5 5
PF DG Min 0.80 0.80 0.80
Max 1.00 1.00 1.00
Number of DGs 1 2 3
The limits of constraint values after installation of DG based photovoltaic source

(PV-DG) and DG based wind turbine source (WT-DG) is shown in Table 1.
3 Whale Optimization Algorithm
Mirjalili developed the WOA approach in 2016 [21] as a novel nature-inspired

heuristic technique to solve problems related to engineering and different mathe-
matical optimization issues. The common behaviors of humpback whales are the
basis of the WOA. This optimization technique is inspired by the bubble net hunting
approach of humpback whales as they follow a circular shaped route for hunting
small fish near the surface. This feeding process is a distinctive behavior of humpback
whales, making this optimization unique among other nature-inspired optimization
methods. To design the mathematical model of the WOA, three steps are involved in
the bubble-net hunting process. All steps are described in detail as follows [21, 22].
A. Encircling Prey
Humpback whales can recognize the location of prey and en-circle them. Since the
position of the optimal design in the search agent is not known in advance, hence, the
WOA algorithm initially assumes that the best position of the target prey is a place
close to the position of the optimal design. Thereafter, the other search agents will
try to update the position towards the best location. This procedure mathematically
is represented by the following equations:
→ −
− →− → →
−
D = C . X ∗ (t) − X (t) (22)
−
→ −
→ −
→− →
X (t + 1) = X ∗ (t) − A . D (23)
where, D is the distance of ith whale to the prey, t indicates the current iteration, A
and C are the coefficient vector. X * is the position vector of the best solution obtained
so far, X is the position vector. It is worth mentioning here that X * should be updated
in each iteration if there is a better solution. In addition the vectors A and C are
calculated from equations:
−
→
A = 2.
a .
r − a (24)
−
→
C = 2.
r (25)
where, a is linearly decreased from 2 to 0 over the course of iterations (in both
exploration and exploitation phases) and r is a random vector generated with uniform
distribution in the interval of 0 and 1. According to Eq. (23) the search agents (whales)
update their position according to the position of the best known solution (prey).
B. Bubble-Net Attacking Method
To model this strategy, the two techniques employed by humpback whales are
presented as follows:
(1) Shrinking Encircling Mechanism
This behavior is achieved by decreasing the value of vector a in Eq. (24). Note that
the fluctuation range of A is also decreased by a. In other words A is a random value
in the interval [−a, a] where a is decreased from 2 to 0 over the course of iterations.
Setting random values for A in [−1, 1], the new position of a search agent can be
defined anywhere in between the original position of the agent and the position of
the current best agent. Figure 1 shows the possible positions from (X, Y ) towards
(X * , Y * ) that can be achieved by 0 ≤ A ≤ 1 in a 2D space.
(2) Spiral Updating of Position
In this process, the separation distance between the whale positioned at (X, Y ) and
prey positioned at (X * , Y * ) is determined. Then, for mimicking the spiral-shaped
movement of humpback whales, a helix condition is composed amongst the whale
location and prey location as:
−
→ −
→ −
→
X (t + 1) = D .ebt . cos(2πl) + X ∗ (t) (26)
where, D determines the location or distance of whales i to the prey, b is a constant

to represent the state of the logarithmic helix, and l is an arbitrary number in the
range −1, and 1. In view of the concurrent swimming of humpback whales around
the prey in a contracting loop and following a helix formed path, the equal likelihood
of choosing either the shrinking surrounding technique or the helix strategy can be
summarized as:
−→∗ −
→− →
−
→ X (t) − A . D i f p<0.5
X (t + 1) = − → bt −
→∗ (27)
D .e . cos(2πl) + X (t) i f p ≥ 0.5
where, p is an arbitrary number having a range between 0 and 1.

C. Search for Prey
An equivalent technique dependent on the heterogeneity of the vector A can be

utilized when searching for the prey (exploration). An arbitrary survey of humpback
whales shows that they are in view of one another and distinguishable. In like manner,
a moving search agent far from a reference whale is expected to be skilled in the
behavior, where |A| > 1. The mathematical model can be expressed as:
→ −
− →− → →
−
D = C . X rand − X (28)
−
→ −
→ −
→− →
X (t + 1) = X rand − A . D (29)
Fig. 1 Exploration mechanism implemented in WOA [21]

where, X rand is an arbitrary position vector (a random whale) which is chosen from
the existing population.
D. The Steps of the WOA
The principal procedures of the WOA approach are as follows. The WOA starts
with an established set of randomly created results. The locations of pursuit agents,
considering an arbitrary determination of the hunt agent or the given best result,
are then updated. For exploration and exploitation, parameter a is reduced from 2
to 0. When |A| > 1, an arbitrary hunt agent is chosen, and when |A| < 1, the best
result is chosen for updating the positions of the pursuit agents as depicted in Fig. 1.
This enhancement strategy can modify the movement among spiral and circular
movements depending on the measure of p.
Finally, the WOA approach is terminated once the termination criterion is satisfied
[22].
The pseudo code of the WOA technique is presented in Algorithm 1 shown below.
Algorithm 1 Pseudo-code of whale optimization algorithm
1. Generate the initial population X i (i = 1, 2, …, np )
2. Evaluate the fitness for each candidate solutions in Xi
3. X * is the best candidate solutions
4. while (t < maximum number of iterations)
5. for i = 1 to np (for each search agent)
6. Update a, A, C, l and p
7. if (p < 0.5)
8. if (|A| < 1)
9. Update the position of the search agent by Eq. (23)
10. else if (|A| ≥ 1)
11. Select a random search agent (X rand )
12. Update the position of the search agent by Eq. (29)
13. end if
14. else if (p ≥ 0.5)
15. Update the position of the current agent by the Eq. (26)
16. end if
17. end for
18. Check if any search agent goes beyond the search space
19. Calculate the fitness of each search agent
20. Update X * if there is a better solution
21. t=t+1
22. end while
23. Return X *
4 Results, Discussion and Comparison
The proposed WOA algorithm is developed using MATLAB 2017.b and simulations
are carried out on a PC that poses Intel i5, 2.7 GHz, and 8 GB RAM.
The first test system used in this paper is a standard IEEE 12-bus radial distribution
system as shown in Fig. 2a which is composed of 12 busses, 11 distribution lines
and branches with an active power load of 435 kW and reactive power load of 405
kVar with nominal voltage of 11 kV.
Fig. 2 Single line diagram of distribution systems. a. IEEE 12-bus, b IEEE 33-bus, c IEEE 69-bus
The second test system is a standard IEEE 33-bus radial distribution system and
composed of 33 busses, 32 distribution lines branches with an active power load of
3715 kW and reactive power load of 2300 kVar with nominal voltage of 12.66 kV
as represented in Fig. 2b.
To demonstrate the applicability of the proposed WOA approach in large-scale
distribution systems, a standard IEEE 69-bus radial distribution system, which is
composed of 69 busses, 68 distribution lines and branches with active power load
of 3791.89 kW and reactive power load of 2694.10 kVar with nominal voltage of
12.66 kV is considered as shown Fig. 2c.
Fig. 3 Convergence curve

of WOA for first case study.
a PV-DG, b WT-DG
The convergence optimization characteristics of the proposed WOA algorithms

for the three considered IEEE distributions systems in the presence of PV-DG and
WT-DG are presented in Fig. 3.
The parameters of all distribution systems before DG integration is represented
in Table 2, and optimization results after integration of PV-DG and WT-DG is
represented in Table 3.
From Table 3, for the first test system (IEEE 12-bus), the best location for instal-
lation of the PV-DG with size 236.60 kW and WT-DG with size 245.10 kW is bus
number 9. It can be observed that the PV-DG is capable of minimizing the APL by
51.9987% whereas the WT-DG by 15.9708%.
For the second test system (IEEE 33-bus), the identified sizes for PV-DGs are
850.3 and 1158.6 kW and for the WT-DG are 828.0 kW with power factor of 0.8944
(lagging) and 1243.3 kW with power factor of 0.8 (lagging). The DGs are connected
at bus number 13 and 30. The results show that APL is minimized by 41.3132%
for PV-DG and 13.9037% for WT-DG.
Table 2 Parameters of distribution systems before DG

Case study PLoss (kW) QLoss (kVar) V min (p.u.) TVV (p.u) TOC (k$)
IEEE 12-bus 20.8311 9.9489 0.9405 0.4466 104.1555
IEEE 33-bus 210.9875 143.1284 0.9038 1.8047 1054.9375
IEEE 69-bus 224.9480 102.1406 0.9092 1.8704 1124.74
Table 3 Optimization results after integration of DG

Parameters PV-DG WT-DG PV-DG WT-DG PV-DG WT-DG
IEEE 12-bus IEEE 33-bus IEEE 69-bus

PLoss (kW) 10.8319 3.3269 87.1657 29.3351 69.4215 4.5367
APL (%) 51.9987 15.9708 41.3132 13.9037 30.8611 2.0168
V min (p.u.) 0.9809 0.9897 0.9685 0.9804 0.9790 0.9943

TVV (p.u) 0.2945 0.0849 0.6775 0.1862 1.0277 0.1087
TVV (%) 65.943 19.010 37.541 10.318 54.946 5.812
TOC (k$) 147.48 117.46 1403.60 1172.28 1402.43 1142.89
TOC (%) 141.599 112.777 133.051 111.123 124.689 101.613
DG Bus 9 9 13–30 13–30 11–21–61 12–21–61
PDG (kW) 236.60 245.10 850.3 828.0 576.1 432.7
1158.6 1243.3 339.3 315.8
1716.8 1662.2
PF DG 1.0000 0.8000 1.0000 0.8944 1.0000 0.8000
1.0000 0.8000 1.0000 0.8000
1.0000 0.8000
For the third test system, the IEEE 69-bus is considered. The three PV-DGs with
the sizes of 576.1, 339.3, and 1716.8 kW are placed at buses 11, 21, and 61 respec-
tively, which minimizes the APL to 30.8611%. Similarly, the WT-DG with sizes
of 432.7, 315.8, and 1662.2 kW are installed at buses 11, 21 and 61 respectively and
lead to minimization of APL to 2.0168%.
Figure 4 indicates the bus voltage profiles for different case studies performed for
all distribution systems. Voltages of all the buses are within allowable limits. The
best profile is justifiably obtained after PV-DG and WT-DG installation.
This is because active power is higher than the reactive power and hence compen-
sation of real power has greater influence in controlling losses and voltage in the
distribution system.
Figure 5 illustrates the impact of DG installation on active power losses for the
three test systems. Analysis of Fig. 5 reveals that after installation of the DG, there is
significant minimization on active power losses. For example, the WT-DG installation
contributes to more minimization of the power losses than the installation of PV-DG.
This is related to the contribution of the reactive power injected from WT-DG.
Figure 6 shows the influence of DGs’ installation on voltage variation on all
three-test systems. In general, the installation of the two types of DGs i.e. PV-DG
and WT-DG will reduce the voltage variation in all the buses. However, from the
(a). IEEE 12-bus. (b). IEEE 33-bus.
(c). IEEE 69-bus.
Fig. 4 Bus voltages profile of standard distribution system

(c). IEEE 69-bus.
Fig. 5 Active power loss of standard distribution system
voltage variation curve, one can suggest that the WT-DG provides more improvement
than PV-DG in terms of system voltage.
Table 4 presents the comparison results of various optimization algorithms against
the proposed WOA algorithm in presence of DG based photovoltaic and wind turbine
sources.
Figures 7 and 8 show the bar chart form representation of TOC and APL for
the three IEEE test scenarios in presence of PV-DG and WT-DG respectively.
The results shown in Figs. 7 and 8 suggest that in all the three IEEE test scenarios,
the proposed WOA technique performed well in maximizing the active power loss
index and total operation cost index up to an optimal value. Furthermore, the overall
objective function gain accomplished by the suggested the WOA exhibits the predom-
inance and favorable circumstances of WOA over the algorithms specified in the
references.
(c). IEEE 69-bus.
Fig. 6 Voltage variation value of standard distribution system
5 Conclusions
This paper has presented optimal integration of renewable energy source based
distributed generations in electrical distribution networks using WOA approach for
techno-economic analysis. In this paper, a multi-objective function has been intro-
duced for optimal integration of solar photovoltaic source and wind turbine based
DG units and tested for three different electrical power distribution networks to
demonstrate the effectiveness of the proposed optimization algorithm.
The optimal integration of PV-DG and WT-DG units has been selected to the
active power loss index, the total voltage variation index and total operation cost index
value while satisfying system operational constraints within a predefined tolerance to
Table 4 Results comparison of various algorithms

Test system Algorithms PLoss (kW) APL (%) TVV (%) TOC (%)
IEEE 12-bus PV-DG PSO [4] 14.7000 70.5676 65.7859 156.454
BA [6] 10.9200 65.5750 66.9727 152.460
SAT [7] 13.6600 52.4216 68.8983 141.937
WOA 10.8319 51.9987 65.9427 141.599
WT-DG PSO [4] 7.9000 37.9241 43.1930 130.339
CS 3.5120 16.8594 16.9727 113.488
SCA 3.3764 16.2085 14.6216 112.967
WOA 3.3269 15.9708 19.0103 112.777
IEEE 33-bus PV-DG BSOA [8] 89.3400 42.3437 41.0040 133.875
FPA [9] 89.2000 42.2774 35.5738 133.822
KHA [10] 87.4260 41.4366 37.4577 133.149
WOA 87.1657 41.3132 37.5409 133.051
WT-DG PSO [5] 39.1000 18.5319 10.6721 114.826
BSOA [8] 31.9800 15.1573 32.6148 112.126
ALOA [11] 30.9251 14.6573 9.2037 111.726
WOA 29.3351 13.9037 10.3175 111.123
IEEE 69-bus PV-DG CDE [12] 69.4360 30.8676 54.9401 124.694
CSFS [13] 69.4300 30.8649 54.9401 124.692
SFSA [14] 69.4280 30.8640 54.9348 124.691
WOA 69.4215 30.8611 54.9455 124.689
WT-DG SA [15] 16.2600 7.2283 15.0075 105.783
BFOA [16] 12.9000 5.7347 25.9356 104.588
GWO [17] 7.3100 3.2496 7.4476 102.600
WOA 4.5367 2.0168 5.8116 101.613
avoid constraints violation. Using the proposed WOA algorithm, the optimal size and
location of renewable energy sources are identified which results in total power loss
reduction and satisfaction of the permissible voltage limits. The simulation results
show that the proposed WOA algorithm provides better results when compared with
other existing optimization algorithms.
Fig. 7 Comparison of WOA

with other algorithms in the
presence PV-DG. a IEEE
12-bus, b IEEE 33-bus,
c IEEE 69-bus
Fig. 8 Comparison of WOA

with other algorithms in the
presence WT-DG. a IEEE
12-bus, b IEEE 33-bus,
c IEEE 69-bus
References
1. Ehsan A, Yang Q (2018) Optimal integration and planning of renewable distributed generation
in the power distribution networks: a review of analytical techniques. Appl Energy 210:44–59
2. Abdmouleh Z, Gastli A, Ben-Brahim L, Haouari M, Al-Emadi NA (2017) Review of optimiza-
tion techniques applied for the integration of distributed generation from renewable energy
sources. Renew Energy 113:266–280
3. Zubo RHA, Mokryani G, Rajamani HH, Aghaei J, Niknam T, Pillai P (2017) Operation
and planning of distribution networks with integration of renewable distributed generators
considering uncertainties: a review. Renew Sustain Energy Rev 72:1177–1198
4. Kayal P, Chanda CK (2013) Placement of wind and solar based DGs in distribution system for
power loss minimization and voltage stability improvement. Int J Electr Power Energy Syst
53:795–809
5. Kaur S, Kumbhar G, Sharma J (2014) A MINLP technique for optimal placement of multiple
DG units in distribution systems. Int J Electr Power Energy Syst 63:609–617
6. Remha S, Chettih S, Arif S (2017) A novel multi-objective bat algorithm for optimal placement
and sizing of distributed generation in radial distributed systems. Adv Electr Electron Eng
15(5):736–746
7. Naik SG, Khatod DK, Sharma MP (2013) Optimal allocation of combined DG and capacitor
for real power loss minimization in distribution networks. Int J Electr Power Energy Syst
53:967–973
8. El-Fergany A (2015) Optimal allocation of multi-type distributed generators using Back-
tracking search optimization algorithm. Int J Electr Power Energy Syst 64:1197–1205
9. Oda ES, Abdelsalam AA, Abdel-Wahab MN, El-Saadawi MM (2017) Distributed generations
planning using flower pollination algorithm for enhancing distribution system voltage stability.
Ain Shams Eng J 8(4):593–603
10. Chithra Devi SA, Lakshminarasimman L, Balamurugan R (2017) Study krill herd algorithm
for multiple DG placement and sizing in a radial distribution system. Eng Sci Technol Int J
20:748–759
11. Ali ES, Elazim SA, Abdelaziz AY (2017) Ant lion optimization algorithm for optimal location
and sizing of renewable distributed generations. Renew Energy 101:1311–1324
12. Kumar S, Mandal KK, Chakraborty N (2019) Optimal DG placement by multi-objective oppo-
sition based chaotic differential evolution for techno-economic analysis. Appl Soft Comput J
78:70–83
13. Nguyen TP, Tran TT, Vo DN (2018) Improved stochastic fractal search algorithm with chaos
for optimal determination of location, size, and quantity of distributed generators in distribution
systems. Neural Comput Appl 30(11):3545–3564
14. Nguyen TP, Ngoc Vo D (2018) A novel stochastic fractal search algorithm for optimal allocation
of distributed generators in radial distribution systems. Appl Soft Comput J 70:773–796
15. Injeti SK, Kumar NP (2013) A novel approach to identify optimal access point and capacity of
multiple DGs in a small, medium and large scale radial distribution systems. Int J Electr Power
Energy Syst 45(1):142–151
16. Mohamed Imran A, Kowsalya M (2014) Optimal size and siting of multiple distributed gener-
ators in distribution system using bacterial foraging optimization. Swarm Evolut Comput
15:58–65
17. Nowdeh SA, Davoodkhani IF, Moghaddam MJH, Seifi Najmi E, Abdelaziz AY, Ahmadi A,
Razavi SE, Gandoman FH (2019) Fuzzy multi-objective placement of renewable energy sources
in distribution system with objective of loss reduction and reliability improvement using a novel
hybrid method. Appl Soft Comput J 77:761–779
18. Nguyen-Phuoc T, Vo-Ngoc D, Tran-The T (2017) Optimal number, location, and size of
distributed generators in distribution systems by symbiotic organism search based method.
Adv Electr Electron Eng 15(5):724–735
19. Deb G, Chakraborty K, Deb S (2019) Spider monkey optimization technique based allocation
of distributed generation for demand side management. Int Trans Electr Energy Syst 29:1–17
20. Imran AM, Kowsalya M (2014) Optimal size and siting of multiple distribution generators in
distribution system using bacterial foraging optimization. Swarm Evolut Comput 15:58–65
21. Mirjalili S, Andrew L (2016) The whale optimization algorithm. Adv Eng Softw 95:51–67
22. Elaziz MA, Mirjalili S (2019) A hyper-heuristic for improving the initial population of whale
optimization algorithm. Knowl Based Syst 172:42–63
Enhancement of Power System Transient
Stability with a Large Penetration
of Solar Photovoltaic Using Facts
Ikram Boucetta, Naimi Djemai, Salhi Ahmed, and Zellouma Laid
Abstract In this paper, a study on the transient stability of power system with a
large amount of solar photovoltaic (PV) penetration was performed on the famous
30 bus IEEE test system, under the Power System Analysis Toolbox (PSAT) simu-
lation software. Based on the critical clearing time (CCT) as a stability index. The
simulation results showed a significant decrease in network stability performance
after solar PV penetration, this degradation of the power system stability is due to
the non-participation of photovoltaic sources in the regulation of the voltage plan as
a consequence of the absence of reactive power, which leads us to have recourse to
the FACTS systems. the simulation results after integration of the two famous types
of FACTS (SVC, STATCOM) clearly showed the improvement of network stability
with a slight superiority in favor of STATCOM.
Keywords Renewable energy · Power system · Photovoltaic · PSAT · CCT ·

FACTS
1 Introduction
The main objective of integrating renewable sources into power system is to

contribute to the reduction of toxic gases emitted by fossil fuel power plants, the main-
tenance of fossil fuel reserves and the reduction of dependence on fossil resources
I. Boucetta (B) · Z. Laid

Echahid Hamma Lakhdar University, LEVRS, Eloued, Algeria
Z. Laid
N. Djemai · S. Ahmed
Mohamed Khider University, LGEB, Biskra, Algeria
S. Ahmed

https://doi.org/10.1007/978-981-15-6403-1_36
534 I. Boucetta et al.
threatened with exhaustion in the short term. The influence of the integration of large
quantities of these renewable sources on the state of the transmission grid is currently
the subject of several scientific publications [1].
In this paper, a study is presented concerning the improvement of the integration
of a renewable source widely known by the scientific community is photovoltaic
solar energy. This research focuses on the impact of this integration on the transient
stability of the power system [2] where the index of this stability is represented by
the maximum time during which the grid can withstand the fault without losing
its stability, this time is known as the critical clearing time (CCT) [3]. The use of
the FACT system to improve network performance in the presence of this type of
renewable source represented by PV source is justified by the opportunity offered
by power electronics [3]. A three-phase short circuit is applied to a 30-bus IEEE test
system; all simulations are performed under the Power System Analysis Toolbox
software (PSAT) [4]. This paper is organized as follows. After a shirt introduction,
transient stability is presented in second section, the facts devices are described
briefly in Sect. 3; also, the aim of Sect. 4 is the description of solar PV penetration.
Then, the test system, results and discussion are presented respectively in Sects. 5
and 6. Finally the paper concludes with a summary in Sect. 7.
2 Transient Stability
A. Definition
Power system stability has been a major concern in the last few decades. Transient
stability of an electrical network is defined as the ability to ensure synchronous
operation of its generators when it is subjected to severe transient disturbances such
as a short circuit affecting a network element or a loss of a significant part of the load
or generation. Thus, practically the entire system remains intact [3, 5]. Consequences
of these defects can be very serious and can lead to the total collapse of the network
(Black-out) (Fig. 1).
B. Swing equation
For a system to be transiently stable during a disturbance, it is necessary for the

rotor angle, to oscillate around an equilibrium point. If the rotor angle increases
indefinitely, the machine is said to be transiently unstable as the machine continues
to accelerate and does not reach a new state of equilibrium. So, for maintain this
equilibrium, the power system must at any time keep one’s balance between all
torques applied on rotor of synchronous generators, as its behavior described by the
swing equation [5].
Enhancement of Power System Transient Stability … 535
Fig. 1 Classification of
power system stability [12]
2
2H d δ(t) D dδ(t)
w pu (t) = Pmpu (t) − Pepu (t) − (1)
wsyn d 2 (t) wsyn dt
where t is time in seconds, P mpu is mechanical power supplied by prime mover minus
mechanical losses, per unit and P epu is electrical power output of the generator plus
electrical losses, per unit. w pu and ws yn represent the per unit frequency and the
synchronous radian frequency respectively, as to δ is rotor position with respect to
synchronously rotating reference in rad. D represents a damping torque anytime the
generator deviates from its synchronous speed. Finally, it’s convenient to work with
a normalized inertia constant, called the H constant [5]. The swing equation is a
second-order differential equation and it is also nonlinear (Fig. 2).
In steady-state operation, all synchronous machines in the system rotate at the
same electrical angular speed. The mechanical torque Tm has the same direction as
the rotation of the generator axis. The electrical torque Te is in the opposite direction
to that of rotation [6].
Pe Electromagnetic power
Pm Mechanical power
W Angular speed.
C. Critical Clearing Time (CCT)
Fig. 2 Illustration of the couples adjusting on the rotor

This is the most decisive parameter in the analysis of the transient stability of an elec-
trical grid, mathematically it is the solution of the second order non-linear differential
equation known as the Swing equation, which physically represents the maximum
time during which our network can withstand a fault (short-circuit, over load, over
current…) without losing its stability [6]. This is the stability assessment criterion
adopted in this article.
3 Flexible AC Transmission System (Facts)
Transient stability is one of the most important key factors during power transfer at
high levels. One of the powerful methods for enhancing the transient stability is to use
flexible AC transmission system (FACTS) devices [7]. The FACTS controllers offer a
great opportunity to regulate the transmission of alternating current (AC), increasing
or diminishing the power flow in specific lines and responding almost instantaneously
to the stability problems. The potential of this technology is based on the possibility
of controlling the route of the power flow and the ability of connecting networks that
are not adequately interconnected, giving the possibility of trading energy between
distant agents [8]. Flexible Alternating Current Transmission System (FACTS) is
static equipment used for the AC transmission of electrical energy. It is meant to
enhance controllability and increase power transfer capability. It is generally power
electronics-based device [8].
A. Classification of Facts Devices
There are different classifications for the FACTS devices can be divided in three
groups, depending on the type of connection to the network:
In this paper, we focus on shunt compensators (SVC, STATCOM) [9] (Fig. 3).
• SSSC: Static Synchronous Series Compensator
• TCSC: Thyristor Controlled Series Capacitor
• SVC: Static Var Compensator
Fig. 3 Block diagram of

FACTS controllers [10]
Fig. 4 SVC connected to a

transmission line [10]
Fig. 5 STATCOM
connected to a transmission
line [10]
• STATCOM: Static Compensator

• UPFC: Unified Power Flow Controller
• IPFC: Inter line Power Flow Compensator (Figs. 4 and 5).
4 Solar PV Penetration
Renewable energy sources, such as photovoltaic (PV) systems, wind turbines, and
fuel cells, are integrated into conventional power systems to address fossil fuel defi-
ciency, intensifying energy demand, and environmental pollution. Among all types
of renewable energy resources, solar PV receives major attention for its promising
energy resources and low-cost installation. The fundamental operation system of
solar PV differs from other generating systems. Solar PV converts sunlight into DC
power using semiconductor solar cells. The DC power is then converted into AC
power through a DC-to-AC converter [11]. PV penetration level is defined as the
ratio of total PV generation to total system generation, as expressed in the following
equation:
T otal P V generation (M W )
P V penetration (%) f or generation based =
T otal generation (M W )
(2)
5 Test System
The simulation analysis was established on the famous 30 bus IEEE test system by
PSAT/MATLAB, which gives access to an extensive library of grid components.
A. Simulation Environment
Several simulation tools were used to analyze the stability of electrical networks,
such as: (MATLAB, Power World, PSAT…, Etc.).
In the study presented, PSAT was chosen as a simulation tool:
• PAST combines traditional statistical testing and data visualization functions that
are generally only found in specialized software.
• PSAT is compatible with the databases of most electrical network analysis
software.
• PSAT is an “Open Source” software that allows for modification as required by
the study [4]. Figure 6 shows the famous 30 bus IEEE test system simulated by
PSAT under MATLAB.
Fig. 6 IEEE 30 bus model with PSAT

Table 1 Technical data of the study model

Number of bus Number of Number of Number of Number of line of
charges generators transformers transmissions
30 21 6 7 33
B. Characteristics of the Staudy Model. Table 1 represents the characteristics of the

Staudy Model (Technical data of the study model).
A. Choice of fault Type
Three-phase short-circuit is chosen as a result of its most dangerous consequences

against power system stability.
B. Analysis Strategy
A comparative study has been carried out with respect to the values of the critical
clearing time corresponding to each simulation. This study is divided into three parts:
Part 1: initial state
• Creating a fault at bus “i”, except the reference bus.
• Calculation of critical clearing time with minimum singular value method, in
order to identify the most sensitive bus in the presence of the fault.
The same study will be performed for all selected buses.
Part 2: penetration of the solar photovoltaic (PV)
Part 3: improvement of transient stability using an SVC, STATCOM.
C. Simulation Results
The performance of the proposed method was tested in the MATLAB/Simulation

environment.
a. Steady State Analysis (Before creating the default)
Steady state analysis of the power grid represents no anomalies were essentially the
voltage profile presented in Fig. 7 shows that the voltage values at all buses are within
their acceptable limits (0.90–1.1 p.u).
Fig. 7 Voltage profile
b. Initial State. We took bus 8 as excemple to show the location of the fault, this
one is represented in Fig. 8.
T Critical clearing time (CCT) (Fig. 9)
According to Fig. 10, we can see that the challenge of just 1 (ms) further of the
critical clearing time that can cause the instability of the system.
The study performed at bus 8 will be done in the same way for all selected buses.
The results are shown in Table 2.
According to Table 2, we find that the critical time varies from one bus to another;
this is due to the ability of each bus to support the fault.
Fig. 8 Fault location at bus 8

Fig. 9 Voltage profile at the presence of a fault at bus 8
Fig. 10 Generator rotor speeds
Table 2 Critical clearing time for different fault location

Bus Bus 2 Bus 4 Bus 5 Bus 8 Bus 12 Bus 17 Bus 18 Bus 27
CCT (ms) 15 19 20 40 20 48 30 20
c. Penetration of the Solar Photovoltaic (PV)
A fault has been considered at bus 18 with integration a solar PV plant. The power
injected by solar PV plant into the power grid around to 50 MW. Figure 11 shows
the Penetration of a solar PV with presence of the fault at bus 18.
Decreased is observed in Fig. 12 when the solar PV is connected to power system
grid. It can be seen that the change of just 1 (ms) can cause the instability of the
power grid. Figure 13 shows the simulation results obtained in MATLAB.
According to Fig. 14, it is clear that the penetration of solar PV has enormously
reduced the CCT and consequently the ability of the electrical network to maintain
its stability during the fault in all the cases studied. This is due to the inability of the
Fig. 11 Penetration of a solar PV with presence of the fault at bus 18
Fig. 12 Rotor speeds corresponding

Fig. 13 Voltage profile with penetration of solar PV at bus 18
Fig. 14 CCT comparison histogram with and without PV integration
PV to participate in the setting of the voltage plan in the absence of reactive energy
in such sources.
Additionally, PV penetration rate is inversely proportional to the corresponding
critical time values. This is shown by the results of the Table 3.
d. Improvement of transient stability using an SVC, STATCOM
The use of the FACTS system is justified to improve the network’s ability to withstand
faults for longer periods of time.
Table 3 Variation in critical clearing time with percentage penetration of solar PV

Bus Bus 2 Bus 4 Bus 5 Bus 8 Bus 12 Bus 18
Active power (MW) 0.2 0.2 0.05 0.05 0.06 0.06
CCT (ms) 3 4 8 9 7 3
PV penetration % 23 23 6 6 7 7
In this section, an SVC and STATCOM respectively has been introduced in order
to checking their influence on solar power penetration in the same localization. After
several simulations important results are achieved:
• SVC
According to Table 4, the integration of the SVC has improved the critical clearing
time of the network. We take bus 4 as an example: from 4 to 90 (ms) (Fig. 15).
Instead of SVC, another FACTS device (STATCOM) of the same type has been
introduced in same bus.
Table 4 CCT for different rates of solar power penetration with and without SVC
Bus Bus 4 Bus 5 Bus 8 Bus 12
CCT (ms)without SVS 19 20 40 20
CCT(ms) with PV without SVC 4 8 9 7
CCT(ms) with PV and SVC 90 60 130 419
Fig. 15 Localization of SVC

• STATCOM
After the integration of FACTS (SVC, STATCOM) it can be seen that a considerable
improvement in the CCT has been noticed, which will provide an opportunity to
integrate more energy from PV sources. Figure 16 shows the location of the SVC
in the test system. Table 5 represents the critical clearing time for different rates of
solar power penetration with and without STATCOM.
STATCOM as shown in the histogram in Fig. 17 shows a modest superiority over
SVC.
Fig. 16 Localization of STATCOM
Fig. 17 CCT comparison

histogram
Table 5 CCT for different rates of solar power penetration with and without STATCOM
Bus Bus 4 Bus 5 Bus 8 Bus 12
CCT (ms) without PV and STATCOM 15 20 40 20
CCT (ms) with PV penetration 4 8 9 7
CCT (ms) with PV and STATCOM 100 123 140 421
7 Conclusion
This paper has mainly focused on the assessment of power system transient stability
by determinate a critical clearing time (CCT) for the several cases by observing the
behavior simulation of a test system during grid faults using a Power System Analysis
Toolbox (PSAT).
According to previously simulations, we find that:
• The penetration of solar PV has reduced the CCT and thus the ability of the
electrical network to maintain its stability during the fault in all the cases studied.
• A considerable improvement in the CCT has been noticed after the integration of
FACTS (SVC, STATCOM).
STATCOM shows a modest superiority over SVC.
• The simulation results clearly showed that there is an opportunity to integrate
more energy from PV sources into the power grid through FACTS.
References
1. Yagami M, Ishikawa S, Ichinohe Y, Misawa K, Tamura J (2015) Transient stability analysis of

power system with photovoltaic systems installed. J Energy and Power Eng 9
2. Remon D, Cantarellas AM, Mauricio JM, Rodriguez P (2017) Power system stability analysis
under increasing penetration of photovoltaic power plants with synchronous power controllers.
IET Renew Power Gener 11:733–741
3. Damor KG, Patel DM (2014) Improving power system transient stability by using facts devices.
Int J Eng Res Technol (IJERT) 3(7). ISSN: 2278-0181
4. Milano F (2006) Power system analysis toolbox documentation for PSAT version 2.0.0 β1,
July 9, 2006
5. Naimi D, Bouktir T, Ahmed S (2013) Improvement of transient stability of algerian power
system network with wind farm. IEEE Conference: irsec’13, At Ouarzazet morroco, March
2013
6. Ayasun S, Liang Y, Nwankpa CO (2006) A sensitivity approach for computation of the prob-
ability density function of critical clearing time and probability of stability in power system
transient stability analysis. Appl Math Comput, 563
7. Damor KG, Patel DM, Agrawal V, Patel HG (2014) Improving power system transient stability
by using facts devices. Int J Eng Res Technol (IJERT) 3(7). ISSN: 2278-0181
8. Abido MA (2009) Power system stability enhancement using facts controllers: a review. Arab
J Sci Eng 34(2):153–172
9. Gupta S, Tripathi RK (2014) Transient stability enhancement of multimachine power system
using robust and novel controller based CSC-STATCOM. Adv Power Electron, 12. Article ID
626731
10. Ramakrishna Rao BT, Chanti P, Lavanya N, Chandra Sekhar S, Mohan Kumar Y (2014) Power
system stability enhancement using fact devices. J Eng Res Appl 4(4)(Version 1):339–344.
ISSN: 2248-9622
11. Kouro S, Leon JI,Vinnikov D, Franquelo LG (2015) Grid-connected photovoltaic systems: an

overview of recent research and emerging PV converter technology. IEEE Ind Electron Mag
9:47–61
12. Kamaruzzaman ZA, A. Mohamed A, Shareef H (2015) Effect of grid- connected photovoltaic
systems on static and dynamic voltage stability with analysis techniques – a review. Przegl˛ad
Elektrotechniczny, ISSN0033-2097, R. 91 NR 6/2015
Minimization of the Energy
Consumption of an Aircraft
Abbes Lounis, Kahina Louadj, and Mohamed Aidene
Abstract In this article, we studied a problem of minimizing an aircraft energy flow,

the model used is obtained from the fundamental principle of dynamics. We solved
this problem with the method of relaxation coupled with the method of shooting.
Keywords Free final time · Shooting method · Pontryaguin principle · Relaxation

method
1 Introduction
Today, one of the major problems of aeronautics is the excessive consumption of

fuel by aircraft. As we all know, excessive fuel consumption has a real impact on
the environment because these planes pollute our atmosphere, increases the costs of
travel and maintenance, so the minimization of Kerosene is a major challenge that
has great interest in our daily life.
Fuel consumption depends on altitude and speed. When climbing, the device
consumes more obviously, downhill, it consumes less. It is therefore realistic to
consume on average compared to our cruise consumption.
When climbing, a fraction of the weight is added to the drag. The resultant is no
longer balanced by the thrust. We must therefore increase the thrust or risk seeing
the speed decrease. The thrust cannot be increase indefinitely, so if we continue to
increase the incline, the speed will decrease. To maintain lift equal to the apparent
weight, it will increase the impact to offset the decrease in speed. If the rate continues
A. Lounis (B) · K. Louadj · M. Aidene

Laboratoire de Conception et Conduite des systèmes de Production, University of Tizi Ouzou (
UMMTO), Tizi-Ouzou, Algeria
K. Louadj
M. Aidene

https://doi.org/10.1007/978-981-15-6403-1_37
550 A. Lounis et al.
to fall, it will reach the incidence of dropping out, and the plane will fall. In descent,
a fraction of the weight is added to the thrust. Maintain a constant speed therefore
requires reducing power. If too accentuates the slope, the trail will not be enough to
balance the resulting «push more weight.” The aircraft will continue to accelerate
and eventually end up in over speed.
In this work, we studied a problem of minimizing the energy of an airplane; the
model used is obtained from the fundamental principle of dynamics. We solved this
problem with the method relaxation coupled with the method of shooting. We applied
these results to a Boeing 747 [1–8].
2 Statement of the Problem
• V: flight speed.
• θ: angle between horizontal and flight path.
• α:Angle of attack (angle between flight path and chord line).
• w: aircraft weight.
• L: Lift, force normal to flight path generated by air acting on aircraft.
• D: Drag, force along flight path generated by air acting on aircraft.
• M: Pitching moment.
• T: Propulsive force supplied by aircraft engine/propeller.
• α T : Angle between thrust and flight path.
To derive the equations of motion, we apply

F = m a
dV
FI I = ma I I = m
dt

FI T = T cosαT − D − wsinθ (1)
dV
T cosαT − D − wsinθ = m (2)
dt
Applying (1) in ⊥ direction to flight path
V2
F⊥ = ma⊥ = m
rc
rc : radius of curature of flight path:
V2
F⊥ = L + T sinαT − wcosθ = m
rc
Minimization of the Energy Consumption of an Aircraft 551
V2
L + T sinαT − wcosθ = m (3)
rc
dθ
F⊥ = ma⊥ = mV
dt

F⊥ = L + T sinαT − wcosθ (4)
dθ
L + T sinαT − wcosθ = mV
dt
We have:
dx dh
= V cosθ = V sinθ (5)
dt dt
The equations are:
⎧
⎪
⎪ = V cosθ
dx
⎪
⎪ dt
⎪
⎨ = V sinθ
dh
dt
dV
= 1
(T cosα T − D − wsinθ ) (6)
⎪
⎪
dt m
⎪
⎪
dθ
= 1
(L + T sinαT − wcosθ )
⎪
⎩ dt Vm
dm
dt
= −Q
The optimal control problem considered is to find the control Q(t) that minimizes
the performance index:
t f
Q(t)dt → Min Q (7)
0
where Q is mass flow the prescribed initial condition at the initial time t 0 : x(0) = x 0 ;
h(0) = h0 ; V(0) = V 0 ; θ (0) = θ 0 ;m(0) = m0 ; and prescribed final condition at the final
time t f : x(t f ) = x tf ; h(t f ) = htf ; V(t f ) = V 0 ; θ (t f ) = θ tf ;m(t f ) = mtf .
The aim of this paper is to determine a trajectory (x(t);Q(t)) leading an initial state
x(0) = x 0 ; h(0) = h0 ; V(0) = V 0 ;θ (0) = θ 0 ;m(0) = m0 ; to the final state:
x(t f ) = x tf ; h(t f ) = htf ; V(t f ) = V 0 ; θ (t f ) = θ tf ;m(t f ) = mtf which minimize a criteria
(7) tf is not fixed (tf is free libre) [4, 2, 3].
We will solve the problem as follows:
t f
Q(t)dt → Min Q
0
⎧
⎪
⎪ = V cosθ
dx
⎪
⎪ dt
⎪
⎨ = V sinθ
dh
dt
dV
= 1
(T cosα T − D − wsinθ ) (8)
⎪
⎪
dt m
⎪
⎪
dθ
= V m (L + T sinαT − wcosθ )
1
⎪
⎩ dt dm
dt
= −Q
With a boundary conditions:

x(0) = x 0 ; h(0) = h0 ; V(0) = V 0 ; θ (0) = θ 0 ;m(0) = m0 .
x(t f ) = x tf ; h(t f ) = htf ; V(t f ) = V 0 ; θ (t f ) = θ tf ;m(t f ) = mtf .
With constraint of control:
−1 ≤ Q(t) ≤ +1
The Hamiltonian H is defined by:
H = p1 (t)V cosθ + p2 (t)V sinθ + p3 (t) m1 (T cosαT − D − mgsinθ)

+ p4 (t) V1m (L + T sinαT − mgcosθ) − p5 (t)Q(t) − Q(t)
The Euler-Lagrange equations are given by:

⎧
⎪
⎪ ẋ = ∂∂ pH1
⎪
⎪
⎪
⎪ ḣ = ∂∂ pH2
⎪
⎪
⎪
⎪ V̇ = ∂∂ pH3
⎪
⎪
⎪
⎪
⎪
⎪ θ̇ = ∂∂ pH4
⎪
⎪
⎪
⎪ ṁ = ∂∂ pH5
⎨
·
p1 = − ∂∂Hx
⎪
⎪ ·
⎪
⎪ = − ∂∂hH
⎪
⎪
p2
⎪
⎪ ·
⎪
⎪ p3 = − ∂∂ HV
⎪
⎪ ·
⎪
⎪ p4 = − ∂∂θH
⎪
⎪
⎪
⎪ ·
⎪
⎪ p5 = − ∂∂mH
⎩
−1 ≤ Q(t) ≤ +1
We have the system of the equations follows:

⎧
⎪
⎪ ẋ = V cos θ
⎪
⎪
⎪
⎪ ḣ = V sin θ
⎪
⎪
⎪
⎪ V̇ = m (T cos αT − D − mg sin θ )
1
⎪
⎪
⎪
⎪ θ̇ = V1m (L + T sin αT − mg cos θ )
⎪
⎪
⎪
⎪ ṁ = −Q
⎪
⎪
⎨ ·
p1 = 0
⎪ ·
⎪
⎪ p2 = 0
⎪
⎪ ·
⎪
⎪ = −[ (t) θ + (t) θ − p4 (t)
(L + α − θ )
⎪
⎪
p 3 p 1 cos p 2 sin 2 T sin T mg cos
⎪
⎪
mV

⎪
⎪ ṗ = −[− p (t)V sin θ + p (t)V cos θ − p3 (t) mg mg
cos θ + p4 (t) mV sin θ
⎪
⎪
4 1
2 m
⎪
⎪
·
p5 = − − pm3 (t) p4 (t)
⎪ 2 (T cos α T − D) − m 2 V (L + T sin α T )
⎪
⎪
⎩
−1 ≤ Q(t) ≤ +1
From the Hamiltonien H, the control Q is given by:
Q(t) = max H Q(t)∈[−1.1] = si gn(− p5 (t) − 1)
3 Shooting Indirect Method
The shooting indirect method is used to obtain the value of p(0) necessary to the
solution of the problem characterized by the Pontryaguin principle. If it is possible,
from the condition of minimization of the Hamiltonian to express the control extremal
function of (x(t); p(t)) then the extremal system is a differential system of the form
ż(t) = G(t, z(t)) where z(t) = (x(t), p(t)). With a numerical integrator from z0 we
obtain:z i∼z0 ∼z(ti ) where the t i , i = 1, 2… are the time moments discretized by the
integrator. But in z0 = (x 0 ; p0 ); the value x 0 is given by the initial condition. Then, by
doing some
variations on p0 , we∼z0 obtain the different z i∼z0 . Which interests us are the
z N ∼z t f (at final time); else z N = (x N∼z0 , p ∼z0
∼z0 ∼z0
N ) and only the x N are significant.
∼P0
Since they depend only on p0 , note that x N . Let G be the implicit function giving
p0 by numerical calculation using an integrator returns x N∼P0 − x f :
G: Rn → Rn and
G( p0 ) = x N∼P0 − x f
Here G is an implicit nonlinear system of n equations and n unknowns: G(p0) =

0 [9–12].
For the solution, we used the Newton’s method. The principle of the Newton’s
method is described as follows: in the k–th step, let p0k be an approximation of the
zero p0 of G; therefore p0 can be written p0 = p0k + p0k and then:
∂G k
0 = G( p0 ) = G p0k + p0k = G p0k + p p0 − p0k + o p0 − p0k
∂ p0 0
This leads to the solution of:

∂G k
−G p0k = p0 p0 − p0k
∂ p0
k
where ∂∂Gp0
p0 is the Jacobian matrix of the application p0 → G( p0 ) computed when
p0 = p0 ; note that the mapping p0 → G( p0 ) is not explicitly known but is known
k
numerically p0 → G( p0 ) is not explicitly known but is known numerically So we

will use a method ofnumerical
derivation based on the finite difference.
k To avoid
the calculation of ∂∂G
p0
p0k it suffices to find an approximation of ∂∂G
p0
p0 , according
to [13], we will use one of the following finite difference approximations.

∂G i k 1 j

j−1
p ≈ [G i ( p0 + h ik e ) − G i p0 +
k
h ik e )
k
∂ p0 j 0 hi j k=1 k=1
Or else
∂G i k 1
p0 ≈ [G i p0 + h i j e j − G i ( p0 )]
∂ p0 j hi j
where the hij are the given discretization step of the i-th equation with respect to the
j-th variable, and ek are the k-th vector of the canonical basis; note that, classically,
we can always choose the values of hij equal each other at a common value h. Let
ij (p0 ; h) be a finite difference approximation, we have
∂G i
lim i j ( p0 , h) = ( p0 ), i, j = 1, 2 . . .
h→0 ∂ p0 j
Let,
( p0 , h) = i j ( p0 , h)
which is an approximation of the Jacobian matrix; then the approximate Newton’s

method can be written as follows:
−1
p0k+1 = p0k − J p0k , h k .G p0k
The problem of convergence of this iterative process is ensured by using a result

of the book of Ortega and Rheinboldt [14]; indeed if the discretization step hij are
small and tend to zero, the convergence is ensured.
A. Transversality condition on p
Generally, when the terminal cost is considered in the cost functional, the functional
to be minimized can be written as follows:
t f

J = φ t0 , t f + L(t, x, u)dt (9)
t0
Let M 0 and M 1 be two subsets of Rn ; then to minimize the cost functional one
should find a trajectory between M 0 and M 1 . Moreover if M 0 and M 1 are two varieties
of Rn having the tangent spaces T x0 M 0 and T x(tf) M 1 respectively x 0 ∈ M 0 and x tf ∈
M 1 ; then the vector p(t) must verify the transversality conditions:
p(0) ⊥ Tx0 M0 (10)

p t f − p 0 ∇x φ t f , x f ⊥Tx f M1 (11)
where p0 is a real such that p0 < 0 leads to the Pontryaguin’s maximum principle and
p0 > 0 leads to the Pontryaguin’s minimum principle. If M 0 = x 0 ; the condition (10)
becomes empty and the variety M 1 can be written as follows:

M1 = x ∈ Rn /F1 (x) = F2 (x) = . . . . = Fq (x) = 0
where F i are functions of class C 1 on Rn ; then the tangent space to M1 at a point x∈

M1 is defined by:

Tx M1 = ν ∈ Rn /∇ Fi (x(t f ))ν = 0i = 1, 2, . . . q
and the condition (11) is written as follows

q

∃ν1 , ν2 , . . . . . . . . . ..νq ∈ R/ p t f = νi ∇x Fi (x(t f )) + + p 0 ∇x φ t f , x f
i=1
where ν i are the Lagrange multipliers.

The transversality condition of Hamiltonian is defined by:

H t f , x f , p t f , p0 , u t f , = 0
corresponding to the fact that the Hamiltonian vanishes at final time.
4 Numerical Solution
For the numerical solution, we used the shooting indirect method to determine a final
time. Then we have to solve the following system:
⎧
⎪
⎪ v̇1 = v3 cosv4
⎪
⎪
⎪
⎪ v̇1 = v3 sinv4
⎪
⎪
⎪
⎪ v̇3 = v5 (T cosαT − D − v5 gsinv4 )
1
⎪
⎪
⎪
⎪ ·
⎪
⎪ v4 = v3 v5 (L
1
+ T sinαT − v5 gcosv4 )
⎪
⎪ ·
⎪
⎪ v5 = −Q
⎪
⎪
⎪
⎪ ·
⎪
⎪ v6 = 0
⎨ ·
v7 = 0
⎪
⎪ ·
⎪
⎪ v8 = −[v6 cosv4 + v7 sinv4 − vvv9 2 (L + T sinαT − v5 gcosv4 )
⎪
⎪ 5 3
⎪
⎪ v5 .g v5 .g
⎪
⎪ v9 = [v6 v3 sinv 4 − v7 v3 cosv4 +v 8 cosv4 − v9 sinv4
⎪
⎪
v5 v5 v3
⎪
⎪ (T − (L + T)
v8 v9
⎪
⎪ v10 = 2 cosα T D)+ 2 T sinα
⎪
⎪
v5 v5 v3
⎪
⎪ μ = sign(−v10 − 1)
⎪
⎪
⎪
⎪ −1 ≤ Q(t) ≤ +1
⎪
⎪
⎩
vi (0) ∈ R, i = 1, . . . .10
Let υ(t, x 0 , h0 ,v0 , θ 0 ,m0 , p1 , p2 , p3 , p4 , p5 ) be the solution of the previous system

at time t with the initial condition υ 1 (0), υ 2 (0), υ 3 (0), υ 4 (0), υ 5 (0), υ 6 (0), υ 7 (0),
υ 8 (0), υ 9 (0), υ 10 (0).
We construct a shooting function which is a nonlinear algebraic equation of the
variable ψ at time t = 0; This shooting function is computed by a numerical procedure
of integration of ordinary differential equation (using Euler method, Runge-Kutta
method,…); the shooting function is defined by:
⎛ ⎞
v1t f , x0 , h 0 , v0 , θ0 , m 0 , p1 , p2 , p3 , p4 , p5 − xt f
⎜v t ,x ,h ,v ,θ ,m , p , p , p , p , p − h ⎟
⎜ 2 f 0 0 0 0 0 1 2 3 4 5 tf ⎟
⎜ ⎟
ϕ(p) = ⎜ v3 t f , x0 , h 0 , v0 , θ0 , m 0 , p1 , p2 , p3 , p4 , p5 − Vt f ⎟
⎜ ⎟
⎝ v4 t f , x0 , h 0 , v0 , θ0 , m 0 , p1 , p2 , p3 , p4 , p5 − θt f ⎠

v5 t f , x0 , h 0 , v0 , θ0 , m 0 , p1 , p2 , p3 , p4 , p5 − m t
The problem to solve is then written: Find p(0) such that ‘(p(0)) gives the desired
value of x(T).
The algorithm for numerical solution of this problem will then be completely
defined if one gives oneself:
1. The integration algorithm of a differential system with initial condition (e.g.,
Euler or Runge-Kutta procedure) to compute the shooting function G (imple-
mented in ‘ode45’ of Matlab which is a method of Runge-Kutta 4/5 with variable
pitch).
2. The solution algorithm G(z) = 0 which in our case uses the method quasi- newton
(implemented in ‘fsolve’ of Matlab).
5 The Steps of the Algorithm (Relaxation Methodcoupled

with Shooting Method)
1. Approximate an initial control u0 (t); t ∈ [t 0 ; t f ], and adjoint state p0 (0).

2. r ← 0
3. while convergence > ε do
• Determine a state x r (t) and adjoint state pr (t) component by component
sequentially to t∈ [t 0 ; t f ] by numerical integration, the increasing time
⎧
⎪
⎪
d xr
= V r cosθ
⎪
⎪ dt
⎪
⎨
dh r
= V r sinθ
dt
dV r
= m r (T cosαT − D − wsinθ )
1
⎪
⎪
dt
⎪
⎪
dθ r
= V r m r (L r
1
+ T sinαT − wcosθ )
⎪
⎩ dt dm
dt
= −Q
x(0) = x 0 ; h(0) = h0 ; V(0) = V 0 ; θ (0) = θ 0 ;m(0) = m0

⎧ dp1r
⎪
⎪ =0
⎪
⎪ dt
⎪
⎪
dp2r
=0
⎪
⎪
dt
⎪
⎨ dp3r pr
= − p1r (t) cos θ r + p2r sin θ − m r V4 r 2 (L + T sin αT − m r g cos θ r )]
r
dt

⎪
r r r
dp4
⎪
⎪ = − − p1r (t)V r sin θ r + p2r (t)V r cos θ r − p3r (t) mm rg cos θ r + p4r (t) mmr Vgr sin θ r
⎪
⎪
dt
⎪
⎪ dp5r
=
p3r
(T cosαT
pr
− D)+ m r 24V r (L + T sinαT )
⎪
⎪ mr 2
⎩ dt
pr (0)
where pr (0) is found by shooting method [15].

• Determine a control ur+1 (t)

u r+1 (t) ← sign − p5r (t) − 1

• Convergence: u r+1 (t) − u r (t)
• Determine a shooting function:

ϕ( p) = x r t ∗ − x ∗
• Solution of the shooting equation is determine by Newton method and determine

a new values of p(0)
p r+1 (0) ← p r (0) + cor r ect i on
• r ← r + 1.
while end
6 Numerical Application
Aircraft Boeing 747, Cruise Stage. The boundary conditions are: x 0 = 12 km; h0
= 6.092 km; V 0 = 920 km/h; θ 0 = 0.0349066 rad, m0 = 90,000 kg. x tf = 775 km,
htf = 12.192 km, V tf = 955 km/h, θ tf = 0.0523599 rad, mtf = 4000 kg. Shooting
Method (Determine final time t f ) Final Time: t f = 0:95 h, iterations Number:
25 iteration, execution time: 2.74 s. Shooting Method (Determine p(0)) execution
time: 1.32 s, Iterations Number: 03. Relaxation method execution time: 5.23 s,
Iteration Number: 03 (Figs. 1, 2, 3, 4, 5 and 6).
Fig. 1 a The position x(t). b Altitude h(t)
7 Conclusion
In this paper, we have solved a problem of minimization problem of an energy flow

of an aircraft with free final time using relaxation method’s coupled with shooting
method’s based Pontryaguin’s maximum principle. We used the shooting indirect
method to find the transversality conditions in both cases where the state is submitted.
We concluded that, in the numerical procedure, the convergence is fast and the
computational time are small.
Fig. 2 a The Speed V (t). b The control Q(t)

Fig. 3 a The adjoint state p1 (t). b The adjoint state p2 (t)

Fig. 4 a The adjoint state p3 (t). b The adjoint state p4 (t)

Fig. 5 a The adjoint state p5 (t). b The mass m(t)

Fig. 6 The angular acceleration θ(t)
References
1. Efstathios Bakolas-Panagiotis Siotras (2013) Optimal synthesis of the Zermelo-Markov-

Dubins, problem in a constant drift field, J Opt Theory Appl 156:469–492 (Springer)
2. Pontryagin et al L (1974) Mathematical theory of optimal processes. Mir, Moscou
3. Lee EB, Markus L (1967) Foundations of optimal control theory. The SIAM series in applied
mathematics. Wiley, New-York, London-Sydney
4. Hagelauer P, Mora Camino F (2014) A soft dynamic programming approach for on-line aircraft
4D-trajectory optimization. LAAS, Toulouse
5. Titouche S, Spiteri P, Messine F, Aidene M (2015) Optimal control of a large thermic. J Control
Syst 25:50–58
6. Sethi SP, Thompson GL (2000) Optimal control theory: applications to management science
and economics, 2nd edn. Springer, Berlin, pp 504
7. Demim F, Nemra A, Louadj K (2016) Solution of an optimal control problem of the robot with
Bellman Principle’s. In: Conférence Internationale en Sciences et Technologies Electriques,
October 26–28, CISTEM’16, Morocco, Marakech
8. Demim F, Nemra A, Louadj K, Hamelain M (2018) Simultaneous localization, mapping, and
path planning for unmanned vehicle using optimal control. Adv Mech Eng 10(1):1–25
9. Hull DG (2003) Optimal control theory for applications. Springer, New York
10. Sethi SP, Thompsonn GL (2000) Optimal control theory, applications to management science
and economics, 2nd edn
11. Sager S, Bock HG, Reinelt G (2009) Direct methods with maximal lower bound for mixed-
integer optimal control problems. Math Program Ser A 118:109–149
12. Ternovskii VV, Khapaev MM (2008) Direct numerical method for solving optimal control
problems. Doklady Math 77(3):428431
13. Trelat E (2005) Contrôle optimal: théorie et applications. Vuibert, collection Mathématiques
Concrètes, France
14. Ortega JM, Rheinboldt WC (1970) Iterative solution of nonlinear equations in several variables.
Academic Press, New York
15. Demim F, Louadj K, Aidene M, Nemra A (2016) Solution of an optimal control problem with
vector control using relaxation method. Autom Control Syst Eng J 16(2) (ISSN 1687-4811,
ICGST LLC, Delaware, USA)
Study to Improve the Technical
Parameters for the Optimising
the Injection of the Photovoltaic Energy
into an Electrical Network via a Line
30 KV
Bahriya Badache, Hocine Labar, and Mounia Samira Kelaiaia
Abstract The aim of this paper is study to improve the technical parameters for the
optimizing the injection of the photovoltaic energy into an electrical network via a
line 30 kV, the study of three variants to improve the integration of the solar energy
of the OUED KEBRITE located in the EST of Algeria power station to the electrical
network. This investigation is based on geological topology of the region essential
to the grid design and the economical payback of the whole project. Concerning the
investment of the construction of the photovoltaic power plant, which will enable it
to produce 15 MWp to be fed into the electricity grid, the 30 kV line used to evacuate
its production has a thermal limit that does not allow all the energy produced to pass
through. There for we have carried out a study to find a solution in order to overcome
the constraint of limiting the energy transit of the 30 kV line.
Keywords Photovoltaic systems · Grid connected · Simulation · Improvement ·

Technical parameters
1 Introduction
In this research, we approached three variants: the rehabilitation of the existing 30 kV

line, the construction of a new line and the construction of an underground cable line
in order to improve the injection of photovoltaic energy into the electricity grid.
Renewable energy is a key challenging problem increasingly gaining attention
in worldwide [1]. Solar energy is the most abundant, inexhaustible and clean of all
renewable energy resources [2]. In the twenty first century the energy demands are
B. Badache (B) · H. Labar · M. S. Kelaiaia

Department of Electrical Engineering, University Badji Mokhtar Annaba, Annaba, Algeria
H. Labar
M. S. Kelaiaia

https://doi.org/10.1007/978-981-15-6403-1_38
566 B. Badache et al.
rapidly growing-up [3]. Renewable energy can reduce greenhouse gas emissions
by providing renewable electricity [4]. The renewable energy sources (RES) are an
important part of the energy system, with the features of wide distribution capability
and great development potential. RES are environmentally friendly and sustainable
compared with conventional energy sources. Therefore, the exploitation and utiliza-
tion of RES has been an important measure in the world to ensure energy security
and to tackle climate change [5]. Renewable energy can supply two-thirds of the
total global energy demand, and contribute to the bulk of the greenhouse gas emis-
sions reduction that is needed between now and 2050 for limiting average global
[6]. Furthermore, the PV panels produce sometimes an excess of power and in other
times a lack of power [3]. The photovoltaic generator is the only direct converter
to transform light into electrical energy [7], a photovoltaic generator is a gener-
ator whose characteristic I = f(U) is highly non-linear. Consequently, for the same
amount of sunlight, the power delivered will be different depending on the load.
A maximum power point tracking (MPPT) controller controls the static converter,
connecting the load (a battery for example) and the photovoltaic panel, in order to
permanently provide the maximum power to the load. The Perturbation and Obser-
vation (P&O) method is the most widely used maximum power search method; it is
an iterative method to obtain the maximum power point (MPP) [8]. The PV system
works around its MPP according the preset voltage step [9]. Photovolatic systems
normally use a MPPT technique to continuously deliver the highest possible power to
the load when variations in the insolation and temperature occur [10]. The principle
of the Perturbation and Observation method consists in disturbing the voltage (VPV)
of the generator with a small amplitude around the initial values and analysing the
behaviour of the resulting power variation (PPV). If a positive voltage increment
(VPV) results in an increase in power (PPV), this means that the operating point is
to the left of the PPM Fig. 1, if, on the contrary, the power decreases, this means
that the system has exceeded the PPM. A similar reasoning can be applied when
the voltage decreases. To ensure the photovoltaic generator operates at optimal effi-
ciency, requires the insertion of static converters (chopper) between the generator and
receiver (load) at a stable (constant) voltage can be optimized by adjusting the duty
cycle α so that the generator can operate at optimal speed. However, for connection
to the grid it requires an inverter converter [8] (Fig. 2).
Fig. 1 Continuation of the

maximum power point
Study to Improve the Technical Parameters … 567
Fig. 2 The real photovoltaic power station production according to the radiation day of 24-10-2018
The photovoltaic power plant is located in the eastern region of Algeria Fig. 3,
was commissioned in 2014.
The plant has a capacity of 15 “Megawatt (MW)” and is built on an area esti-
mated at 30 “hectares (ha)”, this first station of its kind in eastern Algeria, will
have photovoltaic solar cells. The Photovoltaic plant is composed of approximately
48.000 photovoltaic modules with a power of 250 “Watt peak (Wp)” in Polycrys-
talline silicon for an actual installation power of 15 “Mega Watt peak (MWp)”. Each
subfield consists of approximately 4.000 photovoltaic modules with an installed
capacity of approximately 01MWp. Each subfield is equipped with two centralized
inverters and a step-up transformer. These inverters, with an output voltage of 315
“Volt (V)”, are connected on the low voltage side via alternating current cables to
the 1.250 “kilo volt ampere (kVA)” step-up transformer, which raises the voltage to
31.5 “kilo volt (kV)”.
Fig. 3 Renewable energies

in Algeria
This power plant will support the electricity grid and fight against power outages,
often recorded during the summer period. The operation will also help to generate new
jobs, particularly in the technical specialities related to this field, such as renewable
energy sectors. The purpose of this work is to examine the impact of the integration
of the plant
Photovoltaic of 15 MWp in Eastern Algeria from Shariket Kahraba wa Taket
Moutadjadid (SKTM) on the distribution network. The evacuation of production to
the 30 kV floor of the 220/90/30 kV substation in EL AOUINET.
The connection mode proposed by the Electricity Distribution Company (SDE),
the plant will flow all its production on the 30 kV busbar on the 30 kV floor of EL
AOUINET substation. The study will determine the optimal injection point on EL
AOUINET transformers [8].
2 Economical Impact of Renewable Energy
The electric power generated by a solar array fluctuates depending on the solar radi-
ation value and temperature [10]. Therefore, a change in the spatial arrangement of
energy generation systems is taking place. Traditional locations for energy generation
(e.g. lignite or hard coal mining areas) may lose their substantial significance in favor
of renewable energy generation locations, if the former do not possess suitable local
conditions for the generation of renewable energy (RE), which leads to economic
losses [11]. Cross-regional connection lines are cost-effective methods of increasing
renewable energy consumption [12]. The incorporation of renewable energy in a
power system is determined by certain factors, including maximum system load,
daily load curve and peak-valley difference power source mix, capacity of cross-
regional lines, as well as power transmission curve and demand side response of the
system Fig. 2, [12].
3 Case Study Modeling and Features
In recent years, the production of photovoltaic (PV) systems and integration into the
electricity grid has become the most important [13]. In many countries, photovoltaic
(PV) solar systems are considered one of the best renewable energy (RE) sources in
terms of installation costs [14].
There has been an increased attention to the photovoltaic (PV) energy systems
during the last decade owing to the many advantages that these systems have such
as: it is a worldwide available energy source, it is pollution free, it has noiseless
operation [15].
The revision of the national program focuses mainly on the development of photo-
voltaic and wind power on a large scale, the introduction of biomass (waste recovery),
cogeneration and geothermal energy and also on the postponement of the develop-
ment of Concentrated Solar Power (CSP) to 2021 [8]. This program underwent a
first phase devoted to carrying out pilot projects and tests of the various available
technologies Fig. 3, during which relevant elements concerning the technological
developments of the sectors under consideration appeared on the energy scene and
led to the revision of this program. In accordance with current regulations, the imple-
mentation of this national program is open to domestic and foreign public and private
sector investors [8].
The investment of the photovoltaic power plant Fig. 3, is planned to produce a
capacity of about 15 MWhp, nevertheless the production of the power plant is limited
to 12 MWp because it transits energy on a 30 kV line whose thermal limit is 120
A/HT, see the characteristics at level Table 1, because of the dilapidated state of
the 30 kV OUED KEBRIT/EL AOUINET line Fig. 4, an available energy of 03
“Mega Watt peak (MWhp)”not produced and not consumed, it is for this reason that
a reliability study of the line is necessary. This line crosses the two lines 220 kV
Table 1 Characteristics of the 30 kV Kebrit/El Aouinet line

Tower type Tower height Conductor cable Type of Transit limit Length
section insulators
Poles 95BS66 Between 10 m et 98.30 mm2 CTV 175 120 A 6 km
and concrete 12 m
Poles
Fig. 4 Representing the different types of supports of the 30 kV line

KHROUB/EL AOUINET and EL AOUINET/AIN BEIDA, the line is crossed on the

30 kV span ex M’DAOUROUCHE.
A. Characteristic of line 30 kV Oued Kebrit/El Aouinet:
B. Simulation and results:
(1) VARIANT N°1: Study for the renovation of the existing line to ensure a higher
thermal limit (Fig. 6).
Fig. 5 The 30 kV line from the photovoltaic power plant to El Aouinet substation
Fig. 6 The line layout on GOOGLE EARTH, option 1

Replacement of the 98.3 mm2 conductive cable by 116 mm2 cable of section type
ALMELEC see length profile shown on the AutoCAD software Fig. 7. Using the
same existing pylons see the layout located on a uneven ground Fig. 6.
This study requires checking the crossing distances or to cross (regulatory vertical
distances) Fig. 8 and calculating the electrical parameters according to the formulae
(1–7).

Dmoy
X 0 = 0, 144 log + 0, 016 (1)
rcond
Fig. 7 The profile along the 30 kV line on AUTOCAD, option 1
Fig. 8 The result of the crossing on NC/EtudLig software, option 1

X L = X0 × L (2)
R = R0 × L (3)
7, 58 × 10−6
b0 = c0 × ω = (4)
Dmoy
log rcond
Y×Z
A =1+ =D (5)
2
B=Z (6)

Y×Z
C = 1+ ×Y (7)
4
C. Simulation and results:

This variant consists in studying the construction of a new 30 kV line, choosing a
new layout shorter than that of the existing line and a level-free ground Fig. 9. Using
larger and more robust supports to ensure proper adjustment of the conductor cables
see the long profile Fig. 10, in order to improve the crossing distance or to cross
(regulatory vertical distance) Fig. 11, which reduces the frequency of aggression of
the latter and reduces the number of its activations and improves its quality of service
parameters by using anti-pollution insulators.

Fig. 10 The profile along the 30 kV line on AUTOCAD, option 2
Fig. 11 The result of the crossing on NC/EtudLig software, option 2
4 Analyze and Discussion
In order to ensure the transit of the available power of the OUED KEBRIT 15 MWp
power plant instead of 12 MWp. The construction of a new line (variant n°2) is the
most economical solution. Variant n°3 consists of the construction of an underground
cable line (Fig. 12) this solution presents a considerable improvement in electrical
parameters and the transit limit is the highest. For the last variant the loss of revenue
will be saved by the cost of maintenance and operation.
From the above, choosing the third variant is necessary Table 2 for the following
reasons:
• Improvement of Service quality standards following the commission of a new
underground line since it is not affected by climatic conditions or external
aggressions.
• Incident recovery: current time, quick access to fault point.
• Aesthetically, the underground line does not hinder aesthetically.
• The underground line also provides the optical fibre link to connect the OUED
KEBRIT power plant to SONELGAZ’s optical fibre network.
• No impact on the environment and health
• Cost is the only disadvantage of this variety, nevertheless the latter will be
saved since the underground line requires no maintenance, even the number of
interventions is very reduced comparing with variants N°1 and 2.
5 Conclusion
The objective of this work is to study the integration of renewable energies into the
networks because the improvement of the voltage of the electricity grid is a concern
of the electricity distributor. In the Medium Voltage (MV) network, the quality of the
voltage will be unacceptable when the voltage drop exceeds 10%. In addition to the
study to improve the integration of solar energy injected by the photovoltaic power
plant into the 30 kV grid. It was necessary to find an answer to the question: what
is the revenue loss of the National Electricity and Gas Company (Sonelgaz) and the
losses suffered caused by the thermal limit of the 30 kV Oued Kebrit/ El Aouinet
line. In this work three (03) cases are considered and studied, using four (4) software:
Table 2 Comparison of the three (03) variants

Variants Advantages Disadvantages
Variant n°1 Section of conductor: 116 mm2 Drawing of the line: Reuse the existing
ALMELEC line layout Length: Number of tower:
120 Longest: 06 km
Works duration: Short (about 12 days Ground: Uneven level, Rough
without refund) Undistributed Energy (MWh): Extended
High Tower: 10 m to 12 m, type 95BS66
Poles and concrete poles
Restoring situation in the event of an
incident: Important time, Long line
Type of Insulator: CTV 175 normal
Line and plant logging: (about 12 days),
Throughout the period of the works
Presentation of the most dangerous risk
to the third part
Showing the effect of electromagnetic
fields from overhead lines on human
health, animals and the environment
The mechanical strength of the existing
towers does not allow the installation of
the guard cable to protect the line
against lightning strikes, which causes
permanent disruption of the network
The line is exposed to adverse weather
conditions, strong winds, lightning,
pollution, etc.
The line does not have an optical fibre
link with the SONELGAZ network
Cost: In order of 25,000,000.00 DA
Variant n°2 Undistributed energy (MWh): low Length: Number of Tower: 84 Shorter:
05 km Drawing of the line: New layout,
Realization of a new line
High tower: 14 m, type Metallic tower Works duration: Long (more than
126BS77 03 months)
Section of conductor: 116 mm2 Important deadlines and times for the
ALMELEC regularization of landowners’ objections
Type of insulator: U80AS The cost of compensating landowners
Anti-pollution for the installation of new towers
Line and plant logging: One day only Restoring situation in the event of an
for connecting conductive cables incident: Less time, shorter line
Cost: in order of 44,000,000.00 DA Presentation of the most dangerous risk
to the third part
(continued)
Table 2 (continued)
Variants Advantages Disadvantages
Showing the effect of electromagnetic
fields from overhead lines on human
health, animals and the environment
The line not protected from adverse
weather conditions
The line does not have an optical fibre
link with the SONELGAZ network
Variant n°3 Construction of an underground cable Cost: In order of 34,000,000.00 DA
line, Shorter: 3.50 km
No implementation constraints such as
oppositions and no line maintenance
Undistributed Energy (MWh): low
Section of conductor: 120 mm2
Armored type
Recovery in the event of an incident:
current time, quick access to the fault
point
Lack of danger to third parties
No impact on the environment and
health
Line and plant logging: One day only
for connecting conductive cables
Capacitive current is not very important
for the 30 kV voltage level
The underground line also provides the
optical fibre link to connect the OUED
KEBRIT power plant to SONELGAZ’s
optical fibre network
The underground line is protected
against adverse weather conditions
Works duration: Long (1.5 months)
MATLAB SIMULINK, AutoCAD, Google Earth, NC/EtudLig, the advantages and

disadvantages of each variant are well addressed. The state of the art of renewable
energies (photovoltaic) plays a major role in such a study. This work is initiated by
a theoretical study of the characteristics of a primary source (sunshine), followed by
an examination of how a photovoltaic system is integrated into the electricity grid.
The aim is therefore to optimize the overall treatment of electrical energy within the
system and in particular, at the generator level, by placing it at all times at its optimal
operating point thanks to a MPPT-type tracking system. In addition, due to their
switching operations, these static converters generate disturbances that can affect the
proper functioning of the generator both electrically and energetically.
On the energy level, the ripples of the voltage or current at the input of some
converters result in an oscillation of the operating point on the characteristic around
the optimal power point, which implies a degradation of the photovoltaic conversion.
Variant n°03 seems to be the best solution to be adopted by Sonelgaz company so
that it can allows a recovery of about 100% from its investment [8].
Annexes:
Data Option N°1:
• Length: 06 km
• X0 = 0.326 “ohrm ()”
• XL = 1.956 “ohm.kilometre (.Km)”
• B0 = 3.51710−6 “siemens (S)”
• A = D = 0.99 + j 1.79 × 10−5
• B = 1.69 + j 1.95
• C = 1.88 × 10−10 + j 2.11 × 10−5
Data Option N°2:

• Length: 05 km
• X0 = 0.326
• XL = 1.63 .Km
• R = 1.41 .Km
• Z = 1.41 + j 1.63
• B0 = 3.51710−6 S
• A = D = 0.99 + j 1.24 × 10−5
• B = 1.41 + j 1.63
• C = −1.09 × 10−10 + j 1.75 × 10−5
Data Option N°3:

• Length: 3.5 km
• R = 0.325 .Km
• C = 0.182 μF.Km
• L = 0.415 mH.Km
References
1. Harrou F, Sun Y, Taghezouit B, Saidi A, Hamlati M-E (2017) Reliable fault detection and
diagnosis of photovoltaic systems based on statistical monitoring approaches. Renew Energy
(Accepted manuscript)
2. Ali MH, Rabhi A, El hajjaji A, Tina GM (2016) Real time fault detection in photo-
voltaic systems. Energy Proced 111:914–923 (Elsevier, Sept 2016, Digests 8th International
Conference Sustainability in Energy and Buildings, SEB-16, Italy, 2016, p 914)
3. Djelailia O, Kelaiaia MS, Labar H, Necaibia S, Merad F (2018) Energy hybridization photo-
voltaic/diesel generator/pump storage hydroelectric management based on online optimal fuel
consumption per kWh. Sustain Cities Soc (Accepted manuscript)
4. Azhgaliyeva D (2019) Energy storage and renewable energy deployement: empirical evidence
from OECD countries. Energy Proced 158:3647–3651(Digests 10th International Conference
Applied Energy “ICAE2018” Hong Kong China, p 1, 2019)
5. Zeng M, Duan J, Wang L, Zhang Y, Xue S (2015) Orderly grid connection of renewable energy
generation in China: management mode, existing problems and solutions. Renew Sustain
Energy Rev 41:14–28
6. Gielen D, Boshell F, Saygin D, Bazilian MD, Wagner N, Gorini R (2019) The role of renewable
energy in the global energy transformation. Energy Strategy Rev 24:38–50
7. Khemliche M, Djeriou S, Latreche S (2012) Diagnostic de Défauts dans le Système Photo-
voltaïque par les Réseaux de Neurones, Diagnostic fault in photovoltaic system for Neurones
networks, SIENR, pp 1–9, Oct 2012 (Digests The 2 nd International Seminar on New and
Renewable Energies Ghardaïa Algérie, p 1, 2012)
8. http://biblio.univ-annaba.dz/ingeniorat/wp-content/uploads/2018/10/Merzougui-Nour-El-
Houda-Badache-Bahriya.pdf, 2018
9. Hocine L, Samira KM (2018) Real time partial shading detection and global maximum power
point tracking applied to outdoor PV panel boost converter. Energy Convers Manag 171:1246–
1254
10. Hua C, Lin J, Shen C (1998) Implementation of a DSP-controlled photovoltaic system with
peak power tracking. IEEE Trans Ind Electron 45(1):1–9
11. Jenniches S (2018) Assessing the regional economic impacts of renewable energy sources.
Renew Sustain Energy Rev 93:35–51
12. Zhang D, You P, Liu F, Zhang Yu, Zhang Ya, Feng C (2018) Regulating cost for renewable
energy integration in power grids. Global Energy Interconnect 1(5):544–551
13. Al-Shetwi AQ, Sujod MZ, Blaabjerg F, Yang Y (2019) Fault ride-through control of grid-
connected photovoltaic power plants. Sol Energy 180:340–350
14. Ahmad NI, Ab-Kadir MZA, Izadi M, Azisa N, Radzi MAM, Zainia NH, Nasir MSM (2017)
Lightning protection on photovoltaic systems (in press)
15. Mellit A, Tina GM, Kalogirou SA (2018) Fault detection and diagnosis methods for
photovoltaic systems. Renew Sustain Energy Rev 91:1–17
Second-Order Super-Twisting Control
of an Autonomous Wind Energy
Conversion System Based on PMSG
for Robustness and Chattering
Elimination
Abderrahmane Abdellah, Djilali Toumi, and M’hamed Larbi
Abstract This article focuses on improving the performance of the autonomous

wind energy conversion system based on a Permanent Magnet Synchronous Gener-
ator (PMSG). There are several problems related to the control of this system, such
as chattering phenomenon, static error and coupling between d-q currents, associated
with the use of conventional regulators like the PI controller. To remedy these prob-
lems, a comparative study is performed between the Super Twisting Sliding Mode
Controller (STSMC) and Adaptive Fuzzy Sliding Mode Controller (AFSMC) that
are used to drive the rotor speed to the optimal speed to get the maximum power.
The simulation study shows that the proposed STSMC is more effective at different
operating conditions.
Keywords Wind turbine · PMSG · Adaptive fuzzy controller · Sliding mode

controller · Super twisting algorithm
1 Introduction
Fossil fuels do much more damage than renewable energy, including air and water
pollution, public health, loss of wildlife and habitat, water use, global warming
emissions. Moreover, the demand for this source of energy continues to increase as
industrial and economic development depends on these energy needs. Faced with
this challenge, we opted for alternative solutions, including the use of renewable
energies.
A. Abdellah (B) · D. Toumi · M. Larbi

Laboratory of Energy Engineering and Computer Engineering (L2GEGI), Ibn Khaldoun
University, Tiaret, Algeria
D. Toumi
M. Larbi

https://doi.org/10.1007/978-981-15-6403-1_39
580 A. Abdellah et al.
However, Wind energy is one of the cleanest and most sustainable ways to produce
energy because it produces no toxic pollution or global warming emissions. The wind
is also abundant and inexhaustible, and affordable, making it a viable and large-scale
alternative to fossil fuels [1].
Currently, the generators most used in the production of wind energy are based
on the Doubly-Fed Induction Generator (DFIG) and PMSG. Faced with the prob-
lems of producing wind energy, the synchronous generator with permanent magnet
has advantages that are among others: elimination of DC excitation system, low
maintenance, …etc.
There are several methods of having been proposed for controlling PMSGs,
including the conventional PI controller may not give a satisfactory performance
under challenging operating conditions like a parameter variation.
To remedy this problem, research has turned to control techniques which ensure
the robustness and improve the operation performance of the PMSG.
In [2–4], the author worked on an induction machine using the PI controller,
Sliding Mode Controller (SMC) and Adaptive Fuzzy Sliding Mode Controller
(AFSMC) in order to compare between them and found that the AFSMC is better
than the others in time response, performances, robust to parameter variations and the
steady-state chattering elimination. However, there is an interesting robust method
that has better performances, it called Super-Twisting Sliding Mode control (STSMC)
that is studied in [5–8], their results show that it gets better performance in the test
of robustness and chattering elimination.
The idea of this article is concerning the comparison between three proposi-
tions of control as follow (PI–PI), (AFSMC–SMC) and (STSMC–SMC) the first
controller corresponded to the speed control and the second to the d-q currents
control. The purpose of this comparison is to see the control performances and
chattering elimination at each proposition.
The paper is planned as follows: in Sect. 2 the studied system modeling including
PMSG and wind turbine is presented. The control of speed and currents of the PMSG
using the three propositions of control is synthesized in Sect. 3. Section 4 presents
the results to compare the three propositions of control. Finally, the main conclusion
is presented in Sect. 5.
2 System Description
The proposed structure of an autonomous wind energy conversion system is

composed of a wind turbine, a PMSG, a PWM (Pulse Width Modulation) converter,
and a battery, as it is shown in Fig. 1.
Second-Order Super-Twisting Control of an Autonomous … 581
Fig. 1 The wind energy conversion system
A. Wind Turbine Model

The power produced by the wind is given by (1):
1
Pw = ρ π R3 V 3 (1)
2
The aerodynamic power captured by the wind turbine is dependent on the
produced power, and it’s given by (2):
1
Paer = ρ π R 3 V 3 C p (λ, β) (2)
2
where, ρ: the air density, R: the turbine radius, V: the wind speed, λ: the tip speed
ratio, β: the pitch angle.
Thus, the general form of the power coefficient C p is [1]:

C2 −C5
C p = C1 − C3 β − C4 e λi + C6 λi (3)
λi
where, C1 = 0.5176, C2 = 116, C3 = 0.4, C5 = 21C6 = 0.0068

and 1
λi
= 1
λ + 0.08β
− 0.035
β3 + 1
The tip speed ratio is given by (4):
Rt
λ= (4)
V
where t denotes the turbine speed.

The Maximum Power Point Tracking (MPPT) algorithm is used to maximize the
wind power extraction. For that, we set the tip speed ratio and the blade pitch angle
to their optimal values λopt = 8.1 and βopt = 0, respectively.
Thus, the optimum speed is given by:
λopt V
opt = (5)
R
The aerodynamic torque expression is given by (6):
Paer
Caer = (6)
t
In order to adapt the speed of the turbine with the generator one, a gearbox is
installed. Therefore, the turbine and the generator speeds are linked by a coefficient
noted G and called the gear ratio, as it is given in (7).
= G · t (7)
Thus, the torque of the generator is presented by (8):
Caer
Cg = (8)
G
B. Model of PMSG
The mathematical model of the PSMG is established under the d and q axis
synchronization reference frame [9].
It is simplified as:
did Rs Lq 1
= − i d + ωe i q + Vd (9)
dt Ld Ld Ld

diq Rs Ld 1 1
= − i q − ωe id + ϕf + Vq (10)
dt Lq Lq Lq Lq
where subscripts d and q refer to the physical quantities that have been transformed
into the d − q synchronous rotating reference frame, Rs is the stator resistance [],
L d and L q are the inductances [H] of the generator on the d and q axis, ϕ f is the
permanent magnetic flux [Wb] and ωe is the electrical rotating speed [rad/s] of the
generator, defined as ωe = p where p is the number of pole pairs and is the
mechanical rotating speed of the generator.
The electromagnetic torque equation of PMSG is described by this equation:
3
Te = iq p L d − L q id + ϕ f = K t iq (11)
2
3
Kt = p L d − L q id + ϕ f (12)
2
The mechanical equation which connects the generator with the wind turbine is
described by:
d
Jeq = Te + Tl − Bm (13)
dt
where Tl is the torque applied by the turbine to the generator. Jeq is the equivalent
moment of inertia; Bm and is the viscous turn coefficient.
Substituting (11) and (12) in (13) yields
d Kt Tl Bm
= iq + − = Ai q + DTl + B (14)
dt Jeq Jeq Jeq
where A = Kt
Jeq
, B = − BJeqm , D = 1
Jeq
.
3 Controller Synthesis
In this section, different types of controllers are examined and presented to regulate
the current and speed.
A. Classical PI Controller
Figure 1 shows the scheme of the generator side control [9]. It consists of two loops,
the first is used to control the d-q axis components of the stator currents of the PMSG.
The outer-loop is the speed control that tracks the optimal speed reference given in
(5) to generate the maximum power from the wind generator (Fig. 2).
B. Adaptive Fuzzy Sliding Mode Controller design
The structure comprises a speed regulation loop by AFSMC which generates the
∗ ∗
current reference Iq∗ . While regulating the currents Id−q imposes the Vd−q commands.
In this work, the regulating of currents Id−q are done by first order SMC. The choice
of surfaces for each loop is described as follows:
(1) The Speed Control of PMSG by AFSMC
The expression of speed error [2, 3]:
e (t) = ∗ (t) − (t) (15)

Fig. 2 Generator side control scheme of the direct drive wind turbine
Considering the following expression of switching surface:
S (t) = h (C e (t) + ė (t)) (16)
where C and h are strictly positive constant. Substituting (14) in (16), the first
derivative of S(t) is obtained as
∗
Ṡ (t) = h C ¨ ∗ − A Iq − D Tl − B
˙ − A i q − D Tl − B + ˙ (17)
During the sliding mode and the steady-state, S(t) = 0 and Ṡ (t) = 0. From (17)
we get the equivalent control law as:
1
Iqeq = ˙ ∗ − D Tl − B ) +
C ( ¨ ∗ − D Tl − B
˙ (18)
(1 + C) A
We must add a discontinuous term to the equivalent control part across the sliding
surface S(t). The discontinuous term is called reaching control part of control effort
and it is given as [3, 4].
Iqr = −(Ah)−1 k (t) sgn (S (t)) (19)

where k(t) > 0

Which give us the Iq∗ output control of the speed regulator.
Iq∗ = Iqeq + Iqr (20)
Stability condition can be obtained from Lyapunov stability theorem as [4],
V̇ (t) = S (t) . Ṡ (t) ≤ −η |S(t)| (21)
where, η is a strictly positive constant.

Substituting (16) and (17) in (21) and considering Ṫl < m, the stability of the
system is guaranteed by the following equation.
K (t) ≥ h m + η (22)
We know that sliding mode techniques can generate undesirable chattering. To

eliminate the chattering phenomenon, we replace a discontinuous “sgn” function by
a fuzzy system. Then, the reaching control can be defined as
Iqr = −(A h)−1 k (t) I f smc (23)
where I f smc is the output fuzzy system:
I f smc = F S MC (S (t),
S (t)) (24)
Iqr e f = Iqeq − (A h)−1 k (t) I f smc (25)
The input membership functions for the switching variable S(t) and the alteration
of the switching variable
S(t) are fuzzy sets negative (N), zero (Z), and positive (P).
The output membership functions for the control variable I f smc are divided into five
fuzzy segments namely negative big (NB), negative medium (NM), negative small
(NS), zero (ZE), positive small (PS), positive medium (PM), positive big (PB) on the
common universe of discourse is taken as [−1 1]. The selected membership function
for the input and output variables have triangular shapes as shown in Figs. 3 and 4.
Corresponding to these Figs, the thickness of the boundary layer can be changed by
modifying the universe of discourse of the fuzzy sets “Z” and “ZE” respectively in
the input and output.
The rule base expressed by linguistic terms is shown in Table 1.
The switching gain is adapted using the following algorithm according to S(t)
k̇ (t) = λk |S (t)| (26)

Fig. 3 Membership
1 N Z P
functions of the inputs (S(t),
Degree of mumbership function
S(t))
0.8
0.6
0.4
0.2
-1 -0.5 0 0.5 1
Fig. 4 Membership
functions of the output 1 NB NM NS ZE PS PM PB
Degree of membership function
(I f smc )
0.8
0.6
0.4
0.2
-1 -0.5 0 0.5 1
Table 1 Fuzzy rules used

dS S
N Z P
N NB NS PM
Z NB ZE PB
P NM PS PB
where, λk is a strictly positive constant. In fact, k(t) acts like an adaptive filter to
minimize the control effort.
(2) The Current Control of PMSG by SMC
The surfaces of the inner loops responsible for controlling the d-q axis stator
currents are described by [10].

S (t) = Iq∗ − Iq
(27)
S (t) = Id∗ − Id
The derivative of the surfaces is given by.

ϕf
Ṡ I q (t) = I˙q∗ + Rs
Lq
Iq + Ld
Lq
Id ωe + Lq
ωe − 1
Lq
Vq
(28)
Ṡ I d (t) = I˙d∗ +
Lq
Rs
Ld
Id − Ld
Id ωe − 1
Ld
Vd
During the sliding mode and the steady-state, we have.

S I dq (t) = 0 and Ṡ I dq (t) = 0
Hence we deduce.
⎧
⎨ V qeq = L q I˙∗ + Rs Iq + L d Id ωe + ϕf
ωe
q Lq Lq Lq
(29)
⎩ V deq = L d I˙∗ + Rs Id − L q Id ωe
d Ld Ld
In this work, the reaching control, it is given as.

Vqr = kiq tanh S I q (t)
(30)
Vdr = kid tanh (S I d (t))
Since the hyperbolic tangent function is continuous and smooth function, the
hyperbolic tangent function tanh(x) can be used to reduce the chattering in the sliding
mode control by using the hyperbolic tangent function tanh(x) instead of the sgn
function sgn(x) [5], where kiq , kid is the positive gain to be set. The outputs control
of the regulators Vq∗ , Vd∗ for the d-q axis stator currents are given by

Vq∗ = Vqeq + Vqr
(31)
Vd∗ = Vdeq + Vdr
The block diagram of the speed regulation by AFSMC and the d-q axis stator
currents regulation by SMC based on PMSG is shown in Fig. 5.
C. Super-Twisting Sliding Mode Controller design
In this section, a second-order SMC controller based on the Super-twisting algorithm,
and first-order SMC d-q axis stator currents controller are designed. The STSMC
consists of two terms: the equivalent control Iqeq and super twisting control Iqr ST .
We use the speed error and the surface that is shown in (15), (16), then during the
sliding mode and the steady-state the equivalent control law is the same as like as
(18). Therefore, the STSMC speed controller can be designed as [6, 8]:
Iq∗ = Iqeq + Iqr ST (32)
Iqr ST = I1 + I2 (33)
Fig. 5 Control system structure
1
I1 = −λ |S (t)| 2 sgn (S (t))
(34)
I˙2 = −γ sgn (S (t))
where λ and γ are the positive constants, in order to ensure the convergence in finite
time, the constants λ and γ can be chosen as follows [7].
⎧ ϕ
⎨λ >
min max (λ−ϕ) (35)
⎩γ2 ≥ 2
4ϕ
(λ−ϕ)
min min
The block diagram of the Super-twisting sliding mode controller is illustrated in
Fig. 6.
Fig. 6 The block diagram of the STSMC

4 Simulation Results and Discussion
The system studied was simulated in Matlab/Simulink environment with the param-
eters of the PMSG and the turbine used in the simulation are shown respectively in
Tables 2 and 3.
The system is initially simulated with the wind speed profile shown in Fig. 7. The
Fig. 8 shows that the overshoot of the AFSMC and STSMC controllers are better
than that of the PI controller and this last shows that there is a static error in the
steady-state, whereas the AFSMC and STSMC follow their reference speed but the
STSMC shows that less chattering than AFSMC and PI controller.
Figures 9 and 10 show respectively the response of the d-q axis stator currents of
the PMSG, which are controlled by the three propositions of control that we studied
in the previous section, are named by (PI–PI), (AFSMC–SMC), (STSMC–SMC).
From these Figs, it can be noted that there are large peaks in the q axis stator current
based on the PI controller and AFSMC at the transient regime and speed variation
(at 1 s). Also, the PI controller shows in d axis stator current a peak at the moment
of speed variation. On the other, hand the STSMC show that it has small peaks in
the q axis and no peaks in the d axis compared to the conventional PI controller
and AFSMC. It means that the AFSMC and STSMC ensure a perfect decoupling
between the two axes.
We notice that there is less chattering in the SMC which has the speed STSMC
compared to PI controller and AFSMC. Also using the harmonic spectrum of one
Table 2 PMSG Parameters

Parameter Value (units)
Rated power 2500 (W)
Rated line current 11 (A)
Rated speed 3000 (rpm)
Mutual inductance in q-axis 7.5 (mH)
Mutual inductance in d-axis 7.5 (mH)
Stator resistance 0.45 ()
Number of pole pairs 3
Permanent magnet flux 0.52 (Wb)
Friction coefficient 0.017 (Nm/rad/s)
Inertia moment 0.00208 (Kg.m2 )
Table 3 Turbine parameters

Parameter Value (units)
Rotor radius 3 (M)
Gear box ratio 5.4
Inertia moment 0.042 (Kg.m2 )
Friction coefficient 0.017 (Nm/rad/s)
Fig. 7 Wind profile 7
6.5
wind (m/s)
5.5
4.5
4
0 0.5 1 1.5 2
time (s)
100
W*
Mechanical speed (rad/s)
W AFSMC
80 W PI
W STSMC 94.85
60 70
94.8
40
65
20 94.75
60 1.5 1.505 1.51
0.005 0.01 0.015 0.02
0
0 0.5 1 1.5 2
time (s)
Fig. 8 Mechanical speed
30
-6 Iq AFSMC
Iq PI
20 -6.5 Iq STSMC
Iq*
current Iq (A)
10 -7
20
10
-7.5
0 0.5 0.55 0
-10
-10
1 1.005
-20
0 0.5 1 1.5 2
Time (s)
Fig. 9 PMSG stator current in the q-axis

4
2
Id AFSMC
3
Id PI
0
2 Id STSMC
Current Id (A)
-2 Id*
1 0.99 1 1.01
-1
-2
-3
0 0.5 1 1.5 2
Time (s)
Fig. 10 PMSG stator current in the d-axis
Fundamental (45.2493Hz) = 11.76 , THD= 2.78%

12
10
8
Mag
0
0 100 200 300 400 500
Frequency (Hz)
Fig. 11 Spectrum harmonic of one phase stator current for PI
phase stator current of the PMSG obtained using the FFT technique for the three
controllers which are presented in Figs. 11, 12 and 13. It can be seen that the THD is
reduced for STSMC (THD = 1.95%) when compared to AFSMC (THD = 2.29%)
and PI controller (THD = 2.78%). Therefore, the STSMC presents a great advantage
to ensure the lifetime of PMSG, it means that it protects the machine against the
over-current and the vibration which are generated by the chattering phenomenon.
In order to test the robustness of the used regulators in speed and currents, the
parameters of the PMSG are modified by doubling the values of Inertia moment and
friction coefficient. The result presented in Fig. 14 shows that STSMC and PI are
robust to the variation of the moment of inertia and viscous friction but the response
time of the STSMC is faster than the PI regulator, and for the AFSMC we notice that
there are a static error and a large uncertainty to external disturbances in the system.
It also found in Figs. 15 and 16 when we varied the parameters of the PMSG that the

12
10
8
Mag
0
0 100 200 300 400 500
Frequency (Hz)
Fig. 12 Spectrum harmonic of one phase stator current for AFSMC

12
10
8
Mag
0
0 100 200 300 400 500
Frequency (Hz)
Fig. 13 Spectrum harmonic of one phase stator current for STSMC
100
Mechanical speed (rad/s)
W*
90 W AFSMC
W PI 94.8
W STSMC
80
94.7
70 94.6
94.5
1.5 1.6 1.7 1.8
60
0 0.5 1 1.5 2
Time (s)
Fig. 14 Mechanical speed with parameters variation

0
Iq AFSMC
Iq PI
Current Iq (A) -5 Iq STSMC
Iq*
Inertia and Friction changing

-10
-15
0 0.5 1 1.5 2
Time (s)
Fig. 15 PMSG stator current in the q-axis with parameters variation
4
Id AFSMC Inertia and Friction changing
3
Id PI
2 Id STSMC
Current Id (A)
Id*
1
-1
-2
-3
0 0.5 1 1.5 2
Time
Fig. 16 PMSG stator current in the d-axis. with parameters variation
chattering phenomenon of the generator with the proposed (STSMC–SMC) is also

less as compared to the other proposed controllers.
5 Conclusion
In this paper, the control of autonomous wind energy conversion system based on
a PMSG has been studied. However, three types of regulators (PI–PI), (AFSMC–
SMC), (STSMC–SMC) to control PMSG has been proposed. Our objective is to
make a comparative study between them concerning the control performance and
chattering elimination. The simulations show that the STSMC has a better perfor-
mance in the transient and steady-state. Moreover, it is robust against parameter
variations compared with the two other regulators. Besides, the STSMC eliminates
the chattering phenomenon in speed and currents and ensures a perfect decoupling
between the two axes that leads to a great advantage by ensuring the lifetime for the
PMSG against vibrations and over-currents. On the other hand, The AFSMC shows
that there is a large uncertainty in the system appears in the form of an undesirable
error-static when we varied the parameters of the PMSG.
References
1. Barkia A et al (2016) A comparative study of PI and Sliding mode controllers for autonomous
wind energy conversion system based on DFIG. In: 2016 17th international conference on
sciences and techniques of automatic control and computer engineering (STA), Sousse, Tunisia,
pp 612–617
2. Boubzizi S, Abid H, Kharrat M et al (2015) Comparison of adaptive input-output linearization
& fuzzy sliding-mode control for induction motor. In: 2015 16th international conference
on sciences and techniques of automatic control and computer engineering (STA), Monastir,
Tunisia, pp 746–751
3. Saghafinia A et al (2015) Adaptive fuzzy sliding-mode control into chattering-free IM drive.
IEEE Trans Ind Appl 51(1):692–701
4. Saghafina A, Ping HW, Uddin MN, Gaied KS (2012) Adaptive fuzzy sliding-mode control
into chattering-free induction motor drive. In: 2012 IEEE industry applications society annual
meeting, Las Vegas, NV, USA, pp 1–8
5. Li Z, Zhou S, Xiao Y, Wang L (2019) Sensorless vector control of permanent magnet
synchronous linear motor based on self-adaptive super-twisting sliding mode controller. IEEE
Access 7:44998–45011
6. Meghni B, Dib D, Azar AT (2017) A second-order sliding mode and fuzzy logic control
to optimal energy management in wind turbine with battery storage. Neural Comput Appl
28(6):1417–1434
7. Azzaoui ME et al (2017) FPGA implementation of super twisting sliding mode control of
the doubly fed induction generator. In: 2017 14th international multi-conference on systems,
signals & devices (SSD), Marrakech, pp 649–654
8. Belabbas B, Allaoui T, Tadjine M, Denai M (2017) high order sliding mode controller simu-
lation by a wind turbine for DFIG protection against overcurrent. Electrotehnica, Electronica,
Automatica 65:142–147
9. Ayadi M, Naifar O, Derbel N (2017) Sensorless control with an adaptive sliding mode observer
for wind PMSG systems. In: 2017 14th international multi-conference on systems, signals &
devices (SSD), Marrakech, pp 33–37
10. Larbi M, Gherabi Z, Doudar K (2017) A robust sensorless control of PMSM based on sliding
mode observer and model reference adaptive system. IJPEDS 8(3):1016
Open-Switch Faults Based
on Six/Five-Leg Reconfigurable
AC-DC-AC Converter
Sahraoui Khaled and Gaoui Bachir
Abstract The purpose of this paper is the study by simulation of a fault-tolerant

control with Pulse Width Modulation (PWM) ac-dc-ac converter supplying a three-
phase rotor field-oriented induction motor. Before fault occurrence, the fault-tolerant
converter operates like a conventional back-to-back six-leg converter and after the
fault; it becomes a five-leg converter. Fault-tolerant topology of ac-dc-ac converter
that without redundancy have been studied and associated with affective and fast
method of fault detection and compensation to guarantee the continuity of service,
in the presence of a possible open circuit failure on the level of one of their legs.
Although of the presence of open circuit on the level of one of the converter legs the
control based on the Zero Sequence Signal (ZSS) assure the service continuity, the
simulation results obtained prove that it is possible to maintain the good performance
of the drive without redundancy leg (6/5 topology).
Keywords Asynchrouns machine · Fault diagnosis · PWM power convertors ·

Service continuity · Zero sequence signal
1 Introduction
Back-to-Back converters are widely used in various industrial areas such as electric
machine drives, uninterruptible power supply, unified power quality conditioner and
grid connected renewable energy systems [1, 2]. A research survey has reported
that the failure rates of these converter are 30%, 26%, 21% and 13% due to their
capacitors, printed circuit boards, semiconductors, and soldering, respectively [2, 3].
Therefore, the reliability and performance of PWM converters have been paid a great
deal of attention. Recently, fault detection and tolerant control techniques for power
S. Khaled (B) · G. Bachir

LACoSERE Laboratory, University of Amar Telidj, 03000, BP 37G Laghouat, Algeria
G. Bachir

https://doi.org/10.1007/978-981-15-6403-1_40
596 S. Khaled and G. Bachir
converters have attracted a lot of interest due to their higher reliability and lower
maintenance.
Recently, the using of power electronic increase dramatically in industrial appli-
cations, in the most industrial applications, induction motors are predominantly
fed from pulse width modulation voltage source inverter (PWM) for variables
speed operation [3]. Voltage three-phase structure static converters are essential
elements of many power electronic systems such as inverters machines alternatives,
uninterruptible power supplies, and active filters [4, 5].
The security of these systems, their reliability, performance, power quality and
continuity of service are today the major concerns in the field of energy [6]. PWM-
induction motors are usually more reliable than those supplied directly online. For
instance, the problem of faults can occur in the machine or in the converter [7–10].
The use of converters has some drawbacks because the power electronic converters
came with an increased possibility of component failures [6]. The failures of a power
converter, whether from drivers, its close control, a controllable power components,
lead to loss total or partial control of currents in phases. These failures can cause
serious system malfunctions. Indeed, it would endanger the system in some cases if
the defect is not detected quickly and compensated [2, 10].
Several researchers in the fault-tolerant converter applications have obtained
promising results by using an additional leg to guarantee the continuity of service, in
[2], the authors studied the fault-tolerant six-leg converter for a wind energy conver-
sion system (WECS) with a doubly-fed induction generator (DFIG). In [11–14], the
authors treated the fault detection, isolation, and compensation as an open-circuit
and short-circuit tolerant motor drive system and discusses fault-tolerant ac-dc-ac
single-phase to three-phase converter.
The approach introduced here minimizes the time interval between the fault occur-
rence and its diagnosis. This paper demonstrates the possibility to detect a faulty
switch in less than 40 µs by using a time criterion instead of voltage criterion. In
this paper, a more general fault-tolerant converter without redundancy is proposed
for the back-to-back three-phase two-level converters based on a 6/5 leg topology.
More general, fault-tolerant converter without redundancy leg is proposed for the
induction motor drives. The studied system is described briefly in Sect. 2. In Sect. 3,
a modeling and control of the two sides of our converter and the induction machine
are devoted. The fault detection and control schemes method are used as in [6, 15, 16]
discussed in Sects. 4 and 5. Finally, fully simulation results are provided in Sect. 6.
The obtained simulation results here prove that the continuity of service is assured
although the presence of faults in the switches.
A technical advantage of the voltage converter PWM from the decoupling capacitor
placed between the network-side converter and the machine side of the converter.
The network side converter provides sinusoidal currents and voltage regulation of the
Open-Switch Faults Based on Six/Five-Leg Reconfigurable … 597
DC bus and the machine side of the converter which assure the control thereof. This
will provide some protection and separate controls for the two sides. By against the
important inconvenience of this voltage converter PWM lies in the switching losses
and reduce the overall life of the system [1, 14] (Figs. 1 and 2).
converter which assure the control thereof. This will provide some protection and
separate controls for the two sides. By against the important inconvenience of this
voltage converter PWM lies in the switching losses and reduce the overall life of the
system [1, 14].
Pulse Width Modulation (PWM)
Electrical
grid
IM
DC-AC Inverter AC-DC Rectifier
Fig. 1 Topology of conventional converter ac-dc-ac PWM
Fig. 2 Control of converter side grid

3 Modeling and Control of Studied System
A. Modeling of converter side grid
A PWM rectifier is simply a voltage inverter used in reverse. It produces a DC voltage

from an AC supply, It is built based on bidirectional semiconductor components.
Also it preferred for frequently applications operating in regenerative mode such as
electric drives especially where the AC rectifier is part of the controller. According
to the closing or opening of the switches, kij branch voltages can be equal to Uc or
0. Introduces another variable S11, S21 and S31 which taking 1 if the switch ki1 is
closed or −1 if the switch opened [17, 18].
⎡ ⎤ ⎡ ⎤⎡ ⎤
V1N 2 −1 −1 S11
⎢ ⎥ U c⎣
⎣ V2N ⎦ = −1 2 −1 ⎦⎣ S21 ⎦ (1)
6
V3N −1 −1 2 S31
The rectified current can be written (Fig. 3)
S11 + 1 S21 + 1 S31 + 1

i rec = ia + ib + ic (2)
2 2 2
Fig. 3 Control of converter side machine

B. Modeling of converter side machine
The inverter is a power electronics converter that allows conversion from contin-
uous to alternative. It works by forced commutation and designed generally based
transistors
⎡ ⎤ ⎡ ⎤⎡ ⎤
V1N 2 −1 −1 f 11
⎢ ⎥ Uc ⎣
⎣ V2N ⎦ = −1 2 −1 ⎦⎣ f 21 ⎦ (3)
6
V3N −1 −1 2 f 31
The input current of inverter can be expressed by
f 11 + 1 f 21 + 1 f 31 + 1
if = i as + i bs + i cs (4)
2 2 2
C. Modeling of the machine
The equations which describe the dynamic behavior of asynchronous machine [6].
Stator voltage equations

Vsd = Rs i sd + dφsd
dt
− ωs φsq
dφsq (5)
Vsq = Rs i sq + dt
− ωs φsd
Rotor voltage equations

Vrd = Rr i rd + dφrd
dt
− (ωs − ωr )φrq
dφrq (6)
Vrq = Rr i rq + dt
− (ωs − ωr )φrd
Stator flux equations

φsd = L s i sd + Mi rd
(7)
φsq = L s i sq + Mi rq
Rotor flux equations

φrd = L r i rd + Mi sd
(8)
φrq = L r i rq + Mi sq
Torque equation
3 pM
Te = (φrd i sq − φrq i sd ) (9)
2L r
4 Fault Tolerant System
In case of fault occurrence in sensors, actuators or other parts of a complex system,

conventional systems may produce undesirable results. In critical applications where
the need for safety, reliability and fault tolerance is high, it is necessary to anticipate
the fault tolerance ability against unpredicted faults to increase the reliability and
availability of the system while providing acceptable performances [10, 13]. These
types of systems are known as Fault-Tolerant Systems (FTS). In other words, an
FTS can maintain stability and acceptable performance after a fault occurrence. The
performance of the system in normal and post-fault conditions may be different.
An FTS may be based on redundancy or not. In an FTS with redundancy, a redun-
dant part will replace the faulty part, after fault detection and upon reconfigura-
tion. In this case, by correctly designing the FTS, the same operation capability can
be obtained. However, in some cases, some degree of performance degradation is
accepted. In these cases, a suitable reconfiguration might be enough to assure the
minimum required performance [10, 18, 19].
After fault detection and isolation, a reconfiguration is necessary. Reconfiguration
is required in both hardware and software parts of the system. Reconfiguration in
the software part consists of modifying the references and the controller character-
istics. The system is composed of three-phase electrical sources and an induction
machine, with an ac-dc-ac fault tolerant converter between them. All different parts
are explained further in detail.
D. 6/5 leg converter topology
The studied converter is shown in Fig. 4 [14]. In normal operation (before the fault
occurrence), the converter is a conventional back-to-back converter with three addi-
tional bidirectional switches (Fig. 4a). These bidirectional switches are used for the
converter reconfiguration after the fault isolation, i.e. before the fault occurrence,
these switches are all off. The converter configuration after a fault in one of the
inverter legs is shown in Fig. 4b. In this case, the fault has been occurred at either
B31 or B32. Depending on the application, each side of the converter can be connected
to a source, load or machine. For example, in drive application, one side might be
connected to a three-phase source and the other will supply a three-phase achine. For
a WECS with DFIG, side 2 will supply the rotor of the machine. In this paper, it is
assumed that side 1 of the converter is connected to a three-phase balanced sinusoidal
source and the second side of the converter is connected to a balanced three-phase
load.
It should be noted that the voltage producing capability in the five-leg converter
is lower than the six-leg converter, meaning that with the same dc-link voltage, it can
produce smaller three-phase voltages at its AC terminals, compared to the six-leg
Fig. 4 6/5 leg fault tolerant converter a pre-fault b post-fault (after reconfiguration)
converter. Also, the current rating in the shared leg of the converter is larger than in
the other legs.
E. Reconfigurable Control of the converter
Before the fault occurrence, the fault-tolerant converter is performing like a conven-
tional six-leg converter and any PWM method can be used for its control. After the
fault detection and reconfiguration, one leg will be shared between the two sides of
the converter and the faulty leg will be disconnected. So, the converter will become
a five-leg structure. Among the several PWM approaches studied in the literature
for this 5- leg post-fault topology, it seems that the suggested method in [15, 16].
Therefore this approach is used in this paper.
In this method, a so-called double zero-sequence injection method is performed.

∗
Voltage reference signals vxi (x ∈ {a, b, c}, i ∈ {1, 2}) for both sides of the converter
are calculated using the appropriate methods depending on the targeted applications.
Then, a Zero Sequence Signal (ZSS) is added to these values to form the modulation
signals (10). ZSS does not change the output line-to-line and phase voltages, therefore
it is used as a degree of freedom to reduce the current harmonics and improve the
dc-bus utilization.
∗
vxi (t) = vxi (t) + vzssi (t) (10)
where vzssi (i ∈ {1, 2}) is the ZSS for side i.

ZSS signal for a three-phase system is computed as follows:
∗ ∗

1 max(vai (t) + vbi (t) + vci∗ (t))+

vZSSi (t) = − ∗ ∗ (11)
2 min(vai (t) + vbi (t) + vci∗ (t))
Since there are 6 v references and only 5 legs, a reduction in the number of voltage
references is required. Reduction of the number of voltage references can be made
by using an inverse lookup table [2, 11], this is realized by adding another ZSS in
accordance to the converter configuration in five-leg mode. The new set of voltage
references assuming that two c legs of the two sides are connected in the five-leg
mode are calculated as:
v A1 = va1 + vc2 ; v B1 = vb1 + vc2

v A2 = va2 + vc1 ; v B2 = vb2 + vc1
vC = vc1 + vc2 (12)
Since the same signal is added to all three reference values of the converter, the
fundamental output voltage will not be affected. Figure 5 shows the principle of this
method. In Fig. 5a the ZSS for a conventional converter is shown. This ZSS injection
will be repeated for both sets of three-phase voltage references at the two sides of
the converter, then by using (11) a new set of five voltage references are produced
and sent to the PWM unit, as shown in Fig. 5b.
F. Voltage reference generation
In fact, in healthy or faulty cases, the controls of both converters are not the same
and must be changed as quickly as possible to avoid any discontinuity or transient in
the currents The control unit of the 6/5 leg converter produces the reference voltages
and sends them to the PWM blocks [21, 22]. For each application case, depending on
the source/load connected at two sides of the converter, there are different methods
for producing the reference voltages.
Fig. 5 a Principle of PWM module for one converter; b for 5-leg converter when B3 legs are shared
5 Fault Detection and Reconfiguration
Fast fault detection is essential for fault-tolerant systems. Here the method detailed in
our earlier contributions is used for detection of fault and its location [20]. Using this
method, it is possible to detect the fault quickly and efficiently. In [20] it is shown that
the fault occurrence in each leg can be diagnosed by comparing the measured and
estimated pole voltages. However, in reality, due to measurement and discretizing
errors, and mainly because of non-ideal behaviors of switches and drivers (turn-off
and turn-on propagation time and dead time generated by the drivers), the voltage
error is not zero during normal operation [23, 24].
To avoid false fault detections, two adjustments are employed to compensate for
the effect of the measurement errors and delays. For this purpose, first, the absolute
value of the error between measured and estimated pole voltage is calculated. Then,
this value applied to a comparator with a threshold value h, to determine if the differ-
ence between the measured and estimated voltages is large enough to be considered
as an error. Then, this signal is applied to an upcounter that computes the number of
pulses while the output of the first comparator is high. In the other words, the output
of the up-counter corresponds to the time during which vknm (measured voltage)
and vknes (estimated voltage) are different [7, 13].
Consequently, the fault occurrence is detected using simultaneously a time crite-
rion and a ‘voltage criterion’. To do this, the up-counter output is applied to a second
comparator with a threshold value of ‘N’. In this way, false fault detection due to
semiconductor switching is avoided and fault can be detected very fast. Figure 6
shows one leg of the converter, while there is an open-circuit fault in the upper
switch. Note that in some cases, based on the direction of the current ik, there might
be a condition that Dk conducts instead of Sk; therefore, in this case, the converter
operates normally and the fault cannot be detected. For example, while Tk = 1 and
ik < 0, Dk is on. Here Tk is the command signal for the upper switch of the leg k.
Tk = 0 indicates that the switch is commanded to be open, whilst Tk = 1 means
that the switch is commanded to be closed. The switch commands in each leg are
complementary.
Figure 7 shows the mentioned fault detection principle. The stateflow diagram of
the fault detection scheme is presented in Fig. 8. For a short-circuit fault of a switch,
the very fast acting fuses isolate the faulty leg. Lets now consider two converter legs
as shown in Fig. 9 and suppose that there is a short circuit fault in Sk. When Tk = 1
the leg k operates normally. However when the Tk = 0, the dc-link is short-circuited.
Figure 9b shows the short circuit path. The short circuit current is limited by only
the short circuit resistance, which is a small value. That is why suitable protection
Fig. 6 One leg of converter

during an open-circuit fault
in the upper switch
Fig. 7 Fault detection scheme
Fig. 8 State-flow diagram of the fault detection scheme
is needed in this case. In this case, if the fuse clearing time is smaller than the fault
detection time, the faulty leg will be isolated by the fuses. If the fault is detected
before the fuses act, the command of the faulty leg switches is set at zero and the
corresponding bidirectional switch will be turned on. In this case, there is still at least
one fuse in the short circuit path; therefore the short circuit will be cleared after fuse
operation (Fig. 9c).
We simulated the system studied by the following test: At first, we start our machine
unloading ref = 100 rad = s then we inserting and removing a torque load 12 N·m
in instants t = 0.3 s and t = 0.7 s respectively, after we change direction of rotation
ref = −100 rad = s at t = 1 s, after that we inserting and removing a torque load
−12 N·m in instants t = 1.4 s and t = 1.7 s respectively.
We considered that the fault has occurred in the converter side machine in the leg
B3 at t = 0.5 s. The fault-tolerant control is activated after 40 µs, by the homopolar
injection method (ZSS) which ensures the continuity of service of our system.
Fig. 9 Equivalent circuit for

the leg ‘k’. a A short circuit
fault in Sk. b Short circuit
path after the fault. c Short
circuit path when the fuse
clearing time is larger than
the fault detection time
The obtained results are presented in Fig. 10. We observed that the converter
operates normally, as a healthy system. DC bus voltage control stays stable (Fig. 11c)
40
0.1*va (V)
20
Va(V), ia(A)
ia (A)
0
-20
-40
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
(a) Voltage and current sources
Rotor speed (rad/s)
100
-100
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Time (s)
(b) Rotor speed
1.5
Rotor flux (Wb)
Phir (Wb))
1 Phiqr (Wb)
Phidr (Wb)
0.5
-0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
(c) Rotor flux
Input voltage rectifier (V)
500
-500
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
(d) Input voltage rectifier
Fig. 10 Simulation results of the healthy system
800
600
Uc (V)
400
200
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
(e) DC bus voltage control
100
Tem (Nm)
50
Tem (N.m)
Temref (Nm)
0
-50
-100
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
(f) Electromagnetic torque
Stator current phase (A)
40
20
-20
-40
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
(g) Stator current phase
500
Output voltage source
inverter (V)
-500
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
(h) Output voltage source inverter
Fig. 10 (continued)
50
V (V), i (A) 0.1*va (V)
ia (A)
a
0
a
-50
0 0.5 1 1.5 2
Time (s)
(a) Voltage and current sources
50
Zoom in V (V), i (A)
0.1*va (V)
a
ia (A)
0
a
-50
0.2 0.3 0.4 0.5 0.6 0.7 0.8
Time (s)
(b) Zoom in voltage and current sources
1400
U (V)
c
1350
0 0.5 1 1.5 2
Time (s)
(c) DC bus voltage control
Output voltage source
1000
inverter (V)
-1000
0 0.5 1 1.5 2
Time (s)
(d) Output voltage source inverter
Fig. 11 Simulation results of topology 6/5 leg fault tolerant without redundancy leg (the fault has
been occurred in B3)
200
Rotorspeed (rad/s)
100
-100
-200
0 0.5 1 1.5 2
Time (s)
(e) Rotor speed
Electromagnetic torque
100
Tem (Nm)
Temref (Nm)
0
(N.m)
-100
-200
0 0.5 1 1.5 2
Time (s)
(f) Electromagnetic torque
1.5
Rotor flux (Wb)
1
Phir (Wb)
Phiqr (Wb)
0.5
Phidr (Wb)
-0.5
0 0.5 1 1.5 2
Time (s)
(j) Rotor flux
Stator current phase (A)
40
20
-20
-40
0 0.5 1 1.5 2
Time (s)
(h) Stator current phase
Fig. 11 (continued)
(we constrained to increase the DC bus voltage to ensure the necessary capacity in
the production of the three alternating voltages).
Figure 11a presents the line current in the input system, we can observe converter
side grid before and after the fault detection and compensation, it is in phase with the
voltage quasi sinusoidal and does not mark a significant peak. The other controlled
quantities suit regularly their references (rotor speed, torque, and flux). Also, we can
note that no noticeable deformation appears on the waveforms, because the time of
fault detection and composition is shorter (40 µs) than the commutation period of
the switches.
The ac drive system configuration employs a five-leg converter and can handle
the failure of all the power switches connected to the positive or negative rail of the
dc-link. The operation of drive in the post-fault is intended for short time periods to
avoid problems related to losses and magnetization. However, the magnetization of
the machine as resulted from its operation in the post-fault condition can be removed
once the drive is repaired by designing a suitable pulse width modulation pattern.
The simulation results have demonstrated the feasibility of the proposed strategy for
improving the reliability of a three-phase ac motor drive system.
7 Conclusion
In this paper, we devoted our study of the topology without redundancy leg converter
“back-to-back”, named in this paper “6/5 leg converter”, with a very fast fault detec-
tion scheme and suited reconfigurable control is studied. Moreover, a specific and
appropriate “fault-tolerant” controller was necessary to guarantee the fast and effi-
cient reconfiguration of the system, not only at the level of the converter topology
but also at the level of its control, which must be modified during the commuta-
tion a 6-leg topology with a 5-leg topology. However, we do not know at which leg
possible defect would occur and which leg would be shared. Similarly, an increase in
the DC bus voltage reference may be required after reconfiguration to maintain the
converter’s rated capacity if system operation requires it. It is, therefore necessary
to dimension accordingly the entire 6-leg “fault-tolerant” converter if it is desired
to guarantee operation in normal mode after defect dimensioning. Finally, we have
proposed methods to reduce the number of voltage sensors required for fault detec-
tion. This ability to reduce the number of voltage sensors reduces the additional cost
of fault tolerance.
References
1. Extremiana G, Abad G, Arza J, Torre I (2013) Rotor flux oriented control of induction machine
based drives with compensation for the variation of all machine parameters. Pol Acad Sci Tech
Sci 61(2):309324
2. Shahbazi JM, Zolghadri MR, Poure P, Saadate S (2013) Wind energy conversion system
based on DFIG with open switch fault tolerant six-legs AC-DC-AC converter. In: 2013 IEEE
international conference on industrial technology (ICIT). IEEE, pp 1656–1661
3. Dehghan SM, Amiri A, Mohamadian M, Andersen MAE (2013) Modular space-vector pulse-
width modulation for nine-switch converters. IET Power Electron 6(3):457467
4. Jones M, Dujic D, Levi E (2008) A performance comparison of PWM techniques for five-leg
VSIs supplying two-motor drives. Proc
5. Frapp E, De Bernardinis A, Bethoux O, Marchand C, Coquery G (2010) Action palliative par le
convertisseur statique en cas de dfaillance dun gnrateur PAC modulaire de puissance, Electron,
puissance du Futur
6. Sahraoui K, Gaoui B, Mokrani L, Belarbi K (2016) Fault tolerant control of redundant topology
of pwm ac-dc-ac converter supplying an electromechanical drive. In: 2016 8th international
conference on modelling, identification and control (ICMIC). IEEE, pp 374–380
7. Jacobina CB, De Freitas IS, Da Silva ERC, Lima AMN, De Araujo Ribeiro RL (2006) Reduced
switch count DC-link ACAC five-leg converter. IEEE Trans Power Electron 21(5):13011309
8. Sae-Kok W, Grant DM, Williams BW (2010) System reconfiguration under open-switch faults
in a doubly fed induction machine. IET Renew Power Gener 4(5):458–470
9. Peng T, Dan H, Yang J, Deng H, Zhu Q (2016) Openswitch fault diagnosis and fault tolerant
for matrix converter with finite control set-model predictive control. IEEE Trans Ind Electron
63(9):59535963
10. Sahraoui K, Kouzi K, Ameur A (2017) A robust sensorless iterated extended Kalman filter for
electromechanical drive state estimation. Electrotehnica, Electronica, Automatica 65(2)
11. Daigavane MB, Vaishnav SR, Shriwastava RG (2015) Sensorless field oriented control of
PMSM drive system for automotive application. In: 2015 7th international conference on
emerging trends in engineering & technology, no 2, pp 106–112
12. Karimi S, Gaillard A, Poure P, Saadate P (2008) FPGA based real-time power converter failure
diagnosis for wind energy conversion systems. In: IEEE transactions on industrial electronics,
vol 55, No 12, pp 4299–4308
13. Shahbazi M, Poure P, Saadate S, Zolghadri MR (2011) Five-leg converter topology for wind
energy conversion system with doubly fed induction generator. Renew Energy 36(11):3187–
3194
14. Jacobina CB, Araujo Ribeiro RL, Lima AMN, da Silva ERC (2003) Fault-tolerant reversible
AC motor drive system. In: IEEE transactions on industry applications vol 39, pp 1077–1084
15. Jones M, Vukosavic SN, Dujic D, Levi E, Wright P (2008) Five-leg inverter PWM technique for
reduced switch count two-motor constant power applications. IET Electr Power Appl 2(5):275–
287
16. Jones M, Dujic D, Levi E (2008) A performance comparison of PWM techniques for five-leg
VSIs supplying two-motor drives. IECON, pp 508–513
17. Rodrigues PLS, Jacobina CB, de Freitas NB (2019) Singlephase universal active power filter
based on acdcac converter with eight controlled switches. IET Power Electron
18. Malinowski A, Yu H (2011) Comparison of embedded system design for industrial applications.
IEEE Trans Industr Inf 7(2):244–254
19. Shahbazi M, Poure P, Saadate S, Zolghadri MR (2013) FPGAbased reconfigurable control
for fault-tolerant back-to-back converter without redundancy. IEEE Trans Ind Electron
60(8):3360–3371
20. Shahbazi M, Saadate S, Poure P, Zolghadri M (2016) Open-circuit switch fault tolerant wind
energy conversion system based on six/five-leg reconfigurable converter. Electr Power Syst
Res 137:104–112. End the bibliography
21. Sahraoui K, Gaoui B, Mokrani L, Belarbi K (2018) Fast detection control for fault-tolerant
converter back-to-back with redundancy leg supplying an induction motor drives. Int J Digit
Signals Smart Syst 2(1):15–35
22. Sahraoui K, Kouzi K, Ameur A (2017) Optimization of MRAS based speed estimation for speed
sensorless control of DSIM via genetic algorithm. Electrotehnica, Electronica, Automatica
65(3)
23. Sahraoui K, Ameur A, Kouzi K (2018) Performance enhancement of sensorless speed control of
DSIM using MRAS and EKF optimized by genetic algorithm. In: 2018 international conference
on applied smart systems (ICASS). IEEE, pp 1–8
24. Sahraoui K, Kouzi K, Ameur A (2018) Neural networks trained with sliding mode control for
DSIM supplied by two voltage inverters on three levels. In: 2018 international conference on
electrical sciences and technologies in maghreb (CISTEM). IEEE, pp 1–6
A Robust Model Predictive Control
for a DC-DC Boost Converter Subject
to Input Saturation: An LMI Approach
O. Hazil, S. Bououden, and M. Chadli
Abstract The design of robust model predictive controller (RMPC) for nonlinear
system under actuator saturation is a challenging topic in the area of systems control.
Yet. A new RMPC algorithm is developed here for control of a DC-DC boost
converter subject to actuator saturation. Based on the saturating actuator property, an
appropriate linear feedback control law is designed that minimises an infinite horizon
cost function within the framework of linear matrix inequalities. In particular, it is
shown that the solution of the optimisation problem can stabilise the system under
actuator nonlinearity. The extensive simulation results are presented to illustrate the
efficiency of the proposed control scheme.
Keywords Rrobust model predictive controller · DC-DC boost converter ·

Actuator saturation · Linear matrix inequalities
1 Introduction
Power converters are used extensively in most of the power supply systems such as
personal computers, laptops, aircrafts and electronic equipment. A DC-DC converter
is a switching circuit, which transform a certain electrical voltage to another level
O. Hazil
Centre de Développement des Energies Renouvelables CDER, B.P. 62, Route de l’observatoire,
Bouzaréah, 16340 Alger, Algérie
Faculty of Technology, Department of Electronics, University of Constantine 1, Campus A.
Hamani, Route Ain El Bey, 2500 Constantine, Algeria
S. Bououden (B)
Department of Industrial Engineering, University Abbes Laghrour, Khenchela, Algeria
M. Chadli
IBISC Lab.-UEVE, University of Paris-Saclay, 40 rue Pelvoux, Evry, France

https://doi.org/10.1007/978-981-15-6403-1_41
616 O. Hazil et al.
of voltage, this is obtained by switches operating (open or closed) at high frequen-

cies; the control objective of such devices is to maintain regulation of the output
voltage at the desired value [1, 2]. Generally, the dynamic behavior of DC-DC
converters is usually described by bilinear models with non-minimum phase, under
saturating control signal. In spite of nonlinearities, the linear feedback control of
power converters based on the small signal control theory proposed by Middlebrook
and Cuk [3] is widely applied at present. For instance, the proportional–integral (PI)
controllers [4–6] and PID controllers [7, 8]. The models used to derive the control
strategy usually neglect the high-frequency ripple considering that the switching time
is small enough [9].
In DC-DC converters, the duty ratio is bounded in the interval [0, 1]. The clas-
sical control techniques applied to power converters do not take into account input
saturation constraints, which can lead to degradation of performance and even to
destabilization of the closed-loop system. In this work, we interest to exploit the
ability of the linear matrix inequality (LMI) based MPC algorithm for the control of
systems with actuator non-linearity, in which a new approach consists of a saturation
model based on a deadzone nonlinearity is used to handle the saturation non-linearity.
At each time step, a minimization problem of an infinite horizon performance index
is solved to obtain a state feedback control law. The stability of the system is guaran-
teed in the sense of Lyapunov, and the adequate state feedback synthesis condition is
obtained by solving an LMI optimisation problem, which directly incorporates input
saturation. By this means a computationally cost effective algorithm is proposed,
which is implementable for a wide range of systems with actuator non-linearity,
This paper is organized as follows. Section 2 gives the averaged model of DC-
DC boost converter and the mathematical formulation of the problems. Section 3
describes the mathematical formulation of the proposed RMPC problem with guar-
anteed stability as an LMI-based optimization problem. Finally we give some
simulation results to test the robustness of the proposed control law in Sect. 4.
2 Averaged Model of Basic PWM Boost Converter
A. State-space of the boost converter
The schematic circuit diagram of boost DC-DC converter under consideration is

shown in Fig. 1, and the relevant control signal.
In Fig. 1, vs is the output voltage and vin the line voltage. The output voltage must
be controlled using the duty ratio d in order to keep the given constant vref value,
the diode D is on inverse polarization, R models the converter load, while C and
L represent, respectively, capacitor and inductor values, the switch sw was a power
transistor controlled by a binary signal ub . The binary signal that triggers on and
off the switches is controlled by a fixed-frequency pulse width modulation (PWM)
circuit (Fig. 2). The constant switching frequency is 1/T s , with T s the switching
period is given by the sum of T on (when ub = 1) and T off (when ub = 0) and the ratio
A Robust Model Predictive Control for a DC-DC … 617
iL
L D
vin sw C
vC R vs
PMW d (t )
Fig. 1 Schematic of the boost converter
u b (t ) v (t )
d (t )
T on T off
Fig. 2 Waveforms of the PWM process
of T on /(T on + T off ) is the duty cycled d(t). Duty cycle is compared with a sawtooth
signal v(t) of amplitude equals to 1. Consequently 0 ≤ d ≤ 1.
We assume that the converter operates in continuous conduction mode (CCM)
and that the inductor current is always larger than zero. So, the studied converter
has two working topologies corresponding to its switch states (Fig. 3). Choosing the
state vector as x = [i L , vC ]T where vc is the capacitor voltage; iL is the inductor
current, the general equation that governs the operation of the boost converter is:

ẋ = Ai x + Bi
(1)
y = Ci x
where: i = 1 to the first configuration described in Fig. 2a and: i = 2 the second

configuration shown in Fig. 2b. Where:
iL iL
L L
vin vC
C
vs vin vC C R vs
R
(a) (b)
Fig. 3 Equivalent circuit of the boost converter. a Switch on. b Switch off
618 O. Hazil et al.

0 0 1/L
A1 = , B1 = , C1 = 0 1
0 −1/RC 0

0 −1/L 1/L
A2 = , B2 = , C2 = 0 1
1/C −1/RC 0
So the combination of state-space representation of mode 1 (on mode) and mode

2 (off mode) induces the following linear state space representation:

ẋ = Ad x + Bd u
(2)
y = Cx
With Ad = d A1 +d A2 ,Bd = d B1 +d B2 ,Cd = dC1 +d C2 and d = 1−d, u =

d. This gives:

0 −(1 − d)/L 1/L
Ad = , Bd = , Cd = [ 0 1 ]
(1 − d)/C 1 − /RC 0
B. Discrete-time state-space model
The MPC is to be fed with a discrete time model which is easily obtained from the
continuous model assuming that the switching period T s is much smaller than the time
constants associated with the circuit, the following discrete-time model is obtained
using the forward Euler approximation, from the continuous time version (2):
x(k + 1) = (I + Ts Ad )x(k) + Ts Bd d(k) (3)
Using this assumption, the discrete-time state-space model of the buck-boost

converter can be written as:

x1 (k + 1) 1 −(1 − d(k))Ts /L x1 (k) Ts /L
= + d(k)
x2 (k + 1) (1 − d(k))Ts /C 1 − (Ts /RC) x2 (k) 0
(4)
The system in (4) is a nonlinear system since there are products of two input
signals. It is therefore mandatory to make some sort of linearization at some point.
We can rewrite the nonlinear model of boost converter as follows:
x(k) = f (x(k), sat(u(k))) (5)
where sat(u(k)) is the saturated control input u(k) of the boost converter, constrained
between 0 and 1 and is defined as:
⎧
⎨0 if u(k) < 0
sat(u(k)) = u(k) if 0 ≤ u(k) ≤ 1 (6)
(0,1) ⎩
1 if u(k) > 1
The total instantaneous quantities can be presented as the sum of the DC and AC
components,
x = x̃ + X (7)
d = d̃ + D (8)
where X and D represent the equilibrium values and x̃ and d̃ are the perturbed values
of state and duty cycle, The linearized averaged model of the DC-DC converter can
be written as
x(k + 1) = A x̃(k) + AX + B sat (K x̃(k) + K X ) (9)

(0,1)
where the matrices A and B are given by:

δ f
1 −(1 − D)Ts /L δ f
Vc Ts /L
A= = ,B = =
δx
x=X (1 − D)Ts /C 1 − Ts /RC δd
x=X −I L Ts /C
d=D d=D
The control input is subject to a no symmetric actuator saturation, we can use

the method proposed in [9] to transform it to a symmetric saturation. The function
sat(u(k)) can be can be written as follows:
⎧
⎨0 if K x̃(k) + K X < 0
sat (K x̃(k) + K X ) = u(k) if 0 ≤ K x̃(k) + K X ≤ 1 (10)
(0,1) ⎩
1 if K x̃(k) + K X > 1
Which can be rearranged in the following form:

⎧
⎨0 if K x̃(k) < −K X
sat (K x̃(k)) = u(k) if K X ≤ K x̃(k) ≤ 1 − K X (11)
(K X,1−K X ) ⎩
1 if K x̃(k) + K X > 1 − K X
Both saturation functions are equivalent if we add KX to (11), that is:
sat (K x̃(k) + K X ) = sat (K x̃(k)) (12)

(0,1) (K X,1−K X )
Assume that the operating point is an equilibrium point which the steady state
part of the model is equal to zero
620 O. Hazil et al.
f (X, U ) = AX + K X = 0 (13)
The steady state control signal is KX = 0.5. The system described in (9) can be
written as follows:
x̃(k + 1) = A x̃(k) + Bsat(δ̃(k)) (14)
where sat(δ̃(k)) = sat (K x̃(k)) the new control input and the new model can
(K X,1−K X )
take into account the symmetric saturation model as in [10]:
−u 0 < u(k) < u 0 ↔ −K X < δ(k) < K X (15)
We defined a new decentralized deadzone nonlinearity
φ̃(k) = sat(δ̃(k)) − δ̃(k) (16)
The closed-loop system is rewritten as:
x̃(k + 1) = (A + B K )x̃(k) + B φ̃(k) (17)
The saturation obtained in (18) is a symmetric saturation and the following lemma
can be used to handle the saturation nonlinearity
Lemma 1 ([11]) Assume that the saturation constraint is defined as (18), let:
φ = sat(δ) − δ (18)
Then, there exists a real number ε ∈ (0, 1) such that
φ T φ ≤ εδ T δ (19)
where,ϕ = [ϕ1 , ϕ2 , . . . , ϕm ]T ∈ R m and ϕi = (i = 1, 2, . . . , m) is the dead-zone

nonlinearity function.
Proof See [12].
Moreover, we have the following technical lemmas for later use.
Lemma 2 (Schur complements) The LMI

Q(x) S(x)
>0 (20)
S(x)T R(x)
In which, Q(x) = Q(x)T , R(x) = R(x)T and S(x) are affine functions of x, and
is equivalent to
R(x) > 0, Q(x) − S(x)R(x)−1 S(x)T > 0 (21)
or, equivalently
Q(x) > 0, R(x) − S(x)Q(x)−1 S(x)T > 0 (22)
Proof See [13].

Lemma 3 For any matrices (or vectors) X and Y with appropriate dimensions, we
have:
X T Y + Y T X ≤ α X T Y + α −1 Y T Y (23)
where α > 0
C. Uncertainty Model
We consider that the load R at the operating point is uncertain or time-varying param-
eter. Then, matrice A depend on such uncertain which have been grouped in a vector
p, and we can express (17) as a function of these parameter
x̃(k + 1) = (A( p) + B K )x̃(k) + B φ̃(k) (24)
Generally, the vector p consists of N uncertain parameters p =(p1 , …, pN ), where

each uncertain parameter pi is bounded between a minimum and a maximum value
p̄i and P i

pi ∈ pi , pi (25)
The admissible values of vector p are constrained in an hyperrectangle in the

parameter space ɌN with L = 2 N Vertices {ʋ1 , …, ʋL }. The images of the matrix
A(p) for each vertex ʋi corresponds to a set {j 1 , …, j L }. The components of the
set {j 1 , …, j L }are the extrema of a convex polytope, noted Co{j 1 , …, j L }, which
contains the images for all admissible values of p if the matrix A(p) depends linearly
on p, that is
L

L
A( p) ∈ Co{ς1 , . . . , ς L } = λi ς1 , λi ≥ 0, λi (26)
i=1 i=1
In this context, we consider that N = 2 and the parameter vector p∈[1/R] where:
1/R ∈ [1/Rmax , 1/Rmin ] (27)
Since the matrice A depend linearly on the uncertain parameter 1/R, we can define
a polytope of L = 2 Vertices that contains all the possible values of the uncertain
622 O. Hazil et al.
matrices. The vertices of the polytopic model are:

1 −(1 − D)Ts /L 1 −(1 − D)Ts /L
A1 = , A2 =
(1 − D)Ts /C 1 − Ts /Rmin C (1 − D)Ts /C 1 − Ts /Rmax C
3 Proposed Robust Model-Based Predictive Control Using

LMIs
In this section, a synthesis approach to robust MPC using LMIS is presented for a
class of systems under saturating control signal in the presence of uncertainties and
then, we incorporate an input constraint.
A. The proposed RMPC
In order to design a state-feedback control law δ̃(k + i|k ) = K x̃(k + i|k ), i ≥ 0 for
(24), one may consider the minimisation problem of the infinite horizon quadratic
performance index as follows:
∞

J (k) = x̃(k + i|k )T Q x̃(k + i|k ) + sat(δ̃(k + i|k ))T Rsat(δ̃(k + i|k )) (28)
i=0
where R(i), Q(i) are two positive definite states and control weights respectively.
Moreover, the input constraints are considered as:
2

δ̃(k + i|k ) ≤ δmax
2
,i ≥ 0 (29)
2
where δmax is the maximum control input.

Let us introduce a quadratic function V (x) = x T Px, P > 0 of the state x(k|k) of the
system (24), with V (0) = 0. At sampling time k, suppose the following inequality is
satisfied
V (k + i + 1|k ) − V (k + i|k ) ≥ −(x̃(k + i|k )T Q x̃(k + i|k )

+ sat(δ̃(k + i|k ))T Rsat(δ̃(k + i|k )) (30)
Summing (30) from i = 0 to i = ∞, we have
x̃(∞|k )T P x̃(∞|k ) − x̃(k|k )T P x̃(k|k ) ≥ −J (31)
If the resulting closed-loop system for (24) is stable, x(∞|k) must be zero and
result in
J ≤ x̃(k|k )T P x̃(k|k ) ≤ −γ (32)
where γ is a positive scalar and is regarded as an upper bound of the objective in (28)
∞

x̃(k + i|k )T Q x̃(k + i|k ) + sat(δ̃(k + i|k ))T Rsat(δ̃(k + i|k )) ≤ γ (33)
i=0
Theorem 1 At each sample time k, consider the discrete-time system (3) and
assume x̃(k|k ) is the measured state of x̃(k). The state feedback control law that
minimizes the upper bound γ on the infinite horizon quadratic performance index
J (k)and

robustly stabilizes
the closed-loop system within an invariant ellipsoid
ε = x̃
x̃ T Q −1 x̃ ≤ 1) , is given by δ̃(k + i|k ) = K x̃(k + i|k ), i ≥ 0, G = γ P−1
and K = Y G −1 Where G,G > 0 and Y are the solutions to the following LMIs:
min γ subjecte to
G,Y,γ ,ξ1 ,ξ2

I x̃(k|k )T
≥0
x̃(k|k ) G
⎡ ⎤
G ∗ ∗ ∗ ∗ ∗
√
⎢ (1 + α1 )(Ai + B K ) G ∗ ∗ ⎥
∗ ∗
⎢ √ ⎥
⎢ QG 0 γI ∗ ⎥
∗ ∗
⎢ √ ⎥
⎢ (1 + α2 )RY γI ⎥≥0
∗ ∗
⎢ 0 0 ⎥
⎢ ⎥
⎢ ε 1 + α1−1 BY 0 0 0 ζ1 I ∗ ⎥
⎣ ⎦

ε 1 + α2−1 Y 0 0 0 0 ζ2 I
2
δmax I Y
≥0
YT G
G − ξ1 I > 0
G − ξ2 I > 0 (34)
The symbol ‘*’ denotes symmetric elements in the matrix

Proof To obtain (34), the modified quadratic function V (x̃) is required to satisfy:
V (k + i + 1|k ) − V (k + i|k ) ≥ −(x̃(k + i|k )T Q x̃(k + i|k )

+ sat(δ̃(k + i|k ))T Rsat(δ̃(k + i|k )) (35)
Now, by substituting the state space (24), in inequality (35) results in
[(A + B K )x̃(k + i|k) + B φ̃(k + i|k)]T P[(A + B K )x̃(k + i|k) + B φ̃(k + i|k)]
+ x̃(k + i|k)T Q x̃(k + i|k) + sat(δ̃(k + i|k ))T Rsat(δ̃(k + i|k )
− x̃(k + i|k)T P x̃(k + i|k) ≤ 0
624 O. Hazil et al.
Defining the function g(x̃, δ̃) as
g(x̃, δ̃) = [(A + B K )x̃(k + i|k) + B φ̃(k + i|k)]T

P[(A + B K )x̃(k + i|k) + B φ̃(k + i|k)]
= [(A + B K )x̃(k + i|k)]T P[(A + B K )x̃(k + i|k)]
+ [(A + B K )x̃(k + i|k)]T P[B φ̃(k + i|k)]
+ [(A + B K )x̃(k + i|k)]P[B φ̃(k + i|k)]T
+ [B φ̃(k + i|k)]T P[B φ̃(k + i|k)]
Applying Lemma 3, the upper bound of g(x̃, δ̃) becomes
g(x̃, δ̃) ≤ (1 + α1 )[(A + B K )x̃(k + i|k)]T P[(A + B K )x̃(k + i|k)]

+ (1 + α1−1 )[B φ̃(k + i|k)]T P[B φ̃(k + i|k)]
Consider P ≤ λ1,max I ≤ μ1 I , then

+ (1 + α1−1 )μ1 B T B φ̃(k + i|k)T φ̃(k + i|k)
The term involving φ̃(k + i|k) in the above equation is bounded as φ̃(k +
i|k)T φ̃(k + i|k) ≤ εδ̃(k + i|k)T δ̃(k + i|k), then

+ ε(1 + α1−1 )μ1 B T B δ̃(k + i|k)T δ̃(k + i|k)
Moreover, we defined the function h(δ̃) as
h(δ̃) = sat(δ̃(k + i|k ))T Rsat(δ̃(k + i|k )

= [sat(δ̃(k + i|k )) − δ̃(k + i|k ) + δ̃(k + i|k )]T
× R[sat(δ̃(k + i|k )) − δ̃(k + i|k ) + δ̃(k + i|k )]
= [φ̃(k + i|k ) + δ̃(k + i|k )]T R[φ̃(k + i|k ) + δ̃(k + i|k )]
= φ̃(k + i|k )T R φ̃(k + i|k ) + φ̃(k + i|k )T R δ̃(k + i|k )
+ φ̃(k + i|k )R δ̃(k + i|k )T + δ̃(k + i|k )T R δ̃(k + i|k )
Applying Lemma 3, the upper bound of h(δ̃) becomes
h(δ̃) ≤ (1 + α2 )δ̃(k + i|k )T R δ̃(k + i|k ) + (1 + α2−1 )φ̃(k + i|k )T R φ̃(k + i|k )
Consider R ≤ λ2,max I ≤ μ2 I ,then

h(δ̃) ≤ (1 + α2 )δ̃(k + i|k )T R δ̃(k + i|k ) + (1 + α2−1 )μ2 φ̃(k + i|k )T φ̃(k + i|k )
The term involving φ̃(k + i|k) in the above equation is bounded as φ̃(k +
i|k)T φ̃(k + i|k) ≤ εδ̃(k + i|k)T δ̃(k + i|k), then
h(δ̃) ≤ (1 + α2 )δ̃(k + i|k )T R δ̃(k + i|k ) + ε(1 + α2−1 )μ2 δ̃(k + i|k )T δ̃(k + i|k )
By replacing g(x̃, δ̃) and h(δ̃) in the inequality (54), the following condition holds
for all i ≥ 0
(1 + α1 )[(A + B K )x̃(k + i|k)]T P[(A + B K )x̃(k + i|k)]

+ ε(1 + α1−1 )μ1 B T B δ̃(k + i|k)T δ̃(k + i|k) + x̃(k + i|k)T Q x̃(k + i|k)
+ (1 + α2 )δ̃(k + i|k )T R δ̃(k + i|k ) + ε(1 + α2−1 )μ2 δ̃(k + i|k )T δ̃(k + i|k )
− x̃(k + i|k)T P x̃(k + i|k) ≤ 0
Consequently, by replacing δ̃(k + i|k ) = K x̃(k + i|k ), the inequality can be

expressed as:
(1 + α1 )[(A + B K )T P[(A + B K )] + ε(1 + α1−1 )μ1 B T B K T K + Q

+ (1 + α2 )K T R K + ε(1 + α2−1 )μ2 K T K − P ≤ 0
Pre- and post-multiplying by G > 0, and substituting G = γ P−1 , Y = KG, 1 =

γ μ1 and 2 = γ μ2 , and then applying Schur complements, (19) becomes
⎡ ⎤
G ∗ ∗ ∗ ∗ ∗
√
⎢ (1 + α1 )(Ai + B K ) G ∗ ∗ ⎥∗ ∗
⎢ √ ⎥
⎢ QG 0 γI ∗ ⎥∗ ∗
⎢ √ ⎥
⎢ (1 + α2 )RY γI ⎥≥0
∗ ∗ (36)
⎢ 0 0 ⎥
⎢ ⎥
⎢ ε 1 + α1−1 BY 0 0 0 ζ1 I ∗ ⎥
⎣ ⎦

ε 1 + α2−1 Y 0 0 0 0 ζ2 I
By applying Schur complements to (47), we derive

I x̃(k|k )T
≥0 (37)
x̃(k|k ) G
B. Input constraint
Physical limitations inherent in the process impose hard constraints on the manipu-
lated variable. The amplitude constraint can be represented by:
626 O. Hazil et al.
2

δ̃(k + i|k ) ≤ δmax
2
,i ≥ 0 (38)
2
Setting K = Y G−1 then δ̃ = K x̃ = Y G −1 and (38) can be written as:

2

δ̃(k + i|k ) =
K x̃(k + i|k )
2
2
2
2
= Y G −1/2 (G 1/2 x̃(k + i|k ))2
2
≤ Y G −1/2 2
(39)
The input two-norm constraint in (39) is rewritten as
Y T G −1 − δmax I ≤ 0 (40)
Using Schur complement, (40) is equivalent to:

δmax
2
I Y
T ≥0 (41)
Y G
4 Simulations and Results
The performances of the proposed robust MPC control design applied to the boost
converter are illustrated through simulations. The parameters nominal values of the
buck-boost converter are given by R = 20 , C = 20 μF, L = 3.5 mH, vs = 24 V,
vin = 12 V, f s = 20 kHz, D = 0.5. We perform two simulation tests to verify the
performance of the proposed control law and its robustness according to the presence
of the input constraint and the parametric uncertainty.
Figure 4 illustrates the transient simulation of the boost converter without any
change in the converter parameters, we assume that the saturation limit is δ max = 0.5.
The waveforms depicted in the Fig. 4 are the duty-cycle d capacitor voltage vc and
inductor current iL . One can conclude that the output voltage settle to their desired
value without any overshoot and the settling time of the circuit as clearly seen in the
graph is t = 80 s. These results demonstrate that our RMPC controller guarantees
better stabilization performance under harder actuator saturation constraint. The best
performances are obtained with controllers parameters α 1 = 0.01, α 2 = 0.211, Q =
0.1, R = 0.001.
In Fig. 5, we take into account not only the actuator saturation but also the load
uncertainty in order to show the robustness of the proposed controller. Figure 5
show that our controller is able to stabilize the system on the desired output voltage
in presence of load change. These results demonstrate the effectiveness of proposed
Fig. 4 Simulated transient Output voltage

of the boost converter under 24.5
saturation constraint δ max = Vs
Vreff
0.5
24
Vs
23.5
23
0 20 40 60 80 100 120 140 160 180 200
Time (s)
Inductor current
3
2.5
1.5
1
IL
0.5
-0.5
-1
-1.5
0 20 40 60 80 100 120 140 160 180 200
Time (s)
control signal
1
0.9
0.8
0.7
0.6
d
0.5
0.4
0.3
0.2
0.1
0
0 20 40 60 80 100 120 140 160 180 200
Time (s)
628 O. Hazil et al.
Fig. 5 Simulated transient Output voltage

of the boost converter under 25
a load step transient R = 20 Vs
Vreff
Ω → R = 10 Ω and
24.5
saturation constraint. δ max =
0.5
24
Vs
23.5
23
22.5
0 50 100 150 200 250 300 350 400
Time (s)
Inductor current
6
3
IL
-1
-2
0 50 100 150 200 250 300 350 400
Time (s)
method and its robustness, in particular its aptitude to quickly compensate the system
dynamic variations even for significant parameters change.
5 Conclusions
In this paper, a RMPC approach based on LMI method has been proposed for the
nonlinear systems subject actuator nonlinearity. The LMI-based RMPC controller is
applied to regulate the output voltage of a boost DC-DC converter; we can extend this
work to other converters (buck converter, buck-boost converter, Cuk converter, etc.).
The saturation effect is represented by a model based on a deadzone nonlinearity. The
state feedback control law is obtained by minimizing the upper bound of the infinite
horizon cost function at each time instant. The internal stability of the closed-loop
system is guaranteed in the sense of Lyapunov. Finally, the simulation results on a
DC-DC converter have demonstrated the effeteness of the proposed control.
References
1. Wu Y et al (2019) A strong robust DC-DC converter of all-digital high-order sliding mode

control for fuel cell power applications. J Power Sources 413:222–232
2. Tian Z, Lyu Z, Yuan J, Wang C (2019) UDE-based sliding mode control of DC–DC power
converters with uncertainties. Control Eng Pract 83:116–128
3. Middlebrook RD, Cuk S (1977) A general unified approach to modeling switching-converters.
Int J Eletron 42:521–550
4. Fakham H et al (2011) Power control design of a battery charger in a hybrid active PV generator
for load-following applications. IEEE Trans Ind Electron 58(1):85–94
5. Eghtedarpour N, Farjah E (2012) Control strategy for distributed integration of photovoltaic
and energy storage systems in dc micro-grids. Renew Energy 45:96–110
6. Elgendy Mohammed A, Zahawi Bashar, Atkinson David J (2012) Assessment of perturb and
observe MPPT algorithm implementation techniques for PV pumping applications. IEEE Trans
Sustain Energy 3(1):21–33
7. Montagner VF, Oliveira RCLF, Leite VJS, Peres PLD (2005) LMI approach for H1 linear
parameter-varying state feedback control. In: IEE Proceedings control theory and applications,
vol 152(2), pp 195–201
8. Shirazi Mariko, Zane Regan, Maksimovic Dragan (2009) An autotuning digital controller
for DC–DC power converters based on online frequency-response measurement. IEEE Trans
Power Electron 24(11):2578–2588
9. Olalla C, Queinnec I, Leyva R, El Aroudi A (2011) Robust optimal control of bilinear DC-DC
converters. Control Eng Pract 19(7):688–699
10. Gomes da Silva JJM, Tarbouriech S (2005) Antiwindup design with guaranteed regions of
stability: an lmi-based approach. In: IEEE transactions on automatic control, vol 50, no 1, pp
106–111
11. Sun L, Wang Y, Feng G (2015) Control design for a class of affine nonlinear descriptor systems
with actuator saturation. IEEE Trans Autom Control 60(8):2195–2200
12. Boyd S, ElGhaoui L, Feron E, Balakrishnan V (1994) Linear matrix inequalities in systems
and control theory. Studies in applied and numerical mathematics, vol 15. SIAM, Philadelphia
13. Sitbon M et al (2015) Disturbance observer based voltage regulation of current-mode-boost-
converter-interfaced photovoltaic generator. In: IEEE transactions on industrial electronics pp
0278–0046 (c)
Strategy for Optimization of Energy
Management Based on Fuzzy Logic
in an FCEV with a Contribution
of a Photovoltaic Source
Said Belhadj, Kaci Ghedamsi, and Zina Larabi
Abstract In this article, we will study how to increase fuel economy, in a fuel
cell/battery (FC + B) and fuel cell/battery/photovoltaic source (FC + B + PV)
configuration in a hybrid vehicle with fuel cell. Hybrid vehicle models (FC + B)
and (FC + B + PV) will be designed under Simulink/Matlab use a new hybrid
approach based on the battery’s soc based on the availability of sunshine so the
availability of the PV source. The results show that the proposed control of this
strategy can meet the energy needs for good optimization in energy management.
Comprehensive comparisons of this energy management strategy based on the control
and monitoring of the fuel consumption of the fuel cell according to the chosen driving
cycles. Therefore, the proposed strategy will provide a novel approach for advanced
of energy management optimization system.
Keywords Hybrid vehicle · Fuel cell · PV source · Battery · Fuel economy ·

Energy management · Fuzzy logic
1 Introduction
The performance of a vehicle and the fuel consumption of hybrid vehicles are signif-
icantly influenced by the strategy of power management and the dimensioning of
components due to the multiplicity of power sources and the difference in their
S. Belhadj (B) · K. Ghedamsi

Laboratoire de Maitrise des Énergies Renouvelables, University of Bejaia, Bejaia, Algeria
K. Ghedamsi
Z. Larabi
Departement d’Electrotechnique, University M. Mammeri of Tizi-Ouzou, Tizi-Ouzou, Algeria
Departement d’Ingenierie et Systemes Electriques, University M. Bougara of Boumerdes,
Boumerdes, Algeria

https://doi.org/10.1007/978-981-15-6403-1_42
632 S. Belhadj et al.
characteristics [1]. ICE technology can be considered quite mature as ICE-equipped

vehicles have been around for more than 100 years. This type of vehicle generates
during their use, two major drawbacks that are air pollution and the greenhouse
effect. Thus, HEVs were developed to overcome the disadvantages of ICE vehi-
cles and pure BEVs [2, 3]. According to [4], Globally, the transportation sector is
the second largest energy consuming sector after the industrial sector and accounts
for 30% of the world’s total delivered energy. Fuel cells convert chemical energy
directly into electrical energy, usually characterized by the electrolyte material used.
Currently, five types of fuel cells are on the market [5]. The benefits of this tech-
nology are high efficiency, almost at partial load, low emissions, no noise and free
adjustable ratio (50 kW–3 MW) of electric and heat generation. A high conversion
efficiency, between 40 and 85% and even more, depending on the type of fuel cell
and the application of use [6]. In this article, in addition to these two sources which
are the fuel cell and the battery, we have added a PV source. The power assignment
between the fuel cell system, the PV source and the auxiliary energy storage device,
which is the battery in our case, is the energy management in the hybrid vehicle,
representing an important technique.
Various energy management control strategies have been used for hybrid vehicles
in recent years. Chen et al. [7] proposed an adaptive control approach with fuzzy
logic parameter tuning (AFLPT) for the energy management of electric vehicles
that are using fuel cell battery hybrid systems. Yu et al. [8] had used an innovative
strategy for optimal power allocation for fuel cell, battery and supercapacitor hybrid
electric vehicle. Zheng et al. [9] had used the Minimum Principle to reduce fuel
consumption by designing an optimal controller in the power management of fuel cell
hybrid vehicles. Kamal et al. [10] have proposed an intelligent energy management
strategy for a hybrid hydraulic-electric vehicle in order to minimize its total energy
consumption. In Refs. [7–12] the proposed control strategies are not sufficiently
detailed and the influence of fuel economy and hybrid vehicle dynamics did not take
into account for a balancing purpose. Regarding the limitations of previous work in
the field of energy management using fuzzy logic controllers to optimize energy in
FCEV; we can mention the decrease in the life of the battery, because of its frequent
solicitations, and the risk of exceeding or not to reach the reference values due to the
use of limiters. In this paper, a hybrid vehicle model is designed and implemented
based on subsystems with FC + B configuration in Simulink/Matlab for simulation
validation. After we has developed a second model by adding a subsystem with a PV
source with its command. In order to improve the fuel economy of hybrid vehicles,
increase the mileage of the continuation of the journey, and increase the lifespan of the
FC, a fuzzy logic control is used to design a relevant strategy for the management of
energy for the FC + B Hybrid Vehicle and FC + B + PV Hybrid Vehicle. According
to the NEDC driving cycle conditions, the proposed control strategy is contrasted
with the power tracking control strategy, which is wide adopted in Simulink in terms
of the indexes of fuel economy and dynamic property, as well as the duration of life
of the FC. This paper is decomposed as follows. First, the Fuel cell/Battery (FC + B)
and Fuel cell/battery/ultra-capacitor (FC + B + PV) structures, dynamic modeling of
the FC, PV system modeling, and the modeling of battery pack were presented. After,
Strategy for Optimization of Energy … 633
Converter
DC-DC-
Energy Management System
Power
Desired
With Fuzzy Logic

Fuel Cell
DC-DC-Bus
DC - AC
Inverter
Reference
Converter
DC-DC-
inputs
PV Source Vehicle
PMSM
SOC Battery
Fig. 1 Global scheme of the FC + B + PV Hybrid Electric Vehicle
Energy Management Strategy for Hybrid Vehicle was explained. Before finishing,
the simulation results were showed. Finally, discussions and conclusions were given
(Fig. 1).
2 Drive Structure of Hybrid Vehicle
A. Fuel cell/battery (FC + B) structure

There is a growing interest in the use of fuel cells for hybrid vehicles, given the
great progress of fuel cell technology. In Fig. 2 is presented an FC + B structure
with its command, it should be noted the wide use of this hybrid structure [13, 14].
Fig. 2 FC + B drive structure of fuel cell electric vehicle

Fig. 3 FC + B + PV drive structure of fuel cell electric vehicle
B. Fuel cell/battery/ultra-capacitor (FC + B + PV) structure

In order to improve the performance of the hybrid vehicle, according to the
structure FC + B, a new structure FC + B + PV is proposed as shown in Fig. 3.
In this structure, the PV source supplies the battery without interruption (except
at night) constituting a significant additional supply, necessary for the phases
of starting, acceleration and rise of the vehicle.
3 Modeling of Hybrid Vehicle
The overall system constituting the vehicle consists of two (FC + B) or three sources
(FC + B + PV) depending on the configuration concerned. The three energy sources
(FC, B, and PV) exchange their power with a DC bus. For the FC and the PV are
connected to the DC bus, each via a unidirectional up converter in the current. The
two inductances added as filters and to respect the alternations of sources. The DC
bus powers the traction motor via a three-phase inverter to convert DC power into
an alternative. This system is controlled by the torque of the electric motor acting
on both rear wheels with high precision and high stability of the HEV, reducing the
clutter caused by mechanical parts such as the mechanical differential and the drive
shafts, releasing the space in the vehicle for the PV, the battery and the hydrogen
tank.
The traction machine used is of the three-phase pmsm type, widely used in trans-
port, which is, apparently, the researchers’ favorite topology, since it presents several
advantages: a good power/weight ratio, high efficiency level, continuous decrease of
PMs price [15].
C. Dynamic modeling of the FC
The fuel cell represents the main source of energy of the vehicle, it is of the PEMFC
type. The voltage of its cell and its total power are given by the equations according
to the references [16, 17].
According to JC Amphlett, the expression of Nernst’s equation is as follows [18]:

E Nernst = 1.229 + 0.8510−3 (TFC − 298.15)4.308510−5 TFC 0.5ln PH2 ln PO2 (1)
The power demand of the vehicle PV will be provided by the FC taking into
account the efficiency of ηFC [19].
There is a relationship of proportionality between the partial pressure and the
molar flow of a gas (hydrogen or oxygen) through a certain valve, the expression of
this relation is given in reference [20].
D. PV system modeling
The equivalent circuit of the PV module is modeled under Matlab/Simulink while

representing the dynamic interactions between the parameters. The PV module is
described using various mathematical equations. The configuration of a PV cell is
an ideal equivalent circuit having a current source parallel to a diode [21–24]. The
ideal solar cell can be designed as a single diode connected in parallel with a light
generated Iph current source. With regard to the electrical characteristics of a solar
cell, it can be described by a four-component electrical circuit only. The resistances
of the circuit are respectively Rs and Rsh which are the series resistance and shunt
of the solar cell. In order to estimate the parameters of the equivalent circuit of a PV
module, a method is developed by Batzelis et al. [25].
E. Modeling of battery pack
In this study, a lithium-ion battery is used to power the vehicle and also to recover
energy storage, and especially to be powered by the PV source especially in the
shutdown phase. In this section, the dynamic model of the lithium-ion battery is
presented [26].
4 Energy Management Strategy for Hybrid Vehicle
Since the hybrid vehicle is non-linear and multivariable, the fuzzy model Logic
Controller (FLC) is more suitable for energy management. A list of IF-THEN rules
links the output to the controller inputs. A degree of adhesion is assigned to the
variables according to the definition of the membership functions [7]. Several studies
have developed controllers based on fuzzy logic to determine power sharing between
different sources, and achieve the maximum in fuel economy and system efficiency.
In this article, a fuzzy logic controller according to the power demanded by the
vehicle and the SOC of the battery is presented (Figs. 4 and 5).
F. Fuzzy logic control for FC+ B
To restart the power of the fuel cell and the battery to meet the demand of the electric
motor, a fuzzy treatment between two inputs that are the requested power and the
state of charge of the battery. The output variable is the reference current of the fuel
cell. When the vehicle is started, the battery drives the vehicle alone. The control
provides for the fuel cell intervention in the particular case of a start where the power
demand is high and the SOC of the battery is at a minimum. The relation (2) defines
the powers concerned.
Fig. 4 Fuzzy logic

controller for power split
Fig. 5 Membership
functions of fuzzy logic
controller
Power Demand
FC Reference 200
Current
Reference Speed
Vehicule Speed [Km/h]

125.05
Vehicule Speed (FC+B)
Vehicule Speed (FC+B+PV)
125
124.95
1125.994 1125.996 1125.998 1126 1126.002 1126.004 1126.006 1126.008 1126.01
Time [S]
Fig. 6 Zoom of comparisons between required speed and actual speed

PFC = 0
(2)
PB = Pm
where PFC is the out power of fuel cell system, Pm is the motor power, PB is the
battery power.
∗
(SOC−SOC )
PB = 0.5(SOC max −SOCmin )
Pc
∗
(3)
2SOC = (SOCmax + SOCmin )
where SOC* is the SOC reference value, SOCmax and SOCmin are the maximum and
minimum SOC values, Pc is the regulation power of SOC.
With regard to the stages where the fuel supply of the vehicle is combined between
the fuel cell and the battery, the SOC is higher than the reference SOC, where the
battery discharges and before reaching this level, the fuel cell is actuated to ensure
the growing demand for power. The discharge power of the battery depends on the
difference between the reference SOC and the real SOC. the power of the fuel cell,
in this case, is expressed as follows:
Pm − PB
PFC = (4)
η1
where η1 is the efficiency of the control of the loop generating the reference current
of the fuel cell.
∫ xi μ(xi )d xi
u0 = (5)
∫ μ(xi )d xi
where x i is the element of a set in the field, μ(x i ) is the membership degree, u0 is the
judgment result of fuzzy output.
G. Fuzzy logic control for FC+ B+ PV

To manage the power supply between the three sources fuel cell, battery and PV
source (FC + B + PV), according to the power demand of the electric motor two
sub-control systems are designed in the fuzzy controller according to the phases of
driving the vehicle. When starting the vehicle, the battery is solicited alone but with
a contribution of the PV source. Depending on the power demand and SOC of the
battery and after preheating, the fuel cell system shall be activated in accordance
with the relationship expressed as follows:

PFC = 0
(6)
PB + PPV η2 = Pm
where PPV is the photovoltaic power, and η2 is the efficiency of the control of the
loop generating the reference current of the fuel cell.
In fuel cell-only fuel-based driving mode, the battery is discharged and the PV
source is disabled because the SOC value of the battery is lower than the battery
reference SOC value and the command of the PV source is disabled and the power
required at operating conditions is within the maximum and minimum power range
of the fuel cell system, so the output power of the fuel cell system must satisfy the
requested power, and charge the battery. In the case of delivering the peak load power
to the electric motor quickly, the fuel cell will provide much of the charge until the
battery that will assist the FC and afterward will take over by activating the PV
source. The relation below defines the balance of the different powers.
PFC η1 = Pm + PB + PPV η2 (7)
For Combined Fuel Cell/Battery Combined Mode of Operation, The relationship

equilibrium is expressed as follows:
PFC η1 + PB = Pm + PPV η2 (8)
In the case where the requested power is greater than the nominal power of the fuel
cell, the latter provides its maximum power and the rest will be provided by the PV
source. If the requested power is still not satisfied, the battery should discharge even
though the SOC is not higher than the SOC. The equilibrium relation is indicated as
follows:

PFC = PFC,max
(9)
PFC η1 + PB + PPV η2 = Pm
In this paper, the development of a SIMULINK model is implemented based on the

FC + B hybrid vehicle system architecture. Sub-models reconfigured to load a model
hybrid vehicle FC + B + PV. In this article, the NEDC cycle is selected as the standard
cycle conditions for testing vehicle performance. The choice of this cycle is chosen
with respect to the parameters of the vehicle, namely the degree of hybridization, the
initial SOC values and the reference SOC, SOC_max and SOC_min. To highlight the
relevance and effectiveness of the FLC intelligent controller placed in the FC + B +
PV configuration system compared to FC + B, we compared the fuel economy indices
according to the dynamic property of the hybrid vehicle. In addition, according to
the NEDC standard cycle conditions chosen, the proposed control strategies contrast
with the power tracking control strategy that is widely adopted in SIMULINK for FC
+ B. The FLC speed curves for FC + B, FLC for FC + B + PV can correspond to the
speed curves required in the NEDC cycle conditions. Therefore, the FLC controller
designed for the FC + B + PV configuration can meet the speed requirements for
all driving phases with standard NEDC cycle conditions. The velocity curves are
illustrated in Fig. 7 (Fig. 6).
For optimum fuel economy values, under different cycle conditions, the FLC for
FC + B has a higher fuel consumption than FLC for FC + B + PV. Even the latter
controller has a better fuel economy compared to other methods already designed.
Therefore, the conclusion can be drawn that the FLC for FC + B + PV has better
performance in terms of fuel economy, and even in terms of fuel cell lifetime, due to
the demolition of fuel cell interventions. The FC. The optimal values of the dynamic
property of the FLC for FC + B + PV are the most efficient than the other methods,
namely acceleration time, distance traveled in 5 s and acceleration time for a distance
of 400 m. However, the FLC for FC + B performs better than the other methods
Reference Speed
Vehicule Speed [Km/h]
Vehicule Speed (FC+B)

100
Vehicule Speed (FC+B+PV)
50
0
0 200 400 600 800 1000 1200 1400
Time [S]
Fig. 7 Comparisons between required speed and actual speed

compared to acceleration time of 100 km/h, an acceleration time of 60–100 km/h,

and maximum speed (Fig. 8).
Figures 9, 11 and 12 show the power curves of the three sources supplying the
hybrid vehicle FC + B + PV. The PV source can provide the remaining power for
the acceleration conditions to lighten the fuel cell system and assist the battery. In
regenerative braking condition, the PV source will either deactivate or charge the
battery according to the real SOC and reference of the battery. Therefore, the PV
source in the FC + B + PV structure can assist in powering and decreases the charges
and fast discharges of the battery especially in phases of the urban journey at reduced
speed. In Fig. 10, the comparison of hydrogen consumption in the NEDC cycle is
shown. The FLC for FC + B + PV has the lowest hydrogen consumption. Therefore,
this method can produce a better fuel economy. Referring to Fig. 11, it represent the
4
10
2
Required Power [W]
-2
-4
0 200 400 600 800 1000 1200
Time [S]
Fig. 8 Required Power for the FCEV
4
10
3
FC's Power [W]
0
0 200 400 600 800 1000 1200
Time [S]
Fig. 9 Power of the fuel cell

FLC for FC+B
FLC for FC+B+PV

Hydrogen consumption [L] 6
0
0 200 400 600 800 1000 1200 1400
Time [S]
Fig. 10 Comparisons of hydrogen consumption
3000
Power of PV Source [W]
2000
1000
0
0 200 400 600 800 1000 1200
Time [S]
Fig. 11 PV power that intervenes in charging the battery in the FCEV stop phases
power of the PV source according to the command used, the PV source directly feeds
the DC-DC bus as the battery and the FC, and it is a continuous supply of the PV
source, which charges the battery in phases of stops of the vehicle.
6 Conclusions
In this paper, a fuzzy logic control method is implemented under Simulink Matlab
environment, used as an energy management optimization strategy of hybrid vehicles
FC + B and hybrid vehicle FC + B + PV for the economy of consumption. Hydrogen
and increase the vehicle’s autonomy and the life of the fuel cell. Using the algorithm,
4
10
Battery Power [W] 2
-2
-4
0 200 400 600 800 1000 1200

Time [S]
Fig. 12 Power of the battery
the parameters’ values of the fuzzy controller and the motor sizing have been inserted
in order to optimize energy management. A particular development on Simulink as
being the second model is implemented based on the architecture of the hybrid
vehicle system FC + B. Then, sub-models are loaded in order to reconfigure, a
model of hybrid vehicle FC + B + PV is developed. The results obtained indicate
that the proposed control strategy at the required power demand for the NEDC driving
cycle. In the different phases of the NEDC cycle, the FLC for FC + B has a lower
consumption (L/100 km). The FLC for FC + B + PV has the lowest consumption
(L/100 km) than other methods. Therefore, the conclusion can be deduced that FLC
for FC + B + PV has better performance in terms of fuel economy. In addition, FLC
for FC + B + PV has better performance in terms of dynamic property under driving
conditions than NEDC cycle standards. Therefore, the proposed strategy gives a new
approach to the advanced hybrid vehicle energy management system.
References
1. Wieczorek M, Lewandowski M (2017) A mathematical representation of an energy manage-

ment strategy for hybrid energy storage system in electric vehicle and real time optimization
using a genetic algorithm. Appl Energy 192:222–233
2. Li Z, Khajepour A, Song JC (2019) A comprehensive review of the key technologies for pure
electric vehicles. Energy 182:824e839
3. Chan CC (2014) Overview of electric, hybrid, and fuel cell vehicles. Encyclopedia of
automotive engineering, Online © 2014. Wiley
4. Atabani AE, Badruddina IA, Mekhilef S, Silitongaa AS (2011) A review on global fuel
economy standards, labels and technologies in the transportation sector. Renew Sustain
Energy Rev 15:4586–4610.http://www.epa.gov/climatechange/emissions/downloads09/GHG
2007-ES508.pdf
5. Alaswad A, Baroutaji A, Achour H, Carton J, Al Makky A, Olabi AG (2016) Developments

in fuel cell technologies in the transport sector. Int J Hydrog Energy 41–37:16499–16508
6. Lai J-S, Ellis MW, Fuel cell power systems and applications. Proc IEEE 105(11):2166–2190.
https://doi.org/10.1109/jproc.2017.2723561
7. Chen J, Xu C, Wu C, Xu W, Adaptive fuzzy logic control of fuel-cell-battery hybrid systems
for electric vehicles. IEEE Trans Ind Inform 14(1):292–300. https://doi.org/10.1109/tii.2016.
2618886
8. Yu Z, Zinger D, Bose A (2011) An innovative optimal power allocation strategy for fuel cell,
battery and supercapacitor hybrid electric vehicle. J Power Sources 196:2351–2359
9. Zheng CH, Kim NW, Cha SW (2012) Optimal control in the power management of fuel cell
hybrid vehicles. Int J Hydrog Energy 37:655–663
10. Kamal E, Adouane L, Intelligent energy management strategy based on artificial neural fuzzy
for hybrid vehicle. IEEE Trans Intell Veh 3(1):112–125. https://doi.org/10.1109/tiv.2017.278
8185
11. Crolla DA, Cao D, The impact of hybrid and electric powertrains on vehicle dynamics, control
systems and energy regeneration. Veh Syst Dyn 50(sup1):95–109. https://doi.org/10.1080/004
23114.2012.676651
12. Chen Z, Mi CC, Xu J, Gong X, You C, Energy management for a power-split plug-in hybrid
electric vehicle based on dynamic programming and neural networks. IEEE Trans Veh Techn
63(4):1567–1580. https://doi.org/10.1109/tvt.2013.2287102
13. Sikkabut S, Mungporn P, Ekkaravarodome C, Bizon N, Tricoli P, Nahid-Mobarakeh B, Thoun-
thong P, Control of high-energy high-power densities storage devices by li-ion battery and super-
capacitor for fuel cell/photovoltaic hybrid power plant for autonomous system applications.
IEEE Trans Ind Appl 52(5):4395–4407. https://doi.org/10.1109/tia.2016.2581138
14. Bizon N, A new topology of fuel cell hybrid power source for efficient operation and
high reliability. J Power Sources 196(6):3260–3270. https://doi.org/10.1016/j.jpowsour.2010.
11.049
15. Zheng CH, Oh CE, Park YI, Cha SW (2012) Fuel economy evaluation of fuel cell hybrid
vehicles based on equivalent fuel consumption. Int J Hydrogen Energy 37:1790–1796
16. Mokrani Z, Rekioua D, Rekioua T (2014) Modeling, control and power management of hybrid
photovoltaic fuel cells with battery bank supplying electric vehicle. Int J Hydrog Energy 35:1–
10
17. Yilmaz M, Limitations/capabilities of electric machine technologies and modeling approaches
for electric motor design and analysis in plug-in electric vehicle applications. Renew Sustain
Energy Rev 52:80–99. https://doi.org/10.1016/j.rser.2015.07.033
18. Rekioua D, Bensmail S, Bettar N (2014) Development of hybrid photovoltaic-fuel cell system
for stand-alone application. J Hydrog Energy 35:1604–1611
19. Garcia P, Torreglosa JP, Fernandez LM, Jurado F (2012) Viability study of a FC/Battery/SC
tramway controlled by equivalent consumption minimization strategy. Int J Hydrog Energy
37(11):9368–9382
20. Barelli L, Bidini G, Ottaviano A, Optimization of a PEMFC/battery pack power system for a
bus application. Appl Energy 97:777–784. https://doi.org/10.1016/j.apenergy.2011.11.043
21. Ramos-Hernanz JA, Campayo JJ, Larranaga J, Zulueta E, Barambones O, Motrico J et al (2012)
Two photovoltaic cell simulation models in Matlab/Simulink. Int J Tech Phys Probl Eng 4(10)
(Number 1)
22. Kane A, Verma V (2013) Performance enhancement of building integrated photovoltaic module
using thermoelectric cooling. Int J Renew Energy Res 3(2)
23. Abdulkadir M, Samosir AS, Yatim AHM (2012) Modeling and simulation based approach of
photovoltaic system in simulink model. ARPN J Eng Appl Sci 7(5)
24. Hayrettin C (2013) Model of a photovoltaic panel emulator in MATLAB–Simulink. Turkish J

Electr Eng Comput Sci 21(2):301–308
25. Batzelis EI, Papathanassiou SA (2015) A method for the analytical extraction of the single-
diode PV model parameters. IEEE Trans Sustain Energy 7(2):504–512. https://doi.org/10.1109/
tste.2015.2503435
26. AlSharif A, Das M (2013) A time-varying transfer function model for modeling the charging
process of a Lithium-ion battery. In: 2013 IEEE energytech. https://doi.org/10.1109/energy
tech.2013.6645288
Modeling and Analysis
of an Electromagnetic Vibration Energy
Harvester for Automotive Suspension
Mustapha Zaouia, Bachir Ouartal, Nacereddine Benamrouche,

and Arezki Fekik
Abstract The aim of this work is to present a model for the study and analysis
of an Electromagnetic vibration energy harvester for automotive suspension. The
Electromagnetic energy harvester (EEH) transforms the kinetic energy due to vibra-
tions created by roads irregularities to electrical energy. The model is based on the
electromagnetic equation solved using finite element method (FEM) coupled to the
mechanical one obtained through the modeling a quarter of the automotive system.
The coupling with mechanical phenomena is carried out principally by magnetic
force. The displacement of EEH moving part is simulated using the Macro-Element
(ME) technique. The cycle of vibration is transmitted to electromagnetic energy
harvester device as vertical displacement. The obtained results are mainly the output
electrical power, the suspension deflection and the induced electromotive force, for
an imposed cycle of vibration. The effect of the weight of the sprung mass on the
induced electromotive force and the electrical power will be presented.
Keywords Electromagnetic energy harvester · Automotive · Energy conversion ·

Finite elements · Macro-Element · Electrical power
1 Introduction
Research in alternative energies field is constantly evolving particularly in auto-

motive applications. The kinetic energies due to the vibrations produced by roads
irregularities can be recuperated and transformed into the electrical energy in order
to improve the efficiency of the automotives particularly hybrids and electric one.
To recover these kinetic energies and transform them in electrical energy, types of
generators are used and integrated in the suspension of automotive. In literature,
M. Zaouia · B. Ouartal · N. Benamrouche

LATAGE Laboratory, Mouloud Mammeri University, B.P17 RP, 15000 Tizi-Ouzou, Algeria
A. Fekik (B)
Faculty of Science and Applied Science, Akli Mohand Oulhadj University, Bouira, Algeria

https://doi.org/10.1007/978-981-15-6403-1_43
646 M. Zaouia et al.
several authors have proposed different studies. The systems based on rotary actu-
ators energy harvesting are used in several domains as civil structures [1, 2] and
automotives [3, 4]. The authors have presented in [4], systems constituted of rotary
and ball-screw mechanism for harvesting the energy of vehicle suspensions.
However, using these actuators requires motion converters which have many
disadvantages. Nowadays, research has focused on the use of electromagnetic actua-
tors based on linear tubular topologies because they do not require a motion converter
[5–8].
The objective of our work is the modelling and analysis of the electromagnetic
energy harvester which operates as a generator. The EEH transforms the kinetic
energy of vibrations caused by several irregular forms of roads into electrical energy
to improve the efficiency of vehicles. This electrical energy recovery will be used
to supply different electrical components constituting the automotives and also to
restore the energy to the batteries. The EEH is a permanent magnets (PM) tubular
linear generator. The permanent magnets are necessary to product a magnetic flux.
The magnets are placed in way to alternate the magnetic polarities on the surface of
the moving part in the axial direction.
The analysis of the EEH is carried out by the application of an electromagnetic
– mechanical coupled model. The developed model is based on electromagnetic
equation expressed in terms of Magnetic Vector Potential (MVP) solved using the
FEM. The mechanical equation is obtained after modelling a quarter of the automo-
tive [8, 9] and solved to calculated the displacement, the suspension deflection and
the velocity. The mechanical system representing the quarter of the automotive is
constituted principally by a sprung mass that represents the mass of chassis, and the
unsprung mass including the mass of the wheel or tyre, the suspension and the brake
system.
The electromagnetic and the mechanical equations are coupled through the
magnetic force and the modified magnetic flux distribution due to the displace-
ment of the moving part. To take into account the moving parts the Macro-Element
(ME) technique is used [10, 11]. The ME technique is based on the resolution of
Laplace equation analyticallay in the unmeshed region which separates the meshed
fixed and moving regions. The ME technique presents advantage which is to manage
variable displacement steps obtained through the mechanical equation and allows us
to compute accurately the quantities such as velocities, displacements of the moving
part without any constraints on the meshed topology. The geometrical characteristics
of the electromagnetic energy harvester are given by [12] and [13]. The developed
electromagnetic—mechanical model is implemented under Matlab environment and
applied to study and analyse the global system of the quarter automotive including
the EEH. The important obtained results are principally, the suspension deflection,
electromotive force and the output electrical power, for an imposed cycle of vibration
introduced as vertical displacement. On the other hand the influence of the sprung
mass that presents the chassis of the automotive on the electromotive force and the
principally output electrical power is studied when varying the values of the sprung
mass.
Modeling and Analysis of an Electromagnetic Vibration … 647
Fig. 1 Model of the quarter

automotive including the
EEH
2 Description of the EEH-Quarter Automotive Model
Figure 1. shows the mechanical system which represents a quarter automotive model
[8, 9] including the EEH. The system is composed by two distinct masses. The sprung
mass represents the mass of chassis, and the unsprung mass represents the mass of
the wheel with the suspension and brake equipment. The road position reference
represents a cycle of vibration.
3 Electromagnetic-Mechanical Modeling
A. Electromagnetic modeling
Using Maxwell equations, governing electromagnetic equation in terms of MVP is
given as follow:
−→
→ σ
∧−
∧ υ∇
∇ A +
DA −
→
= J s + υpm ∇ ∧ bm (1)
r Dt

−
→ N
Js= i · n (2)
Sc
−
→ −
→
DA ∂A →
−
= − v ∧ B (3)
Dt ∂t
−
→
where υ is the magnetic reluctivity of the material, υ0 the vacuum reluctivity, A the
−
→
MVP, σ the electrical conductivity. B the flux density, bm remanent induction of
−
→
the permanent magnet, J s is the current density. Sc and N are the cross-sectional
area of the coil, and the number of turns respectively. n is the unit vector along the
direction of the exciting current i and v is velocity of the moving part.
The Eq. (1) is discredised in space using the Galerkin finite element method. When
the displacement is implicitly taken into account using ME technique, the Eq. (1) is
written in space in the 2D cylindrical coordinates (r, ϕ, z) and in time as follows:
¨ ¨
ME
∂αi ∂ A ∂αi ∂ A dr dz σ ∂A αi ∂A
υ + + αi dr dz υo dΓME
∂r ∂r ∂z ∂z r r ∂t r ∂n
Ω Ω ΓME

¨ ¨
αi ∂ A ∂bm z ∂bm r
− υ dΓ = (αi Js )dr dz + υpm · αi − dr dz
r ∂n ∂r ∂z
Γ ∩ΓME Ωs ΩPM
(4)
where, A = r Aϕ is the modified magnetic vector potential, bm z and bm r are respec-

tively the axial and radial components of the remanent induction, ΓME the Macro-
Element boundary and Γ the boundary surrounding the meshed regions area Ω.
The conductors and permanent magnet regions are respectively Ωs and ΩPM . The
weighted function is αi . The MVP is given by means of the approximation func-
tions α j (r, z) and α ME
j (r, z) for each finite element j respectively of the meshed node
regions and the unmeshed Macro-Element boundary nodes:
⎧ nn
⎪
⎪
⎪ α j (r, z)A j in the meshed regions
⎨ j=1
A(r, z) = nME (5)
⎪
⎪
t
⎪
⎩ α ME
j (r, z)A j in the Macro-Element
j=1
With nn and n ME
t are the numbers of the meshed region nodes and Macro-Elements
boundary nodes respectively.
The contribution of the Macro-Element to the global stiffness matrix is considered
by the term of the matrix SiME
j and is given by [10] and [11] as:
αiME ∂α ME
j
j =
SiME ΓME (6)
r ∂n
ΓME
z
Stainless
Steel shaft r
Steel spacer
Magnet
Coil
wc
Nylon Spacer
Tube support
ri
re
Symmetry axis
(a) (b)
Fig. 2 Geometry of the Electromagnetic energy harvester (a) and study domain (b)
αiME (r2 , z) α ME
b
j (r 2 , z)
SiME = dz
j
0 r 2 ∂r

b ME
αi (r1 , z) α MEj (r 1 , z)
− dz (7)
0 r1 ∂r
r1 and r2 are the boundaries of the Macro-Element (Fig. 4).

The introduction of the approximation functions given by Eq. (5) in the MVP for
each node, leads to the following algebraic equation system:

Δt [M] + [M ME ] At+Δt = Δt.({F} + {G}) − [K ] At+Δt − At (8)
where: Where Δ t is the step time, [M ME ] = υ0 SiME

j is the global ME Matrix.
¨
Mi j = υ ∇.α j dr dz
i ∇.α (9)
r
Ω
¨
Fi = (αi Js ) dr dz (10)
Ω
Fig. 3 Electrical circuit Energy

model of the EEH
regeneration
Rc
i
RL
E VL
Lc
dr dz
Ki j = σ αi α j (11)
r
Ωs
¨
∂αi ∂αi
Gi = υpm bmz − bmr dr dz (12)
∂r ∂z
ΩPM
B. Electrical Circuit Model of the Electromagnetic Energy Harvester

The equivalent electrical circuit of electromagnetic energy harvester is given by
Fig. 3, where Rc and L c are the resistance and inductance of the coil respectively,
and R L is the external resistance characterising the energy regeneration.
The electrical load connected to EEH determines the relationship between the
circuit current and induced electromotive force E when the load is R L . The relation
is given by:
di
E = (Rc + R L )i + L c (13)
dt
The maximum output electrical power occurs as the external resistance is set
equal to the internal resistance. The output electrical power is given by the following
equation:
E2
pe = (14)
Rc + R L
C. Mechanical Equation
The mechanical system representing the quarter of automobile, given by Fig. 1, is

governed by the following equation system:
Table 1 Mechanical
Quantity/ Symbol Value
parameters of quarter model
automotive Sprung mass (m s ) 290 kg
Unsprung mass (m u ) 59 kg
Spring stiffness (ks ) 16,000 N/m
Tyre stiffness (ku ) 190,000 N/m
Damping coefficient (bs ) 1000 N/m/s

m s z̈ s = −ks (z s − z u ) − bs (ż s − ż u ) + Fa
(15)
m u z̈ u = ks (z s − z u ) + bs (ż s − ż u ) − ku (z u − zr ) − Fa
where, ż s , ż u and z̈ s , z̈ u are the velocities and accelerations of sprung mass and
unsprung mass respectively. (z s − z u ) and (z u − zr ) are the deflection of suspension
and tyre or wheel respectively. ks is spring stiffness, ku Tyre stiffness and bs is the
damping coefficient. The mechanical parameters are presented in Table 1.
The Lorentz low is used to calculate the force exerted on the mover due to current
flowing through the coils. The magnetic force is given by:

−
→
Fa = N i d l ∧ B r (16)
−
→
where, d l is the tangential direction to the coil turns, and B r the magnetic flux
density through the coil in the radial direction.
Due to the axial symmetry of the EEH, the device can be simplified to a 2D
axial symmetry model in cylindrical coordinates. Only one fourth of the structure
is studied. The geometrical and the study domain of the electromagnetic energy
harvester are presented in Fig. 2 [12, 13]. The permanent magnets used are of
NdFeB type with axially magnetized. The dimensions and characteristics are given
by Table 2.
Figure 4. shows the finite element mesh including the Macro-Element of the
studied domain. The homogeneous Dirichlet condition of the MVP is imposed at the
boundary (Γ − ΓME ) and at the symmetry axis.
The Figs. 5 and 6 show respectively the calculated and given measured values of
radial and axial flux density at 3.8 mm above moving part.
The model is applied to analyse the harvester electromagnetic energy. The rema-
nent flux density of the permanent magnet of NdFeB is 1.2 T and the relative perme-
ability is unity. The steel magnetic permeability is considered as infinite [12]. Figure 7
Table 2 Dimensions and

Quantity/ Symbol Value (m)
characteristics of EEH
Diameter of stainless steel shaft 0.02
Diameter of steel spacer 0.1
Diameter of permanent magnet 0.1
Permanent magnet width 0.025
Steel width 0.025
Inter. Diameter of coil 0.1
Outer diameter of coil 0.155
Inter. diameter of tube support 0.155
Inter. diameter of tube support 0.180
Air gap 0.005
Coil width (wc ) 0.05
Number of turns of each coil (N) 690
Fig. 4 Finite element mesh

including the Macro-Element
Fig. 5 Radial flux density at 3.8 mm above moving part: Calculated and given measured values
Fig. 6 Axial flux density at 3.8 mm above moving part: Calculated and given measured values
Fig. 7 Road position reference versus time
presents wave form of the imposed reference cycle of vibration Z r . This vibra-
tion describes the imposed road irregularities reference position versus time. The
maximum amplitude is of 0.05 m.
Figure 8 presents the suspension deflection from equilibrium, which is the differ-
ence between the position of the sprung and the unsprung mass for the considered
Fig. 8 Suspension deflection versus time
cycle of vibration. Figure 9 depicts the magnetic force versus time corresponding to
the imposed cycle of vibration.
Fig. 9 Magnetic force versus time
Fig. 10 Electromotive force versus time

Figure 10 gives the induced electromotive force versus time corresponding to the
cycle of vibration. The electromotive force depends of the relative velocity between
the moving part and the coils of the harvester electromagnetic energy. The relative
velocity corresponds also to the difference between the sprung and the unsprung
mass velocity.
The obtained output electrical power versus time is shown in Fig. 11. It is computed
using the Eq. (14) governing the equivalent electrical circuit given in Fig. 3. The peak
value of the electrical power corresponding to the imposed cycle of vibrations is of
68.73 W.
In Fig. 12 and Fig. 13, one shows respectively the electromotive force versus
time and the output electrical power versus time for several values of sprung mass
of the automotive. By visualizing the curves of Figs. 12 and 13, one notices that the
weight of the sprung mass has a great influence on the electromotive force and output
electrical power.
Fig. 11 Output electrical power versus time
Fig. 12 Electromotive force versus time for different values of sprung mass
Fig. 13 Electrical power versus time for different values of sprung mass
5 Conclusion
In this work we have presented a model to analyse and study a harvester electro-
magnetic energy for automotive suspensions. This device is used to recover and
transform the kinetic energy, due to the vibration of road irregularities, to the elec-
trical one. The model is based on the coupling of the electromagnetic equation and
mechanical one obtained by considering the model of quarter of the automotive. The
developed model is implemented under Matlab. The principal results as the suspen-
sion deflection, electromotive force and the output electrical power are obtained after
application of the model for an imposed cycle of vibration. We notice that the effect
of the variation of the sprung mass has a great influence on the electromotive force
and electrical power. This type of device is very interesting to recover the kinetic
energy due to the vibrations created by the irregularities of the roads, and transform it
into electrical energy in order to improve the efficiency of automobiles, particularly
electric and hybrid ones.
References
1. Cassidy IL, Scruggs JT, Behrens S (2011) Design of electromagnetic energy harvesters for
large-scale structural vibration applications. In: Proceedings of. SPIE, vol 7977, pp 79770P
2. Tang X, Zuo L (2011) Simulation and experiment validation of simultaneous vibration control
and energy harvesting from buildings via tuned mass dampers,” presented at the American
Control Conference, San Francisco, CA, USA, Jun. 29–Jul. 1
3. Gupta A, Jendrzejczyk JA, Mulcahy TM, Hull JR (2006) Design of electromagnetic shock
absorbers. Int J Mech Mater Des 3:285–291
4. Kawamoto Y, Suda Y, Inoue H, Kondo T (2008) Electro-mechanical suspension system
considering energy consumption and vehicle maneuver. Vehicle Syst Dyn 46:1053–1063
5. Weeks DA, Bresie DA, Beno JH, Guenin AM (1999) The design of an electromagnetic
linear actuator for an active suspension. In: Proceedings of the SAE steering and suspension
technology Symposium, Detroit, MI, pp 1–11
6. Lequesne B (1996) Permanent magnet linear motors for short strokes. IEEE Trans Ind Applicat
32(1):161–168
7. Murty BV, Henry RR (1992) Active vehicle suspension with brushless dynamoelectric Actuator.
U.S. Patent 5 091 679, 25 Feb 1992
8. Martins I, Esteves J, Marques GD, Pinada Silva F (2006) Permanent-Magnets linear actuators
applicability in automobile active suspensions. IEEE Trans Vehicular Technol 55(1):86–94
9. Chaves M, Maia J, Esteves J (2008) Analysis of an electromagnetic automobile suspension
system. In: Proc. ICEM08, Portugal, Paper ID 1285
10. Mohellebi H, Latrèche ME, Féliachi M (1998) Coupled axisymmetric analytical and finite
element analysis of induction devices having moving parts’. IEEE Tran Magn 34(5):3308–3310
11. Azzouz F, Bendjima B, Féliachi M (1999) Application of macro-element and finite element
coupling for the behaviors analysis of magnetoforming systems. IEEE Tran Magn 35(3):1845–
1848
12. Baker NJ (2003) Linear generators for direct drive marine renewable energy converters. PHD.
Thesis, University of Durham, U.K, July 2003
13. Baker NJ, Mueller MA, Spooner E (2004) Permanent magnet air-cored tubular linear gener-
ator for marine energy converters. In: Second international conference on power electronics,
machines and drives, PEMD (Conf. Publ. No. 498) vol 2, pp 862–867
Experimental and Numerical Study
of Hybrid PV/Thermal Solar Collector
Provided with Self Ventilation
and Tracking Structure
Mohamed El-Amine Slimani, Rabah Sellami, Achour Mahrane,

and Madjid Amirat
Abstract This paper presents the results of an experimental and numerical study of
a hybrid photovoltaic/thermal (PV/T) solar air collector provided with fins, assisted
PV ventilation, and tracking support. This work aims to study the effective influence
of the various key elements (fins, assisted PV ventilation, and tracking support) on
the thermal and electrical behaviors of a hybrid PV/T air collector. The addition
of fins placed in the dynamic air vein aims to improve the evacuated heat from
the PV panel. The purpose of the assisted PV ventilation system is to control the
flow of air according to the solar radiation incident on the PV panel. A comparison
between simulation and experimental was done in order to validate the numerical
model developed for this study. The results show that the three elements provided
to PV/T collector give interesting results. The outlet air temperature from the solar
device (23 °C) is adequate for residential applications (building air conditioning).
Electrical, thermal, and overall average efficiencies have reached significant values
of around 13.5, 30, and 70%, respectively.
Keywords Fins · PV/T hybrid collector · Modeling · Energy efficiency · Solar

thermal conversion · Solar photovoltaic conversion
1 Introduction
The conventional hybrid photovoltaic/thermal air solar collectors have low thermal
and electrical performance. To increase the performance of these hybrid collectors,
various methods have been studied by several researchers [1–7].
M. E.-A. Slimani (B) · M. Amirat

Theoretical and Applied Fluid Mechanics Laboratory, Department of Energetic and Fluid
Mechanics, University of Science and Technology Houari Boumediene (USTHB), 16111 Bab
Ezzouar, Algiers, Algeria
R. Sellami · A. Mahrane
Unité de Développement des Equipements Solaires (UDES), EPST Centre de Développement des
Energies Renouvelables (CDER), 42415 Tipaza, Algeria

https://doi.org/10.1007/978-981-15-6403-1_44
660 M. E.-A. Slimani et al.
In the study of Mortezapour et al. [8] a comparison between two hybrid collectors,
one of which the PV module cells are covered with a layer of glass and for the other;
one PV module face is covered with a glass layer, and the other side is covered
with a layer of Tedlar. They noticed that the thermal efficiency of the second hybrid
collector is the highest.
Srinivas and Jayaraj [2] designed and studied a hybrid solar collector with double
air passage with fins in order to evaluate its thermal and electrical performance. This
study has shown that the loss of electrical energy production due to the glazing is
compensated by the thermal gain of the collector; which makes this hybrid collector
very useful for energy conversion. Mojumder et al. [9] studied two types of hybrid
solar collectors to see the effects of solar collector fins on system performance. They
used three algorithms SVM, FFA, and WT. Thin flat metal fins were introduced as
a heat sink in the collector. Othman et al. [4] have developed several prototypes of
hybrid PV/T collectors; a double-pass PV/T collector with fins (in two configurations
with and without concentration), and a PV/T collector with a grooved absorber, and a
PV/T collector with rectangular tunnel collector absorber. Each design has its specific
way to cool the solar cells to increase their yield. Shan et al. [10] did a study on a
PV/T solar collector equipped with baffles using several working heat transfer fluids
such as air, water, and refrigerant R410a. The results of the simulation show the
influence of the meteorological parameters and the evaporation temperature of the
coolant on the performance of the PV/T solar collector.
Hybrid PV/T collectors can be used in several application fields; in residential and
agricultural activities [11–18]. From the previous literature, several investigations
have been carried out on the PV/T collectors design and structure for a single or dual
heat transfer fluid but without combined fins and tracking structure.
This work aims to improve the thermal and electrical performance of a hybrid
PV/T collector provided with fins, assisted PV ventilation, and tracking support. This
system is to be integrated into building heat and air conditioning in winter season. A
numerical model was developed and experimentally validated for this purpose.
The hybrid PV/T collector presented for this study based on a monocrystalline
photovoltaic module (JT-185 M). The PV/T solar collector composed mainly of the
following components; a cover glass, PV cells immersed in a layer of EVA polymer,
a protective layer of Tedlar below the PV cells, a metal plate below the Tedlar sheet,
another metal plate placed on the thermal insulation of the hybrid collector, and
Baffles that are stuck between the two metal plates. A Thermal insulation, to mini-
mize thermal losses on the side and rear sides of the hybrid solar collector. The air
circulates in the sub-channels formed by the baffles to allow the heat transfer fluid to
convey the thermal energy produced. The experimental prototype is provided with
assisted PV ventilation system for PV cooling and tracking support to track the
Experimental and Numerical Study of Hybrid PV/Thermal Solar … 661
Outlet Air
Glass
PV cells
PV panel Tedlar
Metal
plates
Insolation
Length, L Cover
Air inlet
Fig. 1 Schematic representation of hybrid PV/T collector with fins
Table 1 The main electrical

Parameter Value
characteristics of the used PV
panel at standard test PV Module type Monocrystalline silicon (JT-185 M)
conditions (STC) [19] Cells number, N 12 × 6
Short-circuit current, Isc,ref 5.76 A
Open-circuit voltage, Voc,ref 43.2 V
Current at MPP, Imp,ref 5.14 A
Voltage at MPP, Vmp,ref 36 V
Maximum power, Pm,ref 185 W
Electrical conversion efficiency, ηref 14.4%
movement of the sun in order to receive maximum solar radiation throughout the
day (Fig. 1).
During the operation of the hybrid collector, the solar energy reaches the outer
surface of the collector in the form of solar radiation. A portion of this radiation is
converted into electrical energy and another part into thermal energy so that the heat
produced by the cells is evacuated through the coolant flowing between the fins and
the two metal plates (upper and lower). The hybrid air collector electrical and design
parameters used in this document are given in Table 1 and Table 2, respectively.
3 Numerical Modeling
During the modeling and for the energy balance equations of the system through each
constitutive layer of the hybrid collector, a few simplifying hypotheses are taken into
account:
Table 2 The main

PV/T geometry
geometrical characteristics of
the PV/T collector Parameter Characters Values
Length L 1.30 [m]
Width l 0.65 [m]
Thickness of fluid vein e 0.02 [m]
Fins geometry
Number of fins N 12
Length L fin 1.30 [m]
Width h 0.02 [m]
Height efin 0.002 [m]
Space between two fins p 0.05 [m]
• The thermo-physical properties of the system components are assumed to be

constant.
• Each layer of the hybrid collector is considered a temperature node (average of
the input to the output).
• Heat losses on the lateral sides of the hybrid collector are negligible compared by
the losses of the front and rear faces of the collector.
The energy conservation principle is used for each element (i) of the PV/T collector
in order to predict the temperature in each layer of the collector.
dTi
Mi Ci
dt
= Qi − Qi (1)
e s
The thermal balance takes into account the phenomena of heat transfer by conduc-
tion, convection, and radiation; interacting at the system level and the thermo-physical
proprieties of each component.
Energy balance equations for glass layers, PV cells, Tedlar, and thermal insulation
are given and detailed by Ref. [20].
The heat transfer balance for the top metal plate is given by the following equation.
dT p1
M p1 .C p1 . = h c, p−t .S. Tt − T p1 + h v, f − p1 .S . T f − T p1
dt
+ h r, p2− p1 .S . T p2 − T p1 + h c,ch− p1 .S . Tch − T p1 (2)
where S = L.l the collector area, S = N .e.L and S = (N + 1).p.L; e is the
thickness of a baffle, and p is the pitch (distance) between two baffles.
The heat transfer balance for the fins is given by the following equation.
dTch
Mch .Cch . = h c,ch− p1 .S . T p1 − Tch + h v, f −ch .Sch . T f − Tch
dt
+ h c,ch− p2. S . T p2 − Tch (3)
where Sch = 2.N.L.h; N is the fins number, and h is fin height.

The heat transfer balance for the coolant fluid (air) is given by the following
equation.
dT f
M f .C f . = h v, f − p1 .S . T p1 − T f + h v, p2− f .S .(T p2 − T f )
dt
+ h v, f −ch .Sch . Tch − T f − Q u,th (4)
The heat transfer balance for the lower metal plate is given by the following
equation.
dT p2
M p2 .C p2 . = h r, p1− p2 .S . T p2 − T p1 + h v,c− p2 .S . T f − T p2
dt

+ h c,c− p2 .S . Tch − T p2 + h c, p2−i .S. Ti − T p2 (5)
The conduction, convection, and radiation heat transfer coefficients at the hybrid
collector layers are given in detail in our previous works [20, 21].
Several correlations have been proposed by the researchers in this field to predict
the variation of the electrical efficiency of the PV module.
The correlation proposed by EVANS [19] given by:

G
ηele = ηref 1 − βT . TC − TC,ref + γ .ln (6)
G ref
G and G ref which are the illumination and the reference illumination (G ref =
1000 W/m2 );
Tc and Tc,ref which are the temperature and the reference temperature of the cells
(Tc,ref = 25 o C);
where β T is the temperature coefficient, and γ is the solar radiation coefficient.
Another rather important parameter that shows the overall energy efficiency of the
PV/T collector is given as the sum of the thermal yield (ηth ) and the equivalent of the
electrical yield in thermal (ηele,th ) using the thermal conversion factor (Cf ) [22, 23].
ηele
ηT = ηth + ηele,th = ηth + (7)
Cf
where:

ηth = ṁC f . T f,O − T f,i (8)
4 Experimental Tests
As presented in Fig. 2, the experimental designed system is mainly composed of

a single pass unglazed PV/T collector provided with fins linked to a PV-assisted
forced air circulation device. in the first step, and for numerical validation, the PV/T
collector is ventilated by a constant airflow assured by a DC electric alimentation
instrument and the PV/T collector has been fixed (without tracking device) on a
properly insulated aluminum housing using polyurethane foam. The air ventilation
system consists of a brushless type DC fan that can be directly coupled to small PV
modules (N*4Wp) for different arrangements.
In order to allow us to evaluate the thermal performance of the designed
collector, a bench of measurements has been set up for the simultaneous measure-
ment of the different parameters such as the temperature of the fluids (air and
water), the temperature of the layers and components of the collector, solar illu-
mination and air velocity. For this purpose, many measuring instruments have
Fig. 2 Photograph of the

designed hybrid PV/T
system
been used (type K thermocouples, hot wire anemometer, pyranometer, and data
acquisition) to conduct simultaneous measurements of all necessary parameters
with a time step of five minutes.
It is important to note that meteorological parameters were collected from the
meteorological station installed at the UDES site.
The mathematical and numerical model developed for this study has been tested
and validated with the experimental data accorded for a typical day (of Oct. 04, 2017).
The comparison between experimental and numerical results is given by Fig. 3.
The mean absolute percent error (MAE) given by Eq. 9 [24] calculates the differ-
ence between the simulated results found by the model used in this study and the
experimental ones in the same operating conditions. The values of the mean absolute
percent error are 3.10% for Outlet air temperature, 6.31% for Tedlar temperature,
4.1% for Fin temperature, and 3.81% for Electrical Power. The calculated MAE
shows that there is a good agreement between the results of the simulation and the
experimental data with a maximum value of 6.31%.
n
100 X sim,i − X exp,i
M AE(%) = (9)
n 1 X exp,i
80
60
T t (°C)
40 Theo
Exp
20
8 9 10 11 12 13 14 15 16 17
50
40
Tfo (°C)
30
Theo
20 Exp
10
8 9 10 11 12 13 14 15 16 17
60
50
Tfin (°C)
40 Theo
30 Exp
20
10
8 9 10 11 12 13 14 15 16 17
Time (hr)
Fig. 3 The experimental and theoretical hourly evolutions of the Tedlar (Tt ), the outlet air (Tfo ),
and the fins (Tfin ) temperatures for the day of 04/10/2017
5 Results and Discussions
The simulations were made under the weather conditions of UDES (Unité de
Développement des Equipements Solaires) located at Bousmail, Tipaza, Algeria for
a cold day of the winter season (Feb. 24, 2018).
The Meteorological data for the test day of Feb. 24, 2018, are given by Fig. 4.
The weather data include solar radiation, ambient temperature, and wind speed.
The evolution compared between the different temperatures of PV/T layers shows
compliance and logic in the temperature distribution of the hybrid collector elements.
Figure 5 shows the temperature evolution for PV cells and outlet air during the test
day. The PV cells’ temperature is relatively stable between 35–45 °C during the
test day. The outlet air temperature stabilizes between 20 and 23 °C. This range
of air temperature is suitable and comfortable for building heating. Stabilization in
temperature is caused by two key factors; PV-assisted ventilation system and solar
tracking system.
The temperature of the solar cells represents the highest temperature in the hybrid
collector; it reaches a value of about 46 °C. This is explained by the fact that the
photovoltaic cell is a heat generator converting the received solar radiation into
electricity and heat.
Figure 6 and shows the evolution of the temperature of the coolant and the cells,
respectively.
1200 24
1000 20
Solar radiation (W/m²)
800 16 Ambient temperature
600 12
G (W/m²)
400 Tamb (°C)

8
200 4
0 0
6
Wind Speed (m/s)
0
08:00 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00
Time (hr) (Feb.24,2018)
Fig. 4 Evolution of solar irradiance, ambient air temperature, and wind speed during the test day
(Feb. 24, 2018)
50
PV Cell
Outlet Air
40
Temperatue (°C)
30
20
10
0
08:00 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00
Time (h)
Fig. 5 Temperature evolution for PV cells and outlet air during the test day
25
23
Heating Range
21
Air temperature (°C)
19
T -T T
fo amb fo
17 T
amb
15
13
11
9
7
5
08:00 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00 18:00
Time (h) (Feb. 24, 2018)
Fig. 6 Temperature evolution of inlet (Tamb ) and outlet (Tfo ) air during the test day
From Fig. 6 it can be noted that for this configuration, the temperature difference
between PV/T collector inlet and outlet is important and appreciable. The temper-
ature difference range would be as high as 10 °C in a relatively clear sky and cold
conditions. This difference is favorable for comfort improvement inside a building
space when exterior temperature (PV/T collector inlet) is not very low. This indicates
that in such cloudy conditions, the operating flow rate is sufficiently low to allow
noticeable heat gain by the air circulating through the collector duct.
Figure 7 illustrates the evolution of electrical and thermal powers produced from
the hybrid PV/T collector.
500
Ele
450 The
400
350
Power (W)
300
250
200
150
100
50
0
08:00 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00 18:00
Time (h)
Fig. 7 Evolution of electrical and thermal powers during the test day
From Fig. 7, the electrical and thermal powers reached important and stabilized
values about 170 W for the electrical power and between 350 and 450 W for the
thermal power.
The fins have a significant effect on the energy performance of a hybrid collector,
resulting in large heat dissipation of the hybrid collector.
Figure 8 presents the evolution of electrical and thermal efficiencies from the
hybrid PV/T collector.
It is noted that under the studied operating conditions the thermal efficiency of
the hybrid collector quickly increases between 08h00 and 09h00 then stabilizes from
09h00 to 17h00 around a value of 32%.
15 40
Electrical efficiency (%)
Thermal efficiency (%)
14 30
13 20
12 10
11 0
08:00 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00
Fig. 8 Evolution over time of thermal and electrical efficiencies

The variation of electrical efficiency is presented in the same figure. It is clearly

perceived from this figure that interesting values of electrical efficiency are obtained
although conditions are stable and very favorable. This fact is explained by the
fact that under these stable conditions (solar radiation and airflow), the PV cell
temperature is also stable (not very high) and still within the acceptable range. These
results can be confirmed by returning in Fig. 5, the PV cell temperature did not exceed
45 °C and still stable between 35 and 45 °C throughout the whole day.
The simultaneous improvement of thermal and electrical efficiencies leads to an
increase in the overall efficiency of the hybrid collector.
6 Conclusion
In this study, the energy performance of a hybrid PV/T collector provided with
fins and equipped with a direct PV air ventilation device and tracking support was
investigated and analyzed, this configuration suitable for winter cold periods to
air conditioning in a building. A mathematical model was established based on
electrical and thermal parameters and characteristics of the solar device system
with the use of an empirical correlation for airflow prediction as a function of solar
irradiance. The numerical model was successfully validated and compared with
experimental data. The evaluation of the energy performance of the PV/T collector
has been carried out using the numerical model of this study. Interesting results are
obtained from this work.
• The validation results given by the proposed model coincides with the experi-
mental ones.
• The solar system provides a suitable and comfortable air temperature (20–23 °C)
for spaces building heating.
• The developed configuration of PV/T collector makes the electrical and thermal
energy efficiencies appreciable and stable.
References
1. Slimani MEA, Amirat M, Bahria S (2015) Study and modeling of heat transfer and energy
performance in a hybrid pv/t collector with double passage of air. Int J Energy Clean Environ
16(1–4):235–245
2. Srinivas M, Jayaraj S (2013) Performance study of a double pass, hybrid-type solar air heater
with slats. Int J Energy Eng 3(4):112
3. Slimani MEA, Amirat M, Bahria S (2015) Analysis of thermal and electrical performance
of a solar PV/T air collector: Energetic study for two configurations. In: 3rd International
conference on control, engineering and information technology, CEIT 2015, pp 1–6
4. Othman MY, Ibrahim A, Jin GL, Ruslan MH, Sopian K (2013) Photovoltaic-thermal (PV/T)
technology–the future energy technology. Renew Energy 49:171–174
5. Boumaaraf B, Touafek K, Ait-cheikh MS, Slimani MEA (2020) Comparison of electrical

and thermal performance evaluation of a classical PV generator and a water glazed hybrid
photovoltaic–thermal collector. Math Comput Simul 167:176–193
6. Sarhaddi F, Farahat S, Ajam H, Behzadmehr A (2010) Exergetic performance assessment of a
solar photovoltaic thermal (PV/T) air collector. Energy Build 42(11):2184–2199
7. Hegazy AA (2000) Comparative study of the performances of four photovoltaic/thermal solar
air collectors. Energy Convers Manag 41(8):861–881
8. Mortezapour H, Ghobadian B, Khoshtaghaza MH, Minaee S (2012) Performance analysis of
a two-way hybrid photovoltaic/thermal solar collector. J Agric Sci Technol 14(4):767–780
9. Mojumder MSS, Uddin MM, Alam I, Enam HK (2011) Study of hybrid photovoltaic thermal
(PV/T) solar system with modification of thin metallic sheet in the air channel. J Energy Technol
Policy 3(5):47–55
10. Shan F, Tang F, Cao L, Fang G (2014) Dynamic characteristics modeling of a hybrid
photovoltaic–thermal solar collector with active cooling in buildings. Energy Build 78:215–221
11. Chaouch WB, Khellaf A, Mediani A, Slimani MEA, Loumani A, Hamid A (2018) Experimental
investigation of an active direct and indirect solar dryer with sensible heat storage for camel
meat drying in Saharan environment. Sol Energy 174:328–341
12. Slimani ME-A (2017) Etude d’un séchoir solaire agricole muni d’un capteur solaire de type”
PV-THERM”-réalisation d’un prototype et caractérisation,” UNIVERSITE DES SCIENCES
ET DE LA TECHNOLOGIE HOUARI BOUMEDIENE
13. Parker GJ (1976) A forced circulation system for solar water heating. Sol Energy 18(5):475–479
14. Gholampour M, Ameri M (2016) Energy and exergy analyses of photovoltaic/thermal flat
transpired collectors: experimental and theoretical study. Appl Energy 164:837–856
15. Boumaaraf B, Boumaaraf H, Slimani ME-A, Tchoketch_Kebir S, Ait-cheikh MS, Touafek
K (2020) Performance evaluation of a locally modified PV module to a PV/T solar collector
under climatic conditions of semi-arid region. Math Comput Simul 167:135–154
16. Slimani ME-A, Sellami R, Mahrane A, Amirat M (2019) Study of hybrid photovoltaic/thermal
collector provided with finned metal plates: a numerical investigation under real operating
conditions. In: 2019 International conference on advanced electrical engineering (ICAEE), pp
1–6
17. Touaba O et al (2020) Experimental investigation of solar water heater equipped with a solar
collector using waste oil as absorber and working fluid. Sol Energy 199:630–644
18. Slimani ME-A, Sellami R, Mahrane A, Amirat M (2019) Experimental study of a Glazed
Bi-Fluid (water/air) solar thermal collector for building integration. In: 2019 International
conference on advanced electrical engineering (ICAEE), pp 1–6
19. Sellami R, Amirat M, Mahrane A, Slimani ME-A, Arbane A, Chekrouni R (2019) Exper-
imental and numerical study of a PV/Thermal collector equipped with a PV-assisted air
circulation system: configuration suitable for building integration. Energy Build 190:216–234
20. Slimani MEA, Amirat M, Bahria S, Kurucz I, Aouli M, Sellami R (2016) Study and modeling
of energy performance of a hybrid photovoltaic/thermal solar collector: configuration suitable
for an indirect solar dryer. Energy Convers Manag 125:209–221
21. Slimani MEA, Amirat M, Kurucz I, Bahria S, Hamidat A, Chaouch WB (2017) A detailed
thermal-electrical model of three photovoltaic/thermal (PV/T) hybrid air collectors and
photovoltaic (PV) module: comparative study under Algiers climatic conditions. Energy
Convers Manag 133:458–476
22. Luque A, Hegedus S (2011) Handbook of photovoltaic science and engineering. Wiley
23. Joshi AS, Tiwari A, Tiwari GN, Dincer I, Reddy BV (2009) Performance evaluation of a
hybrid photovoltaic thermal (PV/T) (glass-to-glass) system. Int J Therm Sci 48(1):154–164
24. Nadji Maachi I, Mokhtari A, Slimani ME-A (2019) The natural lighting for energy saving and
visual comfort in collective housing: a case study in the Algerian building context. J Build
Eng 24:100760
Maximum Power Point Tracking
(MPPT) for a PV System Based
on Artificial Neural Network ANN
and Comparison with P&O Algorithm
Aicha Djalab, Nassim Sabri, and Ali Teta
Abstract For the optimal operation of photovoltaic system, The MPPT (Maximum
Power Point Tracking) control unit is an essential part for the photovoltaic system.
In addition to the protection function, this command ensures the continuation of
the maximum power point (MPPT) and allows the PV generator to deliver its
maximum power regardless of the variation in climatic conditions (sunshine and)
temperature). This work intends to provide an artificial neural network (ANN)
maximum power point tracking (MPPT) method which is fast and precise in finding
and tracking the maximum power point (MPP) in photovoltaic (PV) applications,
under rapidly changing of solar irradiation, and the P&O algorithm. ANN and P&O
MPPT algorithms, and other components of the MPPT control system which are
PV module and DC-DC boost converter, are simulated in MATLAB/ Simulink, we
used in The proposed ANN two inputs which are irradiation and ambient tempera-
ture, and one output is the optimum voltage of the PV system. The proposed ANN
was analyzed under different irradiation conditions. The response of the proposed
ANN for MPPT controllers found to be lesser oscillation at MPP and faster tracking
response compared with the P&O algorithm. Comparisons of MPPT with P&O algo-
rithm and without MPPT tracker are also shown in this paper. It is demonstrated that
the neural network based MPPT tracking require less time and provide more accurate
results than the P&O algorithm based MPPT.
Keywords Photovoltaic system · Tracking the maximum power point MPPT ·

Artificial neural network (ANN) · The P&O algorithm
A. Djalab (B) · A. Teta

Applied Automation and Industrial Diagnostics Laboratory, Djelfa University, Djelfa, Algeria
A. Teta
N. Sabri
Department of Electrical Engineering, Medea University, Medea, Algeria

https://doi.org/10.1007/978-981-15-6403-1_45
672 A. Djalab et al.
1 Introduction
Currently, looking for alternative energy resources has become essential and a signif-
icant growth worldwide., due to increasing demand for energy and depletion of fossil
fuel reserves, in order to find sustainable energy resources, various researches have
been carried out, mainly on renewable energy sources [1].
There are many sources to produce, but there are also constraints related to
its production, such as the effect of pollution and global warming… etc. These
constraints lead research towards the development of renewable and non-polluting
energy sources; the use of renewable energy such as solar energy has shown that
they could contribute on a large scale to find a solution to the above-mentioned
problems [2].
Renewable energy means that the energy of the Sun, wind, Earth, water or biomass
heat is regarded as a clean alternative energy source. Among the different types
of renewable energy, solar energy has been exploited to produce calories, thermal
energy and lastly the electricity which is the main subject of this article. However,
the characteristics of the outputs of PV systems are not linear and change according
to the temperature and irradiation, so a MPPT controller is needed to extract the
maximum power to the terminals of PVG. Therefore, the MPPT techniques are used
to maintain the PV works at its MPPT [3].
Several MPPT techniques have been proposed in the literature; For example,
Perturb and Observe (P&O) method [4, 5] incremental conductance (INC) method
and fuzzy logic based methods [6], etc.
Many MPPT algorithms have been developed so far, and these MPPT algorithms
vary in application, complexity, precision, sensors, cost, popularity, etc. Among them,
P&O method has drawn significant attention of the researchers due to its simplicity.
However, its deviation from MPP under rapidly changing of solar irradiation, and
oscillation problem in steady states is unavoidable [7].
New approach in MPPT systems is to apply artificial intelligent systems, such as
fuzzy logic and artificial neural network (ANN) as MPPT controller or as a part of
integrated MPPT system [7]. These methods are fast in tracking the MPP, and give
more stable response in compare with conventional algorithms.
This paper presents a PV model with MATLAB/SIMULINK, and tracks the
MPP using an Artificial Neural Network (ANN) and the P&O technique., And the
comparison between them.
The ANN uses the irradiance and temperature as inputs. The PV model was
simulated to generate training and testing data for PV module at various conditions
to train and test the proposed neural net-work.
We started this work by PV module modeling is described in Sect. 2, followed by
DC-DC converter in Sect. 3 and maximum power point tracking (MPPT) in Sect. 3.
Section 4 includes simulation results, and finally a conclusion is given in Sect. 5.
Maximum Power Point Tracking (MPPT) for a PV System … 673
2 Photovoltaic System Modeling and Simulation
Photovoltaic system consists of four blocks as shown in Fig. 1. The first block is
the source of energy (solar panel), the second block is a static converter DC-DC, the
third block represents the load and the fourth block represents the control system.
The main role of the static converter is to an impedance matching so that the Panel
delivers maximum energy.
To obtain a desired output such as the power, the output current and output voltage
[3, 8], the photovoltaic system which consists of a set of basic photovoltaic cells can
be connected in series and/or parallel, where the Photovoltaic cells are the main
components of the module.
A. Model of a Photovoltaic Cell
The simplest circuit of the photovoltaic cell model is represented by a current source
in parallel with an ideal diode, which will be adopted in this study [9].
The equivalent circuit of a solar cell is given in Fig. 2. This equivalent circuit is
composed of a current source controlled models which the photovoltaic effect (the
generated current is controlled by the Sun’s rays). The diode represents the effect of
the junction semiconductor of the cell. Two resistors (Rs, Rsh) represent the series
resistor Rs and a shunt resistor Rsh represent the power losses by connections in
series and in parallel [8].
The basic equations that express the I-V characteristic of the photovoltaic model
are given in the following equations [9, 10]:
Iph = Id + Ish + I (1)
Fig. 1 Photovoltaic system

Fig. 2 Equivalent circuit of a solar cell
The current in the diode Id is given by:

V + Rs I
Id = I0 exp −1 (2)
Vt a
The current in the RP resistance is given by:

V + Rs I
Ish = (3)
Rsh
From Eq. (1), we obtain the expression of current I:
I = Iph − Id − Ish (4)
Replacing (4) in the Eqs. (2) and (3), the characteristic equation becomes:

V + Rs I V + Rs I
I = Iph − I0 exp −1 − (5)
Vt a Rsh
where the junction thermal voltage Vt is defined by
Ns K T
Vt = (6)
q
where:
I Output current [A],
Id Current of parallel diode [A],
Iph Photo current; [A],
Imp Current corresponds to maximum power point,
Vmp Voltage corresponds to maximum power point,
Ish Shunt current; [A],
Table 1 Ideality factor (A)

Technology Ideality factor
Si-mono 1.2
Si-poly 1.3
a-Si-H 1.8
a-Si-H tandem 3.3
a-Si-H triple 5
cdTe 1.5
CTs 1.5
AsGa 1.3
Isc Short-circuit current [A],

Io Saturation current of the diode [A],
Voc PV module open-circuit voltage [V],
V The cell voltage [V],
Rs The resistance series cell [],
Rsh The shunt resistance cell [],
Vt The thermal voltage of the module [V],
Ns The number of cells connected in series,
T The temperature of the cell [°K],
q Electron’s charge e = 1.6 *10−19 C,
K The Boltzmann constant (1.3854*10−23 J°K−1 ),
a Diode quality (or ideality) factor, it can be classified by the PV technology
according to Table 1 [11].
Figure 3 shows a nonlinear characteristic of the photovoltaic cell. This charac-
teristic varies with the change in metrological terms. As the optimal power point
varies broadly according to weather conditions, a power converter switch should be
controlled by a specific algorithm to track the maximum power point [12].
B. DC-DC Boost Converter
DC-DC converters which convert a DC voltage from one level in input to another
DC level at its output are composed of switching devices, inductors and capacitors.
The switch is usually an IGBT or MOSFET. A PWM signal driven at the gate of
either of them, switches it on and off. In MPPT systems, this signal is controlled by
MPPT controller.
DC-DC converter used in this work is a boost converter which steps up the input
voltage to a desired voltage at its output. Input voltage of the boost is output voltage of
the PV module. Figure 4 shows the electrical circuit of the DC-DC boost converter [1].
The duty cycle, D = ton /T where T is the period equal to ton + toff . The relation
between the input and output voltage is given from.
1
vo = vi (7)
1− D
Fig. 3 I-V and P-V characteristics of a photovoltaic cell
Fig. 4 DC-DC boost converter
The input voltage of the converter is the solar module output voltage that is changes
all the time. However, its output voltage must be kept at a desired value. This is done
by controlling the duty ratio, so that the operating point of the PV system can be
adjusted to realize MPPT algorithm [13].
C. PV Array Characteristics
The characteristics of the PV array used in this paper are presented in Table 2.
Table 2 Electrical
Electrical Characteristics
characteristics of the kc 130
ght PV module Power at maximum power point (Pmpp) 130 W
Optimal operation voltage (Vmpp) 17.6 V
Current at maximum power point (Impp) 7.39 A
Open circuit voltage (Voc) 21.9 V
Short circuit current (Isc) 8.02 A
Number of cells connected in series 36
Number of cells connected in parallel 1
3 Maximum Power Point Traking Contrellers
The photovoltaic system to work at maximum power points of their characteristics,

there are specific laws that meet this need. This command is named in the “Maximum
Power Point Tracking” (MPPT) literature. The principle of these commands is to
seek the maximum power point (MPP) by keeping a good adaptation between the
generator and the load to ensure the transfer of maximum power. The technique of
control so to act on the duty cycle in an automatic way to bring the point of operation
of the generator at its optimum value whatever weather instabilities or brutal changes
in load [4]. For such reason, three MPPT control techniques will be discussed.
A. P&O Controller
The perturbation and observation (P&O) algorithm is probably the most frequently
used in practice, mainly due to its easy implementation [4].
As the name suggests it is based on the perturbation of system by the increase or
decrease in Vref where acting directly on the duty cycle of the converter DC-DC,
then observation of the effect on the output power of the panel. If the current value of
the power P(k) panel is greater than the previous value P(k-1) is then retains the same
direction of previous disturbance or we reverse disruption of the previous cycle. The
Fig. 5 shows the flowchart of this algorithm.
B. MPPT Method Based on Artificial Neural Network—ANN
Artificial neural networks (ANNs) are a family of statistical learning models inspired
by biological neural networks (the central nervous systems of animals, in particular
the brain) and are used to estimate or approximate functions that can depend on a
large number of inputs and are generally unknown. Artificial neural networks are
generally presented as systems of interconnected “neurons” which send messages
to each other. The connections have numeric weights that can be tuned based on
experience, making neural nets adaptive to inputs and capable of learning [14]. The
ANN model is modeled in MATLAB/Simulink environment.
In our work, Neural network used is multilayer perceptron (MLP) type of archi-
tecture. with two neurons in input layer, five neurons in output layer and one in output
layer is constructed. The model of network is as below (Fig. 6).
Start
Sense V(k),I(k)
Yes
P(k)-P(k-1)=0
No
No Yes
P(k)-P(k-1)>0
No No Yes
Yes V(k)-V(k-1)>0 V(k)-V(k-1)>0
Decrease Vref Increase Vref Decrease Vref Increase Vref
Return
Fig. 5 Flowchart perturbation and observation algorithm
Fig. 6 Architecture of a typical ANN

Best Training Performance is 2.3274e-05 at epoch 119

10 2
Train
Best
Mean Squared Error (mse)
10 0
10 -2
10 -4
0 20 40 60 80 100
119 Epochs
Fig. 7 Training performance curve
The inputs to neural network are temperature and irradiance and the output is
the voltage at the maximum power point. In hidden layer the basis function is
weighted linear sum and activation function tansig. For output layer basis function
is linear weighted sum whereas activation function is Linear. The algorithm used
is Levenberg-Marquardt, and performance function of the network is mean square
error (MSE) which is given by Eq. (8) as below:

N
E= (t (k) − O(k))2 (8)
k=1
The network which was fully trained with the lowest error is capable to be used in
the testing process. Performance of ANN to minimize the MSE error is shown in the
training performance curve of Fig. 7. Network was trained until it achieved a very
small MSE typically 2.3274e-05 which reached after 119 epochs. which implies that
the network parameters (weight and bias) are optimized (Fig. 7).
In this section, we begin by assessing the photovoltaic system by simulation with

MATLAB/Simulink simulation tool. Then, two further MPPT methods are studied:
150
ANN / G=1000w/m2, T=25°C

100
Power (w)
130 P&O / G=1000w/m2, T=25°C
120
110
50 100
90
80
0.01 0.02 0.03 0.04 0.05 0.06 0.07
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
Fig. 8 PV power by P&O, and ANN method in standard conditions
the method (P&O), and the ANN method. The two systems are simulated under
standard environmental conditions and many changes of irradiance conditions.
A. Operation Under Constant and Uniform Irradiance
In this test PV system was simulated with uniform irradiation on all the solar cells.
The irradiance and temperature are held constant. It takes the values of the standard
conditions: the temperature T = 25 °C and irradiance G = 1000 W/m2 on the PV
module. Only one MPP occurs in this case (130 W).
Figure 8 shows that P&O take more time to converge to MPP estimated at 6.5 ms.
However, the ANN method track the MPP with fast time response estimated at
2.5 ms. Moreover, sustained oscillations are present in PV output power of P&O
around MPP.
Figures 8 show the results of the output power under standard conditions.
B. Operation Under Partial Shading Conditions
In order to evaluate the performance of the P&O and ANN MPPT under changing of
irradiation, they will be tested under a constant temperature of 25 °C and changing
irradiance. the signal presented in Fig. 9 is applied as irradiation to the PV module
in SIMULINK. Figure 10 shows the simulation results for P&O and ANN methods.
As Fig. 10 depicts, ANN method can rapidly track the MPP under rapidly changing
irradiation, while although the advantage of the P&O simple but it causes an oscilla-
tion around the MPP and has a quite low response under sudden change in irradiance
leading to the power loss in steady state.
Finally, it is obvious through the simulation that the ANN gives better results
compared to the P&O under changing irradiance.
1000
800
Irradince (W/m2)
600
400
200
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
Fig. 9 partial shading conditions
150
G = 1000 W/m2
G = 850 W/m2 ANN / T=25°C
100 P&O / T=25°C

Power (w)
100
80
G = 650 W/m2
60
50 G = 420 W/m2
40
20
0.18 0.2 0.22 0.24 0.26 0.28 0.3
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
Fig. 10 PV power by P&O, and ANN method under partial shading conditions
5 Conclusion
The optimal operation of a photovoltaic system is depending on the MPPT controllers.

These controllers are designed to pursue the MPP and minimize the error between
the operation power and the maximum power. In this article, we have described
the main elements of the PV system. Then, we illustrated the principle of a two
MPPT techniques namely P&O, and ANN. Finally, we finished by a simulation of
the different techniques.
Simulation results for ANN and P&O methods presented in this paper, show
that ANN method is very fast and precise in finding and tracking the MPP in case
of changing of solar irradiation. Furthermore, this method can stably extract the
maximum power point under changing of solar irradiation. On the contrary, P&O
method has high oscillation around MPP under slowly changing of solar irradiation,
which leads to high power loss in long term. From this results, we can to say that the
ANN method is the better method compared to the P&O.
References
1. Yongchang Y, Chuanan Y (2012) Implementation of a MPPT controller based on AVR Mega16

for photovoltaic systems. Int Conf Futur Electr Power Energy Syst 17:241–248
2. Safari A, Mekhilef S (2010) Simulation and hardware implementation of incremental conduc-
tance MPPT with direct control method using cuk converter. IEEE Tran Ind Electro. 58(4),
1154–1161
3. Bensaci W (2012) Modélisation et simulation d’un système Photovoltaïque adapté par une
commande MPPT, Thèse de doctorat“, Université Kasdi Merbah–Ouargla
4. Abbes H, Abid H, Loukil K, Toumi A, Abid M (2013) Comparative study of five algorithms of
MPPT control for a photovoltaic system (CIER’13) Sousse, Tunisie. ISSN 2356-5608 –2013
5. Bun L, (2011) Détection et Localisation de Défauts pour un Système PV. Université de Grenoble
Thèse de Doctorat en Génie Electrique
6. Rezgui W, Mouss LH, Mouss MD (2013) Modeling of a photovoltaic field in malfunctioning.
In: The Proceedings of the 2013 IEEE CoDIT (international conference on control, decision
and information technologies), pp 788-793. Hammamet Tunisia. (06–08 May 2013)
7. Khanaki R, Radzi M, Hamiruce M (2013) Marhaban comparison of ANN and P&O MPPT
methods for PV applications under changing solar irradiation. In: IEEE conference on clean
energy and technology (CEAT)
8. Drir N, Barazane L, Loudini M (2013) Fuzzy logic for tracking maximum power point of
photovoltaic generator. Revue des Energies Renouvelables 16(1): 1–9
9. Shareef H, Mohamed A, Mutlag AM (2014) A current control strategy for a grid connected PV
system using fuzzy logic controller. In: IEEE International Conference on Industrial Technology
(ICIT), pp. 890–894. Busan, Korea (February 2014)
10. Abbes H, Abid H, Loukil K (2015) An improved MPPT incremental conductance algorithm
using T-S Fuzzy system for photovoltaic panel. IJRER 5(1)
11. Bellia H, Youcef R, Fatima M (2014) A detailed modeling of photovoltaic module using
MATLAB. NRIAG J Astron Geophys
12. Harsha P, Dhanya PM, Karthika K (2013) Simulation and proposed hardware implementation
of MP controller for a solar PV system. Int J Adv Electr Electron Eng (IJAEEE) 2(3)
13. Abdelkader HI, Hatata AY, Hasan MS (2015) Developing intelligent MPPT for PV systems
based on ANN and P&O algorithms. Int J Sci Eng Res 6(2). ISSN 2229-5518
14. Mathur D (2014) Maximum power point tracking with artificial neural network. Int J Emerg
Sci Eng (IJESE) 2(3). ISSN: 2319–6378
Development of a Stand-Alone
Connected PV System Based
on Packe U Cell Inverter
Khaled Rayane, Mohamed Bougrine, Atallah Benalia, and Kamel Guesmi
Abstract In this paper, we investigate a competitive Multilevel Inverter (MLI)

named by Packed U Cell (PUC). This converter is capable to generate more levels
with fewer components compared with existing MLI topologies. The system consists
of a PV arrays supply connected to a single-phase PUC5 inverter. A boost converter
is employed to extract the Maximum Power (MPP) controlled by an MPP tracking
(MPPT) algorithm. Then, A nonlinear control strategy with Pulse Width Modulation
(PWM) is used to control the 5-level PUC (PUC5) to ensure good tracking and high-
quality convention. Some simulations were performed in MATLAB/SIMULINK to
demonstrate the effectiveness of this topology with the proposed control scheme.
Keywords Photovoltaic · MPPT · Packed U cell (PUC) · PUC5 · Multilevel

converters
1 Introduction
The use of renewable energy has increased significantly in the last decade in order
to reduce environmental pollution and satisfy the dramatically increased energy
demand. Among others, photovoltaic energy became more favorable due to its
simplicity of allocation, high dependability and lower cost. However, due to their
DC nature, photovoltaic generators require the use of an inverter to convert DC to
K. Rayane · M. Bougrine · A. Benalia

LACoSERE Lab, University of Laghouat Laghouat, Laghouat, Algeria
M. Bougrine
A. Benalia
K. Guesmi (B)
Department of Electrical Engineering, University of Djalfa, Djelfa, Algeria

https://doi.org/10.1007/978-981-15-6403-1_46
684 K. Rayane et al.
Table 1 Components counts of different multilevel topologies

Inverter DC source Clamped diode Flying capacitor Active switch Total
NPC 1 10 0 12 23
FC 1 0 6 12 19
CHB 2 0 0 8 10
PUC7 1 0 1 6 8
AC. Conventional inverters have many drawbacks, such the high power losses and
harmonics. In this context, multilevel topologies are becoming preferable for their
high-quality conversion and low harmonics content that can provide [1–3].
Drawbacks of multilevel inverters are not trifling, the most salient one is their topo-
logical complexity. The most common structures developed during the last decades
are the Neutral point clamped (NPC), Cascade H Bridge (CHB), Flying capacitors
(FCI) [4–9]. Recently a new topology has been introduced by [10] called Packed U
Cells (PUC) that can generate more levels with lower components requirement.
PUC can generate up to 7-level voltage with 6 active switches, one capacitor
and one DC source [11, 12]. Compared to other topologies, the PUC7 decreases
significantly the required components for the same number of levels (see Table 1). In
spite of these, the PUC7 has a big flaw, it requires a complicated control, and more
than that, the obtained wave lacks symmetry, which is caused by the uncontrollability
of the capacitor voltage. To solve this problem a 5-level PUC inverter has been
proposed [13–15] were in this case, the capacitor voltage should be stabilized at E/2
instead of E/3, where E is the DC source voltage. This will give some redundant
states that make the capacitor voltage perfectly controllable.
In this work, we aim to study the PUC5 converter when supplied by a PV system.
First. The studied system composed of PV arrays supplies an AC load through PUC5
inverter. A DC-DC boost converter is employed to extract the maximum power from
the solar panels. Perturb and Observe (P&O) is used to control the boost converter to
track the maximum power point [16, 17]. Then, a nonlinear feedback control based
on a linearizing PI regulator is designed to control the PUC5 inverter to convert DC
to AC with an attempt to provide high-quality conversion with low harmonics.
The rest of the paper will be organized as follows. Section 2. Will briefly describe
the PV system used for this work. Section 3 presents the PUC5 with a description
of its state space averaged model. Section 4 discusses the nonlinear control schemes
applied to the PUC5. Section 5 provides some simulation results. Finally, we achieve
the paper with a conclusion.
2 The Proposed Power Converter System
The PV system proposed in this work is constructed of three main parts, the first
part is the PV array composed of three Sun Power SPR-305-WHT PV modules in
series and two parallel strings. The second part is the DC/DC boost converter used
Development of a Stand-Alone Connected PV System … 685
Fig. 1 Equivalent Circuit of

PV cell
to extract the maximum power from the PV array. The third part is a PUC5 inverter
to convert DC power to AC in order to supply an AC load.
A. PV Array Modeling
A photovoltaic cell converts solar light into electrical energy. When photons hit the
surface of the photovoltaic cell an electron is released from the atom, which results
in generating a flow of electricity. The equivalent circuit of the PV cell is shown in
Fig. 1. The output current of photovoltaic generator (PVG) is given by
(V +I Rs ) V + IR
s
I = I ph − Io e ηK T /q − 1 − (1)
Rsh
where I ph is the photo-current, I 0 the saturated current voltage across the diode, q
the electron charge, T the junction temperature, K is the Boltzman constant, ï the
ideality factor, and Rs and Rsh are respectively the series and shunt cell resistors.
B. Perturb and Observe (P&O) Algorithm
This part explains the proposed algorithm that controls the boost converter to extract
the maximum power from the PV array, P&O is the most common technique for
achieving the Max Power Point Tracking (MPPT) due to its simplicity and effi-
ciency. Using the measurement of the PV voltage/current and comparing it with the
previous one. If the power increase the perturbation direction take the same direction.
Otherwise, the perturbation takes the reverse direction. The flowchart of the P&O
algorithm is shown in Fig. 2.
3 Five-Level Packed U Cell Converter
Single-phase PUC inverter is shown in Fig. 3 is constructed of six switches, one DC

source cell (E, S1 , S 1 ), one capacitor cell (C, S2 , S 2 ) and two additional switches
(S p , S p ) to generate the negative output voltage. There are eight switching possibil-
ities where the corresponding output voltages are listed in Table 2. In fact, the PUC
Fig. 2 The flow chart of P&O control
is capable of generating seven levels in output voltage, this can be achieved by regu-
lating the voltage across the capacitor to E/3. However, this configuration will cause
the loss of capacitor voltage contractility. To overcome this problem its preferable to
control the capacitor voltage at E/2, this configuration will generate only five levels,
but gives us a redundancies that give different dynamics for the capacitor voltage
(see Table 2).
C. Dynamical Model
By using Kirchhoff law the average mathematical model of PUC5 is expressed as
follow
d V2 1
= (u 2 − u 1 )
dt C
vo = V1 u p − u 1 + V2 (u 1 − u 2 )
Fig. 3 The PUC inverter
Table 2 Output voltage of all switching states

State Sp S1 S2 vo V̇2
1 0 0 0 0 0
2 0 0 1 −V2 –
3 0 1 0 V2 − V1 +
4 0 1 1 −V1 0
5 1 0 0 V1 0
6 1 0 1 V1 − V2 +
7 1 1 0 V2 –
8 1 1 1 0 0
di o 1
= (vo − Ri o ) (2)
dt L
where V1 , V2 and vo are the source, capacitor and output T-average voltages respec-
tively. i o is the load current. C, L and R are the cell capacitor, load inductor and load
resistance respectively. u 1 , u 2 and u p are the T-average inputs, which stand for the
duty cycles of the switches S1 , S2 and S p .
It can be seen from Table 2 S p and S p are responsible of the positive and negative
path of the current and for that they are considered known variables that depend on
the positive and negative cycle of the voltage reference.S p is expressed according to
vor e f as follows

1 vor e f ≥ 0
Sp = (3)
0 vor e f ≤ 0
4 Control Design of PUC5
Figure 4 illustrate the proposed closed-loop control of the PUC5 inverter. The control
strategy consists of two sensors and two linearizing PI regulator for the capacitor
voltage and load current. The two controllers are designed to achieve the desired
current and stabilize the capacitor voltage at half the DC Source. The average model
is given as:
x2
ẋ1 = (u 2 − u 1 ) = w1
C
vo = E u p − u 1 + x1 (u 2 − u 1 ) = w2 (4)
where x1 = V2 , x2 = i o and

w1 = K p x1r e f + x1 + K i x1r e f + x1
w2 = vor e f (5)
We recall that our objective is to track a desired reference of current, using Eq. (2)
we define vor e f as:
vor e f = Ri o + Lw3 (6)
Fig. 4 Control strategy of PUC5

where w3 is a PI action of the form

x2r e f
w3 = + K p x2r e f + x2 + K i x2r e f + x2 (7)
dt
From Eq. (4) we define the input vector as follow:

x2 −1
u1 − xC2 w1
= C (8)
u2 x1 V1 − x1 w2 + V1 u p
In order to examine our proposed scheme, simulation has been performed in

MATLAB/SIMULINK environment. The parameters used in the simulation are listed
on the Table 3.
The PV system application for PUC5 converter in standalone operation is shown
in Fig. 5. The PV system consists of three series and three parallel PV modules of
type Sun Power SPR-305-WHT. The characteristics of the used PV array are shown
in Fig. 6. In order to evaluate the MPPT and the proposed control for the PUC5, three
irradiation steps change was made, begins with 1KW/m2 until t = 2 s were than it
to become equal to 300 W/m2 , another change occur at t = 4 s were the irradiation
becomes equal to 700 W/m2 as is shown in Fig. 7.
The PV voltage and current behavior are depicted in Fig. 8 and Fig. 9, respectively.
were the PV power is shown in Fig. 10, from those results we could say that the P&O
algorithm to obtain MPP has achieved its goal to extract the maximum power from
the PV panel for all the three irradiation scenarios.
Now in order to evaluate the performance of PUC5 a simulation was performed
for the proposed control that has been discussed in Sect. 4. Figure 11 shows the
voltages of the capacitors, it is clear that the capacitor voltage V2 maintains tracking
its reference value (V1/2) with fast response. The load voltage and current wave are
depicted in Figs. 12 and 13. Using the zoom effect it can be clearly seen that the
output voltage is constituted of five levels for every different DC bus voltage case.
Table 3 Parameters used in

Parameters Values
the simulation
Fundamental frequency 50 Hz
Load Inductance 1 mH
Load resistor 15
DC capacitance 100 µF
PWM switching frequency 10 kHz
Fig. 5 PUC5 Supplied by MPPT PV System for stand-alone application
The load current in Fig. 13 takes a sinusoidal waveform following perfectly its
reference i o = i dc sin(wt) with every change condition. Such results show high-
quality conversion with low harmonic contents, which proves the PUC5 can be a
challenging topology with high potentials in photovoltaic application uses.
6 Conclusion
In this work, we discuss a recently developed topology known as Packed U Cell

(PUC) converter. This has a high potential of use on renewable energy conver-
sion such as photovoltaic applications. In order to examine this topology a photo-
voltaic system based on PUC5 has been designed. The system consists of a boost
converter connecting the PV arrays to a single-phase 5-level PUC (PUC5) inverter
to supply an AC load. We proposed a control strategy for the PUC5 to provide high-
quality convergence with low harmonic contents harnessing the five levels voltage it
provides. Therefore, there is no need to use bulky output filters. Simulation results
have proved the effectiveness of this topology with the employed control strategy
making it suitable for photovoltaic applications.
Fig. 6 PV array characteristics of sun power SPR-305-WHT
Fig. 7 Irradiation levels

Fig. 8 PV voltage
Fig. 9 PV current
Fig. 10 PV power
Fig. 11 DC source and capacitor

Fig. 12 Evolution of the output voltage
Fig. 13 Evolution of the output current
References
1. Lai J-S, Peng FZ (1995) Multilevel converters-a new breed of power converters. In: IAS’95
conference record of the 1995 IEEE industry applications conference thirtieth IAS annual
meeting, vol 3, pp 2348–2356. IEEE
2. Al-Haddad K, Ounejjar Y, Gregoire L-A (2016) Multilevel electric power converter. May 3
2016, US Patent 9,331,599
3. Abu-Rub H, Malinowski M, Al-Haddad K (2014) Power electronics for renewable energy
systems, transportation and industrial applications. Wiley
4. Kouro S, Malinowski M, Gopakumar K, Pou J, Franquelo LG, Wu, Rodriguez J, Pe´rez MA,
Leon JI (2010) Recent advances and industrial applications of multilevel converters. IEEE Tran
Ind Electron 57(8): 2553–2580
5. Rodr´ıguez J, Bernet S, Wu B, Pontt JO, Kouro S (2007) Multi-level voltage-source-converter
topologies for industrial medium-voltage drives. IEEE Tran Ind Electron 54(6): 2930–2945
6. Benmiloud M, Benalia A, Defoort M, Djemai M (2016) On the limit cycle stabilization of a
Dc/Dc three-cell converter. Control Eng Pract 49:29–41
7. Benmansour K, Benalia A, Djemai M, de Leon J (2007) Hybrid control of a multicellular
converter. Nonlinear Anal Hybrid Syst 1(1):16–29
8. Rodriguez J, Lai J-S, Peng FZ (2002) Multilevel inverters: a survey of topologies, controls,
and applications. IEEE Trans Indus Electro 49(4):724–738
9. Rodriguez J, Franquelo LG, Kouro S, Leon JI, Portillo RC, Prats MAM, Perez MA
(2009) Multilevel converters: an enabling technology for high-power applications. Proc IEEE
97(11):1786–1817
10. Ounejjar Y, Al-Haddad K, Gregoire L-A (2011) Packed u cells multilevel converter topology:
theoretical study and experimental validation. IEEE Trans Ind Electron 58(4):1294–1306
11. Trabelsi M, Bayhan S, Ghazi KA, Abu-Rub H, Ben-Brahim L (2016) Finite-control-set model
predictive control for grid-connected packed- u-cells multilevel inverter. IEEE Trans Ind
Electron 63(11):7286–7295
12. Sheir A, Orabi M, Ahmed ME, Iqbal A, Youssef M (2014) A high efficiency single-phase
multilevel packed u cell inverter for photovoltaic applications. In: 2014 IEEE 36th International
Telecommunications Energy Conference (INTELEC), pp 1–6. IEEE
13. Vahedi H, Labbe´ P-A, Al-Haddad K (2016) Sensor-less five-level packed u-cell (puc5) inverter
operating in stand-alone and grid- connected modes. IEEE Trans Ind Inf 12(1): 361–370
14. Vahedi H, Al-Haddad K (2016) Puc5 inverter-a promising topology for single-phase and
three-phase applications. In: IECON 2016-42nd Annual Conference of the IEEE Industrial
Electronics Society, pp 6522–6527. IEEE
15. Abarzadeh M, Vahedi H, Al-Haddad K (2019) Fast sensor-less voltage balancing and capacitor
size reduction in puc5 converter using novel modulation method. IEEE Tran Ind Inf
16. Ahmed J, Salam Z (2015) An improved perturb and observe (p&o) maximum power point
tracking (mppt) algorithm for higher efficiency. Appl Energy 150:97–108
17. Oubbati B, Boutoubat M, Belkheiri M, Rabhi A (2018) Extremum seeking and p&o control
strategies for achieving the maximum power for a pv array. In: International Conference in
Artificial Intelligence in Renewable Energetic Systems, pp 233–241. Springer
A Novel Hybrid Photovoltaic/Thermal
Bi-Fluid (Air/Water) Solar Collector:
An Experimental Investigation
Mohamed El-Amine Slimani, Rabah Sellami, Mohammed Said,

and Amina Bouderbal
Abstract This work deals with the construction, and the experimentation of a Bi-
fluid (air/water) hybrid photovoltaic/thermal solar collector intended for the residen-
tial sector. The hybrid PV/T collector consists mainly of à 185Wp mono-crystalline
photovoltaic module, a copper radiator for the circulation of water, and a ribbed metal
absorber with fins. The copper radiator and the absorber were glued to the backside of
the PV module. This structure was mounted in an insulated box with reservation of a
rectangular duct for the air circulation between the absorber and the insulation panel
of the box. An experimental bench was set up to evaluate the energy performance of
the designed collector under real operating conditions. The obtained results showed
that the proposed configuration of the bi-fluid PV/T collector both allows the air to
be heated to temperatures appropriate for the intended application (space heating)
and to recover additional calories by a second fluid (water) at appropriate tempera-
tures (40–45 °C). Hot water can be used for a double purpose; a direct consumption
in a building, and to prolong the production of hot air by a simple recirculation of
the stored hot water in the radiator. The average of thermal, electrical, and overall
energy efficiencies evaluated for this configuration are very important with values of
56.96%, 13.6%, and 94.53%, respectively.
Keywords Photovoltaic cell · Hybrid solar collector · Bi-fluid · Energy

performance · Photovoltaic and thermal conversion
M. E.-A. Slimani (B)

Theoretical and Applied Fluid Mechanics Laboratory, Department of Energetic and Fluid
Mechanics, University of Science and Technology Houari Boumediene (USTHB), 16111 Bab
R. Sellami
Unité de Développement des Equipements Solaires (UDES), EPST Centre de Développement des
Energies Renouvelables (CDER), 42415 Tipaza, Algeria
M. Said · A. Bouderbal
Department of Energetic and Fluid Mechanics, Faculty of Physics, University of Science and
Technology Houari Boumediene (USTHB), 16111 Algiers, Algeria

https://doi.org/10.1007/978-981-15-6403-1_47
1 Introduction
The overall performance of a photovoltaic module is strongly influenced by the

solar radiation received on the plane of the latter, the ambient temperature, and the
wind speed. In fact, the operating temperature of the module is the most important
parameter which plays a key role in the photovoltaic conversion process and which
thus influences its electrical efficiency. The rise in the operating temperature of the
photovoltaic module is due to the part of the solar radiation absorbed by the cells
and converted into heat. This undesirable effect can be reduced or eliminated by
the introduction of a suitable heat extraction device using one or two heat transfer
fluids to maintain, at the same time, the electrical efficiency of the PV module at a
minimum reasonable level and recover from low-temperature heat. This combination
of thermal and photovoltaic effects in the same device was introduced for the first
time by Wolf, Kern and Russell in the 70s [1] of the last century giving rise to hybrid
collects or PV/T type collectors.
A hybrid photovoltaic/thermal (PV/T) collector is the combination of photovoltaic
module and solar thermal collector to produce both electricity and heat at the same
time. About 15–20% of the solar radiation received by the photovoltaic panels is
converted into electrical output, and the rest of the solar energy ends up heating the
PV cells [2].
Several kinds of research have been carried out in this field in order to study and
improve the performance of hybrid PV/T system [3–8]. Browne et al. [9] compared
the performance of a hybrid air collector to a thermal collector. They found that
the PV/T air collector can produce a significant amount of energy on an equivalent
surface. Geographical location and catchment area have a strong influence on the
productivity of the hybrid component.
The study of a PV/T water-based collector with glazed and unglazed to evaluate
its performance has been the subject of several studies [10–13]. The results showed
that glazed collectors show an increase in thermal efficiency and a reduction in
electrical performance. And amorphous cells are the most suitable for thermal appli-
cations because they are less sensitive to temperature variations. Sarhaddi et al. [14]
Contributed to the development of a simulation program for the parametric study of
a PV/T air system in order to improve its thermal and electrical efficiency. Elsafi and
Gandhidasan [15] conducted a comparative study of a dual-pass PV/T air collector,
with and without concentrator and fins. They found the following points; perfor-
mance has improved by using fins, aluminum fins are more efficient than copper or
nickel, and rectangular fins are more efficient. Su et al. [16] developed a dynamic
model for a PV/T water system using the thermosiphon effect. The results showed
that the connection mode of the PV/T system has a significant effect on the energy
outputs. Hybrid PV/T collectors can be used in several sectors; especially in building
and agricultural activities [8, 17–21].
In the previous literature, several investigations have been carried out on the PV/T
sensors for a single heat transfer fluid (either water or air).
A Novel Hybrid Photovoltaic/Thermal Bi-Fluid … 699
The aims of this work are the realization and the study of a hybrid photo-
voltaic/thermal bi-fluid solar collector intended for the residential sector. The main
objective of this work is to offer users a hybrid collector configuration that will
achieve high overall efficiency (Bi-fluid) and extend its usage time by integrating
a heat storage device. For this purpose, this paper is comprised of several sections:
system description; construction and experimental work; energy performance equa-
tions; experimental results and discussions; and finally a conclusion as a summary
part of the work.
2 Structure and Description
The hybrid PV/T collector presented for this study based on a monocrystalline
photovoltaic module (JT-185 M). The PV/T solar collector composed mainly of the
following components; a cover glass, PV cells immersed in a layer of EVA polymer,
a protective layer of Tedlar below the PV cells, a metal plate below the Tedlar sheet,
copper tubes bonded to the back of the module (on the top metal plate), another
metal plate placed on the thermal insulation of the hybrid collector, and Baffles that
are stuck between the two metal plates. A thermal insulator to minimize heat losses
on the side and rear sides of the hybrid solar collector. The air circulates in the sub-
channels formed by the baffles to allow the heat transfer fluid to convey the thermal
energy produced. The experimental prototype can be provided with an assisted PV
ventilation system. Figure 1 represents a prototype of the studied hybrid collector
and schematic design of its component.
During the operation of the hybrid collector, the solar energy reaches the outer
surface of the collector in the form of solar radiation. A portion of this radiation is
converted into electrical energy and another part into thermal energy so that the heat
produced by the cells is evacuated through the coolant flowing between the fins and
the two metal plates (upper and lower). The hybrid air collector electrical and design
parameters used in this document are given in Table 1.
3 Construction and Experimentation
A. Experimental design
As presented in Fig. 1, the experimental designed system is mainly composed of an

unglazed PV/T collector provided with fins and tube radiator. The tow fluids follow
with forced circulation. The air in PV/T collector is ventilated by a constant airflow
assured by a DC electric alimentation instrument and the water circulate through the
radiator by using AC pump PV/T collector has been fixed on a properly insulated
aluminum housing using polyurethane foam. The air ventilation system consists of
a brushless type DC fan that can be directly coupled to small PV modules.
(a) (b)
PV Cells
Copper tube
Fin
(c)
Fig. 1 Designed hybrid PV/T collector: a front view; b back view; c schematic section with fins
and copper tubes
The realization of the PV/T solar collector took place in several stages taking into
account the dimensions and the arrangement of the PV cells in the used photovoltaic
module. It is, in fact, a module of 72 PV cells arranged in six parallel rows of twelve
cells in series. For this purpose, it was decided to make a radiator consisting of 12
tubes in parallel so that each row of cells is cooled by two tubes as shown in Fig. 2.
Once the radiator was completed, further work was started as follows:
Table 1 The main design

Parameter Value
parameter and characteristics
of the hybrid PV/T collector PV module type Monocrystalline silicon
Cells number, N 12 × 6
Short-circuit current, Isc,r e f 5.76 A
Open-circuit voltage, Voc,r e f 43.2 V
Current at MPP, Imp,r e f 5.14 A
Voltage at MPP, Vmp,r e f 36 V
Maximum power, Pm,r e f 185 W
Electrical conversion efficiency, ηr e f 14.4%
PV/T geometry
Parameter Characters Values
Length L 1.60 (m)
Width l 0.75 (m)
Thickness of fluid vein e 0.04 (m)
Fins geometry
Number of fins N 6
Length L f in 1.60 (m)
Width h 0.04 (m)
Thickness e f in 0.002 (m)
Radiator geometry
Number of tubes N 12
Tubes length L tube 1.62 (m)
Outside diameter Do 0.012 (m)
Tubes length Di 0.01 (m)
• Fixing the copper radiator on the back of the PV module by using a polyester resin
reinforced with aluminum filings. The filings are used for a dual purpose which
is to ensure both the reinforcement of the resin and to ensure that good thermal
contact is done between the tubes and the surface of the PV module.
• Making and bonding the fins on the radiator tubes and on the back of the PV
module using an epoxy type resin. It is important to note that these fins have a
dual role to play. They make it possible to extract both the heat transmitted by the
copper tubes and the rear surface of the PV cells. This heat is then transferred to
the air flowing through the PV/T collector.
• Characterization by IR thermography of the reliability of the method of fixing
the radiator and the fins on the rear part of the PV module. For this purpose,
IR images have been taken before and after the passage of cold water inside the
radiator tubes as shown in Fig. 3.
• Installation of the fan at the end of the collector isolation box to allow control of
the airflow circulating in forced convection. In order to guarantee a homogeneous
Fig. 2 The different stages of construction and assembly of the PV/T collector: a radiator design;
b fixation of the radiator on the back of the PV module; c bonding the fins on the radiator tubes and
on the back of the PV module
distribution of air in the air circulation channel, a flow control grille (perforated
grid with different opening diameters) is installed at the outlet of the channel.
• Manufacture of a rectangular shaped tank for water storage. This rectangular
shape of the tank was chosen for purely aesthetic reasons adapted to the realized
system.
• Installation of the water circulation pump, the flow control device, and the thermal
insulation of the piping. As well as the installation of the various measuring
instruments.
B. Instruments of measure
In order to allow us to evaluate the thermal performance of the collector realized

under real operating conditions, a bench of measurements has been set up for the
simultaneous measurement of the different parameters such as the temperature of the
T ambiante (0C)
solar Radiation (W/m 2 )
9
1000
20 Wind Speed (m/s)
8
19 800 7
6
18 600
5
17 4
400
3
16
200 2
15 1
0
0
14
0
0
:0
:0
:0
:0
:0
:0
:0
:0
06
08
10
12
14
16
18
20
Time (h)
Fig. 3 Evolution of solar irradiance, ambient air temperature, and wind speed during the test day
(May. 14, 2019)
fluids (air and water), the temperature of the layers and components of the collector,
solar illumination and air velocity. For this purpose, many measuring instruments
have been used (type K thermocouples, hot wire anemometer, pyranometer, data
acquisition) to conduct simultaneous measurements of all necessary parameters with
a time step of five minutes.
It is important to note that meteorological parameters were collected from the
meteorological station installed at the UDES site.
4 Energy Performance
The thermal power produced from the hybrid PV/T bi-fluid collector by its tow fluids
(air and water) can be calculated using the following equation.

Q th = ṁ a × C p,a Ta,o − Ta,i + ṁ w × C p,w Tw,o − Tw,i (1)
The electrical power generated by a PV or PV/T collector is calculated using the

following correlation.

G
Q ele = G · A · ηref 1 − βT · TC − TC,ref + γ · ln (2)
G ref
where G and G ref 2 which are the illumination and the reference illumination
G ref = 1000 W /m ;
Tc and Tc,ref which are the temperature and the reference temperature of the cells
Tc,ref = 25 ◦ C ;
β T is the temperature coefficient, and γ is the solar radiation coefficient.
The following relations give the thermal and the electrical efficiencies, respec-
tively, this later is calculated from the correlation proposed by Evans [22].
Q th
ηth = (3)
G·S

Q ele G
ηele = = ηref 1 − βT · TC − TC,ref + γ · ln (4)
G·S G ref
Another rather important parameter that shows the overall energy efficiency of
the PV/T collector is given as the sum
of the thermal yield (ηth ) and the equivalent of
the electrical efficiency in thermal ηele,th using the thermal conversion factor (C f ).
ηele
η O = ηth + (5)
Cf
The experimentation was made under the weather conditions of UDES (Unité de
Développement des Equipements Solaires) located at Bousmail, Tipaza, for the test
day of 14 May 2019.
Table 2 gives the imposed operating conditions used for the experimentation step,
including water and air flows and the volume of the storage tank.
The Meteorological data for the test day of 14 May 2019 are given by Fig. 3. The
weather data include solar radiation, ambient temperature, and wind speed.
The evolution between the different key temperatures of PV/T collector that are;
the PV cells temperature, Inlet and outlet water temperatures, inlet and outlet air
temperatures. The evolution shows compliance and logic in the temperature distri-
bution in different PV/T collector elements (Fig. 4). All temperatures of the collector
being higher than the ambient temperature.
The temperature of the solar cells shows the highest temperature in the hybrid
collector; it reaches a value of about 47 °C at 13h30. This is explained by the fact
that the photovoltaic cell is a generator converting the received solar radiation into
electricity and heat at the same time. The temperatures of the collector elements
increase as a function of time until solar noon, then decrease. The temperature of
the fluid at the output of the collector passes through a maximum of 42.5 °C at
14h00 for water, and 32.5 °C for air. This range of temperature is very acceptable
and comfortable for residential and building applications.
Table 2 Gives the imposed

Water flow (l/min) Airflow (m3 /s) Storage volume (l)
operating conditions
1 0.003535 90
Fig. 4 Temperature T O-w

evolution during the test day
45
T Cell
T i-w
40
T amb
Temperature (oC)
T O-a
35
30
25
20
15
6
:4
:1
:3
:0
:2
:4
:1
:3
04
07
09
12
14
16
19
21
Time (h)
From Fig. 4 it can be noted that for this configuration, the temperature difference
between the air inlet and air outlet is important and appreciable. The temperature
difference range would be as high as 12 °C in relatively clear sky.
Figure 5 shows the evolutions of the electrical and thermal powers for the
considered day test.
The graphs in Fig. 5 show, respectively, the evolution of the overall thermal and
electric powers for the test day of 14/05/2019 as a function of time. We observe
that these energies grow until reaching maximum values at solar noon. Note that the
maximum value of the thermal power reaches 736.91 W and that of the electric power
Overall Thermal Power (W)
800
600
400
200
-200
06:00 08:00 10:00 12:00 14:00 16:00 18:00 20:00
140
Electrical Power (W)
120
100
80
60
40
20
0
06:00 08:00 10:00 12:00 14:00 16:00 18:00 20:00
Time (h)
Fig. 5 Evolution of electrical and thermal powers during the test day
133.09 W. It is noted that the evolution of the electric power is substantially related
to the meteorological conditions. The thermal power as for it; is naturally influenced
by the weather conditions but also parametric. The evolution of the thermal power is
related and influenced by the storage volume because by increasing it, the temperature
difference of the coolant (water) increases and consequently the thermal power.
Figures 6, 7, and 8 illustrate the evolution of electrical, thermal, and overall energy
efficiencies, respectively, during the test day.
At the beginning of the day, the thermal efficiency is zero because of the absence
of solar radiation. Then, it increases to relatively stable values between 60 and 70%
and then decreases rapidly towards the end of the day until it is canceled. Peaks can
be observed due to disturbances of solar radiation. An average value of 56.96% was
noted.
The electrical efficiency at the beginning of the day recorded its maximum with
a value of 15.03%. This efficiency decreases when the solar illumination increases
until solar noon with a minimum value of 13%, and this is explained by the increase
in the temperature of the module.
Fig. 6 Evolution of the 100

thermal efficiency during the
Thermal Effeciency, η th (%)
test day 80
60
40
20
0
06:00 08:00 10:00 12:00 14:00 16:00 18:00
Time (h)
Fig. 7 Evolution of the

electrical efficiency during 15,0
Electrical Effeciency, ηele(%)
the test day

14,5
14,0
13,5
13,0
12,5
06:00 08:00 10:00 12:00 14:00 16:00 18:00 20:00
Time (h)
Overall Energy Effeciency ηg (%)

overall energy efficiency
120
during the test day
100
80
60
40
20
0
06:00 08:00 10:00 12:00 14:00 16:00 18:00
Time
It is observed that the overall energy efficiency increases to its maximum value
(about 110%). The yield remains almost stable, then they start to decrease rapidly
towards the end of the day. The peaks are due to weather disturbances (cloudy
passage). The average value of the overall efficiency calculated is 94.53%.
6 Conclusion
The temperatures and the energy performance of an unglazed hybrid PV/T bi-fluid
(air/water) solar collector provided with fins, radiators, and storage tank were investi-
gated and analyzed. The developed configuration of the hybrid collector is designed
to be suitable for winter cold periods to providing both hot water and air condi-
tioning in a building. An experimental bench was set up and the energy performance
of the solar device was evaluated for typical test day under a mild Mediterranean
climate. The results obtained showed that the proposed configuration of the bi-fluid
PV/T collector both allows the air to be heated to temperatures appropriate for the
intended application (space heating or drying) and to recover additional calories by a
second fluid (water) at appropriate temperatures (40–45 °C). Hot water can be used
for a double purpose; a direct consumption in a building (domestic hot water), and
to prolong the production of hot air by a simple recirculation of the stored hot water
in the radiator and thanks to the fins which promote the thermal exchange between
water and the air. The average of thermal, electrical, and overall energy efficiencies
evaluated for this configuration are very appreciable with values of 56.96%, 13.6%,
and 94.53%, respectively.
References
1. Kern MC Jr, Russell EC (1978) Combined photovoltaic and thermal hybrid collector systems.
In: The 13th IEEE photovoltaic specialists, 5–8 June 1978, vol. 79, pp COO-4577-3; CONF-
780619-24
2. Slimani MEA, Amirat M, Kurucz I, Bahria S, Hamidat A, Chaouch WB (2017) A detailed

thermal-electrical model of three photovoltaic/thermal (PV/T) hybrid air collectors and photo-
voltaic (PV) module: Comparative study under Algiers climatic conditions. Energy Convers
Manag 133:458–476
performance in a hybrid Pv/T collector with double passage of air. Int J Energy Clean Environ
16(1–4):235–245
4. Srinivas M, Jayaraj S (2013) Performance study of a double pass, hybrid-type solar air heater
with slats. Int J Energy Eng 3(4):112
5. Slimani MEA, Amirat M, Bahria S (2015) Analysis of thermal and electrical performance of a
solar PV/T air collector: Energetic study for two configurations. In: 3rd international conference
on control, engineering and information technology, CEIT 2015, pp 1–6
6. Othman MY, Ibrahim A, Jin GL, Ruslan MH, Sopian K (2013) Photovoltaic-thermal (PV/T)
technology – the future energy technology. Renew. Energy 49:171–174
7. Boumaaraf B, Touafek K, Ait-cheikh MS, Slimani MEA (2020) Comparison of electrical
and thermal performance evaluation of a classical PV generator and a water glazed hybrid
photovoltaic–thermal collector. Math Comput Simul 167:176–193
8. Gholampour M, Ameri M (2016) Energy and exergy analyses of Photovoltaic/Thermal flat
transpired collectors: experimental and theoretical study. Appl Energy 164:837–856
9. Browne MC, Norton B, McCormack SJ (2016) Heat retention of a photovoltaic/thermal
collector with PCM. Sol Energy 133:533–548
10. Chow TT, Pei G, Fong KF, Lin Z, Chan ALS, Ji J (2009) Energy and exergy analysis of
photovoltaic–thermal collector with and without glass cover. Appl Energy 86(3):310–316
11. Fujisawa T, Tani T (1997) Annual exergy evaluation on photovoltaic-thermal hybrid collector.
Sol Energy Mater Sol Cells 47(1–4):135–148
12. Fraisse G, Ménézo C, Johannes K (2007) Energy performance of water hybrid PV/T collectors
applied to combisystems of Direct Solar Floor type. Sol Energy 81(11):1426–1438
13. Boumaaraf B, Boumaaraf H, Slimani ME-A, Tchoketch_Kebir S, Ait-cheikh MS, Touafek K
(2020) Performance evaluation of a locally modified PV module to a PV/T solar collector under
climatic conditions of semi-arid region. Math Comput Simul 167:135–154
14. Sarhaddi F, Farahat S, Ajam H, Behzadmehr A, Mahdavi Adeli M (2010) An improved
thermal and electrical model for a solar photovoltaic thermal (PV/T) air collector. Appl Energy
87(7):2328–2339
15. Elsafi AM, Gandhidasan P (2015) Comparative study of double-pass flat and compound
parabolic concentrated photovoltaic–thermal systems with and without fins. Energy Convers
Manag 98:59–68
16. Su D, Jia Y, Huang X, Alva G, Tang Y, Fang G (2016) Dynamic performance analysis of
photovoltaic–thermal solar collector with dual channels for different fluids. Energy Convers
Manag 120:13–24
17. Slimani ME-A (2017) Etude d’un séchoir solaire agricole muni d’un capteur solaire de type”
PV-THERM”-réalisation d’un prototype et caractérisation. Universite Des Sciences Et De La
Technologie Houari Boumediene
18. Othman MY, Hamid SA, Tabook MAS, Sopian K, Roslan MH, Ibarahim Z (2016) Performance
analysis of PV/T Combi with water and air heating system: an experimental study. Renew
Energy 86:716–722
19. Xiang B et al (2018) A novel hybrid energy system combined with solar-road and soil-
regenerator: Sensitivity analysis and optimization. Renew Energy 129:419–430
20. Sathe TM, Dhoble AS (2017) A review on recent advancements in photovoltaic thermal
techniques. Renew Sustain Energy Rev 76:645–672
21. Slimani ME-A, Sellami R, Mahrane A, Amirat M (2019) Study of hybrid photovoltaic/thermal
collector provided with finned metal plates: a numerical investigation under real operating
conditions. In: 2019 international conference on advanced electrical engineering, ICAEE’19,
pp 1–6
22. Sellami R, Amirat M, Mahrane A, Slimani ME-A, Arbane A, Chekrouni R (2019) Experimental
and numerical study of a PV/Thermal collector equipped with a PV-assisted air circulation
system: Configuration suitable for building integration. Energy Build 190:216–234
A DC/DC Buck Converter Voltage
Regulation Using an Adaptive Fuzzy Fast
Terminal Synergetic Control
Noureddine Hamouda and Badreddine Babes
Abstract In this paper, an adaptive fuzzy fast terminal synergetic voltage regu-
lation for DC/DC buck converter is designed based on recently developed syner-
getic theory and a terminal attractor method. The advantages of presented syner-
getic control include the characteristics of finite time convergence, insensitive to
parameters variation and chattering free phenomena. Rendering the design more
robust, fuzzy logic systems are used to approximate the unknown parameters in
the proposed controller without calling upon usual model linearization and simpli-
fications. Taking the DC/DC buck converter in continuous conduction mode as an
example, the algorithm of proposed synergetic control is analyzed in detail. All the
simulation results demonstrate the effectiveness and the high dynamic capability of
the proposed AF-FTSC control technique over the FTSC strategy.
Keywords Component · Synergetic control · Fuzzy logic system · Terminal

technique · Finite time convergence · DC/DC buck converter
1 Introduction
Many nonlinear control approaches such as variable structure technique has attracted
community attention in the last decades. Due to its inherent robust features, this
control technique has been extensively used in many nonlinear control systems
such as voltage regulation of DC/DC buck converter. Problems of variable structure
control are the chattering phenomenon, steady state errors caused by high frequency
switching actions, and the fact that it needs fairly high bandwidths for the controller,
which makes digital control realization unfeasible [1]. On the others hand, this article
introduces the fast terminal synergetic control (FTSC) method which is utilized on
N. Hamouda · B. Babes (B)

Research Center in Industrial Technologies (CRTI), Algiers, Algeria
N. Hamouda

https://doi.org/10.1007/978-981-15-6403-1_48
712 N. Hamouda and B. Babes
the DC/DC buck converter voltage control scheme. This approach has numerous
advantages such as: it’s excellent suitable to digital control, it functions at fixed
switching frequency which reduces the burden of filtering design, and it shows low
chattering as well as the inherent robustness features of sliding mode control [2].
However, the robustness of synergetic control structure is not assured pending tran-
sient action when system states trajectories have not achieved the attractor. Therefore,
a new technique to avoid this difficulty intrinsic to fast terminal synergetic control,
an adaptive fuzzy fast terminal synergetic control (AF-FTSC) structure is suggested
thus reducing the reaching phase.
The rest of the manuscript is structured as follows. In Sect. 2, the fundamental
concepts related to synergetic control theory and fuzzy logic systems are reviewed
briefly. In Sect. 3, the basic model of the DC/DC buck converter in continuous
conduction mode (CCM) is introduced. In Sect. 4, the adaptive fuzzy fast terminal
synergetic control (AF-FTSC) is developed for the DC/DC buck converter system. In
Sect. 5, simulation is performed on a DC/DC buck converter to verify the effectiveness
and the applicability of the proposed AF-FTSC control technique over the FTSC
strategy. In Sect. 6, the conclusions are given.
2 Preliminaries
2.1 The Basis of Synergetic Control (SC) Strategy
In this section, we briefly introduce the fundamentals of synergetic control synthesis

for a nonlinear dynamic system expressed as
ẋ = f (x, u, t) (1)
Basically, the synergetic control design structure begins with the definition of a
macro-variable σ as a function of the state variables [3]
σ = σ (x, t) (2)
The control purpose is to oblige the system to work on a pre-chosen manifold = 0.

Design a synergistic controller that would force the system states to exponentially
approach the preferred manifold with an evolution constraint which can be declared as
τ σ̇ + σ = 0 (3)
where τ is a positive constant which imposes the designer select speed convergence
to the specified attractor.
A DC/DC Buck Converter Voltage … 713
Taking into account the chain of differentiation which is given by
dσ
σ̇ = · ẋ (4)
dx
Substituting (1) and (3) into (4) gives (5)
dσ
τ f (x, u, t) + σ = 0 (5)
dx
Resolving (5) for u provides the synergetic control law as
u = g(x, σ (x, t), τ, t) (6)
The obtained control law (6) forces state trajectories to satisfy (3). Appropriate
choice of the macro-variable (2) and judicious manifolds guarantees the performance
and the pursued stability [4].
2.2 The Basis of Fuzzy Logic Systems
It is known that fuzzy logic systems (FLSs) are universal estimators and have excep-
tional functions in controller design and identification. A type-1 FLS involves four
main parts: fuzzifier, rule base, inference engine and defuzzifier. The main feature of
the FLS is the aptitude to express the human-like thinking and deal with the experi-
ence effectively through the conditional fuzzy If-Then rules, which can be formulated
as
Rl : If x1 is F1l , x2 is F2l , . . . , xn is Fnl

Then y1 is G l1 , y2 is G l2 , . . . , ym is G lm (7)
where l = 1, 2,…, L, with L indicating the total number of rules, x = [x1 , x2 , . . . , xn ]T

and y = [yl , y2 , . . . , ym ]T represent the input and output parameter vectors of fuzzy
logic system, respectively. F il and Gjl are the linguistic variables of the fuzzy sets,
described by their membership function vectors μ Fil (xi ) and μG lj (y j ). Using the
singleton fuzzifier, product inference engine, and centre-average defuzzifier, the
inferred output of the FLS can be expressed as [5]
L n
l=1 μ
i=1 Fi l (x i ) y
l
j
y j (x) = = θ jT ξ(x), j = 1, 2, . . . , m. (8)
L n
and:
n
i=1 μ Fil (xi )
ξi (x) = L (9)
n
i=1 μ Fil (xi )
l=1
where y lj is the point in V j at which μG lj (y j ) achieves its maximum value (assuming

μG lj (y j ) = 1).
T
θ j = y 1j , y 2j , . . . , y Lj ∈ R L is called the adjustable parameter vector and
1 T
ξ(x) = ξ (x), ξ 2 (x), . . . , ξ L (x) ∈ R L is the vector of fuzzy primary function or
the antecedent function vector.
T
ξ(x) = ξ 1 (x), ξ 2 (x), . . . , ξ L (x) ∈ R L (10)
If y lj are preferred as the free parameter, the above fuzzy system (7) becomes an
adaptive fuzzy system. Then, it can be rewritten as
y(x) = θ T ξ(x) (11)
where θ is (L × m) matrix, θ j and indicates the (L × 1) j-th column of the matrix θ .
3 Modeling of the DC/DC Buck Converter
The corresponding circuit of the DC/DC buck converter is displayed in Fig. 1. The
positive constants R, L, C, V in and V o are respectively load resistor, self-inductor,
capacitor, input voltage and output voltage.
The dynamic equations of a DC/DC buck converter in the continuous conduction
mode (CCM) are expressed as in [6]
⎧ di 1 Vin
⎪
⎨
L
= − Vo + u
dt L L (12)
⎪
⎩ d Vo = 1 i − 1 V
L o
dt C RC
Fig. 1 The DC/DC buck iL

converter
Q L
+
Vin _ D C R Vo
u
where iL denotes the inductor current and u is the duty ratio of the DC/DC buck
converter. Choosing the output voltage and its derivative as system state variables,
we get

x1 = Vo
(13)
x2 = x1 = d Vo
dt
The state space average model of the DC/DC buck converter can be rewritten as
fellow

ẋ1 = x2
(14)
ẋ2 = − LC
x1
− RC
x2
+ LC
Vin
u
4 Adaptive Fuzzy Fast Terminal Synergetic Control

(AF-FTSC) Design
If we suppose the reference tracking voltage to be V ref , then the tracking error and
its derivative are as follows
e = x1 − Vr e f (15)
ė = x2 − V̇r e f (16)
To allow the application of AF-FTSC algorithm to a DC/DC buck converter, a

special variable called the nonlinear macro-variable function needs to be introduced
as [7]
σ = ė + αe + βe p/q (17)
where α and β are the macro-variable constants, p and q are positive odd numbers, that
assure the following condition 1 < p/q < 2. A constraint imposing desired dynamics
to the macro-variable is selected as
τ σ̇ + σ = 0 (18)
where τ is a positive constant which imposes the designer decide speed conver-
gence to the preferred attractor. When the macro-variable reaches σ = 0, the system
dynamics (17) can be expressed as follows
ė + αe + βe p/q = 0 (19)
On integrating (19), and through setting suitable parameters α, β, p and q, the

convergence time of the system from any given initial condition e(0) = 0 to the
attractor σ = 0 is finite-time t r ≥ 0 and is given by

p α|e(0)|1−( p/q) + β
tr = ln (20)
α( p − q) β
The time derivative of σ can be rewritten as

d p/q
σ̇ = ë + α ė + β e (21)
dt
Combining (18) and (21), leads to (22)
p ( p/q)−1 1
ë + α ė + β e ė = − σ (22)
q τ
For the controller design we set the following simplification
Vin x1 x2
g(x) = , and f (x) = − −
LC LC RC
Solving for the terminal synergetic control algorithm uFTSC , guides to (23)

1 1 p
u F T SC = V̈r e f − f (x) + σ − α ė − β e( p/q)−1 ė (23)
g(x) τ q
Generally, the nonlinear system functions ƒ(x) and g(x) are difficult to determine
precisely. So, a fuzzy logic system using the universal approximation theorem is
used to estimate the nonlinear and unknown functions f (x) and g(x) in (23). Let
fˆ( x|θ f ) and ĝ( x|θg ) be the fuzzy approximates of the vector functions ƒ(x) and
g(x), respectively. Then, the controller (23) is improved as follows

1 ˆ 1 p ( p/q)−1
u F T SC = V̈r e f − f ( x|θ f ) − σ − α ė − β e ė (24)
ĝ( x|θg ) τ q
where
fˆ( x|θ f ) = θ Tf ξ f (x) (25)
ĝ( x|θg ) = θgT ξg (x) (26)
where ξ f (x) and ξg (x) are the fuzzy basis functions vector, θ f and θg are the
parameters vector of fuzzy system.
Define the optimal estimation parameters of fuzzy system as

ˆ
θ ∗f = arg min sup f x|θ f − f (x) (27)
θf ∈ f x∈R n

θg∗
= arg min sup ĝ x|θg − g(x) (28)
θg ∈ g x∈R n
where Ω f and Ω g represent constraint sets for θ f and θ g , respectively.

Defining the minimum estimation error as

ε = fˆ x|θ ∗f − f (x) + ĝ x|θg∗ − g(x) u (29)
Substituting (29) into (18), the dynamic of macro-variable is obtained as follows

1
σ̇ = fˆ x|θ ∗f − fˆ x|θ f + ĝ x|θg∗ − ĝ x|θg u − σ + ε (30)
τ
Leading after some simple manipulations to
∗T 1
σ̇ = θ ∗T
f − θ f ξ(x) + θg − θg ξ(x)u − σ + ε
T T
(31)
τ
Or in a more compact form
1
σ̇ = θ̃ Tf ξ(x) + θ̃gT ξ(x)u − σ +ε (32)
τ
where θ̃ f = θ ∗f − θ f and θ̃g = θg∗ − θg

Define the Lyapunov function candidate V as

1 2 1 1
V = σ + θ̃ Tf θ̃ f + θ̃gT θ̃g (33)
2 r1 r2
where r 1 and r 2 are positive constants.

The time derivative of V can be written as
1 T˙
θ̃ θ̃ f + θ̃gT θ̃˙g
1
V̇ = σ σ̇ + (34)
r1 f r2
Substituting (32) into (34), we can obtain (35)

V̇ = σ θ̃ f ξ(x) + θ̃g ξ(x)u − σ + ε + θ̃ Tf θ̃˙ f + θ̃gT θ̃˙g
T T 1 1 1
(35)
τ r1 r2
Fig. 2 Block diagram for proposed AF-FTSC controller
In which we make use of θ̃˙ f = θ̇ f and θ̃˙g = θ̇g to obtain (36)
1 1 1
V̇ = − |σ | + σ ε + θ̃ Tf r1 σ ξ(x) + θ̇ f + θ̃gT r2 σ ξ(x)u + θ̇g (36)
τs r1 r2
To make V̇ ≤ 0, we choose the adaptive laws as
θ̇ f = −r1 σ ξ(x) (37)
θ̇g = −r2 σ ξ(x)u (38)
Then (36) becomes
1 1
V̇ ≤ − |σ | + σ ε ≤ − |σ | ≤ 0 (39)
τ τ
The block diagram of the designed AF-FTSC controller is presented in Fig. 2.
In this section, some computer simulations are conducted using MATLAB/Simulink

software to show the control performance of the designed AF-FTSC control technique
over the FTSC strategy [7]. The DC-DC buck converter utilized in simulation study
has the following component values: L = 5 mH, C = 1100 µF, R = 20–5 , V in =
50 V. The main parameters used in designing controllers are: α = 100, β = 150, p/q
= 7/9, τ = 0.005, r 1 = 2.106 and r 2 = 3.5.105 .
40
Vref
Output voltage (V)

30
20
FTSC
AF-FTSC
10
0
0 0.01 0.02 0.03 0.04 0.05 0.06
Time (s)
(a)
8
Inductor current (A)
AF-FTSC
6
FTSC
4
iL
2
0
0 0.01 0.02 0.03 0.04 0.05 0.06
Time (s)
(b)
Fig. 3 Simulation results of for a nominal start-up response
40
Decreasing of the load
Output voltage (V)
30
20
Increasing of the load FTSC
AF-FTSC
10
0
0 0.01 0.02 0.03 0.04 0.05 0.06
Time (s)
Fig. 4 Simulation results for a tep-load response
The start-up transient response of V o and iL during the changes in the desired
output voltage from 0 to 30 V is shown in Fig. 3a, b, respectively. It is observed that
the proposed AF-FTSC algorithm has better performance in terms of the settling
time relatively to the FTSC controller.
40
FTSC
AF-FTSC
Output voltage (V)

30
Vref
20
10
0
0 0.01 0.02 0.03 0.04 0.05 0.06
Time (s)
(a)
40
FTSC
AF-FTSC
Output voltage (V)
30 Vref
20
10
0
0 0.01 0.02 0.03 0.04 0.05 0.06
Time (s)
(b)
40
Output voltage (V)
30
20
Vref
FTSC
10
AF-FTSC
0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
Time (s)
(c)
Fig. 5 Simulation results of output voltage for a reference trajectory tracking, b sinusoidal-wave
tracking, c triangular-wave tracking
Figure 4 shows the simulation responses of V o when the load resistance steps
from nominal value of 20 to 5 and vice versa. It is evident from Fig. 4 that under
load resistance changes the AF-FTSC algorithm can respond faster than the FTSC
algorithm. Also, due to the independence of AF-FTSC control law to load resistance
R, the proposed controller shows a robust behavior against drastic changes of the
load resistance.
Figures 5a–c depicts the output voltages tracking performance of the proposed
AF-FTSC and FTSC controllers.
It is intuitive to see that under reference changes the proposed AF-FTSC controller
can respond faster than the FTSC controller. Also, the above simulation comparisons
effectively confirm the advantage of the proposed AF-FTSC algorithm over FTSC
algorithm.
6 Conclusions
In this manuscript, an AF-FTSC control law has been developed and implemented to
deal with the voltage tracking problems of the DC/DC buck converter. The proposed
control algorithm has been designed by using techniques of the synergetic theory of
control and the terminal attractor technique, which assures finite time convergence
of the output voltage to the desired voltage.
The unknown parameters in the proposed controller can be approximated precisely
by using the fuzzy system with suitable adaptive laws. The stability of the system is
also proved using direct Lyapunov theorem.
Simulation results demonstrate the robustness and effectiveness of the proposed
AF-FTSC control technique over the FTSC strategy.
References
1. Son YD, Heo TW, Santi E, Monti A (2004) Synergetic control approach for induction motor
speed control. In: The 30th annual conference of the IEEE industrial electronics society, 2–6
November 2004, Busan, Korea
2. KolesnikovA, Veselov G, Kolesnikov A, Monti A, Ponci F, Santi E, Dougal R (2002) Synergetic
synthesis of DC-DC boost converter controller: theory and experimental analysis. In: IEEE
APEC 2002, 2 DilZ, pp 409–415
3. Santi E, Monti A, Li D, Proddutur K, Dougal RA (2003) Synergetic control for DC-DC boost
converter: implementation options. IEEE Trans Ind Appl 39(6):1803–1813
4. Bouchama Z, Essounbouli N et al (2016) Reaching phase free adaptive fuzzy synergetic power
system stabilizer. Electr Power Energy Syst 77:43–49
5. Yoo BK, Ham WC (2000) Adaptive control of robot manipulator using fuzzy compensator. IEEE
Trans Fuzzy Syst 8:186–199
6. Nettari Y, Kurt S (2018) Design of a new non-singular robust control using synergetic theory
for DC-DC buck converter. Electrica 18(2):292–299
7. Zerroug N, Harmas MN, Benaggoune S, Bouchama Z, Zehar K (2018) DSP-based imple-
mentation of fast terminal synergetic control for a DC–DC buck converter. J Frankl Inst
355:2329–2343
Study of Electrical Field Distribution
in the Insulation of High-Voltage Cables
Rouini Abdelghani, Kouzou Abdellah, and Larbi Messaouda
Abstract Insulations are the most important part of high voltage equipment such
as cables and machines. Therefore the study of the condition and mechanism of
failure of high voltage insulations is important. This paper studies the electrical
constraints within an XLPE insulated cable containing micro-cavities. The purpose
of this work is to determine by simulation the distortions caused by micro-cavities
on the electric field distribution in the insulating layer of cable. The calculation of
these constraints is performed by the resolution of LAPLACE equation using Finite
Difference Method (FDM) because of its simplicity in the symmetrical and two-
dimensional systems, which corresponds perfectly in the case of cables which can
be considered as cylindrical systems with axial symmetry.
Keywords High-voltage cables · LAPLACE equation · Finite difference method

(FDM) · Air microcavities
1 Introduction
The chemically cross linked polyethylene PRC is the most used material in the
industry of electric cables (high voltage). However, the crosslinking processes applied
to polyethylene to obtain the PRC all lead to the formation of microcavities within
the material [1, 2]. The sizes of these defects vary from 1 to 20 μm and they are
concentrated in the part located at three quarter of the thickness of the cable insulation
starting from the core [3]. The heterogeneity of the dielectric medium that presents
R. Abdelghani (B) · K. Abdellah

Electrical Engineering Department, Applied Automation and Industrial Diagnostic Laboratory,
Djelfa, Algeria
K. Abdellah
L. Messaouda
Department of Science and Technology, University of Batna, Batna, Algeria

https://doi.org/10.1007/978-981-15-6403-1_49
724 R. Abdelghani et al.
Table 1 Characteristic of the

Cable features Value for PRC
cable
Type 185 mm2 , Al, 30/50 kV
Ray of the core 7.65 mm
Thickness of the semi-layer 1.0 mm
conductor on the core
Thickness of the insulation 11.3 mm, PRC
Thickness of the semi-layer 1.2 mm
conductive on the insulation
Metal screen made of copper 0.8 mm
wire
PVC protective sheath 5.3 mm
Dielectric loss factor 4.10-3
Electrical conductivity 10–12 ( cm) -1
Thermal conductivity 0.286 W/m °C
Volume heat 2.08 J/cm3 °C
the portion containing the microcavities leads to the distortion in the lines of electric
field in the vicinity of these defects. When the intensity of the electric field inside the
microcavities reaches the dielectric strength of the gas, a partial electric discharge
arises. The magnitude of these distortions depends on the permittivities of the two
environments, the shape of the microcavity and its position relative to the cable core
[4, 5]. In the industry, cable breakdown tests are necessary, however during their
manufacturing, it is difficult to avoid the penetration of foreign particles within the
insulation (defects), and we can hardly realize such case to study the influence of
these particles. The size of the fault, its nature (permittivity) and its position in the
cable insulation as well as the effects of the electromechanical pressure and the
temperature, considerably influence the behavior of the insulation of the cable and
thus limiting its lifetime. The purpose of this article is based on the study of the effect
of the presence of an air cavity, its position and size on the distribution of the electric
field, the electromechanical pressure and the dielectric losses in an insulator. PRC
type used in the insulation of medium voltage cables (18/30 kV) manufactured by
ENICAB-Biskra. This cable has the characteristics listed in Table 1.
2 Influence of Microcavities on the Distribution

of the Electric Field
The effects of a permittivity cavity ε2 on a host dielectric of ε1 permittivity are first

evaluated in terms of potential perturbation [5–10]. We will determine for the case
of a spherical cavity the expressions of the potential and the electric field in the
insulating medium and in the cavity. Assuming that the density of free charges ρ in
Study of Electrical Field Distribution … 725
the host dielectric is small enough to be neglected, the Poisson equation describing
the potential in both media is reduced to the Laplace equation such that:
∇2 V = 0 (1)
Given the cylindrical shape of the cable, the lines of the electric field are radial
and divergent, on the other hand the dimensions of the cavity are quite small in front
of the thickness of the insulation [11–18]. Given the symmetry of the system at turn
of the cable axis (Z), the Eq. (1) can be expressed by the following form:

∂ r2 ∂V 1 ∂ ∂V
+ sin θ =0 (2)
∂r ∂r sin θ ∂θ ∂θ

The general solution of (2) is of the form V (r, θ ) = f (r )g(θ ) With f (r ) An r n
n
and g(θ ) = Pn (cos θ ) or Pn is the Legendre polynomial. This solution is valid in
both environments. The expression of the potential for the case of a spherical cavity
is of the form:

Bi
Vi = Ai + 2 cos θ (3)
r
The index i takes the values 1 or 2 respectively corresponding to the insulator

and the cavity. The integration constants Ai , Bi are determined by the boundary
conditions. The expressions of the potential in the insulator and in the spherical
cavity are obtained as follows:

ε1 − ε2 r23
V1 = E 0 r + · 2 cos θ (4)
2 ε1 − ε2 r
3 ε1
V2 = E 0 r cos θ (5)
2 ε1 + ε2
∂ V2 3 ε1
E2 = = E0 (6)
∂x 2ε1 + ε2
or:
r 2 : Is the radius of the cavity.
r: Is the position of a given point in the insulation relative to the center of the
cavity.
It is noted that the potential V 2 inside the cavity depends on E 0 (rcosθ), Knowing
that:

Rt
V1 ln ri
E0 = = (7)
ri ln Rt
rc
where V 1 is the conductor voltage, Rt is the outer radius of the cable, r c is the radius
of the conductive core and r i is a given radial position.
It follows that the intensity of the proportional E 2 field at E 0 depends on the values
of the permittivities ε1 and ε2 of the two environments [2].
2.1 Approximation of the Electrical Potential at Each Node
Generally, when dealing with a problem with two environments, it is better to use the
finite difference method, especially if the two environments have a symmetry and a
fairly simple geometry. This method consists in replacing the continuous medium, in
which the equation to be studied is applicable at all points by a set of points to which
the discretized equation applies (Fig. 1) [5, 7–10]. For such reason this method has
been chosen to be used in this article.
The equation of the potential obtained by the series development of TAYLOR:
1 (2·ri +h r )

ri2 · h 2t · h r2 2 2 V (i, j + 1) + 2·ri ·h r2

V (i + 1, j)+
V (i, j) = 2 2 2 . ri 1·h t 2·(ri +h r ) (8)
2. ri · h t · h r r 2 ·h 2t
V (i, j − 1) + 2·ri ·h r2
V (i − 1, j)
i
This formula is applicable for each node in both environments [7, 8].
Fig. 1 Discritization of the

cavity
2.2 Approximation of the Electric Field in Each Node
Since the electric field is deduced by calculation of the potential gradient, it is given
by:

grad V = − ∂∂rV · er − r1 · ∂V
· eθ
E = ∂θ (9)
Er · er + E θ · eθ
Thus, electric field module will be:

E= Er2 + E θ2 (10)
However, the radial E r and tangential E θ components of the electric field will be
written as follows:

Er (i, j) = 2h1 r [V (i + 1, j) − V (i − 1, j)]
E= (11)
E θ (i, j) = 2h1t ri [V (i, j + 1) − V (i, j − 1)]
The module in each point will be given by:

E(i, j) = Er2 (i, j) + E θ2 (i, j) (12)
2.3 Approximation of Electrostatic Pressure and Dielectric

Losses
The electromechanical pressure and the dielectric losses are connected directly to
the electric field by the following expressions [19–21]
1
P(i, j) = ε0 εr E 2 (i, j) (13)
2
Pdiel (i, j) = ωε0 εr tgδ E 2 (i, j) + σ E 2 (i, j) (14)
P(i, j): is the electromechanical pressure (N/m2 ) in the point (i, j).
Pdiel (i, j): is the dielectric loss (W/m3 ) in the point (i, j).
3 Results and Interpretations
3.1 Influence of Air Cavity on Electric Field Distribution
In Fig. 2 we have presented the radial distribution of the electric field according
to the radius of the cable. This distribution is marked by increased distortion at the
cavity-insulator interface, which can be explained by the accumulation of polarization
charges that strengthens the field inside the cavity and limits it in the insulator. This
distortion quickly recovers over a distance of about three times the diameter of the
cavity. It is found that the distribution of the electric field which passes through the
center of the cavity and on the walls follows the same pace with values much higher
than the center of the cavity. Figure 3 show the values of the electric field according
to the position of a1μm radius of cavity. The field in the cavity center has a value
of 8.5 kV/mm, it decreases depending the position of the cavity and reaches a value
of 3.5 kV/mm in a cavity near the sheath. In Fig. 4 we presented the values of the
electric field in terms of the radius of a cavity located at a distance of 8.35 mm
from the conductive core. Note that the electric field in the center of the cavity is
proportional to its radius.
On the line passing through the walls

4.5 On the line passing through the center of the cavity
Case of a homogeneous insulation
4
Electric field E (KV/mm)
3.5
2.5
1.5
17.305 17.31 17.315 17.32 17.325 17.33 17.335 17.34
radial position ri (mm)
Fig. 2 Distribution of electric field under the influence of a cavity located 17.325 mm from the
center of the conductive core
9
8 On the line passing through the center of the cavity
7
Electric field E (KV/mm)
1
8 10 12 14 16 18 20
Fig. 3 Electric field inside the cavity according to its position
Fig. 4 Electric field inside the cavity according to its radius

3.2 Influence of the Air Cavity on the Electromechanical

Pressure Distribution
We elaborate in Fig. 5 the electromechanical pressure in accordance with the radius

of the cable. The electromechanical pressure in the center of the cavity is more intense
than in the homogeneous case because of the permittivity changing, and on the other
hand the electromechanical pressure in the walls of the cavity is greater than in the
cavity center because of the accumulation of the charges on the walls of the cavity,
which exerts an additional pressure on its walls.
It is noted in Fig. 6 that the electromechanical pressure is greater in the case
of a cavity close to the conducting core, and each time the distance between the
core and the cavity is enlarged the electromechanical pressure decreases, since the
Electromechanical pressure is proportional to the electric field.
It can be seen in Fig. 7 that the electromechanical pressure increases as the radius
of the cavity increases because of the charges in the walls which are in a very large
quantity as the size of the cavity increases.

160
electrostatic pressure P (N/m2)
140
120
100
80
60
40
20
17.305 17.31 17.315 17.32 17.325 17.33 17.335 17.34
Fig. 5 Distribution of the electromechanical pressure under the influence of a cavity located
17.325 mm from the center of the conductive core
700
On the line passing through the center of the cavity
600
electrostatic pressure P (N/m 2)
500
400
300
200
100
0
8 10 12 14 16 18 20
Fig. 6 Influence the position of the cavity on the electromechanical pressure
Fig. 7 Influence of cavity radius on electromechanical pressure

450
dielectric losses Pdié l (W/m3)
350
300
250
200
150
100
50
17.305 17.31 17.315 17.32 17.325 17.33 17.335 17.34
Fig. 8 Distribution of dielectric losses under the influence of a cavity located at 17.325 mm center
of the conductive core
3.3 Influence of the Air Cavity on the Dielectric Loss

Distribution
We have shown in Fig. 8 the distribution of dielectric losses according to the radial
position. It is noted that the dielectric losses on the line passing through the walls of
the cavity are very important as the line passing through the center of the cavity. In
Fig. 9 we present the distribution of the dielectric losses within the position of the
cavity, it is found that the dielectric losses rises to an important stage in the case of a
cavity (1 μm) close to the conductive core, and drops each time the cavity approaches
the outer sheath. Figure 10 shows the dielectric losses depending the radius of the
cavity, a linear increase is observed because of the charges in the walls, which are in
a very large quantity as the size of the cavity increases.
4 Conclusion
The main goal of this work is the study of the behavior of insulators used in unipolar
high voltage (HV) cables in the presence of cavities. We found that defects disturb
the electric field within the insulation and strongly depend on their positions and
sizes. This field disturbance causes a dielectric losses increasing within the insulator
and leads to the acceleration of the degradation thereof. This disturbance does not
1600
Case of homogeneous insulation
dielectric losses Pdiél (W/m 3)
1200
1000
800
600
400
200
0
8 10 12 14 16 18 20
Fig. 9 Influence of the position of the cavity on the distribution of dielectric losses
Fig. 10 Influence of the radius of the cavity on the distribution of dielectric losses
exceed a certain zone called zone of influence, this allowed us to reduce the size of
the calculation procedure, and reduce the execution time.
References
1. Harisha KS, Gouthami N, Harshitha V, Madhu C (2016) Partial Discharge Analysis of a Solid
Dielectric Using MATLAB Simulink. Int J Innov Res Electr Electron Instrum Control Eng 4(6)
2. Gupta K, Yadav NK, Rattewal PK (2015) Partial discharge within a spherical cavity in solid
dielectric material. Int Res J Eng Technol (IRJET) 02(08)
3. Ramachandra B, Manohara HC (2016) PD analysis in cylindrical void with respect to geometry
of the void. Int J Innov Res Electr Electron Instrum Control Eng 4(10)
4. Srinivasa DM, Harish BN, Harisha KS (2017) Analysis study on partial discharge magnitudes
to the parallel and perpendicular axis of a cylindrical cavity. Int J Eng Trends Technol (IJETT)
45(7)
5. Hadjadj M, Mokhtari B, Mahi D (2014) Prediction of the physical parameters change inside
aspherical cavity located in a material XLPE of a medium voltage cable by non-stationary
modelling. J Electr Eng 14(1):184–190
6. Karmakar S, Sabat A (2011) Simulation of partial discharge in high voltage power equipment.
Int J Electr Eng Inform 3
7. Arief YZ, Izzati WA, Adzis Z (2012) Modeling of partial discharge mechanisms in solid
dielectric material. Int J Eng Innov Technol (IJEIT) 1(4)
8. Sharma P, Bhanddakkar A (2015) Simulation model of partial discharge in power equipment
international. J Electr Electron Res 3(1):149–155
9. Manani HT, Dudani KK (2014) Analysis of partial discharge signal by FDTD technique. Int J
Adv Eng Res Dev (IJAERD) 1(3)
10. Khan Taimur, Haleem Adnan (2018) A novel method for insulation testing of high voltage
electrical equipment. Int J Eng Works 5(2):32–36
11. Kumar CS, Ramachandra B (2016) Comparison of PD activity in cylindrical and cubical void
using MATLAB simulink. Int J Innov Res Electr Electron Instrum Control Eng 4(10)
12. Illias HA, Lee ZH, Bakar AHA, Mokhlis H (2012) Distribution of electric field in medium
voltage cable joint geometry. In: 2012 IEEE international conference on condition monitoring
and diagnosis, 23–27 September 2012, Bali, Indonesia
13. Illias H, Chen G, Lewin PL (2011) Modeling of partial discharge activity in spherical cavities
within a dielectric material. In: IEEE electrical insulation magazine, vol 27, pp 38–45
14. Illias HA, Chen G, Lewin PL (2012) Partial discharge within a spherical cavity in a dielectric
material as a function of cavity size and material temperature. In: IET science, measurement
and technology, vol 6, pp 52–62
15. Agarawal SK, Mittal LK, Jafri H (2017) Simulation of partial discharge in high voltage power
equipment. Int J Electron Electr Comput Syst IJEECS 6(8)
16. Callender G, Rapisarda P (2016) Investigating the dependence of partial discharge activity on
applied field structure. In: IEEE electrical insulation
17. Dabbak SZ, Illias HA, Ang BC (2017) Effect of surface discharges on different polymer
dielectric materials under high field stress. In: IEEE transactions on dielectrics and electrical
insulation
18. Altay Ö, Kalenderli Ö, Merev A (2009) Preliminary partial discharge measurements with a
computer aided partial discharge detection system. In: ELECO
19. Bhandakkar J (2017) Analysis and simulation of partial discharge for different insulation
material. Int J Innov Eng
20. Illias H, Chen G, Lewin PL (2011) Modeling of partial discharge activity in spherical cavities
within a dielectric material. In: IEEE electrical insulation
21. Tanmaneeprasert T, Lewin PL (2017) Analysis of degradation mechanisms of silicone
insulation containing a spherical cavity using partial discharge detection. In: Conference (EIC)
Self Tuning Fuzzy Maximum Power
Tracking Control of PMSG Wind Energy
Conversion System
Chaib Housseyn, Mihoub Youcef, and Hassaine Said
Abstract This paper proposes a chart to compare several ways to control the Wind
Energy Conversion System (WECS) connected to the grid, using permanent magnet
synchronous generator (PMSG). To overcome the limitation of classical fixed gain
based proportional-integral (PI) witch requires knowledge of system parameters, a
Fuzzy Logic Controller (FLC) is witch doesn’t rely on mathematical model and
integrates human experience. In this study, we used several methods to control the
Maximum Power Point Tracking algorithm (MPPT): PI controller, FLC and Self
Tuning Fuzzy Logic Control (ST-FLC). The system contains two controllers for
the machine side (MSC) and grid side (GSC) converters. On the machine side, the
rotation speed control is obtained by assuring the algorithm MPPT to follow the
maximum power of the wind. In the grid side, the DC voltage and (d q) currents
control is used to ensure active and reactive power. The system analysis is real-
ized in volatile wind conditions using the Matlab/simulink® with SimPowerSystems
block. Obtained results show that fuzzy self-tuning technique appears superior to the
others, as it guarantees in general very good performances in the set-point and load
disturbance step responses. It requires a modest implementation effort; therefore its
practical implementation in industrial environments appears to be very promising.
Keywords Wind energy conversion system (WECS) · Maximum power point

tracking (MPPT) · Permanent magnet synchronous generator (PMSG) ·
Proportional integral (PI) controller · Fuzzy logic controller (FLC) · Self-tuning
fuzzy logic controller (ST-FLC)
C. Housseyn (B) · M. Youcef · H. Said

Laboratory of Energy Engineering and Computer Engineering (L2GEGI), University of Tiaret,
Tiaret, Algeria
M. Youcef
H. Said

https://doi.org/10.1007/978-981-15-6403-1_50
736 C. Housseyn et al.
1 Introduction
Winds are a renewable energy resource for electricity generation because of their
productive capacity compared to other types of renewable energy resources. It is
harmless to the environment and is abundant in nature. Thus, wind energy can be used
by mechanically converting it into electrical energy using wind turbines (WTs) [1].
The development of wind turbine concepts and the creation of several different
types of wind generators is a consequence of the fast-paced evolution of wind-transfer
technology. There is two types of wind turbines, fixed and variable speed. Fixed-speed
wind turbine is based on a standard squirrel-cage induction generator (SCIG) that
uses a multi-stage gearbox. Variable-speed wind turbines rely on two different types
of generators. The first type is a doubly-fed induction generator (DFIG) that uses
a multi-stage gearbox. The second type is a direct-drive generator or a low-speed
high-torque synchronous generator without gear [2].
In recent years, wind turbines have developed to use a mixed type with gearbox
and PMSG giving a solution for modern wind systems due to the different advantages
to other generators [3, 4].
Control of wind turbines must be permanent and continuous so as to maintain the
output power when the wind speed is varying. PI classical control requires knowledge
of system parameters [5–7], and presents several problems when switching to low
frequency, as well as a slow response. To overcome this limitation; several advanced
control methods based on artificial intelligence control have been proposed such as
fuzzy logic [8], genetic algorithms [9, 10] and neural networks [10, 11].
FLC based control systems offer much-improved performance to run the wind
system at maximum power under changing environmental conditions and uncertain-
ties [12]. The main interest of Fuzzy Logic Controllers is their ability to integrate
experience, intuition, and heuristics into the system instead of relying on mathemat-
ical models, making them more efficient in system applications especially where
existing models are poorly defined and insufficiently reliable [13]. Fuzzy logic
controllers are suitable for wind systems characterized by high non-linearity [14].
A new technique ST-FLC, in which the controller gain is continuously, adjusted
using the gain update factor is proposed in this study. We focused only on setting the
output scale factor gain (Gu), considering it as equivalent to the gain of the controller.
The highest priority is given to setting the Gu because of its considerable influence
on the performance and stability of the system self tuned by a gain update factor β
determined from a mathematical algorithm [15].
Updating instantaneously the output scale is the major contribution to improve the
FLC in witch this gains is fixed. The purpose of this article is to present a compara-
tive study of control for a maximum power tracking strategy for variable speed wind
turbines between a classical PI, FLC and ST-FLC controllers. The proposed algo-
rithms are applied to control the PMSG rotation speed which regulates the output
voltage to the predetermined value to extract the maximum possible wind power.
The paper is classified into five sections: Sect. 2 details the dynamic modelling
of each element of WECS, Sect. 3 describes the MSC design, Sect. 4 describes the
Self Tuning Fuzzy Maximum Power Tracking Control … 737
GSC design, Sect. 5 presents the simulation results, and final section dedicates to the
conclusion and perspectives.
2 Wind Energy Conversion System Description
Figure 1 shows the system structure of a variable speed wind turbine based on a
PMSG, two full-size back-to-back PWM (pulse width modulation) converters and a
power grid.
The modeling of each system component is presented as following.
2.1 Wind Turbine
The mechanical power extracted from the wind can be expressed as follows
pm = 1/2ρπ R 2 VW3 C P (λ, B) (1)
where Pm is the extracted power from the wind, ρ is the air density (Kg/m3 ), R is the
blade radius (m), V w is the wind speed (m/s) and C p (λ, B) is the power coefficient
and λ is the tip speed ration. The term λ is defined as
ωm · R
λ= (2)
VW
The blade pitch angle B is always constant during MPPT control (B = 0), ωm is
the mechanical speed of the turbine. The input torque for the generator is obtained
from:
Pm
Tm = (3)
ωm
Fig. 1 Wind energy conversion system

2.2 PMSG Model
The generator is modeled by the following voltage equations in the rotor reference
frame (dq-axes):
dφsd
Vsd = Rs · Isd + + ωe · φsq (4)
dt
dφsq
Vsq = Rs · Isq + + ωe · φsd (5)
dt
where isd and isq are the currents in the (dq) reference frame, Rs is the stator resistance,
ωe is the electrical rotating speed, ϕ sd and ϕ sq are the d and q-axes stator flux linkages
given by:
ϕsd = −L d · i sd + φm (6)
ϕsq = −L q · i sq (7)
where L d and L q are the inductances in the (dq)-axis and φ m is the rotor magnetic
flux produced by the generator.
The expression of the electromagnetic torque T e is provided by:
3
Te = P L d − L q i sd i sq + φm i sq (8)
2
where, P is the number oh the pole pairs.
The mechanical equation which connects the generator with the wind turbine is
described by:
dwg
jeq = T e + Tl − Bm wg (9)
dt
where T l is the torque applied by the turbine to the generator. jeq is the equivalent
moment of inertia; and Bm is the viscous turn coefficient.
2.3 DC Line Modeling
The dc line is the capacitor located between the convertors; its function is to keep
the DC voltage stable, the model of the dc line is expressed as
Cd Vdc
= is − i g (10)
dt
where is the DC line voltage, is the current of DC line at the generator side and ig
the courant of DC line at the grid side, C is the capacity of DC-line capacitor.
Let us suppose that the convertor is ideal does not consume power; the voltage of
the capacitor is written as
d Vdc Ps
C = − ig (11)
dt Vdc
where, Ps is the output power of generator. We can write it as:
Ps = 3Vs i s = Vdc i dc (12)
We rephrase it as:
3
Ps = E max Imax = Vdc i dc (13)
2
where V s is the voltage of generator side and E max is its amplitude, I max is the
amplitude of the generator current, and idc is the current through the capacitor.
2.4 Grid Model
The dynamic model of the grid connection when selecting a reference frame rotating
synchronously with the grid voltage space vector is
L g di gd
u gd = −R g i gd − + wgr L g i gq + egd (14)
dt
L g di gq
u gq = −R g i gq − + wgr L g i gd (15)
dt
where Rg and L g are the resistance and inductance respectively of the filtre, this latter
is located between the converter and the grid. ugd and ugq are the inverter voltage
components, wgr is the electrical angular velocity of the grid.
3 Control of the Machine Side Converter
To control the PMSG while the wind speed is variable, we use the MPPT algorithm. It
consists in operating the PMSG at the speed corresponding to the point of maximum
power when the wind speed is lower than the nominal wind speed as shown in Fig. 2.
Fig. 2 Diagram of proposed control system of the machine side converter
3.1 Proportional-Integral Control (PI)
The mathematical model of the control of the mechanical speed of PMSG performed
by the PI controller is given by the following relationship:

Ki
Iq∗ (s) = Kp + × ωm∗ (s) − ωm (s) (16)
S
where ω*m is the reference speed given by MPPT and ωm is the speed of measured
PMSG, K p is the proportional gain and K i is the integral gain.
3.2 Fuzzy Logic Control Design
The FLC code consists of three parts: preprocessing, fuzzy rules and interface
engine, and post-processing. In the preprocessing part, the input variables of the
fuzzy controller are speed error (e) and change of error (e). The input scale and
input error scale gain factors are Ge and Ge respectively. The final step of the FLC
system is post-processing. The output signal I q is multiplied by the scaling factors
du(t) as shown in Fig. 3.
Fig. 3 FLC control structure
e(t) = ωm∗ (t) − ωm (t) (17)
e(t) − e(t − 1)
e(t) = (18)
tsamp
Iq∗ = Iq∗ (t − 1) + du(t) (19)
Using Mamdani’s method, the fuzzy input variables are fuzzified into suitable
linguistic values, then processed in the fuzzy set region which includes membership
function where an appropriate fuzzy output is acquired using fuzzy rules, and there-
after the fuzzy output is converted into a crisp value in the defuzzification using the
centroid method, also known as the center of gravity method specified by:

n
Ci .μi
i=1
U = (20)
n
μi
i=1
where C i is discrete element of an output fuzzy set, μi is its membership function

and n is the number of fuzzy rules.
Once the inputs are converted into linguistic variables the FLC gives a manageable
change in the reference voltage to reach the maximum power, based on the rules
displayed in Table 1.
Table 1 Set of fuzzy rules

dU E
NB NS Z PS PB
E NB NB NB NS NS Z
NS NB NS NS Z PS
Z NS NS Z PS PS
PS NS Z PS PS PB
PB Z PS PS PB PB
Fig. 4 Inputs/Output membership functions of the F-PI controller
The membership functions for the inputs and the output variables are triangular
and trapezoidal membership functions named: Negative Big (NB), Negative Small
(NS), Zero (Z), Positive Small (PS), and Positive Big (PB) (Fig. 4).
3.3 Self-tuning Fuzzy Logic Controller
ST-FLC is used to adjust instantaneously the output scale factor (Gu ), according
to the speed input change error (Ge ). The proposed automatic tuning mechanism
focused on the on-line adjustment of the fuzzy system output as a function of the shift
error. The block diagram of the proposed ST-FLC is shown in Fig. 5. The proposed
self tuning bloc calculation used a simple algorithm reducing the system complexity
and the computational burden.
The ST-FLC scaling factors (Ge , Ge and Gu ) are related to the inputs and outputs
(e, e, u) in the following equations:
Fig. 5 Self-tuning fuzzy logic controller structure
e = G e × e. (21)
e = Ge × e (22)
u = (β ∗ G e ) × u (23)
where:

1
β = K1 + |e| (24)
m
The variable β formulated based on expert Knowledge of the system according

to this concept: ((If the system is moving faster towards its desired operating-point
(small Δe), then output action (Δu) needs to be reduced (reduce Gu ) to prevent big
overshoot and/or undershoot. In contrast, if the system is rapidly moving away from
the desired operating-point (big e), then output action (u) needs be increased
(increase Gu ) for limiting these deviations for a faster recovery of the system to its
desired operating point)). In other words, if the value of e is small, then Gu need
to be reduced and if e is big, then Gu need to be increased. Hence, the β value is
formulated based on this concept by adding the e to the fraction (1/m) to avoid
lower gain multiplication (Gu ) when the e is very small. Lower multiplication of
gain may result in oscillation and not stable condition during steady state operation.
The value of m is chosen to be equal to the number of uniform input (e and e) fuzzy
partition (number of MFs). The value of K1 is chosen to make the possible variation
in β which set to 4 based on tuning process [15]. The ther fuzzy parameters were
kept unchanged, similar to the standard fixed parameters FLC.
4 Control of the Grid Side Converter
The grid side converter control has to adjust the DC (Direct Current) line voltage to
the desired value, and ensure that the output voltage converter has a similar amplitude
Fig. 6 Diagram of the loop control of DC line
and frequency with the grid, but the main objective is to control the active and reactive
power.
4.1 DC Line Control
The control block diagram of the DC-line is presented in Fig. 6.

The current loop is considered to be fast than the voltage loop and supposed equal
to 1. The PI controller is designed with the pole placement method.
4.2 Control of Active and Reactive Power
Then active and reactive power are transmitted through the grid side converter, they
are expressed as follow:
3
P= u gd i gd (25)
2
3
Q= u gd i gq (26)
2
To control active and reactive power is achieved by acting on the direct and
quadrature current as shown in Fig. 7.
The first channel controls the DC voltage which is used to provide the d-axis
current reference for the active power control. This assures that all the power coming
from the PMSG converter is instantly transferred to the grid through the converter.
The second channel controls the reactive power by setting the q-axis current refer-
ence to the current control loop similar to the previous one. The current controllers
will provide a voltage reference for the converter that is compensated by adding
rotational EMF (Electromotive Force) compensation terms. All controllers are PI
and are tuned using the pole placement method.
Fig. 7 Bloc diagram of the proposed control system of the grid side converter
To test the performances of the proposed control system, several simulation test
with a variable wind speed profile indicated in Fig. 9 were developed under
Matlab/simulink® with SimPowerSystems block.
The used wind turbine and the PMSG parameters are given respectively in Tables 2
and 3. The main measurements include speed, torque, three-phase voltage, and stator
current.
The speed response of the ST-FLC controller has better tracking characteristics
than other controllers, as shown in Fig. 8, in which the speed monitoring controller
operates in a critical situation, the response time of the PMSG speed is very short,
there is no overshoot and no steady-state error as shown in Fig. 9.
There is a slight difference in response time to reach the reference DC voltage
between the ST-FLC and FLC, But there is a lag of about a classical PI control near
to 20 ms without no steady-state error as shown in Fig. 10.
Table 2 wind turbine

Parameter Value Units
parameters
Air density ρ 1.22 Kg/m3
Gear ration Gr 5.4
Rotor radius R 3 M
Maximum power coefficien C p 0.48
Equivalent inertia (turbine + generator) 0.042 Kg m2
Table 3 PMSG parameter

Parameter Value Units
Rated power 2.5 kW
Rated line courant 11 A
Rated speed 3000 r/m
Mutual inductance in d-axis L d 7.5 mH
Mutual inductance in q-axis L q 7.5 mH
Stator resistance Rs 0.45
Number of pole pairs P 3
Permanent magnet flux ψf 0.52 Wb
viscous friction Bm 0.017 Nm/rad/s
Fig. 8 Wind speed profile
Figure 11 shows the active power of the grid and the active power of PMSG. Note
that the response time for the ST-FLC and FLC controllers is small compared to
the Classical PI. The difference between the active power of the grid and the active
power of the PMSG is due to the resistance of the filter which causes the voltage to
decrease in the steady-state.
The comparison between the three controllers at the estimated wind speed of
6 m/s is shown in Table 4. We notice that the ST-FLC controller has a short response
time compared to the other controllers. The ST-FLC and FLC controllers have no
overshoot compared to the PI. Finally, we can say that the ST-FLC controller has a
smaller rise time and a settling time than the last two controllers.
Fig. 9 PMSG speed controlled by different methods
Fig. 10 DC voltage controlled by different methods
6 Conclusion
In this paper, a comparison between classical and fuzzy based strategies to control
WECS using PMSG and integrated into the grid. FLC offer much-improved perfor-
mance. It is efficient where the model is poorly defined and can use human experience.
The purpose of this study is to test a two advanced control of the MPPT of variable
speed wind turbines: FLC using the error and the change of error to increase or
decrease the control using fixed gain scale factors and ST-FLC where the output gain
is instantaneously calculated according to the system conditions.
Fig. 11 Power of grid and power of PMSG controlled by different methods
Table 4 Comparison
PI FLC ST-FLC
performances at rated wind
speed for the different Rising time (s) 0.017 0.017 0.016
controllers Settling time (s) 0.020 0.019 0.017
Overshoot (%) 2.03 0 0
Steady state error (%) 0.57 0 0
Simulation results confirm that FLC performance is better compared with a classic
PI controller for response speed and ability to track maximum wind power. The ST-
FLC strategy, using a simple bloc calculation, guarantees robust performances and
gives satisfactory results compared to other strategies used.
Further studies include the testing of ST-FLC control algorithms on the full struc-
ture of WECS to control DC voltage. Finally, the developed simulation algorithms
can be validated through an experimental implementation.
References
1. Polinder H, Ferreira JA, Jensen BB, Abrahamsen AB, Atallah K, McMahon RA (2013) Trends
in wind turbine generator systems. IEEE J Emerg Sel Top Power Electron 1(3):174–185
2. Chen Z, Li H (2008) Overview of different wind generator systems and their comparisons. IET
Renew Power Gener 2(2):123–138
3. Slah H, Mehdi D, Lassaad S (2016) Advanced control of a PMSG wind turbine. IJMNTA
05(01):1–10
4. Durán MJ, Kouro S, Wu B, Levi E, Barrero F, Alepuz S, Six-phase PMSG wind energy
conversion system based on medium-voltage multilevel converter, p 10
5. Wei C, Zhang Z, Qiao W, Qu L (2016) An adaptive network-based reinforcement learning

method for MPPT control of PMSG wind energy conversion systems. IEEE Trans Power
Electron 31(11):7837–7848
6. Liu J, Zhou F, Zhao C, Wang Z (2019) A PI-type sliding mode controller design for PMSG-based
wind turbine. Complexity 2019:1–12
7. Tripathi SM, Tiwari AN, Singh D (2016) Optimum design of proportional-integral controllers in
grid-integrated PMSG-based wind energy conversion system: optimum design of PI controllers.
Int Trans Electr Energy Syst 26(5):1006–1031
8. Suganthi L, Iniyan S, Samuel AA (2015) Applications of fuzzy logic in renewable energy
systems – a review. Renew Sustain Energy Rev 48:585–607
9. Daraban S, Petreus D, Morel C (2014) A novel MPPT (maximum power point tracking) algo-
rithm based on a modified genetic algorithm specialized on tracking the global maximum power
point in photovoltaic systems affected by partial shading. Energy 74:374–388
10. Chavero-Navarrete E, Trejo-Perea M, Jáuregui-Correa JC, Carrillo-Serrano RV, Ríos-Moreno
JG (2019) Expert control systems for maximum power point tracking in a wind turbine with
PMSG: state of the art. Appl Sci 9(12):2469
11. Karrari M, Rosehart W, Malik OP (2008) Hierarchal control system for a variable speed cage
machine wind generation unit using neural networks. Asian J Control 7(3):286–295
12. Asri A, Mihoub Y, Hassaine S, Logerais P, Allaoui T (2019) Intelligent maximum power
tracking control of a PMSG wind energy conversion system. Asian J Control asjc.2090
13. Harrabi N, Kharrat M, Aitouche A, Souissi M (2018) Control strategies for the grid side
converter in a wind generation system based on a fuzzy approach. Int J Appl Math Comput Sci
28(2):323–333
14. Farh HM, Eltamaly AM (2013) Fuzzy logic control of wind energy conversion system. J Renew
Sustain Energy 5(2):023125
15. Farah N et al (2019) A novel self-tuning fuzzy logic controller based induction motor drive
system: an experimental approach. IEEE Access 7:68172–68184
A Backstepping Controller
for Interleaved Boost DC–DC Converter
Improving Fuel Cell Voltage Regulation
Ali Dali, Samir Abdelmalek, and Maamar Bettayeb
Abstract The aim of the present work is to design a robust nonlinear controller
which ensures a regulated output voltage for a Proton-Exchange Membrane Fuel
Cell (PEMFC) fed interleaved boost DC–DC converter. In this regard, a particle
swarm-based optimization algorithm is used for tuning parameters of the nonlinear
controller allowing greater flexibility in guaranteeing system specifications compared
to the classical Proportional Integral (PI) controller. The effectiveness of the proposed
controller for the studied system was validated under various operating conditions
of the PEMFC and load perturbations using Sim Power System (SPS) toolbox and
MATLAB/Simulink environment. Simulations and comparison of results show that
the proposed robust nonlinear controller offers several advantages for smoother
tracking, smaller overshoot, faster response and better stability.
Keywords Interleaved boost converter · Robustness · Conventional PI · Fuel cell ·

Nonlinear controller
A. Dali (B)
Centre de Développement des Energies Renouvelables, CDER, BP 62 Route de l’Observatoire,
16340 Bouzaréah, Algiers, Algeria
S. Abdelmalek
Department of Electrical Engineering, Faculty of Technology, University of Medea, Medea,
Algeria
M. Bettayeb
Department of Electrical and Computer Engineering, University of Sharjah, United Arab
Emirates, and Center of Excellence in Intelligent Engineering Systems (CEIES), King Abdulaziz
University, Jeddah, Kingdom of Saudi Arabia

https://doi.org/10.1007/978-981-15-6403-1_51
752 A. Dali et al.
1 Introduction
In the past few decades, Fuel Cells (FCs) have emerged as a potential alternative
solution to fossil fuel based energy sources [1]. In addition, they are considered as
one of the most promising technology for electrical power generation [2]. The Fuel
cell is capable of converting chemical energy directly into electrical energy, without
any intermediate step. Therefore, the electrical efficiency of FCs is considerably
higher than the other traditional sources for generating electrical power.
Fuel cells are widely used in several applications for military, automotive and low
power generation as they offer many advantages such as low operating temperature,
fast response to load changes, high power density, low corrosion, very low noise
and environmental friendliness [1–3]. In general, DC–DC converters are employed
to adapt the voltage and current levels between sources and loads while guarantee a
low power loss in the power generation systems [4, 5]. The DC–DC converter is an
essential part of the fuel cell power generation to control the delivered power/voltage.
Incorporating PEMFC dynamics with the classical DC/DC converter makes the
voltage regulation task quite challenging [6]. In this context, in order to deal with
this drawback, the interleaved boost DC–DC converter can be used [7]. This kind of
converters has the advantages of reducing the inductors size, increasing the power
level and minimizing the current stress on the power switches [8]. With regard to
the literature, different kinds of control strategies are widely employed for inter-
leaved boost DC–DC converters, such as adaptive controller [9], robust control [10],
nonlinear model predictive control (MPC) [11], state feedback control [12], sliding
mode fuzzy PID controller [12], robust adaptive neural network control [13], adap-
tive sliding mode controller [14] and intelligent controllers [15]. However, among
the aforementioned control techniques, the problem of robust nonlinear controller
design based on optimization techniques was not studied up to now. Therefore, this
work aims to propose a robust nonlinear control law based on the Particle Swarm
Optimization (PSO) algorithm.
The novelty and highlights of the paper can be stated as follows: (i) Using a
new architecture which is based on an interleaved boost DC–DC converter, (ii) A
Backstepping controller design for output voltage regulation under varying loading
conditions, (iii) Inclusion of PSO algorithm for tuning of the proposed controller
parameters and iv) Validation and comparison of the proposed controller design with
those of the conventional PI controller in Sim Power System (SPS) toolbox and
MATLAB/Simulink environment.
The remaining sections are organized as follows. First, Sect. 2 introduces the
model of the interleaved boost DC–DC converter. Section 3 then presents the fuel
cell mathematical model, followed by the design of the nonlinear controller in Sect. 4.
The parameters optimization using the PSO algorithm for the robust nonlinear and
traditional PI controllers is deployed in Sect. 5. The simulation results with robustness
tests are presented in Sect. 6. The last section summarizes the key aspects of this
paper and presents some conclusions.
A Backstepping Controller for Interleaved … 753
2 Interleaved Boost DC–DC Converter
Figure 1 shows the schematic of an interleaved boost DC–DC converter (IBC), the
main advantage of the IBC is the reduction of the current ripples compared to the
classical boost converter. where V c is the output voltage and V fc is the Fuel Cell (FC)
source voltage.
The average model of the interleaved boost DC–DC converter is the same as the
traditional boost DC–DC converter with except that it contains two current inductors
[16]. The dynamics of the converter are represented by:
⎧
⎪ d IL
⎨ L 1 dt 1 = V f c − Vc (1 − D1 )
dI
L 2 L 2 = V f c − Vc (1 − D2 ) (1)
⎪
⎩ C d Vdtc = I (1 − D ) + I (1 − D ) − I
dt L1 1 L2 2 R
such that
⎧
⎨ I f c = IL1 + IL2
D = D2 = 1 − u (2)
⎩ 1
L1 = L2 = L
where L and C are respectively the inductance and the capacitor, D1 and D2 are the
duty cycles of each phase (which one is shifted by half of period from the other).
I L 1 and I L 2 are the inductances currents, I f c is the fuel cell current, I R = V c /R is the
resistor (R) load current and Ic is the capacitor current. Then, the average model of
the system (3) may be expressed as follows

dIfc
dt
= L2 V f c − L2 Vc u
dVf c (3)
dt
= C2 I f c u − RC
1
Vc
Fig. 1 Voltage regulation of the fuel cell interleaved boost DC converter

754 A. Dali et al.
where u is the average duty cycle.

The system state space vector χ = [Vc , I f c , − Vc ]T includes the measurable
output voltage (Vc ) and the fuel cell current (I f c ), the third state is introduced to
improve the robustness of the controller design, the state vector of the augmented
model is then described as:
χ̇(t) = Aχ(t) + Bu(t) + (4)
where the associated matrices are given by:

⎡ ⎤ ⎧ 2 T
0 0 0 ⎪ ⎨ B = − L x2 C x1 0
1
A = ⎣ 0 − RC
1
0 ⎦, χ = [x1 x2 x3 ]T
⎪
0 1 0 ⎩ = L2 V f c 0 0 T
3 Fuel Cell Mathematical Model
This section develops a mathematical model that represents a fuel cell stack by a
controlled voltage source in series with a constant resistance as illustrated in Fig. 2.
The controlled voltage source (E) is expressed as

Ifc
E = E oc − N A0 ln Io
1
Td
s 3+1 (5)
V f c = E − Rohm I f c
where,
Fig. 2 Simplified fuel cell stack model

E oc open circuit voltage (V);

N number of cells;
Ao Tafel slope (V);
Io exchange current (A);
Td the response time (at 95% of the final value);
Rohm internal resistance ();
I fc fuel cell current (A);
Vfc fuel cell voltage (V).
For more details, authors can read Ref. [17].
4 Backstepping Controller Design
A Backstepping controller is designed for satisfying a good trajectory tracking of

the desired voltage (V cref ) by acting on the duty cycle (u) of the interleaved boost
DC–DC converter switch. The important steps to design the Backstepping controller
are described in this section. Let us define the following tracking errors:
⎧
⎨ e1 = e2 dt
e = Vcr e f − Vc (6)
⎩ 2
e3 = I f cd − I f c
with V̇cr e f = 0.
The control law is designed in three steps. A first Lyapunov function (V 1 ) and its
derivative with respect to time is defined respectively by:

V1 = 21 e12
(7)
V̇1 = e1 e2
Then, a new variable e2d is introduced for stabilizing e1 :
e2d = −k1 e1 (8)
where, k 1 is a positive constant.

Then, V 1 is a negative definite function of e1 :
V̇1 = −k1 e12 < 0, k1 ∈ ∗+ (9)
In the following step, a second augmented Lyapunov function is selected as

follows:
1
V2 = V1 + (e2 − e2d )2 (10)
2
756 A. Dali et al.
Differentiating V 2 with some mathematical manipulations yield to,

2u 1
V̇2 = −k1 e12 + (e2 − e2d ) e1 − ė2d − Ifc + Vc (11)
C RC
such that:

C 1
I f cd = e1 − ė2d + Vc + k2 (e2 − e2d ) (12)
2u RC
Based on the Eqs. (11) and (12) and with a mathematical development, one can
get:
V̇2 = −k1 e12 − k2 (e2 − e2d )2 < 0, (k1 , k2 ) ∈ ∗+ (13)
Finally, with the same procedure we choose the third augmented Lyapunov
positive function described as follows:
1L 2
V3 = V2 + e (14)
2C 3
Using Eqs. (9) and (13), with some developments V 3 can be derived as follows:
L
V̇3 = −k1 e12 − k2 (e2 − e2d )2 − k3 e32 < 0, k3 ∈ ∗+ (15)
C
Then, the control input (u) is given by:

L 2
u= − I˙f cd + V f c − k3 e3 (16)
2(Vcd − e2d ) L
The PSO is an optimization technique inspired from the social behavior of biolog-
ical organisms, such as birds flock to a food source with the help of a combination
expressing the influence on the position of the body through the historical positions
of itself and its neighbors. This algorithm has been developed to solve nonlinear
equations. The basic idea is to build a population of particles and to move them
according to specific criteria to achieve the final goal. The movement of a particle
can be represented as follows [18–21]:
X j (t + 1) = X j (t) + V j (t + 1) (17)
V j (t + 1) = ω ∗ V j (t) + C1r1 (ρi − X j (t))

+ C2 r2 (ρg − X j (t)) (18)
Table 1 PSO parameters

PSO parameters O-PI O-NL
C 1, C 2 3.0, 1.0 3.0, 1.0
ω 0.8 0.8
Population size 10 20
Maximum iteration 20 100
Table 2 Search range of the

Controller parameters Min Max
parameters
K = [k 1 k 2 k 3 ] [10−1 , 1, 1] [200 100 1000]
[K p K i ] [0, 0] [10 10]
Table 3 Controllers
Controller Parameters
parameters
O-PI K p = 0.077, K i = 5.65
O-LQRO-LQR K = [137.16 1.57 728.57]
For each particle X j , the speed V j is calculated as a linear combination of three

elements: its own speed, its deviation from its ρi neighborhood, and compared to
the best ρg particle. Then, the current position of the particle is updated through the
coefficients ω, C 1 and C 2 to get the new position. The effectiveness of the method
focuses in particular on the randomness of the last two coefficients (r 1 and r 2 ).
The PSO technique is used to get the optimal controller parameters by minimizing
the following objective function
tf
e= Vcr e f (t) − Vc (t)dt (19)
0
Equation (19) represents the performance index used for each controller technique
(the Backstepping and the conventional PI). The PSO parameters for each controller
and the search range are listed in Tables 1 and 2, respectively.
The optimization results for each controller are listed in Table 3.
In order to verify the advantages of the designed nonlinear controller that adopts
the improved tracking performance in the PEMFC control system, the compar-
ison between the optimized nonlinear (O-NL) controller and the optimized conven-
tional PI (O-PI) controller on the dynamic output characteristic of PEMFC will be
discussed. The interleaved DC–DC boost converter parameters used in the simulation
are listed in Table 4.
758 A. Dali et al.
Table 4 DC–DC interleaved

Parameters
boost converter
L 500 μH
C 1000 μF
R 50
F 100 Khz
Simulations are performed for two different cases for different situations including
normal operating conditions and a load variation.
Test 1: The reference voltage (V cref ) changes as follows:
⎧
⎨ Vcr e f0 − 10%Vcr e f0 0 ≤ t ≤ 0.1 s
Vcr e f = Vcr e f0 0.1 s ≤ t ≤ 0.2 s
⎩
Vcr e f0 + 10%Vcr e f0 0.2 s ≤ t ≤ 0.3 s
with Vcr e f 0 = 75.

Figures 4 and 5 present a comparative study between the proposed controller and
the optimized PI controller. It is clear that V c (O-NL) reaches the desired set point
(V cref ) in the settling time of 0.05 s. In addition, the O-NL controller presents a
minimal tracking error, a good transition response and a very fast system reaction
against set point change. Therefore, one can deduce that the proposed approach is
better than the O-PI control approach.
The reference tracking capability of the proposed COSF controller is investigated.
The obtained results are compared with those O-PI and O-LQR controllers with the
same conditions. The responses of the three controllers are given in Figs. 3 and 4.
The results show that the proposed COSF controller has better performance in terms
of the settling time relative to the O-PI and the O-LQR controllers. It can be concluded
Fig. 3 Fuel cell and load

current
Fig. 4 Voltage tracking
that the proposed COSF controller ensures better robustness and has good satisfactory
response during disturbances and abrupt changes [22–24] (Fig. 5).
Test 2: in this test, a load variation of −50% −10% is considered at t = 0.05 s, then
the system response is given in Figs. 5 and 6. The results confirm that the output
response of the O-PI controller is affected seriously by the load resistance variations.
It can be concluded that the proposed controller ensures better robustness and has
satisfactory response in the presence of load variation. The proposed O-NL controller
has higher accuracy and reliability in tracking reference also in the presence of load
variation.
Figure 7 shows the values of performance indexes of each controller. This figure
confirm the simulation results carried out for the Test 1. It is evident that performance
Fig. 5 Fuel cell and load

currents response to load
variation
760 A. Dali et al.
Fig. 6 Voltage tracking

response to load current
variation
Fig. 7 Performance index
tf

index e = ∫Vcr e f (t) − Vc (t)dt with the proposed controller is smaller than
0
O-PI controller.
6 Conclusion
In this work, a robust nonlinear control design based on the PSO algorithm is
proposed. The main goal of the control was to ensure a well-regulated output voltage
for a Proton-exchange membrane fuel cell fed interleaved boost DC–DC converter.
The simulations and comparison of results showed the superior control performance
and the advantage of the proposed nonlinear compared to the optimized PI controller
in terms of tracking, smaller overshoot, faster response and better stability. In addi-
tion, the designed nonlinear controller is simple to implement in real time. As a part
of this works, it will be extended with T-S fuzzy technique to treat the problem of
fault diagnosis and fault tolerant control for various class of faults (actuator, sensor,
uncertainties, system, …) based on references [25–30].
References
1. Bankupalli P, Ghosh S, Kumar L (2018) Fractional order modeling and two loop control of
PEM fuel cell for voltage regulation considering both source and load perturbations. Int J
Hydrog Energy 43(12):6294–6309
2. Grötsch M, Mangold M, Kienle A (2009) Analysis of the coupling behavior of pem fuel cells
and dc- - dc converters. Energies 2(1):71–96
3. Li Q, Chen W, Wang Y (2011) Parameter identification for pem fuel-cell mechanism model
based on effective informed adaptive particle swarm optimization. IEEE Trans Ind Electron
58(6):2410–2419
4. Dali A, Abdelmalek S, Bettayeb M (2018) A new combined observer-state feedback (cosf)
controller of PWM buck converter. In: CISTEM
5. Abdelmalek S, Dali A, Bettayeb M (2018) An improved observer-based integral state feedback
(OISF) control strategy of flyback converter for photovoltaic systems. In: CISTEM
6. Choe S-Y, Lee J-G, Ahn J-W (2007) Integrated modeling and control of a pem fuel cell power
system with a pwm dc/dc converter. J Power Sources 164(2):614–623
7. Choe S-Y, Lee J-G, Ahn J-W (2014) Analysis, design and performance of a zero-current-
switching pulse-width-modulation interleaved boost dc/dc converter. IET Power Electron
7(9):2437–2445
8. Somkun S, Sirisamphanwong C, Sukchai S (2015) A dsp-based interleaved boost dc–dc
converter for fuel cell applications. Int J Hydrog Energy 40(19):6391–6404
9. Su J-T, Liu D-M, Liu C-W (2009) An adaptive control method for two-phase dc/dc converter.
In: International conference on power electronics and drive systems, pp 288–293
10. Ramakumar R, Chiradeja P (2004) Distributed generation and renewable energy systems. In:
2002 37th intersociety energy conversion engineering conference, pp 716–724
11. Kirubakaran A, Jain S, Nema RK (2009) The pem fuel cell system with dc/dc boost converter:
design, modeling and simulation. Int J Recent Trends Eng 1(3):157
12. Guo L, Hung JY, Nelms RM (2011) A dsp-based interleaved boost dc–dc converter for fuel cell
applications. Int J Hydrog Energy; Comparative evaluation of sliding mode fuzzy controller
and PID controller for a boost converter. Electr Power Syst Res 81(1):99–106
13. Abbaspour A, Khalilnejad A, Chen Z (2016) Robust adaptive neural network control for PEM
fuel cell. J Hydrog Energy 41(44):20 385–20 395
14. Romero A, Martínez-Salamero L, Valderrama H, Pallás O, Alarcón E (1998) General purpose
sliding-mode controller for bidirectional switching converters. In: Proceedings of the 1998
IEEE international symposium on circuits and systems. ISCAS’98, vol 6. IEEE, pp 466–469
15. Navarro-López EM, Cortés D, Castro C (2009) Design of practical sliding-mode controllers
with constant switching frequency for power converters. Electr Power Syst Res 79(5):796–802
16. Habiba M, Khoucha F, Abdelghani H (2017) Ga-based robust LQR controller for interleaved
boost dc–dc converter improving fuel cell voltage regulation. Electr Power Syst Res 152:438–
456
17. Souleman NM, Olivier T, Dessaint L-A (2009) A generic fuel cell model for the simulation of
fuel cell vehicles. Electr Power Syst Res 1722–1729
18. Dali A, Bouharchouche A, Diaf S (2015) Parameter identification of photovoltaic cell/module
using genetic algorithm (ga) and particle swarm optimization (PSO). In: 2015 3rd international
conference on control, engineering & information technology (CEIT). IEEE, pp 1–6
762 A. Dali et al.
19. Rezazi S, Hanini S, Si-Moussa C, Abdelmalek S (2016) Modeling and optimization of the
operating conditions of marrubium vulgare l. essential oil extraction process: kinetic parameters
estimation through genetic algorithms. J Essent Oil Bear Plants 19(4):843–853
20. Abdelmalek S, Belmili H (2015) A new robust h control power. In: Handbook of research on
advanced intelligent control engineering and automation, pp 601–623
21. Baadji B, Bentarzi H, Bakdi A (2019) Comprehensive learning bat algorithm for optimal
coordinated tuning of power system stabilizers and static VAR compensator in power systems.
Eng Optim. https://doi.org/10.1080/0305215X.2019.1677635
22. Bakdi A, Bounoua W, Mekhilef S, Halabi LM (2019) Nonparametric Kullback-divergence-
PCA for intelligent mismatch detection and power quality monitoring in grid-connected rooftop
PV. Energy. https://doi.org/10.1016/j.energy.2019.116366
23. Abdelmalek S, Rezazi S, Bakdi A, Bettayeb M (2019) Voltage dips effects detection and
compensation for doubly-fed induction generator based wind energy conversion system. Revue
Roumaine Des Sciences Techniques-Serie Electrotechnique Et Energetique 64(3):199–204
24. Bakdi A, Kouadri A, Mekhilef S (2019) A data-driven algorithm for online detection of compo-
nent and system faults in modern wind turbines at different operating zones. Renew Sustain
Energy Rev 103:546–555
25. Abdelmalek S, Barazane L, Larabi A (2016) Fault diagnosis for a doubly fed induction
generator. Revue Roumaine Des Sciences Techniques-Serie Electrotechnique Et Energetique
61(2):159–163
26. Bendoukha S, Abdelmalek S, Abdelmalek S (2018) A new combined actuator fault estimation
and accommodation for linear parameter varying system subject to simultaneous and multiple
faults: an LMIs approach. Soft Comput 1–14
27. Abdelmalek S, Barazane L, Larabi A (2017) An advanced robust fault-tolerant tracking control
for a doubly fed induction generator with actuator faults. Turk J Electr Eng Comput Sci
25(2):1346–1357
28. Abdelmalek S, Barazane L, Larabi A, Belmili H (2015) Contributions to diagnosis and fault
tolerant control based on proportional integral observer: application to a doubly-fed induction
generator. In: 2015 4th international conference on electrical engineering (ICEE). IEEE, pp
1–5
29. Abdelmalek S, Barazane L, Larabi A, Bettayeb M (2016) A novel scheme for current sensor
faults diagnosis in the stator of a dfig described by a TS fuzzy model. Measurement 91:680–691
30. Abdelmalek S, Azar AT (2018) A novel actuator fault-tolerant control strategy of dfig-based
wind turbines using takagi-sugeno multiple models. Int J Control Autom Syst 16(3):1415–1424
The Impact of the Solar Radiation Profile
on Sizing and Performance
of Photovoltaic Systems, Case Study
Tamanrasset, Algeria
Madjid Chikh, Aicha Degla, and Achour Mahrane
Abstract The sizing of photovoltaic installations requires a correct control of the

solar radiation profile; incorrect sizing leads to significant additional costs. To over-
come this problem, a comparative conceptual analysis of two design techniques based
on different solar radiation profiles was required. In this paper, we have defined two
different sizing techniques based on two solar radiation profiles; the first one is
based on hourly step and the second with daily step model. The sizing results were
established and analyzed in detail, and followed by a comparative analysis of the
performance of the photovoltaic systems under study. This analysis procedure is
applied to Tamanrasset region located in the south of Algeria.
Keywords Solar · Radiation · Sizing of photovoltaic system · Hourly profile ·

Daily profile
1 Introduction
Sizing and simulation of stand-alone photovoltaic (PV) systems requires a good

knowledge of the solar system variability at a given site, including the global solar
radiation models received on the PV module level.
In order to design stand-alone PV systems, several techniques have been devel-
oped and described in the scientific literature. Among these sizing techniques, there
M. Chikh (B) · A. Mahrane

UDES, Solar Equipment Development Unit, Route Nationale n°11, BP386, 42400 Bousmail,
Tipaza, Algeria
A. Mahrane
A. Degla
Renewable Energy Development Center, CDER, BP 62 Route de l’observatoire Bouzareah, 16000
Algiers, Algeria

https://doi.org/10.1007/978-981-15-6403-1_52
764 M. Chikh et al.
are two main techniques of major importance widely used by designers and users
of PV installations.
The first method requires a solar profile radiation with hourly time step variation;
the second one considers the amount of solar energy accumulated over a day.
However, these two techniques, which certainly lead to different sizes and dimen-
sions of the PV system, have very different torque values (Ps , C s ), [Ps : Peak power
of the PV system and C s : nominal battery capacity].
The objective of this work is to highlight the impact of these solar radiation profiles
on the design of PV systems and to see how the variability of this solar energy will
affect the performance and reliability of these PV systems. This analysis procedure
is applied to Tamanrasset region located in the south of Algeria.
Tamanrasset is an oasis city located in the south of Algeria, in the Ahaggar Moun-
tains. It has an altitude of 1320 m. Tamanrasset has a warm desert climate, with very
hot summers and mild winters. There is very little rain throughout the year, although
occasional rain falls at the end of summer because it is located in the northern
extension of the inter-tropical convergence zone.
2 Global Solar Radiation Estimation for Photovoltaic

Modules
2.1 Global Solar Radiation with Hourly Steps
The general problem when calculating hourly radiation on sloping surfaces when
the global radiation on a horizontal surface is the only known. For that purpose, we
require directions from which direct and diffuse components reach the surface in
question. Incident solar radiation is the sum of a set of radiation fluxes including
direct radiation, the three components of diffuse sky radiation and radiation reflected
from various surfaces as viewed from the sloped surface. The total incident radiation
on this area can be expressed as follows:
IT = IT,b + IT,d,iso + IT,d,cs + IT,d,hz + IT,r e f l (1)
where: iso, cs, hz and refl refer to isotropic, circumsolar, horizon and reflected
radiation respectively.
The global radiation on sloping surface is composed of three elements: Direct,
diffuse and reflected from the ground.
The geometric conversion factor Rb is the ratio of direct hourly (or instantaneous)
radiation on an inclined surface to radiation on a horizontal surface:
Rb = Ib,T /Ib = cosθ/cosθ Z (2)

The Impact of the Solar Radiation Profile on Sizing … 765
Fig. 1 Direct, diffuse and

reflected radiation of the
ground on a sloped surface
Radiation reflected from the ground is assumed diffuse and obtained by [1].

Ig,T = Iρg (1 − cos β) 2 (3)
where ρg is the ground albedo, and β is the slope angle.

Diffuse radiation is difficult for modelling, as its spatial distribution is usually
unknowns and time-dependent. Adding direct radiation and reflection terms from
the ground to the anisotropic model modified by Reindl, see Fig. 1, results in an
HDKR model (Hay, Davies, Klucher, and Reindl model) which can compute the
overall hourly solar radiation on a sloped surface Eq. (4).

(1 + cos β)
IT = (Ib + Id Ai )Rb + Id (1 − A I ) × 1 + f sin β 2
3
2

1 − cos β
+ Iρg (4)
2
where Ai is the Anisotropy Index.
2.2 Global Solar Radiation with Daily Steps
The global solar radiation on a sloped surface is evaluated by using corresponding

components calculated for a horizontal surface. The monthly daily average of the
diffuse radiation, the monthly daily average of direct radiation and the reflected
component from the ground are used to calculate the daily average global solar
radiation on a sloped surface is given by [2];
766 M. Chikh et al.
Hg = Hb Rb + Hd Rd + Hρ ((1 + cos β)/2) (5)
where:

Hd H = 0.741 − 0.552 n N
2 3
− 0.208 n N + 0.099 n N (6)
The average monthly direct solar radiation may be calculated using this equation:
Hb = H − Hd (7)
Rb is the ratio of direct radiation on sloped surface to direct radiation on a

horizontal surface, which can be calculated from [3]:
cos(ϕ − β)∂ sin ωs + (π/180)ωs sin(ϕ − β) sin ∂

Rb = (8)
cos ϕ cos ∂ sin ωs + (π/180)ωs sin ϕ sin ∂
ωs is the hourly angle of sunrise for a sloped surface and can be determined as
follows:
−1

cos (− tan ϕ tan ∂

ωs = min (9)
cos−1 (− tan(ϕ − β) tan ∂
The conversion factor of the diffuse radiation Rd can be determined using the
following equations [4].
1
Rd = (1 + cos β) (10)
2
The total solar radiation on a sloped surface was calculated using the angle of
inclination β from 0° to 90° with an interval of 1 to determine the optimal angle of
inclination. The angle at which H g produced the maximum value was considered the
optimal angle of inclination. This procedure was followed to obtain an optimal angle
of inclination for each month of the year, for the four seasons of the year, every six
months and for a full year.
3 Sizing Study of a Stand-Alone PV System
The main parts of the PV system are shown in Fig. 2. The system is controlled as
follows [5]: the battery imposes its voltage on the system, storage in batteries has
priority, and the battery is controlled between two state of charge limits: maximum
SOCmax and minimum SOCmin, corresponding respectively to the maximum Vbh
and min Vbl voltages of the battery.
Fig. 2 Main parts of the photovoltaic system
3.1 Description of the Sizing Method Based on the Hourly

Model of Solar Radiation
The electrical energy consumption profile used is as shown in Fig. 3. It is almost the
same for every day in the year, and is consistent with the profile usually found in
isolated sites.
In other words, there is a shortage when the power demand is higher than the power
supplied by the PV module and the battery is at the lowest level, so the management
forbids any energy supply to users. The design criteria selected are the shortage and
the cost of installation. At each moment when this occurs, the simulation program
counts it as a shortage time. The problem consists in determining the peak power Pc
of the PV field and the nominal capacity C s of the electrochemical storage system,
corresponding to the minimum total cost of the installation. The method adopted to
Fig. 3 Consomption profile

of a group of houses located
in Tamanrasset
768 M. Chikh et al.
perform the sizing of the PV system components considers that the peak power Pc
and the nominal capacity C s are variable, the other subsystems remaining at constant
cost. For this purpose, the overall cost of the installation of a PV system can be
expressed as follows:
C T = A Pc + B Q s + C T0 (11)
A: cost per PV peak watt (e/Wp), B: cost per kilowatt-hour of the battery (e/kWh),
Q s : the energy amount in the battery (kWh), C T0 is the sum of the costs of all other
subsystems, design studies, on-site system installation (e).
The simulation of the PV system operation allowed us to determine a series of
(Pc , C s ) corresponding to a certain number of shortages αe fixed in advance. Figure 4
shows the overall cost of the PV installation, the optimal point of (Pc , C s ), giving the
optimal size of the PV field and the nominal capacity of the electrochemical storage
battery, corresponds to the estimated minimum cost for the entire PV system.
This PV sizing procedure gives the optimal size of the PV field and battery capacity
and is repeated for several values of the number of shortages, namely 0, 100, 500
and 1000 h of shortages. Table 1 shows the optimized values of the PV system and
the corresponding minimum cost.
Fig. 4 The analytical method for determining the minimum cost
Table 1 Optimal values of the PV system sizing for different number of shortage hours
Nbre of shortage hours Peak power (kWp) Battery capacity (Ah) Minimal cost of PV
αe system (ke)
0 15.4 2300 72.6
100 10.3 2000 56.3
500 9.9 1200 42.0
1000 8.3 1000 35.0
3.2 Description of the Sizing Method Based on the Daily

Model of Solar Radiation
As measured data on diffuse solar radiation are not readily available for the study
site, well-known and generally adapted correlation methods were used to estimate the
components of daily diffuse solar radiation on a horizontal surface. Then, different
components of solar radiation on a sloped surface are determined in order to estimate
the overall solar radiation on that surface. Finally, the total radiation on different
sloped surfaces were obtained to estimate the optimal angle of inclination.
The PV field size is characterized by its peak power Pc (W c ) calculated on the
basis of a consumption profile (see Fig. 2) and by the global average daily radiation
taking into account a K r ratio (quality factor). This factor depends on multitude of
elements: sensor distribution, orientation and inclination of PV panels, quality of
diodes, cables and inverter… It generally ranges between 0.55 and 0.65 for well-
designed conventional installations, according to the International Energy Agency
(IEA). The peak power of the PV system is given by the following expression:
D E × HST C
Pc = (12)
Hg × K r
DE is the daily energy demand (wh); HSTC: solar radiation at STC conditions
(1000 w/m2 ); Hg: The average daily radiation, from the PV module of the worst
month (wh/m2 ) and estimated by the Eq. (5). The nominal storage capacity of the
battery is calculated from the number of days of system autonomy with a preset depth
of discharge (DOD = 80%). This capacity is given by the following expression:
As × D E
Cb = (13)
D O D × Vb
As : Autonomy of the storage system (days) and V b is the battery voltage (V), Table 2
shows the sizes of the PV system and the costs of the PV installation for 03 values
of the quality factor K r .
Table 2 Sizing of the PV

K r factor Peak power Battery capacity Cost of PV
system with a daily solar
(kWp) (Ah) system (ke)
radiation profile
0.55 23.3 4000 118.4
0.60 21.3 4000 114.0
0.65 19.7 4000 110.4
770 M. Chikh et al.
4 Sensitivity Analysis
4.1 Hourly Time Step Model
Figure 5 shows the variation of the installation cost as a function of the allowable
shortage duration.
Table 3 presents the optimal sizing and costs corresponding to different levels of
shortages.

C T is the absolute savings achieved if a certain shortage;
C T /C T represents
relative value of the economy. We can see the slope of the cost increasing as well as
the number of shortage hours decreases; the increase is sudden between 0 and 100 h
of shortage. This is perfectly illustrated in Table 4; we can observe that while the
cost of erasing shortage hour increases with decreasing shortage duration, it suddenly
increases when we move from 100 to 0 h of shortage. Table 5 shows the relative and
absolute savings corresponding to some levels of shortage, taking as a reference the
level α e = 100 h. The ratio α e /(α e , 100) corresponds to the factor of increase (or
decrease) of the shortage compared to the reference level 100 h shortage. It can be
seen that around this reference level, an increase in investment of 29% decreases
the shortage by a factor of 100 and a decrease of 25% leads to an increase in the
shortage by a factor of 5. The dimensioning corresponding to this shortage level could
serve as a realistic reference case for choosing the definitive size of the installation.
Fig. 5 Installation cost

versus the number of
shortage hours
Table 3 Investment costs and savings related to some levels of shortages

C T
Nbre of shortage hours αe PC (kWp) Cb (Ah) CT (ke) CT CT %
1 15.4 2300 72.6 0.00 0.00

100 10.3 2000 56.3 16.30 22.45
500 9.9 1200 42.0 30.60 42.15
1000 8.3 1000 35.0 37.60 51.79
Table 4 Erasing cost of the shortage time

Shortage interval Cost of erasing shortage hours (ke/h)
0–100 0.16
100–500 0.04
500–1000 0.02
Table 5 Relative and absolute economies related to some levels of shortage

C T αe
Shortage hours αe (h) CT (ke)
CT CT % αe,100
1 72.6 16.30 28.95 1/100

100 56.3 0.00 0.00 1
500 42.0 −14.30 −25.40 5
1000 35.0 −21.30 −37.83 10
Indeed, some users could admit 100 h of shortage; they correspond to 16 min of daily
shortage.
4.2 Daily Time Step Model
Figure 6 shows the variation of the installation cost as a function of the quality
factor K r .
It can be seen that the slope of the cost curve remains almost constant regardless
of the K r value; reflecting a linear variation in the system’s investment cost.
Table 6 shows the dimensions, costs and savings corresponding to different Kr
values. We notice that by increasing the value of K r , the savings become more
important; passing from K r = 0.6–0.65, the savings double in size.
Fig. 6 Installation cost

versus the quality factor K r
772 M. Chikh et al.
Table 6 Investment costs and savings related to some Kr values

C T
K r factor PC (kWp) C b (Ah) C T (ke)
CT CT %
0.55 23.3 4000 118.4 0.0 0.00

0.60 21.3 4000 114.0 4.4 3.86
0.65 19.7 4000 110.4 8.0 7.25
Table 7 Main reference

Configuration PC (kWp) C b (Ah) C T (ke)
configurations selected
Hourly profile of solar 10.3 2000 56.3
radiation
Daily profile of solar 19.7 4000 110.4
radiation
This means that more K r increases, savings become more significant; however,
according to the study done by [6], if K r is greater than 0.65 it does not mean that the
system is efficient. There is risk of under-designing i.e. energy demand may not be
met at every moment. However, it appears that a K r value closer to 0.65 characterizes
an optimized system.
Finally, the PV system performance analysis will be performed for the two struc-
tures based on two profiles (hourly and daily) of solar radiation profiles. The two
selected configurations are presented in Table 7.
5 Performance Analysis
In this section, we will see how the solar radiation profile can influence the perfor-
mance of PV systems. By using performance indices and selected indicator coef-
ficients published in IEA PVPQS Task2 [7, 8], we will proceed to quantify the
performance gaps between the two design techniques already described above. The
performance indicators generally used in the evaluation of a stand-alone PV are:
• The efficiency of the PV field related to the energy produced:
Ya = (E P V p,d )/P0 (h/KWp,d ) (14)
• The reference efficiency related to received energy of PV modules:
24
Yr = IT /IT,ST C (KWh/KW) (15)
1
• The final efficiency of the energy used by the consumer:

Y f = E P Vu,d /P0 KWh/KWp,d (16)
• The solar radiation pick-up losses by the PV array

L c = Yr − Ya KWh/KWp,d (17)
• System losses caused by the different equipment composing the PV system.

L S = Ya − Y f KWh/KWp,d (18)
• The Performance ratio
P R = Y f /Yr (19)
A stand-alone PV system, which is not working properly, will present a lower PR;
however, a higher PR value does not mean the system is working correctly. To have
an overview of the system’s use of the potential of solar energy, a new factor has been
implemented. It has been defined as a ratio of the PV power used by the load to the
PV power produced by the PV field. This factor is called UF or the use factor. It also
reflects the notion of self-consumption. It is calculated by the following expression:
E P Vu,d
UF = (20)
E P V p,d
More this factor is high, by reaching the unit; more the system presents very low
energy losses, therefore better performances.
If the UF factor can provide an idea of the overall system, it seems necessary to
define another factor to evaluate the performance of the PV field. It is the factor of
production that is the ratio between two performance indices Y a and Y r .
774 M. Chikh et al.
5.1 Variation of the PR Performance Ratio as a Function

of the UF Use Factor
Figure 7 shows the correlation of PR and UF for the two reference sizing cases (100 h
shortage and K r = 0.65). PR is more or less a linear function of UF. As the UF is
higher, the system uses the maximum of its solar potential. It is clear that the tendency
curve for the PR variation for the 100 h shortage case (hourly solar radiation profile)
has a higher directional coefficient than the K r case = 0.65 (daily solar radiation
profile); this shows that the sizing based on an hourly solar profile is more or less
consistent.
Fig. 7 Variation of PR as a function of UF
Fig. 8 Number of occurrences and cumulative frequencies of the PR performance ratio

Fig. 9 Variation of the PR

performance ratio as a
function of solar radiation
Figure 8 illustrates histograms variation of occurrence in defined PR ranges and

corresponding cumulative frequencies. We observe that more than 68% of the occur-
rence number registered for 100-hour of shortage is in a PR range between 0.6 and
0.8. These values provide information on the correct performance of the PV system.
On the other hand, for the reference system (K r = 0.65), 45% of recorded values
are in the PR range (0.4–0.5). This indicates that the PV system is oversized, causing
additional costs in the initial investment capital.
5.2 Variation of the PR Performance Ratio as a Function

of the Global Solar Radiation
Figure 9 shows that the PR is clearly dependent on the amount of energy received in
terms of PV modules. For the case of 100 h shortage, more solar radiation increases,
more the PR follows an increasing curve, whereas for the case of K r = 0.65, we can
see the opposite, which means that the system is oversized. In other words, not all
PV production was consumed; a significant part remained unusable.
It also shows how the PV system for the 100-hour shortage reference case
efficiently uses its potential for solar energy.
5.3 Calculation of Performance Gaps
To clearly demonstrate the effect of both solar radiation profiles on the performance
of PV systems, we have chosen to quantify this influence by identifying the gaps
between the PV system performances of the two sizing techniques.
776 M. Chikh et al.
Fig. 10 Variation in PR/PF system yield deviations
5.4 Gaps in PR/PF System Performance
Figure 10 shows the system’s performance gaps for each month of the year. There are
positive and significant differences over all months, exceptionally the one recorded in
February, which remains relatively low. It should also be noted that a peak difference
is reached in August, which is estimated at almost 35%, a significant value. For the
other months, the differences range from about 12 to 22%.
5.5 Differences in System Losses Ls
A similar observation is seen in Fig. 11, where the gaps between the losses in the
system caused by both reference systems are displayed (Case of 100 h shortage and
case of K r = 0.65). The negative peak is reached in August and it is estimated at
almost 02 h equivalent to the maximum energy produced by the system.
It is clear that these differences result in a gain, including energy produced by the
PV system and subsequently used by the consumer.
6 Conclusion
This paper highlights the effect of the solar radiation profile on the design and even
the simulation of energy systems, particularly PV systems. The most commonly used
profile in the design of PV systems is the daily solar radiation profile, it is found that
installations made are frequently oversized, generating significant cost.
Fig. 11 Differences in losses of the system Ls
To overcome this problem, we have established a comparative conceptual analysis

of two design techniques based on different solar radiation profiles. In our case, we
have established two sizing techniques based on two profiles of solar radiation, one
with hourly steps and the other based on a daily model.
The sizing results were established and interpreted in a detailed manner, followed
by a comparative analysis of the performance of the PV systems studied. The first
sizing technique using the hourly solar radiation model is based on accounting for
the number of shortage hours determined over a one-year period. The design results
showed that the sensitivity of the system for 100 h of shortage is the most important
with the lowest initial investment cost.
Similarly, the sizing technique using the daily profile of solar radiation is based
on the establishment of a quality factor noted K r . For this purpose, we have taken
three values of this factor, namely K r = 0.55, 0.60 and 0.65. The sizing results for
this technique revealed that for K r = 0.65, the sensitivity of the system is better.
The simulation of the PV system established for both sizing has shown that the
sizing based on the hourly model of solar radiation has better performance. The PR
performance ratio per 100 h of shortage (hourly solar radiation profile) changes in a
way that is consistent with the use of solar energy potential. The system’s efficiency
for 100 h of shortage is significantly higher compared to that of the daily profile
(K r = 0.65).
In general, the sizing technique based on the time profile of solar radiation is
the most consistent, with a low initial investment cost and a minimum of energy
produced by the PV field that is not used by the consumer. Considering all this,
modelling solar radiation at the hourly scale and even at the minute scale is essential
for the evaluation of solar energy systems.
778 M. Chikh et al.
Acknowledgments The National Office of Meteorology located in Tamanrasset, south of Algeria,

sponsored this study.
References
1. Aras H, Balli O, Hepbasli A (2006) Estimating the horizontal diffuse solar radiation over the
Central Anatolia Region of Turkey. Energy Convers Manag 47:2240–2249
2. El-Sebaii A, Al-Hazmi F, Al-Ghamdi A, Yaghmour SJ (2010) Global, direct and diffuse solar
radiation on horizontal and tilted surfaces in Jeddah, Saudi Arabia. Appl Energy 87:568–76
3. Duffie JA, Beckman WA (2013) Solar engineering of thermal processes. John Wiley & Sons
Inc, Hoboken
4. Jiang Y (2009) Estimation of monthly mean daily diffuse radiation in China. Appl Energy
86:1458–1464
5. Chikh M (1994) Analytical and conceptual study of power PV systems in Algeria, magister
thesis, CDER 1994
6. Chikh M, Mahrane A, Chikouche A (2008) A proposal for simulation and performance eval-
uation of standalone PV systems. In: 23rd European photovoltaic solar energy conference and
exhibition, 1–5 September 2008, Valencia, Spain, pp 3579–3584
7. Analysis of photovoltaic systems. Report IEA- PVPS T2-01 (2000)
8. Operational performance, reliability and promotion of photovoltaic systems. Report IEA-PVPS
T2-03 (2002)
Investigation and Prototyping
Implementation of a Novel Solar Water
Collector Based on Used Engine Oil
as HTF
Oussama Touaba, Salah Mohamed AitCheikh, Mohamed El-Amine Slimani,

Ahmed Bouraiou, and Abderrezzaq Ziane
Abstract This work is based purely on experimental results, where a Water Heater
Collector with Storage Tank (WHCST) was realized using used engine oil as a
basic material in the principle operating. After applying several experiments on the
type of used oil type S15 W40t was confirmed that it can be a good absorber of
sunlight and heat transfer simultaneously and it can be used in heating purposes.
Through experimental thermal properties of the proposed thermal collector. this
water heat system provides a significant efficiency since the temperature reaches
values greater than 40 °C for a 50 L water tank in less than four operating hours
under a moderate sunlight (less than 700 W/m2 ) and external temperatures less than
18 °C (In February).
Keywords Solar water heater · Solar energy · Waste engine oil · Solar thermal
collector
O. Touaba (B) · A. Bouraiou · A. Ziane

Unité de Recherche en Energies Renouvelables en Milieu Saharien (URERMS), 01000 Adrar,
Algeria
A. Bouraiou
A. Ziane
O. Touaba
Ecole Nationale Polytechnique d’Alger, 16000 Alger, Algeria
S. M. AitCheikh
Laboratoire des Dispositif de Communication et de Conversion Photovoltaïque, Ecole Nationale
Polytechnique d’Alger, 16000 Algiers, Algeria
M. E.-A. Slimani
Departement of Energetic and Fluid Mechanics, Faculty of Physics, University of Science and
Technology Houari Boumediene (USTHB), 16111 Algiers, Algeria

https://doi.org/10.1007/978-981-15-6403-1_53
780 O. Touaba et al.
1 Introduction
The renewable energy sector became an important sector due to GHG emission
reduction and decreasing fossil sources dependence [1]. Besides, imported fossil
fuels account for more than 95% of their primary energy sources in the world [2],
resulting in high electricity bills, energy system foreclosures, economies of scale,
low economies and low energy security. In the interest of sustainable development
to minimize the impact of pollution on the environment, we proposed using used
motor oils in heating system.
The RE application based on solar energy like photovoltaic and thermal is widely
used in the world in recent years [3, 4]. Solar heating is one of this energy appli-
cation that uses thermal energy from sunlight to heat the water or other fluids via a
collector [5–7]. Several studies concerning various kind of solar heater system were
performed in different locations [6–13]. The improvement and development of this
solar converter is to be exploited in several application and fields; in housing and
agricultural activities [3, 4, 14–18].
The solar water heater is generally composed of one or more solar collectors,
a storage tank and a heat transfer fluid circulation system which transfers the heat
generated by the solar collectors to the storage tank as presented in Fig. 1.
In this paper, novel solar water heater employing waste engine oil as absorber and
heat transfer fluid for water heating, is described.
This device is based on two systems which are thermal and photovoltaic, we will
focus on the characteristics of the thermal panel part.
The work is organization as follow:
– Materials and methods: proposed prototype and operation Mode and mathematical
formulation of energy performance.
– Experimental measures and results: Contain the Platform measurement descrip-
tion and heat thermal panel characteristic, discussion results of power and thermal
efficiency.
Fig. 1 Individual domestic

solar water heater
Investigation and Prototyping Implementation … 781
Fig. 2 Diagram illustrating of all system components
2 Materials and Methods
2.1 Prototype and Operation Mode
In this work, it is a matter of combining photovoltaic system and a thermodynamic

system to release a heating system which is mounted on a structure of aluminum.
The basic components of the system components presented in this study are
illustrated in Fig. 2.
The prototype thermodynamic collector proposed in this study consists of two
plates of tempered glass in which circulates the liquid (used oil) between them and
the distance between the two plates is 5 mm, an electric pump transfers the hot oil to
the water tank, and this system behaves like a greenhouse that absorbs and increase
the heat that converted in high thermal energy.
The heat obtained is transmitted through pipeline containing a hot fluid which is,
in our case, the used oil engine to other liquid (water). The heat exchanger made
from copper, and it is responsible of heat exchange inside the tank storage.
For this, the completion of a prototype has yielded considerable results; so many
experiments have been done on the prototype as shown in Fig. 3.in order to evaluate
this idea and the possibility of its future use in the industrial field.
The design parameters of the prototype are summarized in Table 1.
2.2 Energy Performance
The performance indices like thermal gain and thermal yield are expressed below.
The useful thermal output of the system is given by following relationship:
Fig. 3 Photograph of the proposed prototype
Table 1 Design parameter of the system

Symbol Characteristic Value
Sc Thermal surface exposed to light 1 m2
Ip Pipes length 400 cm
Se Heat exchanger surface 0.17 m2
Ve Volume of oil in the exchanger 0.3 L
Vt Volume of oil in pipes with 0.4 L
thermal insulation
Vc Volume of oil in heat collector exposed to light 3.4 L
Vo Total oil volume 4.1 L
Twt mwt Cpw

Qth = mwt Cpw = Tw,j − Tw,i (1)
t tj − ti
The instantaneous thermal efficiency is calculated using the following relation-

ship:

dqth Tw,j − Tw,i
ηth = = mwt Cpw (2)
Gi .A Gi .A tj − ti
While the overall thermal efficiency is calculated using the following relationship:
tf
dqth dt
ti
ηth = tf (3)
A ti Gi dt
Table 2 Monthly Average

G (KWh/m2 ) TAverage (°C)
day weather parameters
(2016) of Tipaza city January 118.55 15.66
February 119.01 15.1
March 157.36 16.12
April 146.55 19.2
May 144.04 20.75
Jun 161.28 26.95
July 166.84 22.5
August 176.29 21.36
September 166.14 26.27
October 153.48 20.5
November 106.85 18.32
December 119.30 16.55
3.1 Experimentation Sites
The experiments tests were held at the Solar Equipment Development Unit UDES
in the northern Algeria in Bouismail city,
Table 2 provides monthly Average day of weather parameters (Solar Irradiation
G, Tmin and maximum Tmax temperatures, relative humidity RH).
The values of this table were calculated from the daily measurements of the
meteorological station along all the days of the year.
3.2 Used Measurement Materials
The data acquisition equipment used in the experimental prototype for several
consecutive days to perform the necessary parameters (temperature, irradiation, level
heating water), and all the measurement which is exploited in this study, is shown in
Fig. 4 and presented as follow:
1. Data acquisition equipment
2. Weather station model meteorological station (NEAL) installed in UDES
3. Thermocouples type k
4. Computer.
Fig. 4 Data acquisition equipment used in the experimental prototype
In the beginning, we made an experimental test on the prototype at September with

a quarterly heat rate about 23 °C.
In order to assess the performance and effectiveness of the proposed solar thermal
collector, the following scenarios and steps have been developed.
• The first step is to test the prototype in typical sun radiation for a full day for
heating the water in order to know the level of storage water temperature and
the variation in the temperature of the used oil (heat transfer fluid) inside in the
thermal collector.
• The second step consists to measure the oil temperature in the inlet and outlet of
the thermal collector and comparing the temperature levels reached by the two
fluids (Storage water and spent oil), under a typical day and then calculate the
thermal efficiency of the prototype thermal collector
• We compare the effectiveness of the system by testing the prototype in two
different days with different levels of solar radiation by monitoring the level
of effectiveness for each day of measurement
4.1 Evolution of Water and Oil Temperatures
In this step, we carried out the tests on the prototype.

We tested the prototype by experimenting with heating water twice in a day, one
in the morning (see Fig. 5) and the afternoon (Fig. 6). This test for order to observe
the variation of the temperature of the water compared to oil temperature and this
happen by the exchange of heat between them.
Output oil temperature Input Oil temperature

Water temperatue Ambient temperature
100 1000
80 800
Temperatue (°C)
Power (W )
60 600
40 400
20 200
0 0
09:54
10:06
10:18
10:30
10:42
10:54
11:06
11:18
11:30
11:42
11:54
12:06
12:18
12:30
12:42
12:51
13:01
Time of day
Fig. 5 Temperature compared to the temperature of heat transfer oil in the morning
Input Oil temperature Output oil temperature

Water temperatue Ambient temperature
Thermal power
70 700
60 600
Temperatue(°C)
Power(W)
50 500
40 400
30 300
20 200
10 100
0 0
13:07
13:19
13:31
13:43
13:55
14:07
14:19
14:31
14:43
14:55
15:07
15:19
15:31
15:43
Time of day
Fig. 6 Water temperature compared to the temperature of heat transfer oil in the afternoon
The observation of the difference between the temperature of the hot oil coming
from the thermodynamic sensor to the water (blue Curve) and the temperature of the
oil coming out of the water after the exchange heat with water (red curve), see Fig. 5.
As shown in Fig. 5 at 9:54, the oil pump is triggered and the prototype begins
to operate. We note that the temperature of the water in the tank (Curve in green)
increases in parallel with that of the oil and that the oil (Curve in red) is heated
because of the heat absorbed by the sun.
When the oil comes out of the water tank (in blue), its temperature is close to
the temperature of the water. We conclude that the oil has to exchange most of its
temperature to water. We also note that the temperature difference between the input
oil and the output oil is 14 °C. Where we observe that the thermal power (Curve in
black) that the oil gives is increases and reaches 700w maximum. This is due to the
large difference between the heat of the water and then the thermal power decreases
to 100 W, due to saturation of the system despite the presence of solar radiation.
As shown in Fig. 6 we observe almost the same results obtained in the morning
experiment of Fig. 5 with a difference in the temperature of water and oil (the temper-
ature of oil and water in the first experiment is greater than the second) the water
temperature has reached to 58 °C and the temperature of oil in the thermal collector
output to 80 °C, but regarding in the experiment of Fig. 6, we note that the water
temperature does not exceed 46 °C and the temperature of oil 65 °C. This is caused
by the power of solar radiation in the period between 10: 00 and 13: 00; of course, the
thermal energy is less in the second experiment than the first experiment as observed
in (Curve in black) so that it is initially at 650 watts and then decreased rapidly to
reach 350 watts within one hour.
From this experience we conclude that the proposed system can be reheat the
water for domestic use twice a day and also it is concluded that the used oil, which is
used in this system, is the good absorbent solar thermal and good heat transfer fluid.
The average efficiency η (%) of the system of the experiment which detailed
in the previous figure (Fig. 7) reach an important value of 65%. To obtain a high
temperature of water heating we increase the space of the thermodynamic capture
exposed to the sun.
Thermal efficiency
90
80
70
Efficiency(%)
60
50
40
30
20
10
0
13:07
13:16
13:25
13:34
13:43
13:52
14:01
14:10
14:19
14:28
14:37
14:46
14:55
15:04
15:13
15:22
15:31
15:40
Time of day
Fig. 7 The evolution of thermal efficiency during the test day

Radiation A Radiation B
Water temperature A Water temperature B
1200 70
1000 60
Temperatue(°C)
50
Radia on(W/m²)
800
40
600
30
400
20
200 10
0 0
Fig. 8 The variation of water heating temperature for two days: (B) a day with low solar irradiation;
(A) a day with high solar irradiation
4.2 Effect of Solar Irradiation on Water Temperature
These measurements allow observing the efficiency of the system in the heating of
water through climate change.
Figure 8. shows that the water temperature (red curve) inside the tank in the
prototype increases when the solar radiation is large (black curve) and we note that
also the value of radiation in the curve exceeds 1000 w/m2 and the water temperature
reached 60 °C.
On the other hand, we took the same heating time for another day, when the solar
radiation is weak as shown in (blue curve) and its value in the rate of 600 w/m2 , in
addition, note that the temperature of water (green curve) inside the tank does not
exceed 43 °C.
5 Conclusion
In this paper, based on the experimental superficial study reached in this work. A
new water heater prototype based on used oil was presented. Several experimental
tests were conducted in the city of Tipasa to show the feasibility of this device.
Focusing on the thermal collector section of the prototype. The results obtained were
so satisfactory that we reached an acceptable heating temperature of 60 °C and 65%
heating efficiency rate, which is a very acceptable result so this innovative system
can be helpful. In the development of the field of solar heaters and conservation of
the environment successfully.
References
1. Sahouane N, Dabou R, Ziane A, Neçaibia A, Bouraiou A, Rouabhia A, Mohammed B

(2019) Energy and economic efficiency performance assessment of a 28 kWp photovoltaic
grid-connected system under desertic weather conditions in Algerian Sahara. Renew Energy
143:1318–1330
2. Bouraiou A, Necaibia A, Boutasseta N, Mekhilef S, Dabou R, Ziane A, Sahouane N, Attoui I,
Mostefaoui M, Touaba O (2020) Status of renewable energy potential and utilization in Algeria.
J Clean Prod 246:119011
3. Boumaaraf B, Boumaaraf H, Slimani ME-A, Tchoketch_Kebir S, Ait-cheikh MS, Touafek K
(2020) Performance evaluation of a locally modified PV module to a PV/T solar collector under
climatic conditions of semi-arid region. Math Comput Simul 167:135–154
4. Sellami R, Amirat M, Mahrane A, Slimani ME-A, Arbane A, Chekrouni R (2019) Experimental
and numerical study of a PV/Thermal collector equipped with a PV-assisted air circulation
system: configuration suitable for building integration. Energy Build 190:216–234
5. Harmim A, Boukar M, Amar M, Haida A (2019) Simulation and experimentation of an inte-
grated collector storage solar water heater designed for integration into building facade. Energy
166:59–71
6. Smyth M et al (2019) Experimental performance characterisation of a hybrid photovoltaic/solar
thermal façade module compared to a flat integrated collector storage solar water heater module.
Renew Energy 137:137–143
7. Li B, Zhai X, Cheng X (2018) Experimental and numerical investigation of a solar
collector/storage system with composite phase change materials. Sol Energy 164:65–76
8. Al-Kayiem HH, Yassen TA (2015) On the natural convection heat transfer in a rectangular
passage solar air heater. Sol Energy 112:310–318
9. Tewari K, Dev R (2019) Exergy, environmental and economic analysis of modified domestic
solar water heater with glass-to-glass PV module. Energy 170:1130–1150
10. Garnier C, Muneer T, Currie J (2018) Numerical and empirical evaluation of a novel building
integrated collector storage solar water heater. Renew Energy 126:281–295
11. Slimani ME-A, Sellami R, Mahrane A, Amirat M (2019) Study of Hybrid Photovoltaic/Thermal
Collector Provided With Finned Metal Plates: A Numerical Investigation under Real Operating
Conditions. In: 2019 International Conference on Advanced Electrical Engineering, ICAEE
2019, pp 1–6
12. Seddegh S, Wang X, Henderson AD, Xing Z (2015) Solar domestic hot water systems using
latent heat energy storage medium: a review. Renew Sustain Energy Rev 49:517–533
13. Touaba O, Cheikh MSA, Slimani MEA, Bouraiou A, Ziane A, Necaibia A, Harmim A (2020)
Experimental investigation of solar water heater equipped with a solar collector using waste
oil as absorber and working fluid. Solar Energy 199:630–644
14. Chaouch WB, Khellaf A, Mediani A, Slimani MEA, Loumani A, Hamid A (2018) Experimental
investigation of an active direct and indirect solar dryer with sensible heat storage for camel
meat drying in Saharan environment. Sol Energy 174:328–341
performance in a hybrid PV/T collector with double passage of Air. Int. J. Energy a Clean
Environ. 16(1–4):235–245
16. Bahria S, Amirat M, Hamidat A, El Ganaoui M, Slimani M (2016) Parametric study of
solar heating and cooling systems in different climates of Algeria – A comparison between
conventional and high-energy-performance buildings. Energy 113:521–535
17. Slimani ME-A, Sellami R, Mahrane A, Amirat M (2019) Experimental study of a glazed bi-fluid
(water/air) solar thermal collector for building integration. In: 2019 international conference
on advanced electrical engineering, ICAEE 2019, pp 1–6
18. Hmida A, Chekir N, Laafer A, Slimani MEA, Brahim A B, (2019) Modeling of cold room driven
by an absorption refrigerator in the south of Tunisia: A detailed energy and thermodynamic
analysis. J Clean Prod 211:1239–1249
Piezoelectric Energy Harvesting Based
Autonomous Vehicles’ Vibrations
Nadjet Zioui, Sousso Kelouwani, and Gilbert Lebrun
Abstract This work aims to analyze the impact of using carbon fibers based mate-
rials in the process of generating electrical power based energy harvesting. The
harvester is a piezoelectric element exposed to vibrations of a vehicle in motion. The
combination of fiber carbon material associated to the piezoelectric transducer are
used as a source to generate electrical energy to power cars onboard wireless sensors.
Many works treated the possibility to implement the vibrations energy harvesting
using piezoelectric components. The approaches are oriented towards flexible piezo-
electric materials subjected to flexion efforts for instance. In this work, we choose to
use a carbon based material to which we fix a mass placed on a piezoelectric material.
This combination is equivalent to a spring mass system. This paper highlights the
advantages of such a solution compared to a classic spring mass system.
Keywords Energy harvesting · Piezoelectric transducer · Carbone fiber material ·

Smart wireless sensors · Autonomous vehicles
1 Introduction and Recent Trends
Many sources of energy are available in nature. Some of them are unused and most
of them wasted every day. For many years, human being has been thinking about
ways to harvest energies. Sun panels are used to generate electrical energy from solar
source. Sun ovens are widely used in some isolated areas. Wind turbines are used to
extract the wind power in order to provide electricity or to pump water. Water flow
is used to drive turbines and generate electrical power. Sun, water, wind, hydrogen,
Bio sources … are all examples of successful attempts to harvest energy instead of
wasting it. Vibrations are practically present in every mechanical system in motion.
Application to onboard sensors energy generation.
N. Zioui (B) · S. Kelouwani · G. Lebrun

Department Mechanical Engineering, Université du Québec à Trois-Rivières, Trois-Rivières, QC,
Canada

https://doi.org/10.1007/978-981-15-6403-1_54
792 N. Zioui et al.
They are a huge potential source of energy. A vehicle in motion vibrates due to several
reasons, but mainly due to the engine and the mechanical structure of the vehicle
itself, as well as the contact with the road. This work addresses the harvesting of
this source of energy and its use to generate electrical power for vehicle’s onboard
wireless sensors. The paper explores more specifically, the use of carbon fiber in the
physical implementation for energy extraction.
Energy harvesting EH is a current interest for lots of industries. With the emerging
wireless technologies, phones, wearable devices, wireless sensors …etc., the need of
reliable and autonomous ways to provide energy, ideally wirelessly, is a fact. Most of
wireless and embedded devices rely on batteries to operate. However, the intelligence
growing and the sophisticated processing and applications that are involved come out
of a cost: energy. Batteries are starting to show limitations in terms of fulfilling smart
devices needs for processing and communicating, and people are starting to consider
alternative energy resources to compensate these limitations: solar chargers, solar
cellphones [1–4] … etc. Even some new industries are emerging using EH. Many
works on the literature highlighted the need of harvesting energy, especially for
wireless sensors applications. Song and Tan [5] treats the estimation of the energy
needed from wireless sensors in several situations, in order to evaluate the need of
energy for EH applications. Ruan et al. [6] focuses on an algorithm that deals with
the mismatch between the sensors demand and the available harvested energy from
plane aisle vibrations, using a piezoelectric device. Jushi et al. [7] presents the wind
prediction for the nearest future in order to harvest wind energy for wireless sensors.
Obaid and Fernando [8] highlights the importance of EH for rural communities.
Piezoelectric materials are very promising as potential source of energy [9]. In fact,
many works based on piezoelectric materials have been successfully conducted and
showed very promising results. Sharma and Baredar [3], Yesner et al. [10] introduced
piezoelectric material in shoes to generate power depending on the applied mass. A
very interesting work in [4] treats the energy harvesting based piezoelectric panels
do generate energy from the floor under the effect of people and cars motion.
In this paper, we briefly present the principle of EH based piezoelectric materials
to put into perspective the implementation solution that we analyze. In the second
section, we introduce a brief abstract of the analysis of energy consumption for
one type of wireless sensors based ZigBee network for design and implementation
purpose. The analysis concerns two parts of the device: the sensing part and the
wireless network part. In the third section, we present the modal vibration overview
available in a car in motion. This leads to the presentation and the design of the
carbon fiber material for the physical implementation. In section four, we recall
the piezoelectric model and highlight the electrical-mechanical equivalence model
method. The fifth section presents the results using the fiber carbon implementation.
We end this paper with a conclusion as a sixth and last section.
Piezoelectric Energy Harvesting … 793
2 Energy Consumption of a Wireless Sensor
One of the most important steps in designing harvesting systems for smart wireless
sensors is the knowledge of the amount of energy needed by the sensor. In fact, there
are several cases to consider while estimating the energy consumption: the sensor
behaving as an end device or as a router. For these categories, many works consider
the situations of the device powering up, transmitting data, receiving data, with or
without acknowledgement requirements. This last configuration is very important. In
fact, one of the important reasons of consuming all the battery energy or other energy
source is the phase where the device is looking for a network to join by iterating a
process of sending data and waiting for a response. Many works investigated the
power consumption for wireless networks. Song and Tan [5] analyzed the energy
consumption of ZigBee based wireless sensors. The paper also presented several
methodologies to simulate and estimate energy consumption for a wireless sensor.
Most of the studies rely mainly on the wireless device needs rather than the wireless
sensors’ needs in terms of energy. In fact, every sensor has its own operating principle,
its own conditioning circuitry and therefore its own needs of energy. This makes it
very challenging or even impossible to estimate with precision the amount of needed
energy for all kind of wireless sensors. In this work, we determine an estimate of
the energy needed for a particular wireless part of the sensor which is ZigBee based
network, according to the analysis of technical documentation and experimental data
for some of them from several manufacturers.
2.1 Consumption Analysis for the Wireless Layer
In order to estimate the energy consumption of the wireless devices, we are inter-
ested in looking into the technical data of a few wireless devices which some have
been experimentally validated. The selected devices are 2.4 GHz ZigBee protocol
devices. We motivate the ZigBee Radiofrequency technology choice by the low
energy consumption and a large range compared to other technologies for the same
amount of exchanged data.
The data that we seek concern the energy consumption of the wireless devices
during their different operating modes. Table 1 summarizes the technical data
collected from different ZigBee wireless based Radiofrequency devices, from
different manufacturers [11–16]. Apart from some odd values, for a typical oper-
ating voltage of 3.3 V, the sleeping mode implies approximately a 1.2 µA current.
The Reception mode needs around 26 mA and the Transmission mode uses around
33 mA. We base our design of the harvester on these collected values. However,
the current might reach much higher values during the particular phase of powering
up. Therefore, we need to consider this important parameter during the components’
choices and the design step as well. We can accomplish this part by adding a properly
sized capacitor.
794 N. Zioui et al.
Table 1 Energy consumption data for different RF ZigBee modules, for the sleeping, transmission
and reception modes
Devices Typical consumption data
(manufacturers) Operating Typical Rx Typical Tx Typical sleep
voltage (V) current (mA) current (mA) current (µA)
Device 1 (M1) 3.3 19 23 2
Device 2 (M2) 2.1–3.6 17 40 2
Device 3 (M3) 2.1–3.6 26 31 1
Device 4 (M4) 2.0–3.6 15.3 17 0.7
Device 5 (M4) 2.0–3.6 35 22 0.7
Device 6 (M4) 2.0–3.6 175 22 0.7
Device 7 (M5) 2.0–3.6 24 34 0.4
Device 8 (M5) 2.0–3.6 27 175 1.3
Device 9 (M6) 3.3 32 175 0.65
Device 10 (M1) 1.87–3.8 9.8 8.2 2.5
Device 11(M7) 2.1–3.6 30 44 1
Device 12 (M7) 2.1–3.6 34 100 2.4
Device 13 (M8) 2.1–3.6 28 67 0.4–225
2.2 The Sensing Layer
The sensing layer is composed of a transducer that generates electrical energy from
variations of force, or vibrations and a conditioner circuitry that is in charge of the
signals conditioning and adaptation. We assume the transducer as an ideal generator
and neglect the dissipations if any. The only part that is supposed to dissipate any
energy is the conditioner. With a good components sizing and design, the conditioning
circuit for the harvester can reach around 90% efficiency with a 100 mA load [17].
We assume for this study that the conditioner board will draw approximately 10%
of the available energy. Finally, the data and analysis show that a typical wireless
sensor needs in terms of energy are about 100 mW for the wireless module and about
20 mW for the harvester device. Which leads to total energy needs of about 120 mW
for the wireless sensor.
3 Vibrations Analysis for Vehicles
There are many works in the literature about the analysis, the modeling and the simu-
lation of the vibrations of vehicles. Most of the papers if not all of them concentrate
the study on the vibrations transmitted to the body based people comfort or for the
improvement of the vehicle’s structure [18]. The works focus mainly on the suspen-
sion response as a main source of vibration, and highlight the first mode of vibration
Fig. 1 Quarter, half and full car model for vehicles
as a main source. Some papers illustrate the analysis of the vibrations based ampli-
tude and frequency, but still consider the second mode and above as noise. In this
work, we make a synthesis of the common results of the modeling and simulation
works, regarding the available vibrations according to the road speed, the frequency
and the constraints directions. The idea from this is to bring out some insights on
the vibrations available as a source for energy harvesting in one hand. On the other
hand, this gives a tool to support the simulations thus the harvester design.
The main models that the researchers use are quarter, half or full models [19]
as illustrated in Fig. 1. Most of the models consider masses, springs and dampers
models, with more or less complex equations. The models are usually based on
Euler and Lagrange formulations, but there are some other available techniques like
pseudo excitation, decoupling and bound graph methods [20–22]. Barbosa [23] uses
a half car model and considers four degrees of freedom. The paper analyzes the
vertical and angular movement of the body. The paper points out some important
frequency analysis facts like the body’s frequencies that are around 1 to 2 Hz while
the suspension’s modes are around 12 Hz, which represents a difference of about
a decade. Another important result is that the frequency response closely depends
on the speed of the vehicle. In fact, the higher the vehicle speed is, the higher the
frequencies are where the peaks occur. [20] uses a decoupling method to highlight
the several modes of a car in terms of vibrations. The paper adds an additional degree
of freedom considering a man’s seat, thus focuses on the comfort in the development
work. The paper uses computation tools in a comparative way and points out five
modes. Three of the modes have values around 1 to 1.5 Hz while two of them
around 11 and 12 Hz, which joins the analysis results of [23]. Nahvi et al. [24]
evaluates the vibrations according to several parameters and road types. The paper
presents results for several speeds. The author underlines the random character of the
vibrations and points out the Gaussian aspect under some considerations such as low
IRI (International Roughness Index). Unlike the other works found in the literature,
[21] highlights an interesting point of the speed variations during the tests instead of
multiple constant speeds tests. Senthil Kumar et al. [18] presents the modal analysis
of the vibrations of a car using numerical methods. On the light of the several works,
796 N. Zioui et al.
we will consider two main frequencies for the car vibrations to be 1 and 12 Hz. We
base our design on these values.
4 Piezoelectric Element Model
There are several modeling methods for the piezoelectric element. In fact, this element
can be studied from the mechanical side approach or from the electrical behavior
side. The piezoelectric element produces mechanical force if an electrical voltage is
applied to it, and inversely, it can produce electrical energy if subjected to a variable
force.
The mechanical model can be established using several methods. We will consider
the electrical-mechanical equivalence in this section.
The electrical model states that the voltage produced by the piezoelectric element
is proportional to the strain applied to it [25].
U = −kx (1)
where:
U the voltage (V)
k the piezoelectric constant (V/m)
x the displacement (m)
The energy conservation states that:
U q = −F x (2)
where:
F the force (N)
q the electric charge (C)
From (1) and (2), it appears that the electrical charge is proportional to the
applied force for thickness mode vibration [25]. Moreover, if the harmonic exci-
tations are considered, then the voltage and the force will be expressed as (3) and
(4), respectively. Equation (5) is then obtained.
k
U =− v (3)
jω
k
F =− I (4)
jω
v the velocity (m/s)

I the electrical current (A)
Fig. 2 Piezoelectric model

U Z e −Z em I
= (5)
F Z em −Z m v
Z e Electrical impedance (V/A)

Z m Mechanical impedance (kg/s)
Z em Electromechanical coupling impedance (Vs/m)
Figure 2 presents the combination of the electrical and electrical-mechanical

equivalent model for the piezoelectric element.
5 Experimentation and Initial Results
One of the advantages in using carbon fiber material is the fact that the vibrations last
longer in the time, which is a huge advantage since the piezoelectric needs varying
constraints in the time. Moreover, the physical implementation is more optimal in
terms of effort direction. In fact, using a classical spring mass system makes it hard to
guarantee the effort’s direction, thus losses of energy is implied. Another advantage
of using carbon material instead of metal spring is the resistance of the material
against fatigue.
Our experimental platform is composed of a carbon fiber slat under a variable
charge due to the car vibrations. The experimental setup is illustrated in Fig. 3. It
has been implemented on a GMC Terrain 2011 car. Figure 4 illustrates the harvester
design that is composed of a ceramic piezoelectric plate associated carbon fiber slat.
The electrical power that is generated from the piezoelectric plate is limited using
the conditioner to a maximum of 5 VDC voltage.
The experiences were run on a flat asphalt road. The harvester is fixed on the front
of the vehicle and wired up to an acquisition card that is connected to a portable
computer.
We lead the experience with several car speeds and observed the generated energy
for each one of them.
798 N. Zioui et al.
Fig. 3 The experimental

setup
Fig. 4 The carbone fiber slat
Figure 5 summarizes the experiment results of the generated voltage wave form
with regard to the several speed values.
The results highlight a potential of energy harvesting using a car in motion vibra-
tions. The generated power depends on the piezoelectric plates that are used and the
characteristics of the material and its flexibility. We intend on conducting comparative
analysis of several types of piezoelectric materials in future works.
6 Conclusion and Future Work
This work presented a design and experimental setup of a harvester that aims to
provide energy to wireless sensors for autonomous vehicles. The work considers a
particular application of a vehicle on an asphalt road. The design of the harvester
is presented as a combination of two components that are the transducer and the
Fig. 5 Output observed with the experimental protocol

conditioner. Experimental setup is used to analyze results depending on the car

speed. The setup considered for the implementation uses a carbon fiber slat based
design in order to improve the results and benefit from the several advantages of
these materials such as the material resistance and flexibility. In future work, we
intend to implement additive and multi modal configurations in order to include all
potential frequencies signals in order to improve the amount of the harvested energy.
We also intend to study the distribution of the results according to several types of
vehicles and roads. Also, and more precisely, we will analyze the impact of each
communication mode of the sensor on the energy consumption, as well as several
types of piezoelectric materials.
References
1. http://downloadcenter.samsung.com/content/UM/200912/20091210082503578/GT-S7550_
UM_France_Fre_Rev.1.1_091105.pdf. Accessed 30 Oct 2019
2. https://www.amazon.ca/s?k=solar+cellphone&__mk_fr_CA=%C3%85M%C3%85%C5%
BD%C3%95%C3%91&ref=nb_sb_noss_2. Accessed 08 July 2019
3. Turkmen AC, Celik C (2018) Energy harvesting with the piezoelectric material integrated shoe.
Energy 150:556_564. Elsevier
4. Elhalwagy AM, Ghoneem MY, Elhadidi M (2017) Feasibility study for using piezoelectric
energy harvesting floor in buildings’ interior spaces. Energy Proc 115:114–126. Elsevier
5. Song J, Tan YK (2012) Energy consumption analysis of zigbee based energy harvesting wireless
sensor networks. In: 2012 IEEE International conference on communication systems (ICCS).
IEEE, pp 468–472
6. Ruan Z, Chew J, Zhu M (2017) Energy-aware approaches for energy harvesting powered
wireless sensor nodes. IEEE Sensors J 17(7):2165–2173
7. Jushi A, Pegatoquet A, Le TN (2016) Wind energy harvesting for autonomous wireless sensor
networks. In: 2016 Euromicro conference on digital system design (DSD). IEEE, pp 301–308
8. Obaid A, Fernando X (2017) Wireless energy harvesting from ambient sources for cognitive
networks in rural communities. In: 2017 IEEE Canada international humanitarian technology
conference (IHTC). IEEE, pp 139–143
9. Sharma PK, Baredar PV (2017) Analysis on piezoelectric energy harvesting small scale
device—a review. J King Saud Univ Sci
10. Yesner G, Jasim A, Wang H, Basily B, Maher A, Safari A (2019) Energy harvesting and
evaluation of a novel piezoelectric bridgetransducer. Sensors Actuators A Phys 285:348_354,
Elsevier
11. https://www.digi.com/pdf/dsxbee_3_zigbee_3:pdf. Accessed 08 July 2019
12. https://www.silabs.com/documents/public/data-sheets/TG-PM-0516-ETRX35x.pdf2f.
Accessed 08 July 2019
13. https://www.silabs.com/documents/public/datasheets/MGM111DataSheet:pdf. Accessed 08
July 2019
14. http://ww1.microchip.com/downloads/en/DeviceDoc/70329b.pdf. Accessed 08 July 2019
15. https://www.nxp.com/docs/en/data-sheet/JN5168-001-MXX.pdf, accessed: 2019-07-08
16. https://radiocrafts.com/uploads/RC2400RC2400HP_ZNMDataSheet:pdf. Accessed 08 July
2019
17. https://www.analog.com/media/en/technical-documentation/datasheets/35881fc.pdf.
Accessed 08 July 2019
18. Senthil Kumar M, Naiju C, Chethan Kumar S, Kurian J (2013) Vibration analysis and improve-
ment of a vehicle chassis structure. In: Applied mechanics and materials, vol 372. Trans Tech
Publ, pp 528–532
800 N. Zioui et al.
19. Raju B, Venkatachalam R (2013) Analysis of vibrations of automobile suspension system using
full-car model. International Journal of Scientific & Engineering Research 4(9):2105–2111
20. Xu B, Bao J, Wu J, Wang H (2015) Model analysis of car five degree of freedom vibration system
based on energy decoupling method. In: 2015 International conference on electromechanical
control technology and transportation. Atlantis Press
21. Guo L-X, Zhang L-P (2010) Vehicle vibration analysis in changeable speeds solved by
pseudoexcitation method. Mathematical problems in engineering
22. Tan D, Wang Q (2016) Modeling and simulation of the vibration characteristics of the in-wheel
motor driving vehicle based on bond graph. Shock Vib
23. Barbosa RS (2012) Vehicle vibration response subjected to longwave measured pavement
irregularity. J Mech Eng Autom 2(2):17–24
24. Nahvi H, Fouladi MH, Nor MM (2009) Evaluation of whole-body vibration and ride comfort
in a passenger car. Int J Acoust Vib 14(3):143–149
25. Staworko M, Uhl T (2008) Modeling and simulation of piezoelectric elements-comparison of
available methods and tools. Mech AGH Univ Sci Technol 27(4)
Faults Diagnosis-Faults Tolerant Control
(FTC)
Fault Detection of Uncertain Systems
Based on Interval Data Driven Approach
Chouaib Chakour
Abstract Kernel principal component analysis (KPCA) is an effective and efficient

data-driven approach for nonlinear processes monitoring. However, the KPCA algo-
rithm cannot be applied for monitoring uncertain processes data that is bounded by
intervals. The key idea of the present work is to propose an extended methodology of
conventional kernel PCA to handle nonlinear interval-valued data. This improves the
FDI decisions by considering the uncertainties in the analysis. The results for apply-
ing this algorithm on the nonlinear processes of the Tennessee Eastman benchmark
show its feasibility and advantageous performances.
Keywords KPCA · Symbolic PCA · Interval data · Fault detection
1 Introduction
Modeling and identification of an uncertain system are mainly concerned with char-
acterizing the unknown system on the basis of measured input–output data in an
uncertain environment. It is of fundamental importance in predictive control, fault
diagnosis, and decision analysis since most real-life systems are associated with
uncertainties due to noises, measurement errors, uncertain physical parameters,
incomplete knowledge, etc. [1]. This gives an uncertain form of input-output data.
The PCA is one of the most known and used linear data driven approach for process
monitoring. Thus a great difficulty in applying, the traditional PCA identification
technique is dealing with those uncertainties.
Each imprecise data can be identified by a couple of values. However, the values
are intervals, delimited by ordered couples of bounds referred to as minimum and
maximum [2]. The measurement data represented by mono-valued data is one of
the admissible measures that belong to a possible range of data. In fact, it is more
C. Chakour (B)
Department of Electronics and Communications, Kasdi Merbah University,
Route de Ghardaia, Ouargla, BP.511 30 000, Algeria

https://doi.org/10.1007/978-981-15-6403-1_55
804 C. Chakour
appropriate to study variation and variability of phenomena that are better described
by interval-valued data. In its standard form, PCA method is not designed to handle
interval type data. Including uncertainty of the sensors measurements in the analysis
requires extending the PCA methodology to the Symbolic Data Analysis (SDA) [8].
The Midpoints-Radii PCA method, a pioneer in symbolic PCA, is one of the best
known examples of this family of methods [5].
For modeling and fault diagnosis of some complicated cases in industrial processes
with nonlinear characteristics, PCA performs poorly due to its assumption that the
processes are linear. In fact, one of the most used nonlinear generalization of the
linear PCA is the kernel PCA method (KPCA) [3, 4]. Actually, extensive studies
have been conducted for effectively identifying linear uncertain systems, especially
with the symbolic PCA techniques [7], but are even non-existent for the nonlinear
case. The motivation of this paper is to construct reliable symbolic kernel PCA
model for the identification of uncertain nonlinear systems. The symbolic kernel
PCA methodology is developed on the basis of midpoints and radii of the uncertain
and/or interval nonlinear process data. The conventional kernel PCA is extended
to handle nonlinear interval-valued data. In other words, the linear symbolic PCA
methodology is developed and extended to the nonlinear case.
This paper is organized as following. In Sect. 2 the modeling of nonlinear interval
data with kernel PCA method is presented. Section 3 gives the application of the
proposed method for fault detection. Experimental results of the proposed KPCA for
fault detection of Tennessee Eastman process are given in Sect. 4. Finally, conclusions
are given Sect. 5.
2 Modeling Uncertainties as Intervals
2.1 Interval Data (Radii and Midpoints)
There are several measurement scales (or types of data), which are simply ways to
categorize different types of variables. The epistemic uncertainty, which is naturally
modeled by intervals, is called incertitude [6].
Let the matrix X ∈ Rn×m be the conventional data matrix of measurements, where
n is the number of samples and m is the number of process variables collected under
normal operating condition. Consider the vector of measurements at instant k:
x(k) = {x1 (k), x2 (k), . . . , xm (k)} (1)
where, x j (k) is the k − th observation of the j − th variable.

Fault Detection of Uncertain Systems Based on Interval Data Driven Approach 805

Let us note that an j − th interval-valued variable x j (k) is just a standard
variable but its values are intervals, delimited by ordered couples of bounds referred
to as minimum and maximum:

x j (k) ≡ x j (k), x j (k) , (2)
where, x j (k) < x j (k).

Indeed, the interval data matrix [X], having 2m columns and n rows, is formalized
as a set of interval-value variables for different instants of observation.
⎛ ⎞
x 1 (1), x 1 (1) · · · · · · x m (1), x m (1)
⎜ x 1 (2), x 1 (2) · · · · · · x m (2), x m (2) ⎟
⎜ ⎟
[X] = ⎜⎜
.. ..
.
.. ⎟
⎟ (3)
⎝ . . ⎠
.
x (n), x (n) · · · . . x (n), x (n)
1 1 m m

The generic interval-valued variable x j can also be expressed by the couple
x cj , x rj , where x cj represents the center of the interval and x rj its radius,
x j (k) + x j (k)
x cj (k) = (4)
2
and,
x j (k) − x j (k)
x rj (k) = (5)
2

The interval x j is given, from its center and radius, as follows:

x j (k) ≡ x cj (k) − x rj (k), x cj (k) + x rj (k) (6)
In fact, the interval data matrix [X] can be written in terms of midpoints and radii:

[X] ≡ Xc , Xr . (7)
2.2 Kernel PCA to Modelling Uncertain Nonlinear Systems
Consider a set of measurements on a system, in good working order, of n observations

{x1 , x2 , . . . , xn } in a given space R of dimension m. The KPCA algorithm allows the
mapping of the data from the input space to a feature space F and the application of
806 C. Chakour
PCA in F. However, the training samples {x1 , …, xn } ∈ Rm mapped to the feature

space F are represented by its images {φ(x1 ), …, φ(xn )}.
xi → φ (xi ) (8)
An uncertain nonlinear system is defined by intervals measurement data

{[x1 ], [x2 ], . . . , [xn ]} ∈ [X]. The interval-valued data [X] is implicitly transformed
into a functional space called feature space F, where the classical PCA is applied. Let
φ be the non-linear transformation of the interval observations [X] to the functional
space F, where for each interval [xi ] an interval image [φ (xi )] is associated.
[xi ] → [φ (xi )] (9)
As indicated before, the interval nonlinear data [xi ] can be written in terms of
midpoints and radii:
[xi ] ≡ xic , xri . (10)
c r
c [xi ] ≡
Then the transformation of each interval observation xi , xi into the fea-
r
ture space F is discribed by its image [φ (xi )] ≡ φ xi , φ xi .

xic → φ xic (11)

xri → φ xri (12)
The PCA is then performed to find the principal components of mapped samples
φ(xi ). Then, the sample covariance matrix in the feature space is given by:
1
n
Cf = φ(xi )φ(xi ) (13)
n i=1
Assuming {φ(x1 ), …, φ(xn )} have been mean-centered, the PCA is performed by

finding the eigenvalues > 0, and eigenvectors Pφ ∈ F\{0} satisfying P = C f P.
An important property of the feature space F is that it allows us to compute the
value of the inner product without having to carry out the nonlinear map φ explicitly.
However, the computational difficulties in the higher dimensional feature space can
be simplified into finding the dot products of vectors in the feature space by using a
kernel function K [10].
K(xi , x j ) =< φ(xi ), φ(x j ) > (14)

where K is the matrix of size n × n and of general term Ki j = k xi , x j .
⎛ ⎞
k (x1 , x1 ) k (x1 , x2 ) · · · k (x1 , xn )
⎜ k (x2 , x1 ) k (x2 , x2 ) · · · k (x2 , xn ) ⎟
⎜ ⎟
K=⎜ .. .. .. .. ⎟ (15)
⎝ . . . . ⎠
k (xn , x1 ) k (xn , x2 ) · · · k (xn , xn )
For interval observations, the centers and the midranges kernel matrices of [X] in
the feature space are defined separately as:
⎛ c c c c ⎞
k x1 , x1 k x1 , x2 · · · k (x1 , xn )
⎜ k xc2 , xc1 k xc2 , xc2 · · · k xc2 , xcn ⎟
⎜ ⎟
Kc = ⎜ .. .. .. .. ⎟ (16)
⎝ . ⎠
c. c c. c c. c
k xn , x1 k xn , x2 · · · k xn , xn
and, ⎛ r r r r ⎞
k x1 , x1 k x1 , x2 · · · k xr1 , xrn
⎜ k xr2 , xr1 k xr2 , xr2 · · · k xr2 , xrn ⎟
⎜ ⎟
K =⎜
r
.. .. .. .. ⎟ (17)
⎝ . ⎠
r. r r. r r. r
k xn , x1 k xn , x2 · · · k xn , xn
The interval covariance matrix, or the intervel kernel matrix, can be expressed in
terms of kernel matrices of midpoints and radii:

[K] ≡ Kc , Kr . (18)
The global uncertain kernel matrix, from which an uncertain kernel PCA model
will be constructed, is given by the hybridization between the midpoints kernel matrix
and the midranges kernel matrix.
Kg = Kc + Kr − 2Kcr Aφ (19)

where Kicrj = k xic , xrj , and Aφ = QPT . So that,
Kcr = Pcr QT (20)
The eigenvalues and eigenvectors decomposition of the matrix Kg will be cal-

culated to obtain a KPCA representation of the nonlinear uncertain data [X]. The
solution is given by the following eigensystem:
Kg Pφg = nφg Pφg (21)
φ φ φ
where g is the eigenvalues matrix of Kg and Pφg = [ p1 , . . . , pn ]T is the corre-
sponding eigenvectors.
808 C. Chakour
Kernel functions that have been used successfully in the literature is the Radial
Basis Functions (RBF):
2
K(xi , x j ) = exp(− xi − x j /2δ 2 ) (22)
where 2δ 2 = c is the width of the gaussian kernel.
3 Fault Detection
The PCA method models the correlations between the historical data when the pro-
cess is in normal operation. A violation of the correlation indicates an unusual situa-
tion because the data do not maintain their normal relations. The quadratic prediction
error is defined as the norm squared of the residual vector [10, 14],

m
SPE(k) = x(k) − x̂(k)2 = (ei (k))2 (23)
i=1
In fact, the interval [SPE] is defined by:

[SPE] ≡ SPE, SPE (24)
where,
SPE(k) = x(k) − x̂(k)2 (25)
and,
SPE(k) = x(k) − x̂(k)2 (26)
The process is considered in normal operation if,
SPE ≤ δ 2 (27)
where δ 2 is the detection threshold of the SPE.

δ 2 = (θ2 /θ1 ) χα2 θ12 /θ2 . (28)
m m
where θ1 = i=l+1 λi , θ2 = i=l+1 λi2 and λi is the eigenvalue of the covariance
matrix.
The KPCA based monitoring method is similar to that using the PCA, the square
of prediction error (SPE) in the feature space F. The SPE indicates the extent to
which each sample conforms to the KPCA model. In the feature space, the SPE is
defined as [15]:

n
l
(x)2 =
SPEφ = φ(x) − φ t2j − t2j (29)
j=1 j=1
where, n is the number of nonzero eigenvalues and:

φT
t = Pl φ(x) (30)

The interval SPEφ , in the feature space, is defined as:

SPEφ ≡ SPEφ , SPEφ (31)
where,
SPEφ (k) = φ(x) − φ
(x)2 (32)
and,
SPEφ (k) = φ(x) − φ
(x)2 (33)
If it is considered that φ(x) is zero mean and normally distributed, the confidence
limit for the SPE can be computed from its approximate distribution:
S P E lim ≈ γ χh2 (34)
where γ = b/2a and h=2a 2/b, a and b are the estimated mean and variance of the
SPE [11, 12].
4 Application to Tennessee Eastman Process
The TEP was developed by Downs and Vogel of the Eastman Company to provide a
realistic simulation for evaluating process control and monitoring methods. There are
five major units in TEP simulation (Fig. 1) a reactor, separator, stripper, condenser,
and a compressor. The process has 12 manipulated variables, 22 continuous process
measurements, and 19 composition measurements sampled less frequently [10].
The collection of uncertain measurements of the different variables from the
system to be monitored, in normal operation, makes it possible to obtain the interval
data matrix. The TEP was run for one hour, and we collected 1000 samples from 22
measurements. Considering an uncertainty δxi of the order of 15% of the measurement
variation range taken for each variable xi . Then we construct a new interval data
matrix of the process, with δxi being the radius of the data intervals (shown in Fig. 2).
In the modeling steps using MRKPCA and KPCA methods, the first 500 samples
of X were utilized to build the KPCA and MRKPCA models, thus, the rest 500
samples are used for the test. In fact, the matrix of interval observations [X] can,
810 C. Chakour
Fig. 1 Tennessee eastman process
Measurement lower bound Measurement upper bound

2830 70
Reactor pressure
2820
Reactor level
2810
65
2800
2790
2780 60
0 200 400 600 800 1000 0 200 400 600 800 1000
70 3360
Stripper pressure
60
Stripper level
3340
50
3320
40
30 3300
0 200 400 600 800 1000 0 200 400 600 800 1000
Observation number Observation number
Fig. 2 Interval data of reactor pressure, Reactor level, Stripper level and Stripper pressure mea-
surements
20
Upper SPE
18
Lower SPE
16 Control limit 95%
14
12
SPE PCA
10
0
0 100 200 300 400 500 600 700 800 900 1000
Observation number
Fig. 3 Evolution of the KPCA SPE for interval data under normal operation
0.025
Lower SPE
Upper SPE
0.02 Control limit 95%
Interval SPE [SPE]
0.015
0.01
0.005
0
0 100 200 300 400 500 600 700 800 900 1000
Observation number
Fig. 4 Evolution of the MRKPCA SPE for interval data under normal operation
therefore, be estimated from the l selected principal components (l = 6) correspond-

ing to the l largest eigenvalues of the covariance matrix. For the MRKPCA, the
covariance matrix is constructed from the Eq. (19).
In order to show that classical KPCA model is not appropriate to monitor uncertain
processes, Fig. 3 shows the monitoring performances of the PCA method when the
process is operating under normal condition. However, the false alarm rate provided
is undesirable. In this case, the detection index SPE based KPCA shows that the
system is faulty, knowing that the system works properly in this simulation part.
812 C. Chakour
MRKPCA SPE in fault-free case
0.02
SPE
0.01
0
0 100 200 300 400 500 600 700 800 900 1000
MRKPCA SPE in the case of fault
0.04
SPE
0.02
0
0 100 200 300 400 500 600 700 800 900 1000
Fig. 5 Evolution of the MRKPCA SPE for inaccurate acquired measurements
Table 1 False Alarms Rate (FAR) of KPCA and MRKPCA methods

Fault detection metric FAR(%) 95% FAR(%) 99%
PCA SPE lower 37 20
PCA SPE upper 34 22
MRPCA SPE lower 09 5
MRPCA SPE upper 11 6
Figure 4 shows that the SPE based MRKPCA method is more robust to false alarms
due to uncertainties.
Figure 5 shows the SPE index based MRPCA of inaccurate acquired measure-
ments, varying within the initial data intervals, in fault-free case and faulty case. It is
well illustrated that the robust MRPCA monitoring model is more robust to uncer-
tainties and sensitive to faulty measurement. The false alarms rate of both MRPCA
and PCA methods is shown in Table 1.
A fault affecting the variable x2 is simulated between the sample 700 and the
sample 1000 with the magnitude of 40% of the variation range of the D feed sensor.
The control limits are calculated at the confidence level of 95%. After time point
k = 700, it was found that the monitoring index, MRPCA-SPE, continuously exceeds
its threshold, which indicates a fault has been successfully detected.
5 Conclusion
In this paper, a new algorithm of kernel PCA (MRKPCA), based on midpoints

and radii, is proposed. Our study shows the advantages of the proposed method,
in comparison with the classical one, to model interval data and their impact on
the fault detection and isolation (FDI) steps. The Simulation results show that the
proposed method is more robust to uncertainty compared to the classical KPCA
method.
References
1. Chena Yu-Wang, Yang Jian-Bo, Pan Chang-Chun, Dong-Ling Xu, Zhou Zhi-Jie (2015) Iden-
tification of uncertain nonlinear systems: constructing belief rule-based models. Knowl-Based
Syst 73:124–133
2. Lauro NC, Verde R, Irpino A (2008) Principal component analysis of symbolic data described
by intervals. In book: Symbolic data analysis and the SODAS software, pp 279–311
3. Kramer MA (1991) Nonlinear principal component analysis using autoassociative neural net-
works. AICHE J 37(2):233–243
4. Scholkopf B, Smola A, Muller K-R (1998) Nonlinear component analysis as a kernel eigenvalue
problem. Neural Comput 10:1299–1319
5. Palumbo F, Lauro NC (2003) A PCA for interval-valued data based on midpoints and radii. In:
Yanai H, Okada A, Shigemasu K, Kano Y, Meulman J (eds) New developments in psychomet-
rics, Tokyo
6. Le-Rademacher J, Billard L (2012) Symbolic covariance principal component analysis and
visualization for interval-valued data. J Comput Graph Stat 21(2):413–432
7. Chakour C, Benyounes A, Boudiaf M (2018) Diagnosis of uncertain nonlinear systems using
interval kernel principal components analysis: application to a weather station. ISA Trans
83:126–141
8. Lauro NC, Verde R, Irpino A (2008) Principal component analysis of symbolic data described
by intervals. In: Symbolic data analysis and the SODAS software, pp 279–311
9. Chakour C, Harkat MF, Djeghaba M (2015) New adaptive kernel principal component analysis
for nonlinear dynamic process monitoring. Appl Math 9(4):1833–1845
10. Chakour C, Harkat M-F, Djeghaba M (2015) Neuronal principal component analysis for nonlin-
ear time-varying processes monitoring. In: Safe process 9th IFAC symposium on fault detection,
supervision and safety of technical processes. September 2–4, Paris
11. Box G (1954) Some theorems on quadratic forms applied in the study of analysis of variance
problems, I. Effect of inequality of variance in the one-way classification. Ann Math Stat
25:290–302
12. Joe Qin S (2003) Statistical process monitoring: basics and beyond. J Chemom 17:480–502
13. Dunia R, Qin SJ, Edgar TF, McAvoy TJ (1996) Identification of faulty sensors using principal
component analysis. AICHE J 42(10):2797–2812
14. Dunia R, Qin SJ (1998) Joint diagnosis of process and sensor faults using principal component
analysis. Control Eng Pract 6(4):457–469
15. Alcala Perez CF (2011) Fault diagnosis with reconstruction-based contribution for statistical
process monitoring. Thesis of the University of Southern California
Discrimination of Unbalanced Supply
and Stator Interturn Faults in Induction
Machines
D. Kouchih, R. Hachelaf, and N. Boumalha
Abstract This paper describes an efficient approach which is used for the discrim-
ination of unbalanced power supply and stator interturn fault in the star connected
induction motors (IM) using space harmonic components. For this purpose, an analyt-
ical method for the modeling of asymmetrical star connected IM is presented using
coupled circuits theory and taking into account the increase of the common point
voltage for unbalanced conditions. The equations which describe the state model
as well as the calculation of machine inductances are presented. The calculation of
inductances is based on the magnetomotive force (MMF) distribution through the
machine air-gap. It will be deduced that the sideband components around the space
harmonics contain useful information and they can be considered as a good inter-
esting tool for the discrimination of the stator asymmetries. Simulation results show
the consistency and the applicability of the proposed approach.
Keywords Component · Discrimination of faults · Unbalanced conditions ·

Interturn faults · Induction machines
1 Introduction
From a number of surveys, the stator faults of IM drives account approximately

40% of all failures and induce severe impacts on performances of induction motors
[1–4]. Faults detection is really important for improved system design, protection,
and conception of fault tolerant control. For this purpose, many techniques have
D. Kouchih (B) · R. Hachelaf

Department of Electronics, University Saad Dahlab, Blida, Algeria
R. Hachelaf
N. Boumalha

https://doi.org/10.1007/978-981-15-6403-1_56
816 D. Kouchih et al.
been proposed in the literature [5, 6]. The well-known method is based on the Motor
Current Signature Analysis (MCSA) [7, 8]. The classical MCSA has been widely used
due to its inherent advantages. In the literature, two kinds of harmonic components
in the stator current spectrum are proposed as faults indicators. The most references
focus on the sideband components around the fundamental of the stator current.
Others references use the sideband components around the space harmonics [9–11].
The principal disadvantage of the MCSA technique is to clearly discriminate the
unbalanced power supply from stator faults and the saturation of magnetic material
where all cases exhibit similar kind of current signatures [12, 13]. In this study, we are
interested to the analysis of the common point voltage of the star connected induction
machines. For the discrimination of stator asymmetries, it will be concluded that the
sideband components around space harmonics are considered as an efficient tool.
For the modeling of asymmetrical IM and the calculation of the common
point voltage, a new approach is adopted using coupled circuits theory. This
approach extracts unbalance signatures on electromagnetic torque, stator current,
rotor current and common point voltage. Simulation results show the consistency
and the applicability of the proposed approach.
2 Calculation of Machine Inductances
2.1 Stator Inductances
The spatial harmonics are caused according to:

(i) The windings distribution in a finite number of slots (distribution harmonics).
(ii) The slots opening (slotting effect) that alter the effective length of the air
gap (permeance harmonics) [14, 15]. The effect of permeance harmonics is
neglected in comparison with distribution harmonics.
To illustrate the calculation of the machine inductances, it is appropriate to
consider the elementary 2-poles, star connected three phase IM. The stator wind-
ings are concentric with consequent poles. Considering initially only one coil with
Nj turns of a generic phase « j » and followed by current ij . The MMF distribution,
through the machine air-gap, is illustrated by Fig. 1.
Fig. 1 MMF distribution of

a generic coil
Discrimination of Unbalanced Supply and Stator Interturn … 817
The harmonic mutual inductance between the two phases « i » and « j » is expressed
4 μ0 r L 1 qj
π
qi
π
L ji h = cosh(α j − αi ) N j sin(hβ j ) Ni sin(hβi ). (1)
π g h2 j=1
2 i=1
2
As a result, the total mutual inductance will be given by
∞
4 μ0 r L 1 π
qj qi
π
Li j = cosh(α j − αi ) N j sin(hβ j ) Ni sin(hβi ). (2)
π g h=1 h 2 j=1
2 i=1
2
For i = j, we obtain the magnetizing inductance. It’s expressed by
∞
4 μ0 r L 1 π
qi qi
π
L ms = Ni sin(hβ j ) Ni sin(hβi ). (3)
π g h=1 h 2 i=1 2 i=1 2
h Harmonic order
αs Angular position which locates any point along the circumference of the air-gap
from a fixed reference.
αj Value of αs in through the center of coil.
μ0 is the air magnetic permeability and g is the air-gap length.
Ni Number of turns for coil b;
L Magnetic length of the rotor;
r Average radius of the air-gap.
The harmonic mutual inductance between phase « j » and the rotor loop of order
« k » is
∞
4 μ0 r L 1 qj
π
L jr = cosh(θk − α j − δ) N j sin(hβ j ). (4)
π g h=1 h 2 j=1
2
θk Angular position of the rotor loop of order « k»;

αr Rotor loop pitch
αr
δ= . (5)
2
θk = θr + (k − 1)αr . (6)
θr Angular position of the first rotor loop.

3 Machine Equations
3.1 Stator Equations
For unbalanced conditions, we employ the known line to line voltages as inputs in
the state model of the machine. The stator voltage equations can be expressed by:
d[Φs ]
[u s ] = [Rs ][i s ] + . (7)
dt
Using phase voltages, the line to line voltages are expressed:
[u s ] = [T ][vn ]. (8)
[u s ] = [u ab u bc u ca ]T
[vn ] = [van vbn vcn ]T
[i s ] = [i as i bs i cs ]T
uab , ubc and uca are the line to line voltages;

van , vbn and vcn are the phase voltages of the network;
ias , ibs and ics are the line currents.
⎡ ⎤
+1 −1 0
[T ] = ⎣ 0 1 −1 ⎦. (9)
−1 0 +1
The matrix of resistances is expressed

⎡ ⎤
ras −r bs 0
[Rs] = ⎣ 0 r bs −r cs ⎦. (10)
−ras 0 r cs
ras , rbs and rcs are the resistances of stator windings.

The stator flux is defined by
[Φs ] = [T ][φs ]. (11)
[φs ] = [φas φbs φcs ]T
[φs ] = [L sr ][ir ] + [L ss ][i s ]. (12)
[L sr ] is the matrix of stator-rotor mutual inductances.

3.2 Rotor Equations
The rotor is modeled by an equivalent circuit containing q magnetically coupled

circuits. Each rotor loop consists of two adjacent bars and the two portions of the
end ring connect them. The rotor currents comprise of the q rotor loop currents plus
a circulating current in one of the end rings, ie . In a motor with a complete end rings,
ie would be equal zero. The rotor voltage equation is expressed:
d[φr ]
[0] = [Rr ][ir ] + . (13)
dt
With
[φr ] = [L r s ][i s ] + [L rr ][ir ]. (14)
[L r s ] = [L sr ]T (15)
[ir ] is the rotor vector current;

[L r s ] is the matrix of rotor-stator mutual inductances;
[Rr ] is the n by n symmetric matrix of the rotor resistances. In the case of healthy
rotor, it can be demonstrated that [16, 17]:
⎡ ⎤
R0 −rb 0 . . 0 −rb
⎢ −r R0 −rb . . 0 0 ⎥
⎢ b ⎥
⎢ . . . . . . . ⎥
⎢ ⎥
⎢ ⎥
[Rr ] = ⎢ . . . . . . . ⎥. (16)
⎢ ⎥
⎢ . . . . . . . ⎥
⎢ ⎥
⎣ 0 0 . . −rb R0 −rb ⎦
−rb 0 . . . −rb R0
With
R0 = 2(rb + re ). (17)
re is the end ring segment resistance and rb is the total bar resistance.
[L rr ] is the qxq symmetric matrix of the rotor inductances. In the case of healthy
rotor, it can be verified that [16, 17]:
⎡ ⎤
L kk + L 0 L km − lb L km . L km L km L km − lb
⎢ L −l L + L L −l . ⎥
⎢ km b kk 0 km b L km L km L km ⎥
⎢ . . . . . . . ⎥
⎢ ⎥
⎢ ⎥
[L rr ] = ⎢ . . . . . . . ⎥. (18)
⎢ ⎥
⎢ . ⎥
⎢ ⎥
⎣ L km L km . . L km − lb L kk + L 0 L km − lb ⎦
L km − lb L km . . . L km − lb L kk + L 0
With
L 0 = 2(lb + le ). (19)
Lkm : Mutual inductance between two healthy loops of orders « k » and « m » .

Lkk is the magnetizing inductance of each healthy loop, lb is the rotor bar leakage
inductance and le is the rotor end ring leakage inductance.
Applying the same method used for the calculation of rotor inductances, we obtain:
μ0 rl
L kk = αr (1 − αr /2π ). (20)
g
μ0 rl
L km = −(αr )2 /2π . (21)

g
The mechanical equation is
dΩ
J = Te − Tl − f v Ω. (22)
dt
J is the inertia of the rotor and the connected load, Te the electromagnetic torque,
Tl the load torque, Ω the mechanical angular speed and f v is the viscose friction
coefficient. The electromagnetic torque is expressed [18–20]
P ∂[L sr ]
Te = [i s ]t [ir ]. (23)
2 ∂θ
P is the number of poles pairs and θ is the electrical angular displacement of the
rotor.
3.4 Calculation of the Common Point Voltage
The machine common point voltage can be calculated using the stator voltages and
phase voltages of the network (Fig. 2).
Ts is the sampling time. The same voltages are expressed in terms of phase voltages
of the network and the common point voltage by
⎧
⎨ vao = van − von
⎪
vbo = vbn − von . (24)
⎪
⎩
vco = vcn − von
The common point voltage is expressed:

v jn − v jo
j=a,b,c
von = . (25)
3
« n » is the reference neutral point of the network.
The stator voltages are calculated using the numerical derivation of the stator flux:
⎧
⎪ Δϕas
⎪
⎪ vao = ras i as + T
⎪
⎪
⎪
⎨
s
Δϕbs
vbo = rbs i bs + . (26)
⎪
⎪ Ts
⎪
⎪
⎪
⎪ Δϕcs
⎩ vco = rcs i cs +
Ts
Ts is the sampling time.
Fig. 2 Star connected IM (a)
Common point
(o)
(b)
von
(c)
Network neutral point (n)

3.5 Modeling of Interturn Fault
To illustrate this interturn fault, we consider the following example, where the coil
U-V has four turns and occupied two slots. When, a short circuit occurred between
the contact points c1 and c2 , three turns in series are obtained. In addition, a new short-
circuited turn which we call the short circuited phase D is created and magnetically
coupled with all the other circuits (Fig. 3)
Generally, when N j f short-circuited turns are created, the fault current reaches
high values causing the destruction of the short-circuited turns and thus their setting
off during a short transitional period. Consequently, the interturn fault causes the
reduction of the number of turns in faulty phase and the alteration of the magnetizing
and mutual inductances. To characterize the Interturn fault, we define the fault factor:
Njf
k= . (27)
Ns
In interturn fault condition, the effective turns by coil can be expressed in terms
of k. Consequently, all the machine inductances will be updated using the effective
turns.
The proposed state model has been implemented in the MATLAB environment. The
induction machine used in this study is three phases 4 kW, 230/400 V, 50 Hz, 14.2/8.2
A, 2840 rpm, 2 poles, 30 rotor bars.
Fig. 3 Short-circuited coil C1

C2
U V
2
10
fs
0
fsh1 fsh2
Current ias (A)
10
-2
10
-4
10
0 500 1000 1500 2000
Frequency (Hz)
Fig. 4 Spectrum analysis of the stator current ias in healthy condition
4.1 Healthy Condition
The induction machine is supplied by balanced three phase power supply 230/400 V,
50 Hz. To show the impact of space harmonics effect, we take h = 55. The following
figure shows the spectrum analysis of the stator current ias (Fig. 4).
In healthy conditions, we note the presence of specific harmonics through the
spectrum analysis of the stator current. These harmonics vary particularly with the
slip of the machine. For the stator current, the space harmonics are characterized by
the frequency [15]

λq
f sh = (1 − s) ± 1 f s . (28)
p
λ = 1, 2, 3, . . . ., q is the number of rotor bars.

For λ = 1 and s = 5.33%, the first space harmonics are: fsh1 = 1370 Hz and
fsh2 = 1470 Hz.
4.2 Unbalanced Supply
Using the National Electrical Manufacturer Association (NEMA) definition of

voltage unbalance [19, 20], we consider a starting up with rated load under balanced
power supply of the studied IM. At time t = 1 s, an unbalance of +15% is occurred
on phases « b » and « c » . Te machine characteristics are shown as follows (Figs. 5,
6, 7, 8 and 9).
When IM operates under unbalanced power supply, a backward rotating field
will be produced. The interaction of the backward rotating field with the forward
one produces pulsating electromagnetic torque, decrease and ripple in speed. The
100
80
Torque (N.m) 60
40
20
-20
-40
0 0.5 1 1.5
Time (s)
Fig. 5 Electromagnetic torque for unbalanced supply
100
50
Current ias (A)
-50
-100
0 0.5 1 1.5
Time(s)
Fig. 6 Stator current ias for unbalanced supply
100 100
50 50
Current ibs (A)
Current ics (A)
0 0
-50 -50
-100 -100
0 0.5 1 1.5 0 0.5 1 1.5
Time (s) Time (s)
Fig. 7 Stator current ias and ibs for unbalanced supply

4
x 10
1
Rotor loop current (A)

0.5
-0.5
-1
0 0.5 1 1.5
Time (s)
Fig. 8 Rotor loop current for unbalanced supply

Common Point Voltage (V)
30
20
10
-10
-20
-30
0 0.5 1 1.5
Time (s)
(a)
2
10
0
10
-2
10
-4
10
0 500 1000 1500 2000
Frequency (Hz)
(b)
Fig. 9 Common point voltage for unbalanced supply: a common point voltage and b spectrum
analysis of common point voltage
common point voltage increases significantly with only the fundamental frequency.
All these characteristics are traduced by severe impacts such as vibrations, noise and
overheating of the machine, so a progressive destruction of this last.
4.3 Interturn Fault Condition
A short circuit of 20% is occurred on phase a. The induction machine is supplied

by balanced three phase power supply 230/400 V, 50 Hz. We simulated a starting
up mode with rated load. The machine characteristics are represented as follows
(Figs. 10, 11, 12, 13 and 14).
For interturn fault, the IM is subject to oscillating torque, increase of the common
point voltage, decrease and ripple speed. Therefore, the interturn fault causes severe
impacts such as vibrations, noise and overheating of the machine. The common
point voltage increases significantly with the fundamental frequency and sideband
components at ±2 f s around the space harmonics as indicated.
100
80
Torque (N.m)
60
40
20
-20
-40
0 0.5 1 1.5
Time (s)
Fig. 10 Electromagnetic torque for interturn fault
100
Current ias (A)
50
-50
-100
0 0.5 1 1.5
Time (s)
Fig. 11 Stator current ias for interturn fault

100 100
Current ics (A)

Current ibs (A)
50 50
0 0
-50 -50
-100 -100
0 0.5 1 1.5 0 0.5 1 1.5
Time (s) Time (s)
Fig. 12 Stator current ias and ibs for interturn fault
x 10
1
Rotor loop current (A)
0.5
-0.5
-1
0 0.5 1 1.5
Time (s)
Fig. 13 Rotor loop current for interturn fault
5 Conclusion
In this paper, a new approach for the discrimination of the stator interturn fault and
unbalanced power supply in star connected induction motors has been developed
using space harmonic components and the common point voltage. An analytical
method for the stator asymmetry is presented using coupled circuits theory. With
this method, different characteristics of unbalanced IM can be highlighted: pulsating
torque; stator and rotor currents, common point voltage and etc. For unbalanced
supply, the common point voltage is characterized only by its fundamental frequency.
The sideband components around the space harmonics at high order are considered
as good indicators of the interturn fault. Simulation results show the consistency and
the applicability of the proposed approach for the discrimination of stator asymme-
tries. Extensive experimental studies are necessary to full assess usefulness of the
proposed approach. This work must be performed to take into account the saturation

50
-50
0 0.5 1 1.5
Time (s)
(a)
2
10
fs
fsh1 fsh2
0
10 fsh1-2fs fsh2+2fs
-2
10
-4
10
0 500 1000 1500 2000
Frequency (Hz)
(b)
Fig. 14 Common point voltage for interturn fault: a common point voltage and b spectrum analysis
of common point voltage
of magnetic material, skin effect, skewed rotor, and slotting effects in the calculation
of the machine inductances.
Simulated Machine Parameters [16]
Stator phase resistance rs = 1.5950
Rotor phase resistance rr = 1.3053
Effective air-gap g = 0.35 mm
Stack length L = 125 mm
Rotor radius r = 37.35 mm
Stator leakage inductance Lls = 0.0040 H
Rotor leakage inductance Llr = 0.0033 H
Drive inertia J = 0.045 kg.m2
Friction coefficient fv = 0.0038 kg.m2 .s−1
Stator phase turns Ns = 124
Rotor bar resistance rb = 3.04E − 4
Rotor end ring segment resistance re = 8.75E − 7
Rotor bar leakage inductance lb = 5.16E − 7 H
End ring leakage inductance le = 1.59E − 9 H
References
1. Bonnett AH, Soukup GC (1988) Analysis of rotor failures in squirrel-cage induction motors.
IEEE Trans Indus Appl 24(6):1124–1130. https://doi.org/10.1109/28.17488
2. Tolyat HA, Lipo TA (1995) Transient analysis of cage induction machines under stator, rotor
bar and end ring faults. IEEE Trans Energy Conv 10(2):241–247. https://doi.org/10.1109/60.
391888
3. Stefan G, Jose MA, Bin L, Thomas GH (2008) A survey on testing and monitoring methods
for stator insulation systems of low-voltage induction machines focusing on turns insulation
problems. IEEE Trans Indus Electron 55(12):4127–4136 https://doi.org/10.1109/tie.2008.200
4665
4. Pinjia Z, Yi D, Thomas GH, Bin L (2011) A Survey of condition monitoring and protection
methods for medium-voltage induction motors. IEEE Trans Ind Appl 47(1):34–46. https://doi.
org/10.1109/tia.2010.2090839
5. Rangarajan MT, Sang BL, Greg CS, Gerald BK, Jiyoon YA, Thomas GH (2007) Survey of
methods for detection of stator-related faults in induction machines. IEEE Trans Industry Appl
43(4):920–933. https://doi.org/10.1109/tia.2007.900448
6. Arfat S, Yadava GS, Bhim S (2005) A review of stator fault monitoring techniques of induction
motors. IEEE Trans Energy Conv 20(1):106–114. https://doi.org/10.1109/tec.2004.837304
7. Benbouzid MH, Vieira M, Theys C (1999) Induction motors faults detection and localization
using stator current advanced signal processing techniques. IEEE Trans Power Elec 14(1):14–
22. https://doi.org/10.1109/63.737588
8. Henao H, Capolino GA, Razik H (2003) Analytical approach of the stator current frequency
harmonics computation for detection of induction machine rotor faults. In: 4th IEEE Interna-
tional symposium on diagnosis for electric machines, power electronics and drives, Atlanta.
https://doi.org/10.1109/DEMPED.2003.1234583, pp. 259–264
9. Fudeh HR, Ong CM (1983) Modeling and analysis of induction machines containing space
harmonics, part III: three-phase cage rotor induction machines. IEEE Trans Apparatus Syst
PAS-102(8):2621–2628. https://doi.org/10.1109/tpas.1983.317783
10. Neto LM, Camacho JR, Salerno CH, Alvarenga BP (1999) Analysis of a three phase induction
machine including space and time harmonic effects: the A, B, C reference frame. IEEE Trans
Energy Conv 14(1):80–85. https://doi.org/10.1109/60.749151
11. Khezzar A, Kaikaa MY, Oumaamar MEK, Boucherma M, Razik H (2009) On the use of slot
harmonics as a potential indicator of rotor bar breakage in the induction machine. IEEE Trans
Ind Electron 56(11):4592–4605. https://doi.org/10.1109/tie.2009.2030819
12. Lashkari N, Poshtan J (2015) Detection and discrimination of stator intertum fault and unbal-
anced supply voltage fault in induction motor using neural network. In: The 6th international
power electronics drive systems and technologies conference, Shahid Beheshti University,
Tehran, Iran. https://doi.org/10.1109/PEDSTC.2015.7093287, pp. 275–280
13. Refaat SS, Abu-Rub H, Saad MS, Aboul-Zahab EM, Iqbal A (2012) Detection, Diagnoses and
Discrimination of Stator Turn to Turn Fault and Unbalanced Supply Voltage Fault for Three
Phase Induction Motors. In: 2012 International conference on power and energy, Kota Kinabalu
Sabah, Malaysia. https://doi.org/10.1109/pecon.2012.6450347, pp. 910–915
14. Nandi S (2004) Modeling of induction machines including stator and rotor slot effects. IEEE
Trans Industry Appl 40(4):1058–1065. https://doi.org/10.1109/tia.2004.830764
15. Kouchih D, Boumalha N, Tadjine M, Boucherit MS (2016) New approach for the modeling of
induction machines operating under unbalanced power system. Int Trans Electric Energy Syst
26(9):1832–1846. https://doi.org/10.1002/etep.2171
16. Houdouin G, Barakat G, Dakyo B, Destobbeleer, E (2003) A winding function theory based
global method for the simulation of faulty induction machines. In: IEEE electric machines and
drives conference, vol. 1, Madison, Wisconsin, USA. https://doi.org/10.1109/iemdc.2003.121
1279, pp 297–303
17. Xiaogang L, Yuefeng L, Toliyat HA, El-Antably A, Lipo TA (1995) Multiple coupled circuit
modeling of induction machines. IEEE Trans Industry Appl 31(2):311–318. https://doi.org/10.
1109/28.370279
18. Krause PC (1987) Analysis of electric machinery. McGraw-Hill Book Company
19. Kandli S, Amol B (2010) Performance analysis of 3-phase asynchronous motor under various
voltage conditions. Archiv Appl Sci Res 2(2):380–387
20. Sobhan J (2013) Performance Evaluation of Three-Phase Induction Motor Fed by unbalanced
voltage using complex voltage unbalance factor. J Eng Interdiscip Res 64(1):31–37
Impact of the Stator Winding Topology
on the Fault Harmonic Components
in Induction Motors
Seddik Tabet, Adel ghoggal, and Salah Eddine Zouzou
Abstract This paper deals with analysis of the impact of the type of the induction
motor’s (IM) stator windings on the fault harmonic components (FHC). The model
is based on the know modified winding function approach (MWFA) to evaluate
the machine inductances. Both broken rotor bars and air-gap eccentricity faults are
considered. The numerical investigations and simulations set conclusions as far as
the winding topology impact is concerned.
Keywords SCIMs · MWFA · PSH · Slot skewing · FFT · Distribution and

winding function · BRB · Eccentricity · Concentric winding at a full step layer ·
Total-step two-ply nested winding · Concentric winding with two short-pitch
layers · Chain winding · Computer simulation
1 Introduction
Three-Phase IMs are widely used for the many industry applications such as mills,
electric heaters, conveyor belts, air conditioners, pumps, etc. This is because of its
robust construction and easy maintenance. For this reason, condition monitoring tech-
niques to detect incipient faults and prevent IMs from serious damages are particular
concern. The published researches on the fault diagnosis show that the majority of
these studies have been focused on detection of the electrical faults, including the
stator windings and the rotor bars faults [1]. As classified, broken rotor bars (BRB)
have various impacts on the motor performance. This paper provides an analytical
model of IM able to take into account various severity degrees of BRB fault. To
achieve such a task, the multiple coupled-circuit model (MCCM) is applied, the
rotor bars skew effect is taken into account by using 2D winding function theory.
S. Tabet (B) · A. ghoggal · S. E. Zouzou

Electrical Engineering Laboratory of Biskra LGEB, University of Biskra, Biskra, Algeria
A. ghoggal

https://doi.org/10.1007/978-981-15-6403-1_57
832 S. Tabet et al.
Table 1 The meaning of

FFT Fast Fourier transformer
letters and symbols
SCIMs Squirrel cage induction motors
MWFA Modified winding function approach
BRB Broken rotor bars
SE Static Eccentricity
DE Dynamic Eccentricity
ME Mixed eccentricity
PSH Principal slot harmonic
S Single layer winding (concentric winding)
T Double-layer lap winding (full pitch)
R Double-layer concentric winding
C Chain winding
E Stator slot
FHC Fault harmonic components
It is known that the rotor slot number and the magneto-motive force (MMF) have a
great impact on the inductances values and shapes and it on the inductances [2]. In
this work, a 2-D model analytical of IM is called up the comparison between various
winding types and their effect on the FHC, simulation results confirm the paper out
comes (Table 1).
2 Analytical Model for MWFA
We extend the approaches in [3] by using the axial dimension. Therefore, the Gauss’s
law, the integral of the magnetic flux density on closed surface S of a cylindrical
volume defined in comparison to the average radius of the air-gap r is null.

Bds = 0 (1)
s
By defining, at any of coordinates (ϕ, z), the effective air-gap function g the
magneto motive force F, the magnetic field intensity H, such as B = μ0 H and
H = Fg , the Eq. (1) can be written as
2π l
F(ϕ, z, θr )
μ0 r dz dϕ = 0 (2)
g(ϕ, z, θr )
0 0
Of the Amper’s law, where l is the length of the air-gap.

Impact of the Stator Winding Topology on the Fault Harmonic … 833

H (ϕ, z, θr )dl = J ds (3)
abcda Ω
J the current density and is a surface enclosed by the closed path abcda.
According to the MMF and the number of turns enclosed by the closed path abcda
and traversed by the same current i, (3) becomes
Fab (0, 0, θr ) + Fbc + Fcd (ϕ, z, θr ) + Fda = n(ϕ, z, θr )i (4)
By taking into account the permeability of the iron is infinity, Fbc and Fda are
null. Where n(ϕ, z, θr ) is called the 2-D spatial winding distribution [4]. The Eq. (4)
gives
Fcd (ϕ, z, θr ) = n(ϕ, z, θr )i − Fab (0, 0, θr ) (5)
We enter the average value of the reverse air gap function g −1 (ϕ, z, θr )
with
2π l
1 1
g −1 (ϕ, z, θr ) = { g −1 (ϕ, z, θr )dz}dϕ (6)
2π l
0 1
And using (2) and (5), the expressions give Fcd (ϕ, z, θr )
1
Fcd (ϕ, z, θr ) = n(ϕ, z, θr ) − n(ϕ, z, θr ) g −1 (ϕ, z, θr )idzdϕ
2πlg −1 (ϕ, z, θr )
(7)
By dividing the members of (7) by the current i, we can be obtained the 2-D
winding function
2π l
1
N (ϕ, z, θr ) = n(ϕ, z, θr ) − n(ϕ, z, θr ) g 1 (ϕ, z, θr )dzdϕ
2πlg (ϕ, z, θr )
1
0 0
(8)
It is to be noticed that this new expression does not hold any restriction as for the
axial uniformity, in particular in term of skewed slots and axial air-gap non uniformity
[5].
834 S. Tabet et al.
3 The Type of Winding
The SCIMs studied in this paper is a three-phase, 4-pole motor. The different structure
of the stator winding is presented in the Figs. 1, 2, 3 and 4. Where only the phase A
is considered.
Fig. 1 Single layer winding (concentric winding)of the stator phase A
Fig. 2 Double-layer lap winding (full pitch) of the stator phase A
Fig. 3 Double-layer concentric winding of the stator phase A
Fig. 4 Chain winding of the stator phase A

A. Single layer winding (concentric winding).

The three planes winding generally used for a q pair of notches per pole and per
phase, which allows to divide a group of reels, belonging to a phase zone. This
winding was used in some engines [6].
B. double-layer lap winding (full pitch).
On the winding diagram in Fig. 2, we can clearly see the phase zones that make
up the half-phases of the winding. all six groups of connected reels inscribed
in the same line and forming the half-phase of the winding have resulting FEM
equal in value and are in phase; for this reason the given winding allows to
connect the half-phases not only in series as shown in Fig. 2, but also in two and
in four parallel groups [6].
C. double-layer concentric winding.
The following stator winding arrangement.
D. chain winding.
The pattern of such windings shown in Fig. 4. Winding by the shape of its frontal
parts is called chain winding [6].
A. The distribution and the winding function
The figures below represent the winding and distribution function of the first stator
phase A. The machine is supposed to be symmetrical. The winding consists of four
coils per pole and per phase and a beam of w = 25 turns per slot. The calculation
gives us an average value of the distribution function <nA> equal to 1.5w in (a), (b),
(d), and equal to 3w in (c) (Fig. 5).
The inductances are the key points in a successful simulation of induction
machines. These inductances are under the influences of different factors, such as
winding distribution, stator and rotor slots, skewed slots of rotor, and asymmetries
caused by eccentricities faults. The Figs. 6 and 8 illustrate the impact of different
winding upon the inductances of the SCIMs.
There are three types of eccentricity fault: SE, DE and ME faults. Eccentricity
fault varies the air gap length and therefore changes the inductances of the motor.
In the SE fault, the air gap length is non-uniform but the minimum angular position
is constant. This fault happens because of the careless in fixing the rotor. Oval rotor
also leads to the SE fault. In the DE fault, the minimum air gap is a function of the
rotor position and rotates around the rotor. This may be arisen from misalignment,
shaft bending, bearing erosion or mechanical resonance at critical speed. This type
of eccentricity fault causes UMP in one direction and this result in shaft bending and
bearing tear. Presence of both SE and DE fault leads to the ME [7]. The Figs. 7 and
9 illustrate the impact of different winding upon the inductances of the SCIMs. The
four cases are considered with taking into account the slot opening and the rotor bars
836 S. Tabet et al.
(a) (b)
2π 2π
φ φ
(c) (d)
2π 2π
φ φ
(a) (b)
1.5W 1.5W
2π 2π
φ φ
-1.5W -1.5W
(c) (d)
3W 1.5W
2π 2π
φ φ
-3W -1.5W
Fig. 5 Turns functions n(φ) and winding functions N(φ) of stator phase a and rotor loop 1
-4
x 10
-4 x 10
4 4
(a) (b)
2 2
L (mH)
L (mH)
0 0
-2 -2
-4 -4
0 2 4 6 0 2 4 6
teta (rad) teta (rad)
-4 -4
x 10 x 10
4 4
(c) (d)
2 2
L (mH)
L (mH)
0 0
-2 -2
-4 -4
0 2 4 6 0 2 4 6
Fig. 6 a The inductance of Single layer winding (concentric winding), b double-layer lap winding
(full pitch), c the inductance of double-layer concentric winding, d the inductance of chain winding
-4 -4
x 10 x 10
4 4
(a) (b)
2 2
L (mH)
L (mH)
0 0
-2 -2
-4 -4
0 2 4 6 0 2 4 6
-4 -4
x 10 x 10
10 4
(c) (d)
2
L (mH)
L (mH)
5
0
0
-2
-5 -4
0 2 4 6 0 2 4 6
Fig. 7 Mutual inductance between rotor loop 1 and stator phases in different winding with 20% SE
and 20% DE a Single layer winding (concentric winding), b double-layer lap winding (full pitch),
c double-layer concentric winding, d the chain winding
skewing. We noticed that inductance (c) is the closest to the sinusoidal (Figs. 8 and
9).
It is now clear that the presence of PSH in line current of three-phase IM is
primarily dependent on the number of rotor slots and number of fundamental pole
pairs of the machine. It was also shown in previous works that only PSH associated
to a nontriplen pole pair can be seen ideally with a balanced power supply and in
symmetrical conditions. For more learning about this problem, the effect of pole
pair and rotor slot numbers on the presence of these harmonics under healthy and
eccentric conditions was mainly studied in [8] (Fig. 10).
B. Bar faults in different types of winding
In this context, the analysis of the simulation tests consists in particular of inspecting
the presence of lines characterizing default in the stator current spectrum. Since the
stator coupling adopted is stationary with free neutral. The Figs. 11, 12, 13 and 14
represent the broken bar default.
C. Evolution of default to different winding
By comparing the types of winding we conclude in broken rotor bars that there is
a small difference in the amplitude, the tendency and harmonic of default. And in
mixed eccentricity we notice that there is a difference in amplitude that we find the
concentric two-ply shortned winding has a weak harmonic compared to the other
838 S. Tabet et al.
-3 -3
x 10 x 10
dL/dteta (mH/rad)
1 1
dL/dteta (mH/rad)
(a) (b)
0.5 0.5
0 0
-0.5 -0.5
-1 -1
0 2 4 6 0 2 4 6
teta(rad) teta (rad)
-3
x 10
-4 x 10
1
dL/dteta(mH/rad)
dL/dteta (mH/rad)
(c) (d)
2 0.5
0 0
-2 -0.5
-4 -1
0 2 4 6 0 2 4 6
Fig. 8 a The inductance derivative of Single layer winding (concentric winding), b the inductance
derivative of double-layer lap winding (full pitch), c the inductance derivative of double-layer
concentric winding, d the inductance derivative of chain winding
dL/dteta (mH/rad)
dL/dteta (mH/rad)
0.05 0.05
(a) (b)
0 0
-0.05 -0.05
0 2 4 6 0 2 4 6
dL/dteta (mH/rad)
dL/dteta (mH/rad)
0.05 0.05
(c) (d)
0 0
-0.05 -0.05
0 2 4 6 0 2 4 6
Fig. 9 Mutual inductance derivative between rotor loop 1 and stator phases in different winding
a Single layer winding (concentric winding), b double-layer lap winding (full pitch), c double-layer
concentric winding, (d) the chain winding with 20% SE and 20% DE
windings and we also note that the amplitude of the harmonic at 20% SE and 30%
DE is large compared to 40% SE and 10% DE (Figs. 15, 16).
Amplitude (dB)
0
Amplitude (dB)
(a) (b)
-100 -100
-200 -200
0 1000 2000 3000 4000 0 1000 2000 3000 4000
Fréquence (Hz) Fréquence (Hz)
0
Amplitude (dB)
Amplitude (dB)
0
(c) (d)
-100 -100
-200 -200
0 1000 2000 3000 4000 0 1000 2000 3000 4000
Fréquence(Hz) Fréquence (Hz)
Fig. 10 Simulated normalized spectra of the stator current at steady state with different winding
a Single layer winding (concentric winding), b double-layer lap winding (full pitch), c double-layer
concentric winding, d chain winding
Amplitude (dB)
Amplitude (dB)
0 0
(a) -44.7382 -45.0724 (b) -45.7759
-45.6603
-50 -50
-100 -100
0 50 100 0 50 100
Fréquency (Hz) Fréquency (Hz)
Amplitude (dB)
Amplitude (dB)
0 0
(c) -45.9572 -44.8983 (d) -43.1445 -44.0548
-50 -50
-100 -100
0 50 100 0 50 100
Fig. 11 Partial break of a bar default
5 Conclusions
In this study, different winding topologys are considered to check the harmonic
content of the stator current of IM suffering from BRB and eccentricity fault. To
somme up no obvious difference can be classified here and the winding types can
not be considered as an influential parameter.
840 S. Tabet et al.
Amplitude (dB)
Amplitude(dB)
0 0
(a) -44.7062 -44.9191 (b) -45.6511 -45.6578
-50 -50
-100 -100
0 50 100 0 50 100
0
Amplitude (dB)
Amplitude (dB)
0 (d) -43.1282
(c) -45.9399 -44.2214
-44.815
-50 -50
-100 -100
0 50 100 0 50 100
Fig. 12 A Broken bar default
0
Amplitude (dB)
0
Amplitude (dB)
(a) -40.4136 -37.2794 (b) -41.614 -39.5883
-50 -50
-100 -100
0 50 100 0 50 100
0
Amplitude (dB)
Amplitude (dB)
0
(c) (d) -38.38 -38.37
-40.99 -38.3918 -20
-50 -40
-60
-100 -80
0 50 100 0 50 100
Fréquncy (Hz) Fréquency (Hz)
Fig. 13 Default of two bars broken

Amplitude (dB)
0
Amplitude (dB)
0 (b) -35.2542
(a) -33.6953 -32.9158 -32.4428
-50 -50
-100 -100
0 50 100 0 50 100
Amplitude (dB)
0
Amplitude (dB)
0 (d) -34.4673
(c) -37.3822 -34.0212 -36.0319
-50 -50
-100 -100
0 50 100 0 50 100
Fréquency (Hz) Fréquecy (Hz)
Fig. 14 Default of three bars broken
-30 -30
S T R C S T R C
Amplitude (dB)
Amplitude (dB)
-35 -35
-40 -40
-45 -45
-50 -50
0 20 40 60 80 100 0 20 40 60 80 100
default level (%) default level (%)
Fig. 15 Evolution of the default to the left (1 − 2 g) and right (1 + 2 g). For different levels of
broken rotor bars
-20 -20
S R C T S R C T
Amplitude (dB)
Amplitude (dB)
40%SE 40%SE
20%SE 20%DE -30 20%SE 20%DE
-30 30%DE 20%SE
20%SE 30%DE
30%DE -40 20%SE 30%DE
20%SE 40%SE
10%DE 40%SE
-40 10%DE 30%DE
-50 40%SE 30%DE
40%SE
10%DE 10%DE
-50 -60
30 40 50 60 70 30 40 50 60 70
default level (%) default level (%)
Fig. 16 Evolution of the default to the left (1 − 2 g) and right (1 + 2 g) for different levels of
mixed eccentricity
Appendix
Machines Parameters
3 phase; p = 3 KW; E = 36; Nb = 28; rs = 1.8; Ws = 25; p = 2; f = 50; J =
0.023976; fv = 0.00014439.
842 S. Tabet et al.
References
1. Ojaghi M, Yazdandoost N (2015) Oil-Whirl fault modeling, simulation and detection in sleeve
bearings of squirrel cage motors. IEEE Trans Energy Conv 30(4):1537–1545
2. Gyftakis KN, Kappatou J (2013) The impact of the rotor slot number on the behavior of the
induction motor. Adv Power Electron
3. Al-Nuaim NA, Toliyat H (1998) A novel method for modeling dynamic air-gap eccentricity in
synchronous machines based on modified winding function theory. IEEE Trans Energy Conv
13(2):156–162
4. Bossio G et al (2004) A 2-D Model of the induction machine. Extension of the modified winding
function Approach. IEEE Trans Energy Conv 19(1):144–150
5. Ghoggal A, Sahraoui M, Aboubou A, Zouzou SE, Razik H (2005) An improved model of the
induction machine dedicated to faults-detection—extension of the modified winding function.
In: Proceeding of IEEE ICIT 2005, Hong-Kong, China, pp 14–17
6. Kotenko G, Piotrrovski M (1979) Machines électrique. Edition MIR; Moscou
7. Faiz J, Ghasemi-Bijan M (2015) Estimation of induction machine inductances using three-
dimensional magnetic equivalent circuit. IET Electric Power Appl
8. Nandi S, Ahmed S, Toliyat H (2001) Detection of rotor slot and other eccentricity related
harmonics in a three-phase induction motor with different rotor cages. IEEE Trans Energy Conv
16(3):253–260
Sensor Fault Detection for Uncertain T-S
DC Model with Descriptor Observer
Approach
Moussaoui Lotfi, Aouaouda Sabrina, and Righi Ines
Abstract In this paper we present a sensor fault detection approach of a DC motor

described by a Takagi-Sugeno (T-S) multimodel with parameter uncertainties based
on a descriptor observer. Indeed the observer is synthesized with a guaranteed H∞
performance. A descriptor design approach is used by considering the sensor fault as
an auxiliary state variable. Then, the Lyapunov stabilization conditions are expressed
as a linear matrix inequality (LMI) formulation.
Keywords Fault detection · Takagi-Sugeno (T-S) model · Sensor faults ·

Descriptor observer
1 Introduction
Diagnostic, monitoring and fault detection methods play an attractive role in indus-
trial processes [1, 2]. Recently, several fault diagnosis methods have been developed
for linear models [3]. However, linear models do not give an accurate representa-
tion of real systems and almost physical dynamic systems cannot be represented
by linear differential equations. Moreover, the increasing demand for consistency,
availability and high system performance has led to the use of nonlinear models to
represent the systems. Hence obtained models are very complex and task of model-
based fault diagnosis becomes more difficult to achieve. However, the complexity
of the nonlinear models obtained present difficulties for the diagnostic and control
based on nonlinear models.
Recently, nonlinear systems described by T-S models have been considered
actively and specially in the fields of control, state estimation and diagnosis of
nonlinear systems. The popularity of T-S modeling framework is due, on the one
M. Lotfi (B) · A. Sabrina · R. Ines

Faculty of Science and Technology, University of Souk Ahras LEER, BP.1553, Souk Ahras
41000, Algeria
A. Sabrina

https://doi.org/10.1007/978-981-15-6403-1_58
844 M. Lotfi et al.
hand, to the property of universal approximation [4] and, on the other hand, to its
ease of manipulation from the mathematical point of view compared to the orig-
inal nonlinear models. In the literature, various diagnostic approaches have been
proposed for systems represented by T-S models [5, 6]. Unfortunately, there is little
literature dealing with the problem of diagnosing sensor or actuator faults when the
system model is subject to uncertainties or time-varying parameters. In general and
in most works, the uncertainties are defined as parameters of the model [7, 8], or as
inputs of the model [9].
Recent work has focused on estimating uncertain T-S system parameters based
on fault diagnosis [10].
In this paper we propose a descriptor observer to estimate the sensors faults of
uncertain T-S model with unmeasurable premise variables.
The observer is realized with a multi-objective requirement for a T-S model
subjected to sensor faults, external disturbances and uncertainties. For the estima-
tion procedure, the idea is to add the sensor faults to the state vector which will be
augmented and this system is called descriptor system [11]. The observer gains are
obtained by minimizing a norm of H∞, and the stability conditions has been devel-
oped by minimizing the Lyapunov quadratic error in the form of a linear Matrix
Inequality (LMI).
The outline of this paper is organized as follows: in Sect. 2, a description of uncer-
tain T-S system is given. In Sect. 3, a robust descriptor observer for fault detection is
given with unmeasurable premise variables. Section 4 presents an example to illus-
trate the effectiveness of the proposed methods. Concluding remarks are given in
Sect. 5.
The change of the system environment or the change of the system itself can cause
errors in the parameters of the model describing this system, this error is called the
uncertainties of the model [12]. The modeling of this uncertainty can be represented
as follows:
ẋ(t) = (A + A)x(t) + (B + B)u(t) (1)
where A, B represents the matrices of the nominal model and A, B the uncertainties
of the system.
The global uncertain continuous T-S model affected by sensor faults f s (t) and
external disturbances d(t) is represented as follows:

r
ẋ(t) = μi (ξ (t))((Ai + Ai )x(t) + (Bi + Bi )u(t) + Bd d(t))
i=1 (2)
y(t) = C x(t) + Ds f s (t)
Sensor Fault Detection for Uncertain T-S DC Model … 845
where x(t) ∈ R n is the state vector, u(t) ∈ R m is the input vector and y(t) ∈ R p is the
output vector. Ai , Bi , C, Bd and Ds are constant real matrices with appropriate dimen-
sions. ζ j ( j = 1, . . . , g) are the premise variables. Mi j (i = 1, . . . , r ; j = 1, . . . , s)
are the fuzzy sets and r is the number of rules.
Where:
wi (ξ (t))
μi (ξ (t)) = r
i=1 wi (ξ (t))

r

wi (ξ (t)) = Mi j ξ j (t)
j=1
r
i=1 μi (ξ (t)) = 1, and μi (ξ (t)) > 0 for i = 1, . . . , r
The uncertainties Ai ∈ Rn×n , Bi ∈ Rn×m are defined by:
Ai = AF A (t)E A
Bi = BF B (t)E B
where A, E A , B, E B : are known constant matrices.

And F A (t), F B (t): are the matrices uncertainties supposed unknown and bounded
verifying the following condition:
(F A (t))T F A (t) ≤ I, (F B (t))T F B (t) ≤ I
In order to estimate state vector x and sensor fault f s , an augmented system is

constructed using the descriptor technique. Then the T-S fuzzy system (2) can be
written as follows:

r
˙ =
Ēx̄(t) μi (ξ(t)) Āi x̄(t) + B̄i u(t) + Nxs + B̄d d(t)
i=1 (3)
y(t) = C¯x̄(t) = C0 x̄(t) + xs
where

x
xs = Ds f s (t), x̄ = ,
xs

I 0 Ai 0 Bi Bd
Ē = n , Āi = , B̄i = , B̄d =
0 0 0 −Ip 0 0

0
N= , C0 = C 0 , C̄ = C Ip
Ip
846 M. Lotfi et al.
Vector xs is considered as an auxiliary state of the augmented system (3). So if

the state estimation of the augmented system (3) exists, then the state estimation of
the original system (3) and the fault estimation exist too.
Based on this model, the following fuzzy observer structure is adopted

r
E ż(t) = μi (ξ(t)) Fi z(t) + Bi u(t)
i=1
ˆ
x̄(t) = z + Ly (4)
ˆ
where z(t) ∈ R n+ p is an auxiliary state vector of observer and x̄(t) ∈ R n+ p is the
state estimation of (3).
Lemma 1 ([13]). If there exist positive definite symmetric matrices P1 and P2 ,

Matrix R, non singular matrix M and scalar γ > 0 such that the following LMI is
satisfied for i = 1, . . . , r
⎡ ⎤
P1 A1 + (P1 A1 )T + Z 1 C + (Z 1 C)T + In ∗ ∗
⎣ Z 1T − P2 C Ai − Z 2 C −Z 2 − Z 2T ∗ ⎦ < 0 (5)
P1 −C T P2 −γ 2 I
Then there exist a fuzzy observer in the form (4) to asymptotically estimate the
state and fault sensor for fuzzy system (2). The observer parameters are given by:

Ai 0
Fi = (6)
−C −I p

0
L= (7)
IP

In + RC R
E= (8)
MC M
where
−1
M = P2−1 Z 2 − C P1−1 Z 1 (9)
R = P1−1 Z 1 M (10)
3 Design of Robust Descriptor Observer

with Un-Measurable Variables for Uncertain T-S System
In this section, we assume that the weighting functions that depend on the system
state, then the T-S uncertain system (3) become:

r
ẋ(t) = μi (x(t))((Ai + Ai )x(t) + (Bi + Bi )u(t)) + Bd d(t)
i=1 (11)
y(t) = C x(t) + Ds f s (t)
We obtain the following equivalent system with the weighting functions depending
on the estimated state.

r
ẋ(t) = μi x̂(t) ((Ai + Ai )x(t) + (Bi + Bi )u(t)) + Bd d(t) + w(t)
i=1 (12)
y(t) = C x(t) + Ds f s (t)
Then global system can be rewritten as follows:

r
ẋ(t) = μi x̂(t) ((Ai + Ai )x(t) + (Bi + Bi )u(t)) + B̄d v(t)
i=1 (13)
y(t) = C x(t) + Ds f s (t)
where: T
B̄d = Bd I and v = d T w T
And the augmented system becomes:

r
E¯x̄(t)
˙ = μi (ξ(t)) Āi x̄(t) + B̄i u(t) + B̄¯ d v(t) + D̄xs (t)
i=1 (14)
y(t) = C¯x̄(t) = C0 x̄(t) + xs (t)

xs (t) = Ds f s (t) ∈ R p , x̄(t)T = x(t)T xs (t)T ∈ R n+ p

In 0 Ai 0 Bi
Ē = , Āi = , B̄i = ,
0 0 0 −Ip 0

B̄d
B̄d = , B̄d = Bd I
0

0
D̄ = , C0 = C 0 , C̄ = C Ip
Ip
Ai = Ai + Ai , Bi = Bi + Bi ,
848 M. Lotfi et al.
Let us consider the following fuzzy

r
E ż(t) = μi (ξ(t)) Fi z(t) + Bi u(t)
i=1
ˆ
x̄(t) = z + Ly (15)
ˆ ∈ R n+ p is the
where z(t) ∈ R n+ p is an auxiliary state vector of observer and x̄(t)
state estimation of (14).
By substituting (14) into differential algebraic equation of the observer (4), we
obtain:

r
E x̄˙ˆ − E L C̄ x̄˙ = μi (ξ (t)) Fi x̄ˆ − LC0 x̄ − L xs + B̄i u (16)
i=1
Subtracting (16) from (14):

r
Ē + E L C̄ x̄˙ − E x̄˙ˆ = B̄¯ d v(t) + μi (ξ (t))
i=1

Āi + Fi LC0 x̄ − Fi x̄ˆ + (N + Fi L)xs + Ai x̄ + Bi u (17)
ˆ and suppose that:

Let the error estimation be defined by ē = x̄ − x̄,
E = Ē + E L C̄ (18)
Fi = Āi + Fi LC0 (19)
N = −Fi L (20)
The error dynamic can then be written as follows:

r
E ē˙ = μi (ξ (t)) Fi ē + B̄¯ d v + Āi x̄ + B̄i u (21)
i=1

r
E ē˙ = μi (ξ (t))Fi ē + M̄i
¯ (22)
i=1

M̄i = B̄¯ d Āi B̄i (23)
⎡ ⎤
v(t)
¯ = ⎣ u(t) ⎦

x̄(t)
By simplifying

r
ē˙ = μi (ξ (t))Si ē + Ḡ¯
¯ (24)
i=1
where:
Si = E −1 Fi (25)
Ḡ¯ = E −1 M̄i
V = ē T P ē > 0 (26)
With P > 0 and a positive scalar γ such the following condition:
V̇ + ē T P ē − γ 2
¯ T
¯ ≤0 (27)
With:

r
ē˙ = μi (ξ (t))Si ē + Ḡ
¯ (28)
i=1
And

r
¯ T Ḡ¯
T
ē˙ T = μi (ξ (t))ē T SiT + (29)
i=1

¯ T Ḡ¯ P ē + ē T P Ḡ¯
T
V̇ = ē T SiT P + P Si ē + ¯ (30)
And from (27), and (30) we have:

¯ T Ḡ¯ P ē + ē T P Ḡ¯
T
V̇ = ē T SiT P + P Si ē + ¯ + ē T ē − γ 2
¯ T
¯ ≤0 (31)
850 M. Lotfi et al.
From which we deduce that

SiT P + P Si + I P Ḡ¯
≤0 (32)
Ḡ¯ P
T
−γ 2 I
⎡ ⎤
sym(Ai P1 ) + sym(Z 1 C) + I ∗ ∗ ∗ ∗ ∗
⎢ Z T − P C A − Z C + I −Z − Z T + I ∗ ∗ ∗ ∗ ⎥
⎢ 1 2 i 2 2 2 ⎥
⎢ ⎥
⎢ Bd P1 −C Bd P2 −γ I ∗
2
∗ ∗ ⎥
⎢ ⎥≤0
⎢ P1 −C P2 0 −γ I ∗ 2
∗ ⎥
⎢ ⎥
⎣ Ai P1 −CAi P2 0 0 −γ 2 I ∗ ⎦
Bi P1 −CBi P2 0 0 0 −γ 2 I
(33)
⎡ ⎤
0 0 0000
⎢ 0 0 0 0 0⎥
⎢ 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 0 0⎥
Q(t) = ⎢ ⎥
⎢ 0 0 0 0 0 0⎥
⎢ ⎥
⎣ Ai P1 −CAi P2 0 0 0 0 ⎦
Bi P1 −CBi P2 0 0 0 0
⎡ ⎤⎡ ⎤T ⎡ ⎤⎡ ⎤T
0 0 0 0 0 0 0 0
⎢ 0 0 ⎥⎢ 0 0 ⎥ ⎢ 0 0 ⎥⎢ 0 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥
−1 ⎢ 0 0 ⎥⎢ 0 0 ⎥ −1 ⎢ 0 0 ⎥⎢ 0 0 ⎥
Q (t) + Q(t) < ε1i ⎢
T
⎥⎢ ⎥ + ε2i ⎢ ⎥⎢ ⎥
⎢ 0 0 ⎥⎢ 0 0 ⎥ ⎢ 0 0 ⎥⎢ 0 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥
⎣ A −CA ⎦⎣ A −CA ⎦ ⎣ 0 0 ⎦⎣ 0 0 ⎦
0 0 0 0 B −CB B −CB
⎡ ⎤⎡ ⎤T ⎡ ⎤⎡ ⎤T
P1 E TA 0 P1 E TA 0 P1 E BT 0 P1 E BT 0
⎢ 0 P E T ⎥⎢ 0 P E T ⎥ ⎢ 0 P E T ⎥⎢ 0 P E T ⎥
⎢ 2 A ⎥⎢ 2 A⎥ ⎢ 2 B ⎥⎢ 2 B⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ 0 0 ⎥⎢ 0 0 ⎥ ⎢ 0 0 ⎥⎢ 0 0 ⎥
+ ε1i ⎢ ⎥⎢ ⎥ + ε2i ⎢ ⎥⎢ ⎥
⎢ 0 0 ⎥⎢ 0 0 ⎥ ⎢ 0 0 ⎥⎢ 0 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥
⎣ 0 0 ⎦⎣ 0 0 ⎦ ⎣ 0 0 ⎦⎣ 0 0 ⎦
0 0 0 0 0 0 0 0
Simplifying and using Schur complement we obtain:

⎡
sym(Ai P1 ) + sym(Z 1 C) + I ∗ ∗ ∗
⎢ Z 1T − P2 C Ai − Z 2 C + I −Z 2 − Z 2T + I ∗ ∗
⎢
⎢ −C Bd P2 −γ 2 I ∗
⎢ Bd P1
⎢
⎢ P1 −C P2 0 −γ 2 I
⎢
⎢ 0 0 0 0
⎢
⎢ 0 0 0 0
⎢
⎣ E A P1 E A P2 0 0
E B P1 E B P2 0 0
⎤
∗ ∗ ∗ ∗
∗ ∗ ∗ ∗ ⎥⎥
∗ ∗ ∗ ∗ ⎥⎥
⎥
∗ ∗ ∗ ∗ ⎥
⎥<0
−γ 2 I + ε1i AAT + CAAT C T ∗ ∗ ∗ ⎥
⎥
0 −γ 2 I + ε2i BB T + CBB T C T ∗ ∗ ⎥⎥
0 0 −ε1i ∗ ⎦
0 0 0 −ε2i
(34)
4 Application to the Mcc
Let consider the uncertain T-S Model of the MCC

2
ẋ(t) = μi (ξ (t))((Ai + Ai )x(t) + Bu(t) + ECr (t))
i=1
y = C x + Ds f s (t)
With:

− Jf 1
z 1max − Jf 1
z 1min
A1 = 1 J ; A2 = 1 J
z
L a 1max
− LRa z
L a 1min
− LRa
1
0 −J 10 80
B = 1 ;E = ;C = ; Ds = ;
La 0 01 08
The parameters of the MCC are given in Table 1, (Figs. 1, 2).

The uncertainties are defined as follow:

0.1
Ai (t) = A(t) = F A (t) 0.1 0.1
0.1
852 M. Lotfi et al.
Table 1 DC Motor
Paramètres Désignation Valeurs
parameters
R Résistance de l’induit 5.438
La Inductance de l’induit 0.704 H
Ka Inductance Mutuelle 0.78 H
J Inertie du rotor 0.0217 kg m2
f Coefficient du frottement 0.00334 N.m.s
Paramètres Désignation Valeurs
R Résistance de l’induit 5.438
La Inductance de l’induit 0.704 H
Ka Inductance Mutuelle 0.78 H
J Inertie du rotor 0.0217 kg m2
f Coefficient du frottement 0.00334 N.m.s
Fig. 1 Speed and current state of MCC with estimated ones
And the sensors faults are:
f 1 (t) = 2sin(5t − 0.4) exp(0, 6t − 14)3s ≤ t ≤ 25s

⎧
⎨ 0.02(t − 1) 3s ≤ t < 10s
f 2 (t) = 0.01(t − 1) 10s ≤ t < 20s
⎩
0 other wise
Fig. 2 Sensors faults estimated with descriptor observer
From the LMI (34), we obtain:

0.0018 0.0110 0.0018 0.0110
P1 = ; P2 =
0.011 0.1449 0.011 0.1449

−1.0264 −113.0056 1.0347 113.0473
Z1 = ; Z2 =
113.0787 −0.5983 −112.9701 1.4548
then, we calculate G and M:

−0..0484 −0.6404
M = 10−3 ;
0.0084 0.0544

−0.5001 −0.0007
G=
0 −0.5
then, we obtain the observer gains:

854 M. Lotfi et al.
⎡ ⎤
−0.1539 47.8802 0 0
⎢ −1.4759 −7.7244 0 0 ⎥
K1 = ⎢⎣ −1
⎥;
0 −1 0 ⎦
0 −1 0 −1
⎡ ⎤
−0.1539 0 0 0
⎢ 0 −7.7244 0 0 ⎥
K2 = ⎢⎣ −1
⎥;
0 −1 0 ⎦
0 −1 0 −1
⎡ ⎤
0.4999 −0.0007 −0.5001 −0.0007
⎢ 0 0.5 0 −0.5 ⎥
E =⎢
⎣ 0 −0.0006
⎥;
0 −0.0006 ⎦
0 0.0001 0 0.0001
⎡ ⎤
0 0
⎢0 0⎥
L=⎢
⎣1
⎥
0⎦
0 1
5 Conclusion
In this paper, a robust fault detection observer has been designed for uncertain contin-
uous T-S models. Robustness with regard to external disturbances is realized through
L2 optimization approach. Sufficient conditions for the existence of the observer are
given in terms of linear matrix inequalities. At last a DC-Motor model is used to
illustrate the effectiveness of the proposed approach.
References
1. Patton R, Frank P, Clark R (1989) Fault diagnosis in dynamic systems: theory and application.
Prentice Hall International
2. Tanaka K, Hori T, Wang HO (2003) A multiple Lyapunov function approach to stabilization
of fuzzy control systems. IEEE Trans Fuzzy Syst 11(4):582–589
3. Aouaouda S, Chadli M, Karimi HR (2014) Robust static output-feedback controller design
against sensor failure for vehicle dynamics. IET Control Theory Appl 8(9):12
4. Ichalal D, Marx B, Maquin D, Ragot J (2011) Fault tolerant control for nonlinear systems
subject to different types of sensor faults. In: PAPYRUS Workshop on fault diagnosis and fault
tolerant control in large scale processing industries, Porticcio, Corse, France, October 2011
5. Isermann R (2007) Fault-diagnosis systems: an introduction from fault detection to fault
tolerance. Springer
6. Zhou B, Li ZY, Lin Z (2013) Observer based output feedback control of linear systems with
input and output delays. Automatica 49(7):2039–2052
7. Kruszewski A, Wang R, Guerra TM (2008) Non-quadratic stabilization conditions for a class

of uncertain nonlinear discrete-time T-S fuzzy models: a new approach. IEEE Trans Autom
Control 53(2):606–611
8. Chadli M, Abdo A, Ding SX (2013) H −/H∞ fault detection filter design for discrete-time
Takagi-Sugeno fuzzy system. Automatica 49(7):1996–2005
9. Yin S, Wang G, Karimi HR (2014) Data-driven design of robust fault detection system for wind
turbines. Mechatronics 24(4):298–306
10. Aouaouda S, Chadli M, Boukhnifer M (2016) Speed sensor fault tolerant controller design for
induction motor drive in EV. Neurocomputing 214:32–43
11. Bouattour M (2010) Contribution au diagnostic et à la commande tolerante aux
fautes des systemes flous de type T-S: approche descripteur. Thèse de doctorat de
L’ÉcoleNationaled’Ingénieurs de Sfax
12. Nagy-Kiss AM, Shutz G, Ragot J (2015) Parameter estimation for uncertain systems based on
fault diagnosis using Takagi-Sugeno model. ISA Trans 56:65–74
13. Haj Brahim I, Bouattour M, Driss M, Chaabane M, Hashim G Sensor faults observer design
with H∞ performance for non-linear T-S systems
14. Chadli M, Karimi HR, Shi P On stability and stabilization of singular uncertain Takagi–Sugeno
fuzzy systems
15. NagyKiss AM, Marx B, Mourot G, Schutz G, Ragot J Observers design for uncertain Takagi–
Sugeno systems with unmeasurable premise variables and unknown inputs. Application to a
wastewater treatment plant
16. Aouaouda S, Chadli M, Bouarar T Sensor fault observer design for uncertain nonlinear systems
17. Bouarar T, Marx B, Maquin D, Ragot J Trajectory tracking fault tolerant controller design for
Takagi-Sugeno systems subject to actuator faults
18. Hu J, Wen C, Li P (2013) Data-driven aided parity space based approach to fast rate residual
generation in non-uniformly sampled systems. In: 10th IEEE international conference on
control and automation (ICCA), China
19. Cherifi A (2017) Contribution à la commande des modèles Takagi-Sugeno: approche non-
quadratique et synthèse -stable. Thèse de doctorat de l’université de Reims Champagne-
Ardenne
20. Ichalal D, Marx B, Ragot J, Maquin D (2014) Fault detection, isolation and estimation for
Takagi-Sugeno nonlinear systems. J Frankl Inst 351(7):3651–3676
Detection of Stator Interturn Fault
in Double Star Induction Motors
R. Hachelaf, N. Boumalha, D. Kouchih, M. Tadjine, and M. S. Boucherit
Abstract In the present paper, the electromagnetic properties of a Double Star

Induction Machine (DSIM) are investigated experimentally by considering a short
circuit in the stator winding. The DSIM consists of two three-phase windings,
which are spatially phase-shifted by 30 electrical degrees. Usually, the study of the
short circuit fault in an induction machine is done by monitoring the stator current
harmonics. In contrast, the present work focuses on the experimental analysis of
flux density. The measurements are performed by applying a short circuit to a single
phase winding. Two flux sensors, which are located in the stator external periphery,
are used to measure flux density. In addition to the experimental tests, a theoretical
study has been carried out to predict the behavior of both flux density and stator
currents.
Keywords Component · Detection · Interturn fault · Double star induction motor
R. Hachelaf (B) · D. Kouchih

Department of Electronics, University Saad Dahlab, Blida1, Algeria
D. Kouchih
R. Hachelaf · N. Boumalha · M. Tadjine · M. S. Boucherit
M. Tadjine
M. S. Boucherit

https://doi.org/10.1007/978-981-15-6403-1_59
858 R. Hachelaf et al.
1 Introduction
The recent increase in both power consumption and the number of high power appli-
cations has led to the use of multi-phase machines including those with more than
three phases. Indeed multi-phase machines offer several advantages including power
segmentation without increasing the current per phase as well as the minimization of
the electromagnetic torque oscillations. As a result, multi-phase machines are often
used in naval propulsion and electric traction. However, some thermal, electrical,
mechanical or environmental factors may result in either stator or rotor faults thus
limiting the machine life span [1].
Thanks to new developments in data measurement equipment and software signal
processing, it is now possible to diagnose faults in electrical machines. New diagnosis
techniques are based on determining the frequency content of some machine signals
in order to evidence the defects.
According to the literature, the main signals used to obtain information on the
health status of a machine are the stator current, the electromagnetic torque, and the
mechanical vibration [2–4].
2 Modeling of a DSIM Short Circuit
Modeling a DSIM subjected to a stator winding inter turn short-circuit starts from
the DSIM model [5–7]. The double-star induction machine considered in the present
study has two systems of three-phase windings (sa1, sb1, sc1) and (sa2, sb2, sc2),
which are phase-shifted by an angle α = 30° (Fig. 1). The two systems are supplied
by with a sinusoidal voltage [8, 9].
A ‘short-circuit between turns’ is caused by an insulation defect between two turns
or more of the same stator phase. The value of the insulation resistance between the
short-circuited turns indicates the severity of the isolation fault. The part of the stator
winding that is short-circuited may be modeled by a resistor rf, which is in parallel
with the coil in default (Fig. 1). The winding of the phase is subdivided into two
parts: healthy (a1) and shorted (a1 ). The fault resistance rf, whose value ranges from
infinity for a healthy machine to zero for a short-circuit, is a global representation of
the local phenomena throughout the coil [10].
We define the parameter μ, which describes the relative number of shorted turns,
through the following formula [11]:

Nf = Na1 = μ NAS1 + NAS1 = μNs (1)
where Ns is the number of turns per phase and Nf the number of turns of the winding
part affected by the fault (sub coil as 2).
The equations governing the electric circuits in both the stator and rotor may be
written as follows:
Detection of Stator Interturn Fault in Double Star … 859
Fig. 1 Winding
sb1
representation of the DSIM
sb2
θ2
θ1
ra
sa2
Ns α
rb
sa1
rf
rc
   

  
sa1 sa1'
Nf
sc1
sc2

Vs Rs 07∗3 IS d φs
= . + (2)
Vr 03∗7 Rr Ir dt φr
With

φs L ss Msr I
= . s (3)
φr MsrT L rr Ir
The stator current and voltage vectors are denoted

T
Vs = vsa1 vsa1 vsb1 vsc1 vsa2 vsb2 vsc2

T
Is = i sa1 i sa1 i sb1 i sc1 i sa2 i sb2 i sc2
While the rotor vector are given by

T T
Vr = vra vr b vr c Ir = ira ir b ir c
The stator resistance matrix is given by:

T
Rs = rs diag (1 − μ) μ 1 1 1 1 1 (4)
Fig. 2 Evolution of 15
electrical and
10
electromechanical
Stator currents
magnitudes with 5% then 5
10% of c-turn circuit
0
-5
-10
-15
1.15 1.2 1.25 1.3 1.35 1.4
T (S)
a) Evolution of the stator currents.

3000
Rotor Current (A)
2000
1000
-1000
-2000
-3000
0 0.5 1 1.5
T (s)
b) Evolution of the rotor currents.
300
200
Torque (N.m)
100
-100
0 0.5 1 1.5
T (s)
c) Torque evolution.
1000
800
Speed (tr/mn)
600
400
200
0
0 0.5 1 1.5
T (s)
d) Evolution of speed.
where μ is the ratio of the number of turns shorted and the total number of the phase
turns
μ= n cc
ns
Ratio of short-circuit
ncc Number of turns short-circuited
ns Total number of stator windings
The rotor resistance matrix is given by:
T
Rr = rr diag 1 1 1 (5)
Inductance matrices are given:

[L ss ] = L ls (1 − μ)2 μ2 1 1 1 1 1
⎡ ⎤
(1 − μ)2 μ(1 − μ) − 1−μ 2
− 1−μ
2 (1
− μ)c1 (1 − μ)c2 (1 − μ)c3
⎢ μ(1 − μ) μ2 − μ2 − μ2 μc1 μc2 μc2 ⎥
⎢ ⎥
⎢ − 1−μ −2μ
−21 ⎥
⎢ 1 c3 c1 c2 ⎥
⎢ 2
μ ⎥
+ ⎢ − 1−μ − − 1
1 c c c ⎥
⎢ 2 2 2 2 3 1
⎥
⎢ (1 − μ)c1 μc1 c3 c2 1 − 21 − 21 ⎥
⎢ ⎥
⎣ (1 − μ)c2 μc2 c1 c3 − 21 1 − 21 ⎦
(1 − μ)c3 μc2 c2 c1 − 21 − 21 1
(6)
⎡ ⎤
L mr + L lr − 21 L mr − 21 L mr
[L rr ] = ⎣ − 2 L mr L mr + L lr − 21 L mr ⎦
1
(7)
− 21 L mr − 21 L mr L mr + L lr
⎡ ⎤
(1 − μ) cos( pθm ) (1 − μ) cos( pθm + 2π ) − μ) cos( pθm − 2π
3 (1 3
)
⎢ μ cos( pθm ) μ cos( pθm + 2π ) μ cos( pθm − 2π ) ⎥
⎢ ⎥
⎢ cos( pθ − 2π ) )
3
+ 2π
3
) ⎥
⎢ m cos( pθ m cos( pθ m ⎥
⎢ 3 3
⎥
[Msr ] = L sr ⎢ cos( pθm + 3 ) 2π
cos( pθm − 3 ) 2π
cos( pθm ) ⎥
⎢ ⎥
⎢ cos( pθm − α) cos( pθm − α + 3 ) 2π
cos( pθm − α − 3 ) ⎥ 2π
⎢ ⎥
⎣ cos( pθm − α − 2π 3
) cos( pθm − α) cos( pθm − α + 2π 3
) ⎦
cos( pθm − α + 2π 3
) cos( pθm − α − 2π 3
) cos( pθm − α)
(8)
constants of these matrices are:

2π 2π
c1 = cos(α), c1 = cos(α + ), c1 = cos(α − )
3 3
Let:

Vs Is φs
V = I = φ= (9)
Vr Ir φr

L ss Msr Rs 07∗3
L= R= (10)
MsrT L rr 03∗7 Rr
Equation (1) becomes:

dL dI
V = R + m I+L (11)
dθm dt
The mechanical equations give:
dθm
Jm + f v m = Ce − Cr (12)
dt
With m = dθdtm
the electromagnetic torque Ce is expressed as:
1 T dL
Ce = I I (13)
2 dθm
These equations can be written in matrix form as follows:

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
V L 010∗1 010∗1 I
⎣ −Cr ⎦ = ⎣ 01∗10 Jm d
0 ⎦. ⎣ m ⎦
dt
0 01∗10 0 1 θm
⎡ ⎤⎡ ⎤
R + m dθdL
010∗1 010∗1 I
⎢ m
⎥
+ ⎣ Ce = 1 IT dL I fv
2 dθm
0 ⎦⎣ m ⎦ (14)
01∗10 −1 0 θm
Or in condensed form as:

•
[U ] = [B][X ] + [A] X (15)
Or
•
X = [A]−1 [U ] − [A]−1 [B][X ] (16)
Numerical integration of the last equation provides the stator current as well as
both speed and position.
To investigate the fault between-turns, simulations are performed using short circuits
μ values ranging from 5 to 10%
We note that the amplitude of the current in phase with a short-circuit current
is higher than the other phases. However, The Current That Run through These
phases is magnified compared to the healthy snack. This amplification depends on
the number of turns shorted. We are seeing an increase in amplitude of the rotor
currents proportional to the defect. On the oscillations of the speed of rotation, it
increases with the number of turns shorted. In terms of torque, its average value is
equivalent to that of the healthy by (DSIM) cons we observe oscillations (noise) that
are gaining importance as and when the short circuit becomes important. The direct
analysis of electrical and electromechanical remains unclear for the study of defects.
In what follows we use the Fast Fourier Transform (FFT) to better quantify the
anomalies introduced by the defects by analyzing the particular lines of frequencies
near the fundamental.
4 Spectrum Analysis of Current
The analysis of stator currents in the frequency domain is the most used because the
spectrum reflects the existence of faults in an electrical machine.
The analysis of the Fig. 3a, b, c shows that the amplitude of the fundamental
frequency increases with the number of turns shorted, ditto for the rest of the
harmonic.
We can thus conclude that the defects of short-circuit between turns is defined by
a strong increase in the values of the harmonic current spokes (k * fs) with k: odd
number.
To validate our theoretical study, we conducted experimental tests on a prototype of

a double star induction machine, as having power 5.5 KW, 6-pole, 230/400 V [11].
The stator of the machine is wound so that all elementary coils are accessible.
This accessibility allows considering different rates of turns shorted.
The machine will have two additional windings in quadrature for measurement
of EMF Images stator flux, the two measuring coils cover the entire periphery of the
stator and the coils surrounding teeth cover a period of electrical machinery, 12 slots
stator (Fig. 4).
Fig. 3 Spectral analysis of 50

stator current with different
ratio of turns shorted X: 49.98
Y: 12.67
-50
-100
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz)
a. FFT spectrum of stator current in healthy mode.

50
X: 49.98
Y: 17.8
X: 150
Y: -35.71
-50
-100
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz)
b. Stator harmonic currents of the

short-circuited (ratio = 5%) phase winding
50
X: 49.98
Y: 20.57
0
X: 150
Y: -21.95
X: 249.9
-50 Y: -58.76
-100
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz)
c. Stator harmonic currents of the
short-circuited (ratio = 10 %) phase winding
Fig. 4 Winding of flux ɲ

density sensors
ɴ
0
0
A. Description of the Experiment Test
The experiments are carried out by supplying the DSIM with a sinusoidal voltage.
Two 3-phase transformers (Y coupled) are used to generate two three-phase volt-
ages shifted by 30 electrical degrees. The DSIM operates under nominal load. Diag-
nostic test is based on analyzing signals delivered by currents and voltages flux
density probes.
A progressive ratio of shorted circuited turns of one phase (phase 1) is considered.
Figure 6 shows clearly, that the magnitude of current increases gradually with the
ratio of short circuit turns, and the degree of distortion from sinusoidal becomes
larger. Changes in magnitude and phase are also observed in stator currents, which
are attributed to reduction in resistance and inductance of the winding.
Figure 6 concerns the spectrum profile of stator currents, before and after the
fault. Comparing Figs. 1 and 2, it is possible to identify that components harmonics
at frequency f = 150 Hz and 250 Hz are related to inter turn short circuit, their
amplitude increase with the ratio of short circuit, peculiarly in the phase concerned
on the fault, when the current amplitude is 10 (Fig. 5).
Moreover, we find that the currents increase especially at the or the phase involved
in the defect, by way of example, the current ratio may reach a figure of three for μ
= 33.33% and maximum of ten compared to the case when comparing healthy.
In general, the defects of short-circuit between turns are manifested by a sharp
increase in the values of harmonic currents rai (k * fs) with k: odd number. The
Table 1 summarize the numerical values of current amplitudes at harmonics and has
obtained from the experimental to try different modes of operation (healthy and with
default).
B. Stator Flux Measurement
Before defect, the shape of flux density voltage locus is a circle Fig. 8. By increasing
the ratio of short circuit, the distortion of the circle becomes larger Figs. 9 and 10. This
deterioration is attached to the distortion of the magnitudes and phases flux density
components. The flux trajectory is very sensible to the fault and to the geometrical

stator current 3
Stator current (A)

2
1
0
-1
-2
-3
-4
0 500 1000 1500 2000
T(s)
a- Stator current in healthy mode.
5
Stator Current (A)
-5
0 500 1000 1500 2000

T(s)
b- Stator currents of the short-circuited (ratio = 16.67%)
10
Stator Current (A)
-5
-10
0 500 1000 1500 2000

T(s)
c- Stator currents of the short-circuited (ratio = 33.33%)
position of the coils under fault relative to the flux sensors. For DSIM used in tests,
the first component of flux density sensor axis is superposed with those of the phase
concerned by the fault.
As we saw in the case of currents, higher harmonics, fundamental and order (3)
flow in the event of default and the harmonic (3) increases more significantly than
the fundamental and the rest harmonics (Fig. 7).
Fig. 6 Spectral analysis of stator current for a a healthy machine, b for machine with 16.67% short
circuited coil turn phase, and for c a machine with 33.33% short circuited coil turn phase
Fig. 7 Experimental signals measured through flux density sensors and corresponding locus
without fault
Table 1 Synthesis of experimental results

Motor states Double star induction machine (DSIM)
Turn short-circuit f s = 50 Hz 3 f s = 150 Hz 5 f s = 25 0 Hz
Magnitude (dB) (μ = 0%) −12.2 −54.2 −51.8
Motor healthy
Magnitude (dB) (μ = 16.67%) −10 −43 −47.4
Motor failure (μ = 33.33%) −9 −35.4 −42.2
Amplitude of stator Courant (μ = 0%) (μ = 16.67%) (μ = 33.33%)
currents (A) Ia1 1 2.4 4.88
Ib1 1 1.6 2.9
Ic1 1 1.24 2.12
Ia2 1.2 1.4 1.6
Ib2 1.2 1.4 1.44
Ic2 1.2 1.32 1.33
Current in the branch 0 13.9 18.29 Current in the branch
short-circuit (A) short-circuit (A)
Fig. 8 Flux density voltages components and corresponding locus (μ = 16.67%)
Fig. 9 Flux density voltages components and corresponding locus (μ = 33.33%)

Fig. 10 Stator flux voltage harmonic spectrum after fault
The experimental behavior of flux density voltages can also provide accurate and
reliable information on the state of the machine.
It is convenient to identify the existence of a defect by analyzing the currents
content harmonic. In the following, an analysis of the flux density signals will be
developed and compared to that based on the study of harmonic currents. The purpose
of this comparison is to evaluate the accuracy detection of both techniques [12].
Previous work has shown that the use of coils to detect cracks bars is more accurate
than that based on the analysis of the stator current.
6 Conclusion
In this paper, the stator winding inter-turn fault is investigated experimentally. Two
detection techniques were tested and compared; the first is based on analysis of
stator harmonic spectrum. The second focuses on the magnetic flux behavior. The
comparison between the two methods showed that the analysis of flux density voltage
appears as an accurate approach in order to detect stator winding fault in DSIM.
References
1. Tessarolo A (2010) Benefits of increasing the number of stator phases in terms of winding
construction technology in high-power electric machines. In: 5th IET international conference
on power electronics, machines and drives (PEMD), pp 1–6
2. Schoen RR, Habetler TG, Kamran F, Bartheld RG (1995) Motor bearing damage detection
using stator current monitoring. IEEE Trans Ind Appl 31(6):1274–1279
3. Dorrell DG, Paterson NC, Watson JF (1996) The causes and quantification of sideband currents
for use in rotor faults detection systems for cage induction motors. In: ICEM’96, Vigo Spain,
vol 3, pp 414–419
4. Benbouzid MH, Vieira M, Theys C (1999) Induction motors faults detection and localization
using stator current advanced signal processing techniques. IEEE Trans Ind Electron 14(1):14–
22
5. da Silva CR, Corea MBR, Jacobina CB, Lima AMN, da Silva ERC, Andriamalala RN, Razik
H (2006) IEEE
6. Nelson RH, Krause PC (1974) Induction machine analysis for arbitrary displacement between
multiple winding sets. IEEE Trans Power Appar Syst PAS-93(3):841–848
7. Apsley JM (2010) Derating of multiphase induction machines due to supply imbalance,” IEEE
Trans Ind Appl 46(2):789–805
8. Hammache H, Moussaoui D, Marouani K, Hamdouche T (2008) Magnetic properties in double
star induction machine. In: 18th International Conference on Electrical Machines, ICEM 2008
9. Kouki H, Fredj MB, Rehaoulia H (2013) Modeling of double star induction machine including
magnetic saturation and skin effect. In: 2013 10th International multi-conference on systems,
signals & devices (SSD) Hammamet, Tunisia, March 18–21, 3013
10. Vaseghi B, Takorabet N, Meibody-Tabar F (2009) Transient finite element analysis of induction
machines with stator winding turn fault. Progr Electromagn Res PIER 95
11. Andriamalala RN, Razik H, Didier G, Sargos FM (2006) A model of dual stator winding
induction machine in case of stator and rotor faults for diagnosis purpose. IEEE. 1-4244-
03650/06/$20.00(c)
12. Dorrell DG, Thomson WT, Roach S (1997) Analysis of airgap flux, current, and vibration
signals as a function of the combination of static and dynamic airgap eccentricity in 3-phase
induction motors. IEEE Trans Ind Appl 33(1):24–34
Fuzzy-Expert System Fault Tolerant
Control of IGBT Open Switch Fault
for RSC in Wind Turbine System
Amina Bouzekri and Tayab Allaoui
Abstract This paper presents hybrid-intelligent Fault Tolerant Control (FTC)

strategy of IGBT open-circuit for Rotor Side Converter (RSC) based on doubly-
fed induction (DFIG) generator wind turbine systems connected to the grid. The
FTC strategy based on Expert System method and combines a simple method detec-
tion based on fuzzy logic inference and uses rotor current average values to detect
the faulty switch in a very short period of time. In addition, following a power switch
failure, the FTC strategy activates the redundant leg and restores the operation of the
converter to save a power quality and ensure the service continuity. The simulation
model was developed in Matlab/Simulink environment and the simulation results
demonstrate the effectiveness of the proposed FTC strategy.
Keywords Component · Renewable energy · Wind turbine · Artificial

intelligence · Expert system · Fault tolerant control · IGBT open switch fault ·
Fuzzy logic · Doubly fed induction generator
1 Introduction
This Nowadays, power electronics play an important role in motor drives, utility
interfaces with renewable energy sources, power transmission (e.g. high-voltage
direct current systems, and flexible alternating current transmission systems), electric
or hybrid electric vehicles and many other applications. It is interesting to note that
the modern technology evolution in power electronics has greatly contributed to the
knowledge of power devices materials, processing, fabrication, packaging, modeling,
diagnosing and simulation.
Therefore, the reliability of electronics power becomes more and more vital,
and should draw more attention. According to a survey, semiconductor failure and
A. Bouzekri (B) · T. Allaoui

Laboratory of Electrical Energy and Computer Engineering (L2GEGI), Ibn Khaldoun University,
Tiaret, Algeria

https://doi.org/10.1007/978-981-15-6403-1_60
872 A. Bouzekri and T. Allaoui
soldering joints failure in power devices take up 34% of power electronic system fail-
ures [1]. A recent questionnaire on industrial electronic power systems also showed
that all the responders consider electronic power reliability as an important issue, and
31% of the responders selected the “semiconductor power device” as the most fragile
component [1, 2]. Therefore it is worth investigating IGBT’s failure and exploring
the solutions to improve the reliability of IGBT power electronic converters.
When a fault occurs in a system, the main problem to be addressed is to raise
an alarm, ideally diagnose what fault has occurred, and then decide how to deal
with it. The problem of detecting a fault, finding the source/location and then taking
appropriate action is the basis of fault diagnosis and fault tolerant solution [3].
Expert systems found broad application in fault diagnosis from their early stages
because an expert system simulates human reasoning about a problem domain,
performs reasoning over representations of human knowledge and solves problems
using heuristic knowledge rather than precisely formulated relationships, in forms
that reflect more accurately the nature of most human knowledge Feigenbaum (1982)
defined an expert system as “an intelligent computer program that uses knowledge
and inference procedures to solve problems that are difficult enough to require signif-
icant human expertise for their solution” [4]. The first diagnostic expert systems for
technical fault diagnosis were developed in the early 1970’s at MIT as is reported by
Scherer and White (1989). Since then numerous systems have been built. Surveys
of the first diagnostic expert systems of technological processes are provided by
Pau (1986), Tzafestas (1989), Scherer and White (1989). After 1995 expert system
applications started to decline as stand-alone systems and their technology embedded
in mainstream of information technology. New expert systems started to combine
symbolic with numerical information or with other artificial intelligence techniques
to produce more effective systems [4–7].
In this paper, a fuzzy-logic-expert system FTC strategy has been proposed to detect
and isolate the faulty switch in the RSC by changing the faulty leg by the redundancy
one in order to realize the safety and reliability of the system production. The main
goal of this paper is to put a spotlight specially in the using of the expert system in
the diagnosis system for a first using in renewable energy production, and we have
talked also briefly on the fuzzy logic detection system and we have disregard here
the stator powers control which are cited and detailed in Ref. [8].
2 Artificial Intelligence and Expert System
The domain of artificial intelligence is huge in breadth and width. While proceeding,
we consider the broadly common and prospering research areas in the domain of AI
[1, 2, 9, 10]:
• Fuzzy logic;
• Neural network;
• Genetic algorithm;
Fuzzy-Expert System Fault Tolerant Control of IGBT … 873
• Expert systems;
• Robotics;
• Natural language processing.
A. Expert system
Expert systems (ES) are one of the prominent research domains of AI. The first
expert system is introduced in the early 70’s by the researchers at Stanford University,
Computer Science Department [11]. According to Efrain Turban (2001), the ES is a
system that uses human knowledge stored inside a computer to solve problems that
requires human expertise to solve. A good ES is a system that can copy the process
of reasoning in a human.
What is meant by expert? An expert is a person that has the expertise and knowl-
edge of his specialized field. Examples of experts are a heart specialist and math-
ematics expert. Through experience, an expert expands his skills to enable him to
solve problems heuristically, efficiently and effectively [11].
Professor Edward Feigenbuan (1983) from Stanford University, a famous
researcher on ES defines ES as [12]:
“It is an intelligent computer program that uses knowledge and reasoning
procedures to solve difficult problems that need certain expertise to solve the
problems.”
B. Comparaison between Expert system & Expert human
Why an expert system needed? Table 1 will answer our query by comparing the
expert system to that of human.
C. Components of Expert Systems
The components of ES include (as shown Fig. 1) [11]:

• Domain expert (Human expert): It’s the individual who is currently expert to solve
the problems.
• Knowledge engineer: It’s the individual who encodes the expert’s knowledge in
a declarative from that can be used by the expert system.
Table 1 Comparisons
Factor Human expert Expert system
between an expert system and
that of a human expert Time (can be obtained) Working days only Anytime
Geography Local Anywhere
Safety Cannot be replaced Can be replaced
Damages Yes No
Speed and efficiency Changes Consistent
Cost High Intermediate
Cited in Ref. [12]
Fig. 1 Expert system components [11]
• Knowledge base: It’s a declarative representation of the expertise the most popular
expert system type based on rules which often in IF……THEN rules.
• Inference Engine: The inference engine is the most important component and is
considered the “brain” of an ES.
• User Interface: It’s the code that controls the dialog between the end-user and the
system.
• End user: It’s the individual who will be consulting with the system to get advice
which would have been provided by the expert.
D. Rules Creation
Rules are divided into two operators [13]:

• IF, called before (a premise or condition); and
• THEN, it is called effect (conclusions or actions).
In general, rules can have a few conditions by relating each condition to the
keywords “AND, OR” or a combination (AND and OR). On the contrary, it is better
to avoid combining both in one rule.
Rule with “AND” and “OR”: The example below shows how a few conditions are
related to “AND, OR”.
⎧
⎪
⎪ IF condition. 1
⎨
AND condition. 2
(1)
⎪
⎪ OR condition. 3
⎩
THEN action
Rule with “AND” or “OR”: The example below shows how a few conditions are
related to AND.
⎧
⎪
⎪ IF condition. 1
⎪
⎪
⎪
⎪ AND IF condition. 2
⎪
⎨
.
(2)
⎪
⎪ .
⎪
⎪
⎪
⎪ AND IF condition. 2
⎪
⎩
THEN action
3 Fault Tolerant Control of IGBT Open Switch Fault

for RSC in Wind Turbine System
Figure 2 shows the FTC strategy proposed in this work. It is simple and allows the
system to operate without interruption after the occurrence of a switch fault. The
proposed strategy uses a redundant-phase as an additional phase leg to replace a
faulty phase leg.
In order to realize this strategy in a RSC drive we need; the three-phase AC rotor
side current values which have been measured by three current sensors, IGBT gate
Fig. 2 Fuzzy-expert system fault tolerant control for RSC open switch fault schema
Fig. 3 Three-phase rotor

current average values of
rotor side converter
drive logic signals, TRIAC gate drive logic signals, these data can be sensed and fed
to a control circuit to embed the diagnostic program.1
The first step is the calculation of the three-phase currents average values of the
RSC side which provide an indication on the status of the system. Under healthy
conditions, these values are ideally equal to zero. However, when a fault occurs in
the converter switch, these values are no longer equal to zero.
A. Fuzzy open switch fault detection methods
Figure 3 represents the average values of three-phase RSC currents. In the healthy
condition, the average values of the currents are almost zero (between ±2 × 10−4 ).
However, under fault conditions, they assume other values. Table 2 shows the mean
values of the three-phase rotor currents codes for the six switches when there are an
open-circuit fault has occurred. In the table, When the average rotor current increases
above the threshold value of 6 × 10−4 A the output of the detector is 1, when it
1 Identify applicable sponsor/s here (sponsors).

Table 2 Average values of three-phase currents under different open switch fault with their rules
in fuzzy logic detection
Ira_avg Ir b_avg Ir c_avg Combination Switch fault
1 (P) −1 (S) − (S) 1 S1 fault
−1 (S) −1 (S) 1 (P) 2 S2 fault
−1 (S) 1 (P) −1 (S) 3 S3 fault
−1 (S) 1 (P) 1 (P) 4 S4 fault
1 (P) 1 (P) −1 (S) 5 S5 fault
1 (P) −1 (S) 1 (P) 6 S6 fault
0 (Z) 0 (Z) 0 (Z) 0 No fault
decreases below the threshold value of −2 × 10−4 A the output of the detector is
−1 and when they assume a values between ±2 × 10−4 A (healthy condition) the
output of the detector is 0.
These reactions of average values of rotor currents are taken as knowledge bases
for the proposed expert system.
The waveforms of the 3-phase current average values for all switches fault are
studied by experience then represented in Table 2.
The fuzzy sets are designated by the labels:
• S: small when the rotor current mean value takes value equal to −1 (negative),
• Z: zero, when the rotor current mean value takes value equal to 0 (zero),
• P: positive, when the rotor current mean value takes value equal to 1 (positive).
The membership functions for the input variables which present the average values
of rotor current Ira , Irb , Irc and the output variable which presents the faulty switch
and rules viewer are shown in Fig. 4. Sugeno-type fuzzy inference system was used
and defuzzification was based on the centre of gravity method.
B. Expert system Fault tolerant control for RSC
After the fault is detected and in order to avoid any subsequent failure in the other
devices, the faulty switch must be isolated and replaced with a healthy one. As
shown Fig. 5, the converter circuit has an additional leg which contains two switches
numbered S7 and S8 . The redundant leg is connected to both the DC side and to the
rotor side phases via a three-phase Triacs circuit (Triac1, Triac2 and Triac3).
The faulty leg is replaced by the healthy leg by changing the three indices (Ind1 ,
Ind2 and Ind3 ) and the status of switches S7 and S8 based on the detected faulty
switch and the status of the six switches (S1 , S2 , …., S6 ). The three indices take the
value 1 when there is no switch fault. Ind1 = 0, if the faulty switch is either the 1st
or the 4th switch, Ind2 = 0 if the faulty switch is either the 2nd or 5th switch and
finally, Ind3 = 0, when the faulty switch is either the 3rd or the 6th switch.
Fig. 4 Membership functions of the input variables and their Output variable
Fig. 5 Redundancy leg of the FTC strategy

Example:
• IF (S-fault = 0). THEN; (ind1 = 1, ind2 = 1, ind3 = 1, S7 = 0 and S8 = 0).
This means that the system works in the healthy condition, no fault has detected
and the additional leg doesn’t work.
• IF (S-fault = 1). THEN; (ind1 = 0, ind2 = 1 ind3 = 1, S7 = S1 and S8 = S2 ).
This means that the faulty switch is IGBT1; in this case the first leg (which contains
IGBT1 and IGBT4) must be changing by the additional leg (which contains IGBT7
and IGBT8). This step is realized by the changing of the Ind1 from 1 to 0 which
permits the changing of the signals switches from the first leg to the additional leg,
and so on (it’s the same idea for the other cases) as shown flowchart in Fig. 6.
Detection and changing of the faulty leg steps;
• Removing the control orders of the two drivers of the faulty leg;
• Triggering the suited bidirectional switch (TRIAC);
Fig. 6 Flowchart of the FTC strategy

• Using the control orders of the faulty leg for the redundant one;
• Stopping the fault detection algorithm.
Figure 7 shows open switch fault detection power converter response, where the fault
has been detected on the second switch for detection time equal to 0.016 s.
Figure 8 shows the three indices responses (ind1 , ind2 , ind3 ); these indices permit
to activate the Triacs to isolate the faulty leg and change it by the additional one.
When the fault had occurred in the second switch; the ind1 changes from 1 status to
0 in order to isolate the second faulty leg from the system.
However, the other indices state in their initial values (which are 1).
Figure 9 shows a succor leg switches signals (S7 , S8 ). These switches had no
signals in the healthy condition once the fault occur in the converter, they takes the
signals of the faulty leg and are working instead of there.
These indices and switches are changing according to the faulty switch detection
via a FTC algorithm strategy to ensure service continuity.
Figure 10 shows another switch faults detection for first, second and third switches
(S1 , S2 , S3 ), the detection time is not the same for the three detection. For example,
time detection in S1 fault equal to 0.013 s (Fig. 10a), for the second fault switch equal
to 0.015 (Fig. 10b) and for the third switch fault equal to 0.005 s (Fig. 10c). As, we
see for the three fault switches, the detection time is less than one period.
Fig. 7 Fuzzy logic open

switch fault detection (for
second switch fault S2)
Fig. 8 Indices ind1, idn2 and ind3

Before the After the Before the After the

fault fault fault fault
Fig. 9 Redundant leg switching signals
Fig. 10 Time detection for three switches: a S1, b S2, c S3
Figure 11 shows the active and reactive powers and the response of the direct and
quadrature rotor currents measurements with their reference values for wind speed
value is about 7 m/s. As we can see, the fault switch is happened at time equal to
1 s, where an impulsive is appeared for a short time in the both of rotor currents and
stator powers responses as a reaction of the FTC strategy which forced the powers
and currents to follow their references after the switch fault.
Figure 12 shows the average rotor current values before the open switch fault and
after the replacement of the faulty leg via the triacs circuit.
5 Conclusion
In this paper, a Fuzzy-Expert system Fault Tolerant Control strategy has been
proposed for open-circuit IGBT fault switch of RSC converter in DFIG based wind
turbine system. The FTC proposed based on the Expert system methodology using
Fuzzy logic in the fault detection method. An experienced engineer creates the algo-
rithm detection and isolation of the faulty switch to transfer his expert to the system
which can use it after by the technician easily.
This algorithm permits to change the faulty leg of the RSC converter by the
additional leg (the fourth leg) and isolates the faulty leg from the system to realize
the reliability and the service continuity.
Fig. 11 a active stator

power; b reactive stator
power; c direct rotor current;
d quadrature rotor current
Fig. 12 Three-phase RSC currents average values
The simulation results were showed that the proposed FTC strategy is very effi-
cient and detects the faulty switch within time less than one period which allows
the deterioration and prevents the deterioration and the shut-down of the electrical
production power.
References
1. Wolfgang E (2007) Examples for failures in power electronics systems. Paper presented at
ECPE tutorial on reliability power electronic system, Nuremberg, Germany, April 2007
2. Fuchs FW (2002) Some diagnosis methods for voltage source inverters in variable speed
drives with induction machines a survey. In: IEEE Industrial electronics conference, Roanoke,
Virginia, USA
3. Halim A, Christopher E, Chee PT (2011) Fault detection and fault-tolerant control using sliding
modes. Adv Ind Control. Springer-Verlag London
4. Leung D, Romagnoli J (2000) Dynamic probabilistic model-based expert system for fault
diagnosis. Comput Chem Eng 24(A11):2473–2492
5. Angeli C (2010) Diagnostic expert systems: from expert’s knowledge to real-time systems.
Adv Knowl Based Syst 1:50–73
6. Debbe K (1991) Expert system for fault diagnosis of VSI AC drives. IEEE. in Proceedings of
IAS conference, pp 368–373
7. Milne RW, Chandrasekaran B (1986) Fault diagnosis and expert system. In: 6eme conference
on expert systems and their applications, AVIGNON, pp 603–613
8. Bouzekri A, Allaoui T, Denai M (2017) Intelligent open switch fault detection for power
converter in wind energy system. Appl Artif Intell 30(A9):886–898
9. Bose BK (2002) Bose-modern power electronics and AC drives, book. Prentice Hall PTR
10. Fuchs FW (2003) Some diagnosis methods for voltage source inverters in variable speed drives
with induction machines a survey. In: The 29th Annual conference of the IEEE industrial
electronics society, vol 2, IECON’03
11. Neamen D (2003) Semiconductor physics and devices, 3rd edn. McGraw-Hill, New York
12. Rashid MH (2013) Power electronics: circuits, devices applications. In: 4th ed. Full eBook
13. Smith RA (1979) Semiconductors, 2nd ed. Cambridge University Press
Multiclass Support Vector Machine
Based Bearing Fault Detection Using
Vibration Signal Analysis
Issam Attoui, Nadir Fergani, Nadir Boutasseta, Brahim Oudjani,

Mohammed Salah Bouakkaz, and Ahmed Bouraiou
Abstract The bearing element is an essential part in different mechanical equipment

and rotating machinery. An unexpected failure of the bearing may cause substan-
tial economic losses. Vibration analysis has been used as one of the most preferred
measured signal to detect the bearing related fault. In this paper, an online method for
automatic rolling-element bearing fault diagnosis using Multiclass Support Vector
Machine MSVM is proposed. Firstly, the Empirical Mode Decomposition EMD tech-
nique is used to decompose the vibration signal into a finite number of stationary
intrinsic mode functions (IMFs), then different parameters are calculated based FFT
algorithm. The frequency analysis results of different EMD intrinsic mode func-
tions for different vibration signals show that these intrinsic mode functions will
change in different frequency bands when a bearing fault occurs. Therefore, to iden-
tify roller bearing fault class, the Locality Sensitivity Discriminant Analysis LSDA
algorithm is used to transform the selected feature to low-dimensional space. The
final features could serve as input vectors of trained MSVM. The analysis results
from roller bearing signals with inner-race, out-race and ball faults show that the
I. Attoui (B) · N. Fergani · N. Boutasseta · B. Oudjani

Research Center in Industrial Technologies CRTI, P.O. Box 64, Cheraga, Algeria
N. Fergani
N. Boutasseta
B. Oudjani
M. S. Bouakkaz
Department of Electrical Engineering, Laboratoire d’électrotechnique de Skikda «LES»,
Université du 20 Août 1955, 21000 Skikda, Algeria
A. Bouraiou
Unité de Recherche en Energie Renouvelables en Milieu Saharien, URERMS, Centre de
Développement des Energies Renouvelables, CDER, 01000 Adrar, Algeria

https://doi.org/10.1007/978-981-15-6403-1_61
886 I. Attoui et al.
proposed diagnostic approach based on the MSVM as a classifier and the EMD as a
signal processing technique to extract different stationary intrinsic mode functions
as features can identify roller bearing fault type and severity patterns accurately and
effectively even the analyzed vibration signal is acquired in the case of varying load
and speed.
Keywords EMD · FFT · Vibration signals · Bearing fault diagnosis · LSDA ·

MSVM
1 Introduction
The usage of a rotating machine as a generator or a motor in almost all forms of

transmission and transformation energy applications validates the significant position
of the control of their normal operations [1–3]. In regular operation, a diversity of
faults is able to occur within induction machine, such as the bearing fault that is
the most frequent faults in induction machines (41%) followed by stator (37%) and
rotor faults (10%). These faults can cause a complete breakdown if its progress is
not detected. The shaft misalignment, rotor eccentricity, and other rotor related faults
are the major sources of the bearing failures. In fact, the detection of bearing faults
is required, firstly to avoid the damage to the rotating machine itself and secondly to
avoid the damage to industrial processes [2, 4–7].
One of the commonly used techniques that offer a good key to bearing faults
diagnosis is the analysis of the vibration signal. In fact, this technique produces very
informative signature depending on the rotating machine health conditions [6, 8–13],
and therefore, the analysis of these informative signatures permits determining the
health condition of different rotating machine components without dismantling [7].
But, an online and automatic detection system are required to detect the rotating
machine faults.
For this purpose, numerous new machine learning techniques such as the artificial
neural network (ANN) [9], support vector machines (SVM) [14], adaptive neural-
fuzzy inference system (ANFIS) [15], etc. are proposed in the literature to auto-
matic categorization and classification of bearing faults. But, for several difficulties,
the use of the vibration signals directly as input to intelligent machine learning
techniques may be expensive, ineffective or even impossible. Therefore, the current
diagnostic techniques frequently pass through many steps that are feature generation,
feature selection or dimensionality reduction, and fault classification. The first step
involves of generating the features related to the fault types and severities to provide
a reduced representation of the input vibration signal before the feature reduction
step. Several studies for vibration monitoring was concentrated on the use of time
domain, frequency domain and time frequency domain techniques.
The time domain methods include statistical methods such as Root mean square
(RMS); Crest factor (Crf), Kurtosis (Kv), Skewness (Sw), Energy (E), Entropy (Ep)
Multiclass Support Vector Machine … 887
and Lempel-Ziv Complexity (LZC) [16–19]. But, because of the presence of non-
linear noises in machinery, these methods can produce the false alarms [20]. The
signal decomposing methods such as empirical mode decomposition EMD, local
characteristic-scale decomposition LCD, wavelet transform associated with various
statistical methods have also been applied to bearing fault diagnosis [11, 21–26].
Using these methods, the performance of the bearing fault diagnosis technique can
be improved. The diagnostic methods frequently use a dimensionality reduction
phase to improve the performance of the input of the classifier tool. In this paper, the
Locality Sensitive Discriminant Analysis LSDA algorithm is used. The contribution
in this paper is the ability of the proposed method to identify the type and severity
of rolling bearing fault even the analyzed vibration signal is acquired in the case of
varying load and speed.
This paper is organized as follows: in Sect. 2, a theoretical background of the
EMD, LSDA and MSVM approaches is presented. In Sect. 3, a description of the
proposed bearing fault diagnosis technique is presented. In Sect. 4, the experimental
bearing faulty data taken from the Case Western Reserve University (CWRU) are
described [27] and also the results of the proposed procedure have been evaluated
and discussed. Finally, the conclusion is given in Sect. 5.
2 Theoretical Background
A. EMD approach
The aim of EMD technique is to find the oscillation modes presented in time scales
that are defined by the interval between local extrema. Every point on the signal where
its second derivative is not-zero, and its derivative is zero represents a local extremum
point. In each local section of a signal, only one global maximum is present. Also,
only one global minimum point is present. Once the time scales are recognized, the
zero averaged Intrinsic mode functions FMIs are sifted out from the signal. The steps
constituting the EMD process are as follows:
(1) Find the local maxima and minima of distorted signal, s(t),
(2) make cubic spline interpolation between the maxima and the minima to
discovery the envelopes eM (t) and em (t), respectively,
(3) Calculate the mean of the envelopes,
e M (t) + em (t)
m(t) = (1)
2
(4) Calculate c1 (t) = s(t) − m(t), c1 (t) is an IMF if the number of local extrema of
c1 (t), is equal to or differs from the number of zero crossings by one, AND the
average of c1 (t) reasonably zero. If c1 (t) is not an IMF, then repeat steps 1–4 on
c1 (t) instead of s(t), until the novel c1 (t) obtained satisfies the conditions of an
IMF,
(5) Compute the residue, r 1 (t) = s(t) – c1 (t), if the residue, r 1 (t), is greater than a
threshold value of error tolerance, then repeat steps 1–6 on r 1 (t), to find the next
IMF and a novel residue.
To avoid ‘over-improving’ of the c1 (t), a suitable stopping criterion, in step5, can
be used to avoid the significant loss of information. The first IMF obtained, consists
of the highest frequency components present in the original signal. The subsequent
IMFs obtained, contain progressively lower frequency components of the signal.
The original signal is reconstructed as,

s(t) = ci (t) + r (t) (2)
n
n: is the number of orthogonal IMFs that are obtained in the iterative manner.
B. LSDA Algorithm
The LSDA algorithm determines the local manifold structure. This can be discov-
ered by finding a projection which maximizes the margin between data points from
dissimilar classes at each local area.
In LSDA, two graphs (within-class graph Gw and between-class graph Gb) are
constructed. For having a “good” map of vector set X ={x 1 , x 2 , … x m } that contains
m given training samples into a lower dimensional subspace Y, the following two
objective functions are optimized:
⎧
⎪
⎨ min (yi − y j ) Ww,i j
2

ij
(3)
⎪
⎩ max (yi − y j ) Wb,i j
2
ij
where, W w,ij are defined as W w,ij = 1 if xi e N w (x j ) or x j eN w (x i ), and 0 otherwise.

W b,ij are defined as W b,ij = 1 if xi e N b (x j ) or x j eN b (xi), and 0 otherwise. The set
N w (x i ) contains the points which are the k nearest neighbors of xi and share the same
label with xi while N b (x i ) contains the neighbors having different labels.
The optimization problem can be reduced to finding:
arg max a T X (αL b + (1 − α)Ww )X T a

a T X Dw X T a = 1 (4)
where α is a suitable constant and 0 ≤ α ≤ 1. The projection vector that minimizes (4)
is given by the maximum eigenvalue solution to the generalized eigenvalue problem:
X (αL b + (1 − α)Ww )X T a = λX Dw X T a (5)

Let the column vector a1 , a2 , …, ad be the solutions of Eq. (5), ordered according
to their eigenvalues, λ1 > … >λd . Thus, the embedding is as follows:
xi → yi = A T xi
A = (a1 , a2 , a3 , . . . , ad ) (6)
where yi is a d-dimensional vector, and A is an n × d matrix.

C. MSVM classifier
The goal of the Multiclass SVM is to give the labels to instances by using support
vector machines, where the labels are drawn from a predetermined set of numerous
elements. The main objective of this technique is to reduce the single multiclass
problem into multiple binary classification problems [28]. Common methods for
such reduction include: Building binary classifiers which distinguish (i) between
one of the labels and the rest (one-versus-all) or (ii) between every pair of classes
(one-versus-one) [28, 29]. Classification of novel examples for the one-versus-all
case is realized by a winner-takes-all strategy, in which the classifier with the highest
output function assigns the class. For the one-versus-one approach, classification is
realized by a max-wins voting plan, in which every classifier assigns the instance to
one of the two classes, then the vote for the assigned class is increased by one vote,
and finally the class with the most votes determines the instance classification.
3 Proposed Fault Diagnosis Techniques
A. Methodology
Figure 1 gives the proposed method used in this work for bearing fault detection
and classification based on the Empirical Mode Decomposition EMD and FFT tech-
niques, and MSVM classifier. The application process of the proposed procedure
that can be split into offline and online computational steps is exposed in Fig. 1.
B. Offline step (Training phase)
Once a measurement space (historical data) that includes different training vibration
signals for the all cases of bearing condition considered in the classification step is
acquired, three main processing sub-steps are conducted as follows:
(1) The d-dimensional vectors of training samples for each case of bearing condition
is used by a combination between Empirical Mode Decomposition EMD and
FFT techniques.
(2) Different features based on the six first intrinsic mode functions (IMFs) of the
EMD and FFT are calculated.
(3) The LSDA is used to transform the selected feature to low-dimensional space.
Fig. 1 Proposed fault

diagnosis technique
Collected Real-time vibration Data
vibration signals signal of bearings acquisition
of bearings (testing samples)
(training
samples)
1-Using training 1-Using testing

samples, the samples, the
EMD and FFT EMD and FFT
techniques are techniques are
applied. applied.
2- the energy, 2- the energy,

max, mean, max, mean,
kurtosis and kurtosis and Fault
entropy values entropy values features
are calculated are calculated mining
for each EMD for each EMD and
intrinsic mode intrinsic mode reductio
function. function n
3- Reduction 3- Reduction
using LSDA using LSDA
4-Training the 4- Bearing fault

MSVM using classification Fault
reduced using trained classifi
training feature MSVM. c-ation
set.
Output
5- Recognize the the
fault type and results
its severity
Training
Diagnostic phase
phase
(4) The MSVM classifier first needs to be trained with vibration data comprising
different health bearing conditions for three types of faults (inner race, outer race
and ball). For the training following conditions are considered. The training at
the beginning, at the intermediate and at the end of the rotational speed of a
range of rotational speed.
C. Online step (Diagnostic phase)
Once the MSVM classifier is trained in the offline step of the proposed procedure
and a measurement space that includes a testing vibration signal is acquired, three
main processing sub-steps are conducted as follows:
(1) The first sub-step is similar to that of the first sub-step in the offline step of the
proposed procedure. The EMD and FFT techniques are applied in the testing.
Then, the different features based on the six first intrinsic mode functions (IMFs)
of the EMD are considered to give the proposed testing features set.
(2) Although the testing feature set is obtained, the optimized MSVM classifier
is used to classify the bearing operating condition for different health bearing
conditions, including three types of faults (inner race, outer race and ball). The
testing data set should be different from that the data set used in the offline step.
(3) Once the bearing operating condition is classified, the bearing fault type and its
severity can be recognized. In addition, the performance of proposed procedure
can be determined during this sub-step by computing the classification accuracy
that was effectively realized using the proposed approach.
4 Description of the Experimental Setup and Results
The proposed method can be evaluated using the test data from Case Western Reserve
University [27]. This test data was collected for normal and faulty bearings for a
variety of fault locations, fault types and fault severities. The faults can be located in
inner raceway, rolling element (ball) and outer raceway for faults ranging from 0.007
inches in diameter to 0.040 inches in diameter. The shaft rotation speed varies from
1720 to 1797 rpm for different loads. Figure 2 shows the system used for measuring
the data that includes a 2 hp Reliance Electric motor, a torque transducer/encoder.
The vibration signal was collected using accelerometers on the experimental test for
two sampling frequencies (12,000 samples/second and 48,000 samples/second).
In this paper, the collected data sets on the drive end of the motor housing taken
for training and testing steps in the proposed procedure process are chosen arbitrarily
for
– Shaft rotation speeds ranging from 1730 to 1797 rpm;
– Fault severities ranging from 0.007 inches in Diameter to 0.028 inches in diameter;
– Sampling frequency at 12000 samples/s;
– The accelerometer is placed in the six o’clock position.
Fig. 2 Experimental test stand [27]
In our experiment, the time length of each sample was 0.1 s. We randomly selected
just 400 acquired samples as the training samples, and 1200 acquired samples as
testing samples. The vibration data used in the training and testing are given in
Table 1.
The number of features selected with the proposed technique is 5 features.
According to the experimental result the average classification accuracy can wait
91.33%.
Table 1 Description of the vibration dataset

Faulty condition Fault diameter Number of training Number of testing Class
samples samples
Healthy – 40 120 1
Inner race 0.007 40 120 2
Inner race 0.014 40 120 3
Inner race 0.021 40 120 4
Ball fault 0.007 40 120 5
Ball fault 0.014 40 120 6
Ball fault 0.021 40 120 7
Outer race 0.007 40 120 8
Outer race 0.014 40 120 9
Outer race 0.021 40 120 10
400 1200
Table 2 Confusion matrices of the proposed method

Predicted class using 05 features
1 2 3 4 5 6 7 8 9 10
Real class 1 90 0 0 0 0 10 0 0 0 0
2 0 116 1 0 0 0 0 0 0 0
3 0 3 115 0 0 0 0 0 5 0
4 0 0 0 120 0 0 0 0 0 0
5 0 0 0 0 103 0 14 0 0 0
6 0 0 0 0 8 110 3 0 2 0
7 30 0 0 0 9 0 97 0 0 0
8 0 0 0 0 0 0 0 115 0 3
9 0 0 4 0 0 0 6 0 113 0
10 0 1 0 0 0 0 0 5 0 117
In the Table 2, each row represents the instances in predicted class while each
column represents the instances in a real class. This Table makes it easy to see if the
system is confusing ten classes. Using only five features, only some classification
errors occurred between some classes. This was understandable, because the some
similarity of feature values which caused confusion between these classes. Therefore,
the proposed method need more feature vector parameters to solve this problem.
5 Conclusions
In this paper, a new procedure based Empirical Mode Decomposition EMD, FFT
algorithm and MSVM classifier as an automated classification system in order to
detect and classify bearing faults is proposed. The experimental data obtained from
the Case Western Reserve University Bearing Data Centre website has been employed
to validate the proposed method. According to this result, we conclude that this
method can be used for the detection and classification of bearing faults for a variety
of bearing conditions, fault locations, fault types and fault severities. This procedure
provides an accurate classification of 91.33% for 5 features. This shows that the
proposed approach is appropriate to serve as an intelligent fault diagnosis system
for on-line industrial applications. Therefore, the proposed method need more infor-
mative feature vector parameters to solve the problem of certain classification errors
occurred between some classes.
References
1. Attoui I, Omeiri A (2014) Contribution to the fault diagnosis of a doubly fed induction generator
for a closed-loop controlled wind turbine system associated with a two-level energy storage
system. Electr Power Compon Syst 42(15):1727–1742
2. Boudiaf A, Moussaoui A, Dahane A, Atoui I (2016) A comparative study of various methods
of bearing faults diagnosis using the Case Western Reserve University data. J Fail Anal Prev
16(2):271–284
3. Attoui I, Omeiri A (2015) Fault diagnosis of an induction generator in a wind energy conversion
system using signal processing techniques. Electr Power Compon Syst 43(20):2262–2275
4. Thorsen OV, Dalva M (1999) Failure identification and analysis for high-voltage induction
motors in the petrochemical industry. IEEE Trans Ind Appl 35(4):810–818
5. Motor Reliability Working Group (1985) Report of large motor reliability survey of industrial
and commercial installations, Part I. IEEE Trans Ind Appl IA-21(4):853–864
6. Garcia-Perez A, Romero-Troncoso RDJ, Cabal-Yepez E, Osornio-Rios RA (2011) The appli-
cation of high-resolution spectral analysis for identifying multiple combined faults in induction
motors. IEEE Trans. Ind Electron 58(5):2002–2010
7. Attoui I, Fergani N, Boutasseta N, Oudjani B, Deliou A (2017) A new time–frequency method
for identification and classification of ball bearing faults. J Sound Vib 397:241–265
8. Lau ECC, Ngan HW (2010) Detection of motor bearing outer raceway defect by wavelet packet
transformed motor current signature analysis. IEEE Trans Instrum Meas 59(10):2683–2690
9. Sperduti A, Starita A (1997) Supervised neural networks for the classification of structures.
IEEE Trans Neural Netw
10. Attoui I, Boutasseta N, Fergani N, Oudjani B, Deliou A (2015) Vibration-based bearing fault
diagnosis by an integrated DWT-FFT approach and an adaptive neuro-fuzzy inference system.
In: 3rd international conference on control, engineering and information technology, CEIT
2015
11. Attoui I, Fergani N, Oudjani B, Deliou A (2017) Rolling bearing fault diagnosis based on
improved complete ensemble empirical mode decomposition. In: 4th international conference
on control engineering and information technology, CEIT 2016
12. Atoui I, Meradi H, Boulkroune R, Saidi R, Grid A (2013) Fault detection and diagnosis in
rotating machinery by vibration monitoring using FFT and Wavelet techniques. In: 2013 8th
international workshop on systems, signal processing and their applications (WoSSPA), pp
401–406
13. Attoui I, Boudiaf A, Fergani N, Oudjani B, Boutasseta N, Deliou A (2015) Vibration-based
gearbox fault diagnosis by DWPT and PCA approaches and an adaptive neuro-fuzzy inference
system. In: 2015 16th international conference on sciences and techniques of automatic control
and computer engineering (STA), pp 234–239
14. Sain SR, Vapnik VN (1996) The nature of statistical learning theory. Technometrics 38(4):409
15. Jang J-SR (1993) ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans Syst
Man Cybern 23(3):665–685
16. Pincus SM (1991) Approximate entropy as a measure of system complexity. Mathematics
88:2297–2301
17. Zhang L, Xiong G, Liu H, Zou H, Guo W (2010) Bearing fault diagnosis using multi-scale
entropy and adaptive neuro-fuzzy inference. Expert Syst Appl 37(8):6077–6085
18. Lempel A, Ziv J (1976) On the complexity of finite sequences. IEEE Trans Inf Theory 22(1):75–
81
19. Yan R, Gao RX (2011) An introduction to complexity measure: Non-linear statistical param-
eters in measurements: Part 35 in a series of tutorials on instrumentation and measurement.
IEEE Instrum Meas Mag 14(5):27–35
20. Soualhi A, Medjaher K, Zerhouni N (2015) Bearing health monitoring based on hilbert-huang
transform, support vector machine, and regression. IEEE Trans Instrum Meas
21. Goumas SK, Zervakis ME, Stavrakakis GS (2002) Classification of washing machines vibration
signals using discrete wavelet analysis for feature extraction. IEEE Trans Instrum Meas
22. Zheng J, Cheng J, Yang Y (2013) A rolling bearing fault diagnosis approach based on LCD
and fuzzy entropy. Mech Mach Theory
23. Kang M, Kim J, Kim J-M (2015) Reliable fault diagnosis for incipient low-speed bearings
using fault feature analysis based on a binary bat algorithm. Inf Sci (Ny) 294:423–438
24. Boškoski P, Gašperin M, Petelin D, Juričić D (2015) Bearing fault prognostics using Rényi
entropy based features and Gaussian process models. Mech Syst Signal Process
25. He D, Li R, Zhu J (2013) Plastic bearing fault diagnosis based on a two-step data mining
approach. IEEE Trans Ind Electron 60(8):3429–3440
26. Seshadrinath J, Singh B, Panigrahi BK (2014) Investigation of vibration signatures for multiple
fault diagnosis in variable frequency drives using complex wavelets. IEEE Trans Power Electron
29(2):936–945
27. C. W. R. University (2015). http://csegroups.case.edu/bearingdatacenter/pages/download-dat
afile
28. Shalev-Shwartz S, Singer Y, Srebro N, Cotter A (2011) Pegasos: primal estimated sub-gradient
solver for SVM. Math Program
29. Hsieh CJ, Chang KW, Lin CJ, Keerthi SS, Sundararajan S (2008) A dual coordinate descent
method for large-scale linear SVM. In: Proceedings of the 25th international conference on
machine learning
Processing Signal Parameters Based
Fuzzy Inference System Classifier
for Analog Circuit Single Parametric
Faults
Imad Laidani and Naceredine Bourouba
Abstract This paper presents a diagnostic method for singular parametric faults
in analog circuits based on the Fourier coefficients and Distortion rate. The Simu-
lation Before Test (SBT) approach has been used. In the proposed method, the DC
component, the distortion rate and the fundamental amplitude extracted from Fourier
analysis of the circuit output voltage signal have been selected to determine the circuit
state conditions. The selection of these parameters provide a better characterizing
mean of the circuit state conditions and contribute in the faults dictionary construction
under faulty and fault free circuit conditions. The faults in concern are of singular
parametric fault type that corresponds to a deviation from ±10 to ±50% with a
10% step variation of the circuit component nominal value. The faults classifica-
tion is carried out using the Fuzzy Inference System (FIS) where the input data are
the aforementioned Fourier parameters. The proposed method is illustrated through
the Sallen-Key band pass filter test circuit. The obtained results show that the faults
ambiguity problem being totally solved which confirm the efficiently of the proposed
method in the diagnosis of the singular parametric faults in analog circuits.
Keywords Fault diagnosis · Fourier series · Fuzzy inference system · Sallen-key

band pass filter
1 Introduction
During the last decade, mixed circuits including analog as well as digital elements
have witnessed an increasing complexity in size and functionality. Because of the
complexity and the miniaturization of these circuits on one hand, and the very limited
access points on the other hand, it has become very difficult to carry out an effective
I. Laidani (B) · N. Bourouba

Faculty of Technology, University of Sétif, Sétif, Algeria
N. Bourouba

https://doi.org/10.1007/978-981-15-6403-1_62
898 I. Laidani and N. Bourouba
examination of the functioning of these devices. Testing and diagnosing this kind
of circuits has become a current topic, which aims at meeting the vital needs of the
electronics industry and facing the problem of failure and improve the reliability of
integrated circuits [1].
In the field of digital electronics, this topic has witnessed a huge progress to a
point where the fields of results have been very fruitful and the process has become
fully automated [2]. Unfortunately, the analog part knew few results only despite the
efforts made by researchers and industrials groups. These efforts have led of course,
to very important findings concerning “the failure detection method” at different
levels of circuit abstraction (primitive, block functioning, etc.). Therefore, lots of
things have to be made because of the diversity of the electrical parameters that
characterize this kind of circuits. Hence, it is necessary to find a systematic approach
to the problem of circuits testing and diagnosis.
Currently, all the research efforts being made have focused mainly on the devel-
opment of effective methodology and algorithms to diagnose analogical faults. The
challenge, to which designers and test engineers are facing in this domain, is to find
more efficient diagnosis approaches. In other words, finding accurate “universal fault
models” at a low cost despite the complexity of the circuit and the limited access to
the extreme point of the circuit becomes a fundamental economic need to the analog
integrated circuits yield [3]. In fact, the analog circuits diagnosis of the faults has
become by force of things the unavoidable master piece of the validation process of
the design and the performance evaluation inside the integrated circuit.
The work we are dealing with in this paper, consists of combining both the series
Fourier technique and distortion rates measurement to build the fault dictionary to
diagnosis the analog circuit, chosen for the purpose of fault detection and isolation.
The test task is affectively carried out from the use of fuzzy inference system classifier.
The fault detection has shown a 100% effectiveness because of the three selected
parameters signature, where two of them are extracted from the Fourier series and
namely are the continuous components and the first harmonic amplitude of the series,
the third parameter is a rate distortion that is caused by the presence of the fault in
the circuit under test (CUT). These selected features of fault signature are used to
avoid the fault that masks each other and hence ambiguity’s group of fault will be
reduced and faults are distinguishable to each other.
The rest of this article is organized as follows: In Sect. 2 a view on the fault
diagnosis approach is given, Sect. 3 discusses the proposed approach to diagnosing
faults. In Sect. 4, the method of classification of faults is discussed. The analog
circuit chosen to be subjected to the test is described in Sect. 5. The construction
of the fault dictionary for the circuit to be tested is detailed in Sect. 6. Application
of fuzziness and interpretation is given in Sect. 7. Section 8 concludes the proposed
fault diagnosis approach.
Processing Signal Parameters Based Fuzzy Inference System … 899
2 A View on the Fault Diagnosis Approach
The methods related to the diagnosis of analog circuits aim at ensuring the detection,
isolation and identification of faults through the electrical parameters evaluation of
these circuits. This assessment is deduced from the corresponding response measures
under faulty and fault-free circuit conditions.
Because of the encountered difficulties in this domain, researchers are trying
incessantly to find a way to solve and overcome the obstacles mentioned in [4] like:
• The lack of reliable and practical method for modeling faults in analog circuits.
• Continuous values of analog component parameters.
• The problem of tolerance and the non-linearity of the components.
• The restricted access to the external test points of the integrated circuit.
Most of the practical methods of diagnosis are based on the simulation and the
fault dictionary approach. Simulation is considered as an indispensable tool in the
test domain and intervenes in the diagnostic-procedure to solve the problems related
to the characteristics of analog circuits such as tolerances, nonlinearity and others.
Two simulation methods make use in this context: namely simulation after test (SAT)
and simulation before test (SBT) [5].
SAT methods focus on parameters identification and fault verification and shown
efficiency for soft fault diagnosis because they are based on linearized network
models. However the main drawback of these methods lays on the problem of param-
eter identification requiring a huge amount of computational time because of the
task complexity and the inability of the test point access in order to test all circuit
components [6].
The SBT method is used to achieve a compromise between coverage test and simu-
lation test based on the construction of fault dictionary. It is a matter of measurement
the circuit parameters with and without the presence of faults from the corresponding
response. Great computing power is needed to build this fault dictionary which is
only a lookup table which lists each faulty case and nominal case for comparison
purpose. The objective is to make a diagnosis clear in its critical aspect of deciding
of fault detection and diagnosis capability.
Several fault construction dictionary approaches and their classification has
emerged aiming at reaching an efficient diagnosis of singular parametric faults in
analog circuits. Some approaches are developed in the frequency domain [4, 7] using
different parameters like faults signatures such as the band width [2, 4] the upper
cut off frequency, peak amplitude, peak frequency, other one consist of the analysis
of faults in the temporal domain [8] settling time, peak amplitude for each time. In
these approaches, fault diagnosis depend on either a techniques based on fuzzy logic
[1] or on neural networks as illustrated in ref [2, 9] or combining both of them to
provide best results for classifier task. These techniques are fed with data that come
from error response or nominal response of both faulty/fault free CUT respectively.
The processing tool of the technique are varying from one research work to another
and could be either wavelet decomposition [9], support vector machine and kernel
estimation [10] and so ones. One can notice that various technique are combined
to each other aiming to achieving a best score in classifier accuracy, effectiveness
improvement through the increasing number of feature extraction and reliability of
the fault diagnosis [11].
Generally the type of faults to diagnose in the analog circuit and the most processed
faults in the diagnosis of analog integrated circuits are classified in 2 groups: the 1st
group is the one related to the faults which totally damage the circuit and makes it out
of service called the group of catastrophic faults. This type of fault usually comes
from manufacturing defects, physical defects, chemical, strange corpse presence
[12], is electrically modeled by short circuits or open circuits. Because of their
easily observable effect at the output of the circuits, they are easy to detect. The
second group of faults is the one that characterizes the malfunction of the circuit,
its imprecision and uncertainty: it is called group of parametric faults. The type
of fault characterizes the deviation that the elements of the circuit submit in their
nominal value in a more or less important manner. In some research work [13], the
fault being detected shows an arbitrary threshold of around 10% from the original
response value. Due to continuous value changes over time, this type of fault is of
great concern to researchers. As far as their detection and their location is concerned
a dictionary of these faults has to be set for diagnosis purpose. However, the diagnosis
accuracy has to be preserved for various parameters of the fault components without
large scale and low efficiency that traditional dictionary require [13, 14]. This makes
it a hot topic of the diagnostic activity.
3 The Proposed Fault Diagnosis Method
The test approach applied in our current work is based on the decomposition of
output periodic voltage signal, according to the Fourier theorem, in a sum including:
a constant term V 0 (average value or continuous component), a sinusoidal term
of frequency f 0 called fundamental or first harmonic and a sequence of sinusoidal
terms of an integral multiple of fundamental frequency called harmonic. Therefore
the Fourier series representation of a periodic signal can be described as:
∞

V (t) = V0 + (an cos(nw0 t) + bn sin(nw0 t)) (1)
n=1
where n is the integer sequence 1, 2, 3…, V 0 , an and bn are known as the Fourier
coefficients and w0 represents the fundamental frequency.
Then, the ratio of harmonic to fundamental amplitude, allows us to introduce the
total harmonic distortion rate (THD), given by the formula:

∞
n=2 Sn2
THD = , where : Sn = an2 + b2n (2)
S1
The extraction of these parameters the continuous component V 0 , the total

harmonic distortion rate (THD) and the fundamental amplitude using the simu-
lator PSPICE are considered as analog circuit signatures to be used as reference for
comparison with those of the same circuit under fault states. These will constitute
fault signatures that serve as input features to the faults classification system.
Another important point in the test approach is the input stimulus selection that
may contribute to the propagation of the effects of faults to the output circuit. The
chosen excitation signal is a square signal of duration T = 0.1 ms (See Fig. 1a)
applied to the input of the circuit. The output response will have a shape as shown in
Fig. 1b. Such an output signal form allows the establishment of a unique signature
extraction, in each fault case, and contributes in a diagnostic process aiming at faults
isolation and location.
a
2
1.5
0.5
0
0 1 2
-4
Time(s) x 10
b
2
-1
-2
0 1 2
-4
Time(s) x 10
Fig. 1 a The signal of excitation. b Output response
4 Faults Classifying Method
The method used to classify the entire faults group of concern has been inspired
from [1, 15], which is basically carried on two essential elements such as the fault
dictionary (FD) preparation and Fault classification implemented in a fuzzy inference
system (FIS).
The FD preparation has been set from results tabulation performed from fault
circuit responses. This consists of the extraction of the three predefined parameters
from the circuit output signal processing. The aim of this task is to create in the
first step fault and fault-free signatures based on these parameters deviations and
nominal values for fault and fault-free models respectively. Each of these signatures
constitutes a fuzzy system input. Then a range of tolerated values for each signal
parameter under fault-free and faulty conditions is defined and fed to the FIS system
as crisp number input.
The FIS is an expert tool that imitates a human decision that is based on uncertainty
premises like low, high, voltage/current cut frequency value, slope of characteristic
rise time, delay time etc.
Based on Fuzzy logic (FL) concept, the FIS produces a conclusion that indicates
the state of the CUT. A fuzzy inference system (FIS) defines a mapping of the input
data vector into a scalar output, using fuzzy rules. The mapping process involves
input/output membership function, FL operators, fuzzy if-then rules, aggregation of
output sets and defuzzification.
A general model of fuzzy inference system (FIS) is given in Fig. 2.
Where we can see few principal components: fuzzifier, inference engine, rule
base, defuzzifier. These main function components can be summarized as such:
• The fuzzifier takes input values and determines the degree to which they belong
to each of the fuzzy sets via a membership functions. This is required in order to
activate rules that are in terms of linguistic variables and is usually constructed
with IF-THEN statements which supported by logical operators (or, and etc.)
evaluate the relationship between the input and the output namely the premise
and the consequent.
• Rule base contains linguistic rule that are processed by experts or extracted from
numeric data. Once the rules have been established the FIS can be viewed as a
system that maps an input vector to an output vector.
• The inference engine defines mapping from input fuzzy sets into output fuzzy sets.
It determines the degree to which the antecedent (premise) is satisfied for each
Fig. 2 Block diagram of a fuzzy inference system

rule. If the antecedent of a given rule has more than one clause, fuzzy operators
are applied to obtain one number that represents the result of the antecedent for
that rule. It is possible that one or more rules may fire at the same time. Outputs
for all rules are than aggregated. During aggregation, fuzzy sets that represent
the output of each rule are combined into a single fuzzy set. Fuzzy rules are fired
in parallel, which is one of the important aspects of an FIS. In this, the order in
which rules are fired does not affect the output.
• The defuzzifier maps output fuzzy set into a crisp number. Given a fuzzy set that
encompasses a range of output values, the defuzzifier returns one number there
by moving from a fuzzy set to a crisp number. Several defuzzification methods
are used in practice including the centroid, maximum, mean of maxima, length
and modified height defuzzifier. The most popular defuzzification method is the
centroid which calculates and returns the centre of gravity of the aggregated fuzzy
set.
• FIS employs rules however, unlike rules in conventional expert systems; a fuzzy
rule localizes a region of space along the function surface instead of isolating a
point on the surface. For a given input, more than one rule may fire. Also, in an
FIS, multiple regions are combined in the output space to produce a composite
region as shown in Fig. 3. The x i and y are input and output linguistic variables
characterized by terms Txi and T y respectively. These correspond effectively to
fuzzy numbers with membership functions µix and µy .
The method of fault classification being the corner stone of our work has been
carried out through the FIS system implemented in a Matlab/Simulink software tool
box. This technique uses two Fourier series first components and THD parameters
as mentioned previously. These are fed through their readings as FIS inputs to be
processed in order to generating the corresponding output in terms of a unique crisp
value of a specific faulty configuration. This set of three input features is used inten-
tionally to dispatching the faults giving same fault effect at the CUT output and that
cluster into a same group for a particular input feature. Hence this task will reduce
the ambiguous faults groups ratio as low as possible.
Fig. 3 Schematic diagram of a fuzzy inference system

5 The Circuit Under Test Description and Simulation
The analog circuit chosen to be subject to the test described in this article is a Sallen-
Key bandpass filter (See Fig. 4), the most commonly used in the field of telecom-
munication. It has been used in many works aiming at improving the detectability of
the defects and consequently contributing to the establishment of better diagnostic.
Its structure is designed from the µ741 operational amplifier developed by Ferranti
firm.
The transfer function of this filter is as follows:
sA0 G1 C1
G(s) =
+ s[G3 C1 + G3 C2 + G1 C1 + G2 C1 (1 − A0 )] + G3 (G1 + G2 )
s2 C1 C2
R5 1 1 1
Where : A0 = 1 + , G1 = , G2 = , G3 = . (3)
R4 R1 R2 R3
Only, the type of defects being the most concerned in this current work is of
singular parametric faults. These affect the passive components, resistors and capac-
itors that construct the analog circuits and particularly the filtering circuit. Since each
component is assumed to be concerned with 10 single possible faults, corresponding
to ±10, ±20, ±30, ±40 and ±50% deviation of the nominal value component.
Whereas the other components remain their nominal values as in the original circuit
Fig. 4 Sallen-key bandpass filter

during the simulation process. The change in the component value of less than ±10%
is taken as the nominal tolerance limit.
The parametric fault effect is observed through the values of the DC component
(S 0 ), the total harmonic distortion rate (THD) and the fundamental amplitude (S 1 ).
The extraction of these parameters characterizing the state of the circuit without and
with fault is carried out from the circuit simulated by PSPICE this has been run under
a periodic square wave excitation stimulus of 10 kHz frequency and 2 V amplitude.
By applying a transient regime analysis, PSPICE software makes it possible to
perform the decomposition of signal to Fourier series and display the coefficients of
this series and the distortion rate as well.
6 The Circuit Under Test Faults Dictionary Construction
The first step of construction faults dictionary is to simulate the fault-free test circuit of
Fig. 5 using PSPICE to form a reference response to which faulty circuit responses are
compared. For this reason, the faults which are parametric in nature and correspond
to deviations of the component nominal values (R and C) from ±10 to ±50%, given
the considerable number of faults to be injected, we are limited to presenting some
cases of faults which are summarized according to their levels of deviation from
the response of the circuit in two possible cases: cases of indistinguishable and
distinguishable faults (See Fig. 5).
As in example of these cases, the selected faults shown in the figure above illustrate
this aspect and drive us to conclude the following assumptions. First, the faults
with a close or a similar level of the output response deviation, such as F18 (R3 −
40%) and F58 (C1 − 40%), may be clustered to a same group of faults. They are
undistinguishable and exhibit ambiguous classification of faults whose clustering
Fig. 5 Reference response

of the circuit (FF) and to fault
condition (F58, F18, F33)
group is called fault group of ambiguity. Secondly, some other faults such as the
F33 (R5 + 20%), and F58 may exhibit an important output response deviation to
each other. They are called distinguishable faults and thus they will be dispatched in
different fault groups.
Due to this situation, the dictionary to build requires more parameters to establish
a better distinction and detection of faults on one hand, and to achieve a better
fault classification and identification approach on the other hand. This approach is
obviously based on the fuzzy logic concept that has proven its efficiency in a lot of
research works [15]. The parameters used are those mentioned previously as fault
signatures (S 0 , S 1 and THD), on the basis of the result of which, a fault dictionary is
created. An index number has assigned to each fault present in the fault dictionary,
the fault free condition is indexed by F00. The fault dictionary for these singular
parametric faults is presented in Table 1.
Once the fault signatures have been collected and organized into a fault dictionary,
one can see similar signatures generated from the use of one or two input features.
They are a sign of the masking faults. For instance, the faults F21 and F22 are masking
each other since they provide same signatures expressed by same value of the first
and the third input features (see Table 1). In another fault case, same signatures have
been found with the first and the second input features for the faults F36 and F38.
So that making the faults of each set dispatched from each other has required the
combination of the input features.
7 The Fuzzy Logic Application and Interpretation
The Mamadani inference system, constructed for fault identification, is shown in

Fig. 6 where each extracted signature for fault diagnosis will be considered as
fuzzy system input. In our proposed approach, we have three signatures namely the
continuous component (S 0 ), the distortion rate (THD) and the fundamental amplitude
(S 1 ).
The number of membership functions for each input signature depends on the
number of predefined fault patterns for the components and another one their nominal
value.
In our study’s case, we have 11 triangular-shaped memberships (trmf), as shown
in Fig. 7a for the continuous component (S 0 ) values for component R1 under faulty
and fault-free conditions. On the other hand, the outputs of the fuzzy system are the
estimated values of the circuit components. Consequently, the number of outputs of
the fuzzy system equals that of the components of the analog circuit. The number
of membership functions for each output will be equal to the number of derivative
deviations for each component fault. The form of the output membership function
is also triangular, an example of an output membership function for a component
fault F52 is shown in Fig. 7b. After the membership functions are set for both the
fuzzy system input and output, they must be connected using rules built using the
rule editor. Here the AND operator is used because the input has three variables. The
Table 1 Fault dictionary

Fault model Fault index Continuous Distortion rate Fundamental Output FIS
component amplitude
FF F00 −0.017 54.52 1.010 0.22
R1 + 10% F01 −0.019 53.04 0.992 0.94
R1 − 10% F02 −0.015 56.08 1.027 1.97
R1 + 20% F03 −0.020 51.63 0.972 3.1
R1 − 20% F04 −0.013 57.68 1.043 3.82
R1 + 30% F05 −0.021 50.30 0.953 5.16
R1 − 30% F06 −0.010 59.27 1.056 5.98
R1 + 40% F07 −0.022 49.05 0.933 6.91
R1 − 40% F08 −0.007 60.81 1.065 8.14
R1 + 50% F09 −0.023 47.89 0.914 8.86
R1 − 50% F10 −0.005 62.23 1.070 10.1
R3 + 10% F11 −0.023 51.98 1.1 11
R3 − 10% F12 −0.012 57.32 0.918 11.8
R3 + 20% F13 −0.029 49.65 1.187 13.2
R3 − 20% F14 −0.006 60.37 0.823 13.9
R3 + 30% F15 −0.034 47.53 1.272 15
R3 − 30% F16 0 63.68 0.725 16.1
R3 + 40% F17 −0.040 45.60 1.354 16.8
R3 − 40% F18 0.006 67.23 0.626 18.2
R3 + 50% F19 −0.046 43.85 1.432 18.9
R3 − 50% F20 0.012 71.01 0.525 20
R2 + 10% F21 −0.017 53.46 1.016 21.1
R2 − 10% F22 −0.017 53.66 1.015 21.8
R2 + 20% F23 −0.018 52.63 1.020 23.2
R2 − 20% F24 −0.017 57.77 0.992 24
R2 + 30% F25 −0.018 51.94 1.024 24.9
R2 − 30% F26 −0.017 60.39 0.978 26.1
R2 + 40% F27 −0.018 51.38 1.027 26.9
R2 − 40% F28 −0.016 64.34 0.956 28.1
R2 + 50% F29 −0.018 50.91 1.029 29
R2 − 50% F30 −0.016 70.84 0.924 29.8
R5 + 10% F31 −0.014 54.43 0.959 31.2
R5 − 10% F32 −0.021 54.63 1.073 31.9
R5 + 20% F33 −0.012 54.33 0.917 33
R5 − 20% F34 −0.025 54.71 1.152 34.1
(continued)
Table 1 (continued)
Fault model Fault index Continuous Distortion rate Fundamental Output FIS
component amplitude
R5 + 30% F35 −0.010 54.24 0.881 34.8
R5 − 30% F36 −0.030 54.78 1.254 36.2
R5 + 40% F37 −0.008 54.16 0.851 36.9
R5 − 40% F38 −0.036 54.78 1.392 38
R5 + 50% F39 −0.006 54.09 0.825 39.1
R5 − 50% F40 −0.045 54.62 1.588 39.8
R4 + 10% F41 −0.014 54.59 1.067 41.2
R4 − 10% F42 −0.02 54.44 0.954 42
R4 + 20% F43 −0.011 54.63 1.124 42.9
R4 − 20% F44 −0.023 54.34 0.898 44.1
R4 + 30% F45 −0.008 54.65 1.182 44.9
R4 − 30% F46 −0.026 54.21 0.843 46.1
R4 + 40% F47 −0.004 54.65 1.240 47
R4 − 40% F48 −0.029 54.06 0.788 47.8
R4 + 50% F49 −0.001 54.62 1.298 49.2
R4 − 50% F50 −0.032 53.89 0.733 49.9
C1 + 10% F51 −0.017 51.50 1.086 51
C1 − 10% F52 −0.018 58.11 0.929 52.1
C1 + 20% F53 −0.016 48.93 1.157 52.8
C1 − 20% F54 −0.019 62.37 0.842 54.2
C1 + 30% F55 −0.015 46.73 1.222 54.9
C1 − 30% F56 −0.019 67.45 0.750 56
C1 + 40% F57 −0.015 44.85 1.282 57.1
C1 − 40% F58 −0.020 73.51 0.653 57.8
C1 + 50% F59 −0.014 43.23 1.338 59.2
C1 − 50% F60 −0.021 80.72 0.551 60
C2 + 10% F61 −0.019 52.48 1.01 60.9
C2 − 10% F62 −0.016 56.65 1.008 62.1
C2 + 20% F63 −0.020 50.51 1.009 62.9
C2 − 20% F64 −0.014 58.79 1.003 64.1
C2 + 30% F65 −0.021 48.64 1.007 65
C2 − 30% F66 −0.013 60.92 0.996 65.8
C2 + 40% F67 −0.021 46.90 1.003 67.2
C2 − 40% F68 −0.011 62.98 0.986 67.9
C2 + 50% F69 −0.022 45.26 0.998 69
C2 − 50% F70 −0.010 64.89 0.973 70.1
Fig. 6 Mamdani fuzzy inference system
defuzzification procedure combines the consequences of the rules used by means of

the center of the surface method. The fuzzy system output is an estimated value for
each component.
Figure 8 shows the rule editor where 71 input membership functions corre-
sponding to 70 fault cases and one fault free case. This operation is succeeded by
IF-THEN rule definitions that connect the input and output variables. Once the rules
are built, validation is done through the rule viewer. As an example, we consider a test
for a dictionary fault with the index value F52. The continuous component value is
−0.018, so for this entry, the triangular membership function is defined as [−0.019,
−0.018, −0.017] while keeping the extracted value in the center. For the value of
distortion rate 58.11, the triangular membership function is defined as [58.10 58.11
58.12]. Meanwhile for the fundamental amplitude value of the 0.929, the triangular
Fig. 7 a Membership functions of S0 for R1. b Output membership function for fault F52
Fig. 8 Fuzzy rules
membership function is defined as [0.928 0.929 0.930] and the output is defined as
[51.5 52 52.5]. The centroid defuzzifier of the surface (COA) produces the output
value 52.1 (See Fig. 9) which is equal to the fault index 52. Likewise, for the other 69
faults, the 69 input membership functions and the output membership function are
determined from the extracted diagnosis parameter values and the fault index. The
FIS outputs shown in Table 1 agree very well with the corresponding fault indices. It
can be noticed that for the continuous component parameter the fault diagnosis has
achieved a 100% of accuracy with an average score of more than 20% of the total
fault models as false fault parameters. This is considered as a better score than those
found in others research works [1, 15].
8 Conclusion
In this paper, we have described an SBT approach to detect the existence of para-
metric faults in analog circuits. The SBT method developed allows us to perform
fault diagnosis of analog electronic circuit. Consequently, a fault dictionary is created
Fig. 9 Input and Output of the FIS for Fault F52
using the values of the DC component, the amplitude of the fundamental and the
distortion rate as signal parameters extracted from the Fourier series signal decom-
position method. These signal elements have constituted input features of the fuzzy
inference system, FIS used as a tool of fault classification. The feasibility of the
proposed method is validated by the Sallen-key bandpass filter. The results of the
simulation show that singular faults have been detected efficiently by the developed
method, since a 100% of fault coverage has been attained. Other important fact comes
from the accuracy obtained with this classifier using the three mentioned above signal
parameters combination. This is confirmed by the faults totally dispatched from each
others on one side and accomplishing nil fault group ambiguity on the other side.
This kind of results will have an amazing effect on fault identification and isolation.
More confirmation of the classifier efficiency in analog circuits fault diagnosis field
is required in the near future. This is driving us to broaden the fault classification
method application for more analog circuits with different type of faults and under
different fault conditions, i.e. multiple faults.
References
1. Kumar A, Singh AP (2013) Fuzzy classifier for fault diagnosis in analog electronic circuits.
ISA Trans 52(6):816–824
2. Manikandan V, Devarajan N (2007) SBT approach towards analog electronic circuit fault
diagnosis. Active Passive Electron Compon 2007
3. Huang K (2011) Modélisation de fautes et diagnostic pour les circuits mixtes/ RF
nanometriques. These de Doctorat
4. Kavithamani A, Manikandan V, Devarajan N (2009) Analog circuit fault diagnosis based on
bandwidth and fuzzy classifier. In: IEEE region 10 conference on TENCON 2009. IEEE
5. Bandler JW, Salama AE (1985) Fault diagnosis of analog circuits. Proc IEEE 73(8)
6. Vasan ASS, Long B, Pecht M (2013) Diagnostics and prognostics method for analog electronic
circuits. IEEE Trans Ind Electron 60(11)
7. Bo F, Peng X, Junjie L, Ming D (2009) Analog circuit fault diagnosis based on neural network
and fuzzy logic. In: 2009 Chinese control and decision conference. IEEE, pp 199–202
8. Bharat Ram R, Prasanna Moorthy V, Devarajan N (2009) Fuzzy based time domain analysis
approach for fault diagnosis of analog electronic circuits. In: 2009 international conference on
control, automation, communication and energy conservation. IEEE, pp 1–6
9. Aminian M, Aminian F (2007) A modular fault-diagnosis system for analog electronic circuits
using neural network with wavelet transform as a preprocessor. IEEE Trans Instrum Meas
56(5):1546–1554
10. Tang J, Lin T, Chen Y (2010) Analog circuit fault diagnosis based on fuzzy support vector
machine and kernel density estimation. In: 2010 3rd international conference on advanced
computer theory and engineering (ICACTE), vol 4. IEEE, pp V4–544
11. Liu H, Chen G, Jiang S, Song G (2008) A survey of feature extraction approaches in analog
circuit fault diagnosis. In: 2008 IEEE pacific-Asia workshop on computational intelligence and
industrial application, vol 2. IEEE, pp 675–680
12. Sienkiewicz M (2010)”Méthodologie de localisation des défauts soft dans les circuits intégrés
mixtes et analogiques par stimulation par faisceau laser: analyse de résultats des techniques
dynamiques paramétriques. Thèse de doctorat
13. Peng Z, Chen J, Dong Y, Zhang B, Qiao S, Ye D, Ran L (2015) Radio frequency beamforming
based on a complex domain frontend. IEEE Trans Microw Theory Tech 64(1):289–298
14. Yang C, Zhang X, He A, Qiu L (2017) Fault diagnosis of analog circuit based on complex
model. In: Conference: 2017 32nd youth academic annual conference of Chinese association
of automation (YAC). IEEE, pp 949–952
15. Kavithamani A, Manikandan V, Devarajan N (2011) Analog circuit fault detection using poles
and fuzzy inference system. Int J Adv Eng Technol IJAET 2:260–269
Line Current Spectrum Analysis
as a Technique to Diagnose
the Demagnetization Fault in Permanent
Magnet Synchronous Motor
Zakaria Gherabi, Djilali Toumi, Noureddine Benouzza,

and Mostefa Becheikh
Abstract This paper deals with the modeling and diagnosis of the demagnetization
fault in the permanents magnets synchronous motors. To do this, a numerical model
based on the finite element method under the Flux 2D software has been developed.
This model makes it possible to accurately describe the exact geometry of the PMSM
and its physical behavior, taking into account all electromagnetic phenomena. In this
model, the demagnetization fault has been introduced by creating a partial fissure on
the permanent magnet surface periphery. To highlight the interest of the employed
model and the efficiency of the line current spectrum analysis technique; a series
of simulations are carried out in the two operating modes: healthy operation mode
(without demagnetization fault) and faulty operation mode (with demagnetization
fault).
Keywords Permanent magnet synchronous motor · Modeling · Diagnosis ·

Demagnetization fault · Finite element method · Flux 2D · Line current spectrum
analysis
Z. Gherabi (B) · N. Benouzza

Electrical Engineering Faculty, LDEE Laboratory, University of Sciences and Technology of Oran
(USTO-MB), Oran, Algeria
N. Benouzza
D. Toumi · M. Becheikh
Electrical Engineering Department, L2GEGI Laboratory, University of Ibn-Khaldoun of Tiaret,
Tiaret, Algeria
M. Becheikh

https://doi.org/10.1007/978-981-15-6403-1_63
916 Z. Gherabi et al.
Nomenclature
PMSM Permanent Magnet Synchronous Motor

FEM Finite Element Method
DMF Demagnetization Fault
FFT Fast Fourier Transform
MCSA Moteur Current Signature Analysis
PSD Power Spectral Density
ASBC Amplitude of Side Band Components
1 Introduction
In recent decades, synchronous motors and more particularly PMSM have been
increasingly used in various industrial applications. This growing interest is due to
several parameters, including its high reliability due to the absence of ring- brush
contact, its robust rotor construction and its high efficiency. These advantages make
this motor a serious competitor to DC motors and asynchronous motors [1–6].
Although these motors are reliable, but in the presence of certain electrical,
thermal, mechanical or environmental constraints they are subject to unexpected
failures. Among these failures, the irreversible demagnetization is one of the most
common problems of this type of motors, due to their relatively high cost, which
represents more than 50% of the global price of the PMSM [7–11]. For this reason,
it is therefore logical from an industrial, scientific and economic point of view to
focus research on the early diagnosis of this type of defect. This detection will
make it possible to preserve the safety of people and property, increase the lifetime
of motors, avoid the shutdown of the whole chain of production and consequently
minimize financial losses.
In order to diagnose the DMF in PMSM, many techniques have been proposed
in various scientific studies, analyzing: three phases current, supply voltages, air
gap flux, electromagnetic torque, vibration, instantaneous power and other types of
signals [11–19].
The air gap flux density analysis, measured by winded sensors installed along the
stator teeth, was provided in [12] to diagnose the DMF in PMSM. This is confirmed by
[13] which shows that the DMF introduces a second order harmonic in the air gap flux
density spectrum. According to [14], the partial DMF increases the electromagnetic
torque ripples of the surface mounted PMSM as well as increasing the amplitude of
side band components at specific frequencies in the torque spectrum. The 1st and 5th
harmonic magnitudes are used in [15], to diagnose the DMF at high speed. On the
other hand, the Hilbert Huang Transform is used to extract the DMF signatures at
low speed.
Line Current Spectrum Analysis as a Technique … 917
In our opinion, the most widely used technique in recent years is the one based on
the stator current spectrum analysis, known as the MCSA, which has the following
main advantages [13, 16–19]:
• Doesn’t need a special location for the current sensors. In fact, it may be positioned
in any position from the supply voltage to the PMSM.
• It has a rich harmonic spectrum. The frequency location of certain harmonics
provides us with important information about the PMSM condition and on
different faulty types.
• Allows the tracking of defect severity by monitoring the magnitude of deferent
signatures.
Many studies [17–19] have shown that the demagnetization fault is manifested by
the appearance of the particular frequency signature in the PMSM current spectrum.
In order to identify these frequency signatures, the PSD estimation by Periodogram
is the most commonly used method in the industry application because it is easy to
program and fast to execute due to the use of fast Fourier transform [17, 18].
This article is arranged as follows: in the Sect. 2, a brief description of the DMF in
the PMSM. Section 3 describes a PMSM model including the DMF using the FEM
under the Flux 2D software. The deferent frequency signatures of the demagnetiza-
tion fault are presented in Sect. 4. The demagnetization fault identification technique
based on the PSD estimation by Periodogram is illustrated in Sect. 5. The simula-
tion results in the various operating modes (healthy and faulty operation mode) are
mentioned in Sect. 6. Finally, a conclusion is given in Sect. 7.
2 Demagnetization Fault in PMSM
PMSM contain a rotor equipped by permanent magnets. These are either on the
rotor surface or integrated. The irreversible demagnetization of these magnets is one
of the most important faults in PMSM. Indeed, the DMF can be complete, partial
asymmetric or symmetric in a certain region of the rotor pole [20–23].
This type of defect can occur during abrupt load applications or in the presence
of a short circuit fault in the stator windings. Once produced, the irreversible reduc-
tion of the permanent magnet remanent induction automatically reduces the overall
performance of the PMSM [19, 23].
In addition to the diminution of the permanent magnet magnetic force due to
the armature reaction or temperature, it is also known that some magnets (such as
Neodymium Iron Boron) can deteriorate by disintegration. The fissures that form
during manufacture can lead to disintegration at high speed. This disintegration
causes the disruption of the air gap flux and the unbalance of the magnetic attraction
between the rotor and the stator, which produces vibrations and stress on the bearings
of the PMSM [21, 23].
3 PMSM Modeling Using FEM
The FEM is the most commonly used method for the modelling of electromagnetic
and mechanical systems such as electrical machines [23–25]. This method makes it
possible to accurately describe the exact geometry of the machine and its physical
behavior, taking into account all internal, electromagnetic, thermal and mechanical
phenomena, with a low level of simplifying assumptions [24, 25].
Numerical approaches are generally presented in generic software form (Flux 2D,
Ansys, Ansoft Maxwell, Magnet…) when the user work consists in describing the
geometry of the structure, carrying out its meshing and informing the characteristics
of the materials [24, 25]. The increased development of powerful computers has
made it possible to constantly reduce the computation time, which can be rather long,
particularly for a fine-meshing or a strong coupling between the different physical
domains [25].
The studied PMSM consists of 6 pole pairs and 36 stator slots. The arrangement
of the stator windings of this motor and its parameters are indicated in the appendix.
To study the impact of the demagnetization fault on the PMSM electromagnetic
performance, this motor has been modeled by the FEM using Flux 2D software.
These FEM modeling is carried out using the following steps:
• Definition of the machine geometry (see Fig. 1a).
• Meshing of the geometry (see Fig. 1b);
• Definition of physical properties: the properties of the materials used for magnets,
stator yoke, and rotor plates are defined;
• Definition of the power supply circuit: allows you to automatically define the
machine power supply voltages for the different operating modes;
• Solver configuration and resolution;
• storage of the obtained results such as: supply voltage, stator currents and rotation
speed.
A DMF may be caused by the induction reaction or the appearance of a fissure
in the magnet surface. This means that the DMF can be classified as a type of fault
that is uniformly distributed over its entire surface, or located at a specific location
of the pole [21]. In order to take into account this non-uniformity of the DMF in
the analytical model, the magnet is considered to be composed of small elements as
shown in Fig. 2.
Figure 3 shows the images of the deferent types of defects studied (partial and
complete). For the partial demagnetization two elements of the magnet (the magnet
divided into 12 elements) were assumed to be demagnetized and for the complete
demagnetization, the magnet is assumed to be totally defective.
Fig. 1 a Structure of the

studied PMSM. b Mesh size
representation
Fig. 2 Subdivision of the permanent magnet to small elements
(a) (b)
(c) (d)
Fig. 3 Types of considered defects, a and b partial DMF, c and d complete DMF
Fig. 4 Ideal stator current
Magnitude (db)
spectrum around the
fundamental frequency in the
presence of the DMF Lower Upper
sideband sideband
(1-1/P)fs (1-2/P)fs fs (1+1/P)fs (1+2/P)fs

Frequency (Hz)
4 DMF Frequency Signatures
Due to the richness of information contained in the current spectrum, several inves-
tigations [17–19] have shown that the tracking of this spectrum makes it possible to
detect and also to identify all types of defects affecting the PMSM. This identifica-
tion is enabled by simply tracking on the position and frequency amplitude of some
harmonics.
Several investigations [17–19] have shown that the DMF is manifested by the
creation of a series of harmonic around the fundamental frequency in the stator
current spectrum, at the frequencies given by the following equation:

k
f dem = 1± fs (1)
P
With: f s is the power frequency, P is the number of poles pairs and k is a positive
integer.
Figure 4, shows an ideal stator current spectrum around the fundamental frequency
in the presence of the DMF, according to the Eq. (1) and with k = 1 and 2.
5 Identification of the DMF Frequency Signature

by Periodogram
Among the spectral analysis methods used to identify the DMF signature, the PSD
estimation by Periodogram is the most common method in the industrial field because
it is [26, 27]:
• Easy programing;
• Short computation time through the application of the FFT algorithm;
• Simple to implement on FPGA-based boards for use in real time.
The PSD estimation is defined as the square of the Fourier transform modulus,
divided by the number of samples of the discrete signal x(n), it is given by [26, 27]:
|X ( f )| 2
Px ( f ) = (2)
N
With, X(f ) is the Discrete Fourier Transform of a signal x(n), as shown in the
following equation:
N −1
−2 j.π.n. f
X( f ) = x(n).e Fe (3)
n=0
With, F e is the Sampling frequency and N the number of samples of the signal
x(n), that’s in our case the stator current.
The simulation is performed on 7 kW PMSM, 6 pole pairs and 36 stator slots. The
arrangement of the stator windings of this motor and its parameters are indicated in
the appendix. The motor is powered by a balanced voltage source and that operate at
the synchronism speed (668 rpm). The different operating modes carried out in this
study are:
• 1st operation mode: Healthy PMSM;
• 2nd operation mode: PMSM in the presence of the DMF.
• Asymmetric partial DMF in a single pole;
• Asymmetrical partial DMF in two poles;
• Complete DMF of a single pole;
• Complete DMF of two poles.
For the magnetic resolution of our model, the choice of the simulation time and
the sampling frequency is dictated by a number of constraints related to computation
time and storage capacity. For example, for a simulation time of 20 s and a sampling
frequency of 1 kHz, the simulation took about 2 days of calculation with a CORE-
i7 computer with 16 Giga of RAM. For this time, the external storage of solution
files is approximately 37.6 Giga. In the light of all these considerations and for
optimal application of the FFT spectral analysis; all simulations are performed with
a simulation time of 10 s and a sampling frequency of 1 kHz, which led to a frequency
resolution of 0.1 Hz.
A. 1st operation mode
Figure 5 shows the electromechanical characteristics of the healthy PMSM under a

load equal to that of the nominal torque. Figure 5a illustrate the PMSM rotational
speed. The speed response has a ripples of ±3 rpm around the synchronous speed
(668 rpm). These ripples are due to the effect of space harmonics. Figure 5b, Illustrate
800
(a) Healthy PMSM
600
Speed (rpm) 671
400
668
200
665
2 2.02 2.04
0
0 2 4 6 8 10
Time (S)
30
22.23 (A) (b) Isa Isb Isc
20
Current (A)
10
-10
-20
-30
2 2.01 2.02 2.03 2.04 2.05
Time (S)
25
fs (c) Healthy PMSM
0
Magnetude(db)
5fs 7fs
-25 3fs
-50
-75
-100
50 100 150 200 250 300 350 400 450 500
Frequency (Hz)
Fig. 5 1st operation mode: a Speed of rotation, b currents of three stator phases, c current spectrum
of the phase a
the waveforms of the stator current absorbed by the simulated motor. This current
is not purely sinusoidal, it has fluctuations, which are due to the effect of space
harmonics. Figure 5c present the stator current spectrum of phase a. It contains in
addition to the fundamental a series of harmonics at frequencies (2k + 1)f s which
represent space harmonics. This analysis is considered as the reference for all the
following operating modes.
25
fs Healthy PMSM
0
Magnetude(db)
-25
-50
-75
-100
0 20 40 60 80 100 120 130
Frequency (Hz)
Fig. 6 1st operation mode: Current spectrum of the phase a in the frequency band [0–130 Hz]
According to the theoretical study, the frequency signatures of the DMF are located
around the fundamental frequency in the stator current spectrum. Therefore, for a
better visibility of the spectrum, the spectral analysis will be performed only on the
frequency band [0–130 Hz] as shown in the Fig. 6.
B. 2nd operation mode
Figure 7a, show the increase in rotational speed fluctuations from ±3 rpm for a healthy
operation to ±11 rpm for the operation in the presence of the demagnetization fault of
a complete pole. Figure 7b, Illustrate the currents absorbed by the three stator phases
in the presence of a complete pole demagnetization fault. Figures 7c–f, respectively
presents the current spectra for different types of DMF, i.e.:
• Asymmetric partial DMF in a single pole;
• Asymmetrical partial DMF in two poles;
• Complete DMF of a single pole;
• Complete DMF of two poles.
Table 1 summarize the magnitude’s variation of DMF harmonics characteristic as
a function of the fault severity.
7 Conclusion
In this paper, a numerical model based on the finite element method under the Flux
2D software is developed to study the influence of the demagnetization fault on
different electromagnetic quantities of permanent magnet synchronous motor. The
DMF is introduced in our model by creating a partial fissure on the permanent magnet
surface periphery.
The simulation results showed that the DMF is manifested by the creation of
a series of harmonics around the fundamental frequency and at fixed frequencies
800
1 partiel Pole 2 partiel Pole 1 complet Pole
600
Speed (rpm) 679
(a)
400
668
200
657
2 2.005 2.01
0
0 2 4 6 8 10
Time (S)
30
22.78 (A) (b) Isa Isb Isc
20
Current (A)
10
-10
-20
-30
2 2.01 2.02 2.03 2.04 2.05
Time (S)
25
(c) fs 1 partial pole
0
Magnetude(db)
-25 (1-1/P)fs (1+1/P)fs

(1-2/P)fs (1+2/P)fs (1+4/P)fs
(1-4/P)fs (1+5/P)fs
(1+3/P)fs
-50 (1-5/P)fs (1-3/P)fs
-75
-100
0 20 40 60 80 100 120 130
Frequency (Hz)
Fig. 7 2nd operation mode: a Speed of rotation, b currents of three stator phases, c–f current spectra
of the phase a for different types of DMF studied
(depending on constant parameters such as: the number of pole pairs and the power
supply frequency). Moreover, the increase of the fault severity is manifested by the
increase of ripples in the temporal form of deferent electromechanical quantities,
and by the increase of the magnitudes of DMF harmonics characteristic.
25
(d) fs 2 partial poles
0
Magnetude(db)
-25
(1-1/P)fs (1+1/P)fs
(1-4/P)fs (1-2/P)fs (1+2/P)fs
(1+3/P)fs (1+5/P)fs
-50 (1-5/P)fs
(1-3/P)fs
(1+4/P)fs
-75
-100
0 20 40 60 80 100 120 130
Frequency (Hz)
25
(e) fs 1 complet pole
0
Magnetude(db)
-25 (1-1/P)fs (1+1/P)fs

(1+4/P)fs
(1-4/P)fs (1-2/P)fs (1+2/P)fs
(1-3/P)fs (1+3/P)fs (1+5/P)fs
(1-5/P)fs
-50
-75
-100
0 20 40 60 80 100 120 130
Frequency (Hz)
25
(f) fs 2 complet poles
0
Magnetude(db)
(1-1/P)fs (1+1/P)fs
-25 (1-4/P)fs (1-2/P)fs (1+2/P)fs (1+4/P)fs
(1+3/P)fs (1+5/P)fs
(1-5/P)fs (1-3/P)fs
-50
-75
-100
0 20 40 60 80 100 120 130
Frequency (Hz)
Fig. 7 (continued)
In general, the simulation results show the interest of the model employed and the
efficiency of the line current spectrum analysis technique for the demagnetization
fault diagnosis.
Table 1 Amplitudes of the harmonics characteristic of the DMF

Type of demagnetization 1 Partial pole 2 Partial pole 1 Complet pole 2 Complet poles
fault
Frequency (Hz) Magnitude (db) Magnitude (db) Magnitude (db) Magnitude (db)
(1 − 5/P)f s −69.3 −66.75 −59.2 −54.65
(1 − 4/P)f s −63.9 −62.65 −59.3 −55.00
(1 − 3/P)f s −61.45 −63.85 −54.15 −51.21
(1 − 2/P)f s −60.85 −59.1 −54.3 −50.57
(1 − 1/P)f s −58.2 −54.65 −45.95 −43.66
(1 + 1/P)f s −57.45 −52.8 −45.87 −42.10
(1 + 2/P)f s −58.95 −55.25 −52.8 −49.37
(1 + 3/P)f s −59.3 −60.6 −51.9 −50.08
(1 + 4/P)f s −60.45 −62.2 −52.85 −50.65
(1 + 5/P)f s −58.9 −59.85 −53.1 −51.51
Appendix A
The arrangement of the three winding’s stator phases, a, b and c of the studied PMSM
is illustrated in the following figure (Fig. 8).
Fig. 8 The arrangement of the windings of the three stator phases a, b and c of the studied PMSM
Appendix B
PMSM Parameters are illustrated in the following table.
Symbol Description Values Units

Pw power at the rated speed 7 kW
n Rated speed 668 rpm
Vn Rated voltage 380 V
In Rated current 24.8 A
fs Rated frequency 66,8 Hz
Rs Stator resistance 0.442
Ls Synchronous inductance 8.6 mH
J Moment of inertia 1.10−3 Kg m2
F Viscous rubbing 1.10−2 Nm/rad/s
go Nominal air gap 2 mm
rs Stator inner radius 61.42 mm
– Stator outer radius 117.7 mm
rr Rotor inner radius 47.71 mm
– Rotor outer radius 54.92 mm
εm Radial PM length 4.5 mm
– Radial tooth length 491.9 mm
– Area of winding turn 3.66 mm
– Area of slot 281.1 mm
ecs Stator core thickness 7.22 mm
l Length of the machine 204.7 mm
P Number of pole pairs 6 –
– Number of slots 36 –
Ns Number of turn per phase 23 –
εs Slot opening 4 °
αr Opening magnets 24.99 °
Br Magnet remanence 1.2 T
References
1. Toliyat HA, Nandi S, Choi S, Meshgin-Kelk H (2012) Electric machines: modeling, condition
monitoring, and fault diagnosis. CRC Press
2. Faiz J, Mazaheri-Tehrani E (2017) Demagnetization modeling and fault diagnosing techniques
in permanent magnet machines under stationary and nonstationary conditions: an overview.
IEEE Trans Ind Appl 53(3):2772–2785
3. Khan MS, Okonkwo UV, Usman A, Rajpurohit BS (2018) Finite element modeling of demag-
netization fault in permanent magnet direct current motors. In: 2018 IEEE power & energy
society general meeting (PESGM), Portland, OR, pp 1–5
4. M’hamed L, Zakaria G, Khireddine D (2017) A Robust sensorless control of PMSM based on
sliding mode observer and model reference adaptive system. Int J Power Electron Drive Syst
(IJPEDS) 8(3):1016–1025
5. Gherabi Z, Benouzza N, Toumi D, Bendiabdellah A (2019) Eccentricity Fault diagnosis
in PMSM using motor current signature analysis. In: Intl Aegean conference on electrical
machines power electronics (ACEMP), Intl conference on optimization of electrical electronic
equipment (OPTIM), Istanbul, Turkey
6. Gherabi Z, Toumi D, Benouzza N, Henini N (2019) Modeling and diagnosis of stator winding
faults in PMSM using motor current signature analysis. In: Intl Aegean conference on electrical
machines power electronics (ACEMP), 2019 Intl conference on optimization of electrical
electronic equipment (OPTIM), Istanbul, Turkey
7. Ginzarly R, Hoblos G, Moubayed N (2015) Faults in hybrid electric vehicles’ permanent magnet
machines. In: 2015 third international conference on technological advances in electrical,
electronics and computer engineering (TAEECE), Beirut, pp 169–174
8. Nejadi-Koti H, Faiz J, Demerdash NAO (2017) Uniform demagnetization fault diagnosis in
permanent magnet synchronous motors by means of cogging torque analysis. In: 2017 IEEE
international electric machines and drives conference (IEMDC), Miami, FL, pp 1–7
9. Goktas T, Lee KW, Zafarani M, Akin B (2016) Analysis of magnet defect faults in permanent
magnet synchronous motors through fluxgate sensors. In: 2016 IEEE applied power electronics
conference and exposition (APEC), Long Beach, CA, pp 2875–2880
10. Goktas T, Zafarani M, Akin B (2016) Discernment of broken magnet and static eccentricity
faults in permanent magnet synchronous motors. IEEE Trans Energy Convers 31(2):578–587
11. Barmpatza AC, Kappatou JC (2017) Demagnetization faults detection in an axial flux
permanent magnet synchronous generator. In: 2017 IEEE 11th international symposium on
diagnostics for electrical machines, power electronics and drives (SDEMPED), Tinos, pp
153–159
12. Da Y, Krishnamurthy M (2010) Novel fault diagnostic technique for permanent magnet
synchronous machines using electromagnetic signature analysis. In: 2010 IEEE vehicle power
and propulsion conference, Lille, pp 1–6
13. Rezig A, Mekideche MR, Djerdir A (2011) Impact of eccentricity and demagnetization faults
on magnetic noise generation in brushless permanent magnet DC motors. J Electr Eng Technol
6(3):356–363
14. Ebrahimi BM, Faiz J (2013) Demagnetization fault diagnosis in surface mounted permanent
magnet synchronous motors. IEEE Trans Magn 49(3):1185–1192
15. De Bisschop J, Vansompel H, Sergeant P, Dupre L (2017) Demagnetization fault detection in
axial flux PM machines by using sensing coils and an analytical model. IEEE Trans Magn
53(6):1–4, Art no. 8203404
16. Jeong H et al (2018) Classification and detection of demagnetization and inter-turn short circuit
faults in IPMSMs by using convolutional neural networks. In: IEEE energy conversion congress
and exposition (ECCE), pp 3249–3254
17. Faiz J, Nejadi-Koti H (2016) Demagnetization fault indexes in permanent magnet synchronous
motors—An overview. IEEE Trans Magn 52(4):1–11, Art no. 8201511
18. Urresty J, Riba J, Romeral L (2013) A back-emf based method to detect magnet failures in
PMSMs. IEEE Trans Magn 49(1):591–598. https://doi.org/10.1109/TMAG.2012.2207731
19. Liu Z, Huang J, He S (2018) Diagnosis of air-gap eccentricity and partial demagnetisation of an
interior permanent magnet synchronous motor based on inverse transient complex inductance
vector theory. IET Electr Power Appl 12(8):1166–1175
20. Fitouri M et al (2018) Effects of design variables on demagnetization fault in permanent magnet
synchronous motor. In: 2018 15th international multi-conference on systems, signals & devices
(SSD), pp 71–76
21. Farooq JA, Djerdir A, Miraoui A (2009) Identification of demagnetization faults in a permanent
magnet synchronous machine by permeance network. COMPEL Int J Comput Math Electr
Electron Eng 28(6):1619–1631
22. Alameh K et al (2015) Fault detection and isolation approach for permanent magnet
synchronous machines. In: 22nd international congress on sound and vibration (ICSV22)
23. Rajagopalan S (2006) Detection of rotor and load faults in brushless dc motors operating under
stationary and non-stationary conditions. Ph.D. dissertation, School of electrical and computer
engineering. Georgia Institute of Technology, USA
24. Bracikowski N (2012) Modélisation multi-physique par modèles à constantes localisées; appli-
cation à une machine synchrone à aimants permanents en vue de son dimensionnement. Diss.
Ecole Centrale de Lille
25. Alameh K (2017) Contribution au diagnostic et a l’analyse de défauts d’une machine synchrone
à aimants permanents. Thèse de doctorat. Normandie
26. Boudinar AH, Benouzza N, Bendiabdellah A (2016) Induction motor bearing fault analysis
using a root-MUSIC method. IEEE Trans Ind Appl 52(5):3851–3860
27. Aimer AF, Boudinar AH, Benouzza N, Bendiabdellah A (2015) Simulation and experimental
study of induction motor broken rotor bars fault diagnosis using stator current spectrogram.
In: 2015 3rd international conference on control, engineering information technology (CEIT),
pp 1–7
Fault-Tolerant Path-Tracking Control
with PID Controller for 4ws4wd Electric
Vehicles
Youcef Zennir, Sami Allou, and J. J. Fernandez-Lozano
Abstract This work deals with the problem of fault path tracking for an electric
vehicle which has four electromechanical wheel Systems. A controller based on tow
PID the first controller for the speed and the second for the steering control to ensure
that the robot follows the trajectory, during the design of this controller the effort
was concentrated on the simplicity and efficiency. The processing and response time
are very essential factors in the process control, which is what motivated the choices
of the structure and the parameters of the different parts. The effectiveness of the
proposed velocity control method is confirmed by numerical simulations.
Keywords Mobile RAMBLER robot · Modelling · PID controller · Electric

vehicle · Fault-tolerant control
1 Introduction
The saturation of the road network in the city centres has become a problem, to resolve
this problem and facilitate mobility it is necessary to create new networks or to use
the available resources in an optimal way, but the construction of new infrastructure
is not realizable all the time, so it is necessary to imagine new transport system is
what motivated the researchers study all aspects of autonomous vehicles this type
of vehicle has a high efficiency per a port to vehicles driven by humans The electric
vehicle (EV) is a relatively new concept in the world of the automotive industry
[1]. Although some companies have based their entire model of cars around being
proactive and using electricity, some also offer hybrid vehicles that work off both
electricity and gas. EV’s get their power from rechargeable batteries installed inside
the vehicle [2–4]. The electric vehicles are a great way to not only save money but
Y. Zennir (B) · S. Allou

Automatic Laboratory of Skikda, université 20 août, 1955 Skikda, Skikda, Algeria
J. J. Fernandez-Lozano
Departamento de Ingeniería de Sistemas y Automática, Escuela de Ingenierías Industriales,
Universidad de Málaga, Calle Dr. Ortiz Ramos, Málaga, Spain

https://doi.org/10.1007/978-981-15-6403-1_64
932 Y. Zennir et al.
also help contribute towards a healthy and stable environment. Many research works
have been conducted to study electric vehicles control [5, 6], deal steering stability
of platoon equipped with cooperative adaptive cruise control (CACC). In this work
we started by the control of a single vehicle, we used tow PID controllers to control
the vehicle one for the steering and the other one for the vehicle speed [7].
2 Modeling of Rambler
The modelling is an important step before the application of the control, the more
the modelling takes into account all the forces applied on the system the more
the results are efficient. The robot that we used is an autonomous electric vehicle
(RAMBLER). It has been developed by the University of Malaga [1], It is a 4-motor-
wheel skid-steer mobile robot, with active suspension. RAMBLER robot uses 4
in-wheel brushless motors. Each motor is controlled by an embedded board that
integrates a microcontroller (μC) and an independent H-bridge power stages [8–10]
(Fig. 1).
A vehicle (RAMBLER) dynamic model has been developed useful for control
design, neglecting some side effects introduced, by suspensions and tire deformation.
Define a fixed reference frame f(x, u) and a moving frame f(z, y) attached to the
vehicle body, with origin at the vehicle center of mass G (see Fig. 2). Let ẋ, ẏ, θ̇ be,
respectively, the longitudinal, lateral, and angular velocity of the vehicle in frame f.
Fig. 1 Electric robot RAMBLER [1]

Fault-Tolerant Path-Tracking Control … 933
Fig. 2 Free-body diagram [7]
The free-body diagram of forces and velocities is shown in Fig. 2, with the vehicle
having instantaneous positive velocity components ẋ and θ̇ and negative velocity ẏ.
Wheels develop tractive forces Fxi and are subject to longitudinal resistance forces
Rxi for i = 1, …, 4. We assume that wheel actuation is equal on each side so as to
reduce longitudinal slip. Thus, it will always be Fx4 = Fx1 and Fx3 = Fx2 . Lateral
forces Fyi act on the wheels as a consequence of lateral skidding. Also, a resistive
moment Mr , around the center of mass is induced in general by the Fyi and Rxi forces.
For a vehicle of mass m and inertia I about its center of mass, the equations of motion
can be written in frame f as:
max , = 2Fx1 + 2Fx2 − Rx (1)
mav , = −Fv (2)
I θ̈ , = 2t(Fx1 + Fx2 ) − Mr (3)
To express the longitudinal resistive force Rx , the lateral resistive force Fy , and
the resistive moment Mr , we should consider how the vehicle gravitational load
mg is shared among the wheels and introduce a Coulomb friction model for the
wheel-ground contact. We have:
b mg
Fx4 = Fx1 = · (4)
a+b 2
a mg
Fx3 = Fx2 = · (5)
a+b 2
where: a: Front wheelbase; b: Rear wheel base.

At low speed, the lateral load transfer due to centrifugal forces on curved paths
can be neglected. In case of hard ground, we can assume that the contact patch
between wheel and ground is rectangular and that the tire vertical load produces
a uniform pressure distribution. In this condition, Rxi = fr Fzi sgn(ẋi ) where fr is
the coefficient of rolling resistance, assumed independent from velocity. The total
longitudinal resistive force is then:

4
mg
Rx = Rxi = fr (sgn(ẋ1 ) + sgn(ẋ2 )) (6)
i=1
2
Introducing a lateral friction coefficient u, the lateral force acting on each wheel
will be:
Fyi = uFzi sgn(ẏ2 ) (7)
The total lateral force is thus

4
mg
Fy = Fyi = u (sgn(ẏ1 ) + sgn(ẏ3 )) (8)
i=1
a+b
While the resistive moment is

Mr = a Fy1 + Fy2 − b Fy3 + Fy4 + t[(Rx2 − Rx3 ) − (Rx1 − Rx4 )]
abmg mg
= (sgn(ẏ1 ) + sgn(ẏ3 )) + fr (sgn(ẋ1 ) + sgn(ẋ2 )) (9)
a+b 2
3 Control Design
Increasingly stringent standards on vehicle emissions and fuel economy have made
hybrid electric vehicles (HEVs) and/or plug-in HEVs (PHEVs) increasingly more
attractive to both the industrial and academic communities Not only various HEVs
and/or PHEVs, such as the Toyota Prius and the Ford Escape, have already entered
mass production, but the tendency toward pure electric ground vehicles (EGVs) has
also been significantly accelerated by various electrified systems and technologies
instead of the conventional mechanical/hydraulic counterparts [11, 12]. For example,
regenerative braking can transfer the vehicle kinetic energy into battery electrical
energy during the deceleration processes and consequently reduce energy loss from
frictional heat in traditional mechanical/hydraulic brakes. In-wheel electrical motor

control is also a novel technology that provides faster and more accurate torque
actuation. Electric differential (ED) is another electrified technology that eliminates
the traditional mechanical differentials (MDs). An ED system can reduce the weight
and friction loss in the drive train by removing MD components and consequently
improve the overall reliability and efficiency during the power transmission. The
working principle of an ED system is similar to that of a traditional MD. When an
EGV is cornering or making a turn, the outer wheels will travel longer paths than the
inner ones, implying that the rotational speed of the outer wheels will be faster than
that of the inner wheels. The resulting speed differences between the inner and outer
wheels are automatically tuned by an ED system [12–15]. Note that an ED system in
an EGV with independently actuated in-wheel motors is essentially a wheel speed
tracking control module, which is different from the direct torque control of in-wheel
motors.
To keep the mobile robot on our desired trajectory and desired velocity it is
necessary to design a regulator which will allow tracking of arbitrary trajectories (xr
(t), yr (t)). The design of controllers which we used is based on PID the first one for
the speed control receives the speed error of the vehicle and the second the speed
tracking, the robot speed measured by a GPS system as shown in Fig. 3, at the output
of the controller, we receive the linear speed and the steering speed necessary for the
robot to remain on the trajectory [16–18].
Fig. 3 Block diagram of robot control diagram

4 Simulink Model
To illustrate the proposed control for RAMBLER, we built a SIMULINK model as

shown in Fig. 4.
The Longitudinal Wheel block implements the longitudinal behavior of an ideal
wheel. You can specify the longitudinal force and rolling resistance calculation
method, and brake type. Use the block in driveline and longitudinal vehicle simu-
lations where low frequency tire-road and braking forces are required to determine
vehicle acceleration, braking, and wheel-rolling resistance. For example, you can use
the block to determine the torque and power requirements for a specified drive cycle
or braking event. The block is not suitable for applications that require combined
lateral slip. The block calculates the inertial response of the wheel subject to: Axle
losses, Brake and drive torque, Tire rolling resistance and ground contact through
the tire-road interface. The input torque is the summation of the applied axle torque,
braking torque, and moment arising from the combined tire torque.
Ti = Ta − Tb + Td (10)
For the moment arising from the combined tire torque, the block implements
tractive wheel forces and rolling resistance with first order dynamics. The rolling
resistance has a time constant parameterized in terms of a relaxation length.
Td = ωReLe + ωRe(FxRe + My) (11)
To calculate the rolling resistance torque, we the method in Stepwise Coast down
Methodology for Measuring Tire Rolling Resistance. The rolling resistance is a
Fig. 4 Simulink block diagram of robot control

Table 1 Simulation parameters 1

If Lock-up condition Friction model
ω = 0 or TS < |Ti + Tf − ωb| Unlocked Tf = Tk
Where
Tk = FC Reff uk tanh[4(−wd )]
Ts = FC Reff us

2 R03 −RI3
Reff =
3 R02 −Ri2

ω = 0 and TS ≥ |Ti + Tf − ωb| Locked Tf = Ts
Table 2 Simulation parameters 2

If Lock-up condition Dynamic model
ω = 0or T S < |T i + T f − ωb| Unlocked ω j = −ωb + T i + T o
ω = 0 and T S ≥ |T i + T f − ωb| Locked ω=0
function of tire pressure, normal force, and velocity. Specifically:

My = Re a + b|Vx | + cV2x FZB pai tanh(4Vx ) (12)
If the brakes are enabled, the block determines the braking locked or unlocked
condition based on an idealized dry clutch friction model. Based on the lock-up
condition, the block implements these friction and dynamic models (Tables 1 and 2).
The equations use these variables (Table 3).
The robot RAMBLER uses the Brake Type Disc, the block implements a disc
brake. A disc brake converts brake cylinder pressure from the brake cylinder into
force. The disc brake applies the force at the brake pad mean radius. The block uses
these equations to calculate brake torque for the disc brake.

μpπB2a Rm Npads
when N = 0
T= 2
μstatic pπB2a Rm Npads (13)
4
when N = 0
(R0 + Ri )
Rm = (14)
2
The equations use these variables (Table 4).
5 Simulation
The simulation of the controller is done by an implementation on model of robot of

the type with the aid of Toolbox of the Simulink, we tested the robot capability of the
reference speed tracking. We have made three simulations with different parameters
shows as follows (Table 5).
Table 3 Description of
Variables Description
parameters
ω Wheel angular velocity
a Velocity independent force component
b Linear velocity force component
c Quadratic velocity force component
Le; My Tire relaxation length; Rolling resistance torque
Ta; Tb Applied axle torque; Braking torque
Td Combined tire torque
Tf; Ti Frictional torque; Net input torque
Tk Kinetic frictional torque
μk Coefficient of kinetic friction
To; Ts Net output torque; Static frictional torque
Fc; Fx Applied clutch force; Longitudinal force
Reff Effective clutch radius
Ro; Ri Annular disk (outer radius; inner radius)
Re Effective tire radius
Vx; Fz Vehicle normal force; Longitudinal axle velocity
A; β Tire pressure and Normal force exponent
pi; μs Tire pressure; Coefficient of static friction
Table 4 Variables of
Variables Description
equations
T Brake torque
P Applied brake pressure
N Wheel speed
Npads Number of brake pads in disc brake assembly
μstatic Effective clutch radius
μ Disc pad-rotor coefficient of kinetic friction
Ba Brake actuator bore diameter
Rm Mean radius of brake pad force application on brake
rotor
R0 Outer radius of brake pad
Ri Inner radius of brake pad
Table 5 Variables of
Simulation Vehicle mass Absolute Kinetic
equations
KG pressure, Pabs friction
[Pa]: coefficient
Simulation 1 1000 1325 0.8
Simulation 2 1500 2000 2
Simulation 3 3000 5000 5
The figures below show the speed of the robot and the evolution of the control
variables according to time.
The figures Figs. 5, 6 and 7 obtained with simulation 1. The figures Figs. 8, 9,
10 and 11 obtained with simulation 2. The figures Figs. 12, 13, 14 and 15 obtained
with simulation 3. With the obtained results we can noted, that our control design is
robust and system is stable. the error is small and vehicle tacking trajectory in real
time. But with high perturbation (variation of parameters) like in simulation 3, we
need amore robust controller like Fractional order PID controller.
Fig. 5 Curve of error (Simulation 1)
Fig. 6 Curve of force (Simulation 1)

Fig. 7 Curve of speed vehicles (Simulation 1)
6 Conclusion
In this paper we proposed path tracking controller for an autonomous vehicle with
four wheels, the controller is based PID, is able to offer more tracking flexibility. After
several testes that we did we noticed that the starting point and the initial orientation
influenced on the trajectory tracking, for this reason we have to choose the initial
condition to decrease the value of the error, the proposed model can be also used in
the platooning projects (train of vehicles leader-follower), with different design of
control based in multi-architecture controller with different type of controller more
Fig. 9 Curve of brake (Simulation 2)
Fig. 10 Curve of acceleration (Simulation 2)
robust and optimizing with PSO, Genetic Algorithm, Wolf algorithm… In the future
works we plan, to integrate the proposed method to achieve a train of vehicles.
General characteristics of vehicles:
• Dimensions: 1.6 m length, 1.2 m width, 0.66 m height.
• Vehicle weight: 370 kg.
942

Y. Zennir et al.
Fig. 13 Curve of acceleration (Simulation 3)
• Payload: >320 kg.

• Distance between the front and rear wheel contact points: From 0.95 to 1 m.
• Wheels: pneumatic tires of 50 cm diameter.
• Suspension: Active/Semi active suspension system. Pneumatic.
• Distance between left and right wheel contact points is L = 1.05 m.
• Maximum linear speed: 80 km/h, 22.2 m/s.
• Battery: (LiFePO4); Range 44,8 < V < 59,2; 160 Ah (4 cells × 40 Ah);
Configuration:16S4P. Temperature: 0 °C < t < 70 °C. Operativo Control: BMS.
Fig. 14 Curve of brake (Simulation 3)
contactor 600A/80eVdc SPST-NO (Normally Open). Charge Control: External,

59,2 V/80 A, 4,7 kW.
• Motors: Brushless DC Electric Motor with Outrunner Configuration.
• Range: 36–72 Vdc, 500–3000 W. Nominal speed: 300–1000 rpm.
Acknowledgments The authors wish to thank the project DPI2015-65186-R, funded by the
Spanish Government, for the availability of the robot RAMBLER for this work.
References
1. https://www.uma.es/robotics-andechatronics/info/107721/FIRST-ROB/, (consulted
20/10/2018)
2. Baturone I, Sánches-Solano FJS, Ollero A (2004) Automatic design of fuzzy controllers for
car-like autonomous robots. IEEE Trans Fuzzy Syst 12(4):447–465
3. García C, Strang T, Lehner A (2008) A broadcast vehicle to vehicle communication system in
railway environments. ISVCS 2008, Dublin, Ireland, p 6
4. Bingyi L, Kejie J, Dong LN, Jianping W, Libing W (2017) A joint control-communication
design for reliable vehicle platooning in hybrid traffic. IEEE Trans Veh Technol 66(10):9394–
9409
5. Fernandes P (2010) Platooning of autonomous vehicles with intervehicle communications in
SUMO traffic simulator. In: 13th IEEE international conference on intelligent transportation
systems, Portugal, pp 1313–1318
6. Cao Y, Liu S, Visual Servo Control for wheeled robot platooning based on homography. In:
6th data driven control and learning system, pp 628–632
7. Messaoudène K, Azouaoui O (2015) Personalized dynamic model for a car-like vehicle “Robu-
car” used in localization. In: IEEE international conference on systems, man, and cybernetic,
pp 2533–2538
8. Campolo C, Mlinaro A, Araniti G, Berthet AO (2017) Better platooning control toward
autonomous driving: an LTE device to-device communication that meets ultralow latency
requirements. IEEE Veh Technol Mag 12(1):30–38
9. Ploeg J, Vande WOuw N, Nijmeijer H (2014) Lp string stability of cascaded systems/application
to vehicle platooning. IEEE Trans Control Syst Technol 22(2):786–793
10. Dumont P-E (2006) Tolérance active aux fautes des systèmes d’instrumentation, vol 1. Thèse
doctorat, université Lille, p 163
11. Bouibed K (2010) Contribution à la gestion de défaillances d’un train de véhicules électriques
légers autonomes, vol 1. Thèse doctorat, université Lille, p 181
12. Al-Mayyahi A (2015) Path tracking of autonomous ground vehicle based on fractional order
PID controller optimized by PSO. In: IEEE 13th international symposium on applied machine
intelligence and informatics, Slovakia, pp 109–114
13. Turki YA, Abdulkareem AA (2012) PSO-based optimum design of PID controller for mobile
robot trajectory tracking. Int J Comput Appl 47(23):30–35
14. Al-Mayyahi A, Wang W, Birch P (2015) Design of fractional-order controller for trajectory
tracking control of a non-holonomic autonomous ground vehicle. Int J Control Autom Electr
Syst 14
15. Al-Mayyahi A, Wang W, Birch P (2014) Adaptive neuro-fuzzy technique for autonomous
ground vehicle navigation. Robotics, pp 349–370
16. Zoleikha AB, Satadru D, Pisu P (2017) Real-time detection and estimation of denial of service
attack in connected vehicle systems. Ieee Trans Intell Transp Syst Xx(X):11
17. Allou S, Zennir Y, Belmeguenai A (2017) Fuzzy logic controller for autonomous vehicle path
tracking. In: 18th international conference on sciences and techniques of automatic control &
computer engineering—STA’2017, Monastir, Tunisia, pp 328–333
18. Zennir Y, Mechhoud E, Seboui A, Bendib R (2017) Multi-controller approach with PSO-
PIλDμ controllers for a Robotic Wrist. In: The 5th international conference on electrical
engineering—boumerdes (ICEE-B), Boumerdes, Algeria, p 7
Discrete Faults Diagnosis
for the Multicellular Converter:
Algebraic Approach
K. Haddadi, N. Gazzam, I. Ameur, and A. Benalia
Abstract This article addresses the diagnosis of cells fault in a multicellular

converter. The idea is based on using the distinguishability criteria for detected the
discrete fault. Necessary and sufficient condition for the detectability of discrete
faults are obtained. Residuals are generated using an algebraic time derivative esti-
mation. The proposed fault detection and diagnosis scheme is applied to the model
of a three-cell converter and the simulation results illustrate the effectiveness of the
proposed approach.
Keywords Switched affine system · Multicellular converter · Faults cells ·

Distinguishability between modes · Algebraic observer
1 Introduction
Modern industrial applications that require higher power level, efficiency and security
have significantly increased. These requirements can be achieved using high voltages
input and multilevel converter. The multicellular converter, called also flying capac-
itor converter, is a multilevel converter designed by the LEEI (Toulouse, France),
leads to a safe series association of cells working in a switching mode [1]. This
converter offers the possibility of reducing the voltage constraints among each cells
which result in lower conduction losses, high switching frequencies and the improve-
ment of the floating capacitors at specific references [2]. The multicell converters
K. Haddadi (B) · N. Gazzam · I. Ameur · A. Benalia

LACoSERE Laboratory, Amar Telidji University, Laghouat, Algeria
N. Gazzam
I. Ameur
A. Benalia

https://doi.org/10.1007/978-981-15-6403-1_65
948 K. Haddadi et al.
can offer a high power level using medium voltage semiconductors and improve the
quality of the output voltage while reducing the voltage constraints. However, the
large number of the semiconductors used in this structure increases the faulty of the
cells [3].
The multicell converter can be represented by a piecewise affine system. Indeed,
based on the theory of the hybrid systems, the instantaneous model can be transformed
to a hybrid model by an appropriate definition of discrete states [4–7]. This new hybrid
representation of the multicell converter offering a better analysis and design tools
for the cells fault diagnosis. Thus, it is necessary to develop a methodology to detect
the faulty on the cells as soon as possible to avoid permanent damage of the overall
system and to adopt a fault tolerant control strategy.
In the literature, there are many fault diagnostic tools where they can be divided
into model-based methods or without model methods, and the latter have seen
increasing interest in recent years. In this work, we consider the model-based method.
The general principle of this approaches is to generate the residuals. In the published
works, there are three kinds of approaches are classically used for residual genera-
tion: parity space approaches [8], identification algorithms [9] and observer- based
approaches [10–12]. In this paper, we focus our attention in the third kind.
In this paper, we propose fault diagnosis of the faults in the cells of the multicellular
converter based on the algebraic approach. The faults distinguishability is based
on the input-output distinguishability of the discrete modes and the residuals are
computed based on the recent robust algebraic differentiation algorithms.
The paper is organized as follows. In Sect. 2, a hybrid model of the multicellular
converter will be presented. Section 3 recall the basic tools in fault diagnosis of
switched linear monovariable systems. Discrete diagnosis of the failure cell using an
algebraic approach followed by the simulation results of three-cell converter will be
devoted in Sect. 5. A conclusion will summarize and discuss the obtained results.
2 Hybrid Modeling of Multicellular Converter
The structure of the serial multicellular converter, which is depicted in Fig. 1,

is composed by n cells, where each cell is composed of pairs of complementary
switches. The instantaneous model of the multicellular converter is given by [13]:

ẋ(t) = A(u(t))x(t) + B(u(t))
(1)
y(t) = C x(t)
where:
Discrete Faults Diagnosis for the Multicellular Converter … 949
Fig. 1 Topology of n-cells multicellular converter associated to a passive R − L load
⎡ u 2 −u 1
⎤
0 ··· 0 c1
⎢ .. .. .. .. ⎥
⎢ . . . . ⎥
A(u) = ⎢
⎢ u n −u n−1 ⎥
⎥ ∈ R n×n ,
⎣ 0 ··· 0 cn−1 ⎦
u 1 −u 2 u n−1 −u n
L
· · · L
− R
L
⎡ ⎤
0
⎢ .. ⎥
⎢ ⎥
B(u) = ⎢ . ⎥ ∈ R n , and C = 0 · · · 0 1 ∈ R 1×n .
⎣ 0 ⎦
E
u
L n

T
The continuous state vector x = vc1 · · · vcn−1 i L ∈ R n corresponds to the
floating capacitor voltages and load current. The elements E, R, L are the input
voltage, load resistor and load inductance respectively. The state of the kth cell is
controlled by the boolean inputs signals u k ∈ {0, 1}, k = 1, . . . , n, where u k =
1 (u k = 0) indicates that the upper switch is closed (open) and the lower switch is
open (closed).
From the model (1), we observe that the converter contains a continuous and
discrete variables. Hence, we can derive a hybrid model for the converter. More
precisely, we derive a piecewise affine PWA model for the converter based on the
possible values of the boolean input. Since u ∈ {0, 1}n , we can distinguish 2n discrete
modes that correspond to all possible configurations of the cells. Based on the hybrid
theory and the associated notations, we can put (1) in the following piecewise affine
systems model:

ẋ(t) = Aσ (t) x(t) + Bσ (t)
(2)
y(t) = C x(t)
where the switching signal σ (t) takes its values in a finite set of discrete modes
σ (t) ∈ Q = {q1 , . . . , q N } where N = 2n and Aq j , Bq j , C are defined by:
⎡ j
u 2 −u 1
j ⎤
0 ··· 0
⎢ . ..
c1
.. ⎥
⎢ . .. ⎥
⎢ . . . . ⎥
Aq j =⎢ ⎥ ∈ R n×n ,
⎢ 0 ···
j
u n −u n−1
j
⎥
⎣ 0 cn−1 ⎦
j j j j
u 1 −u 2 u −u
· · · n−1L n − RL
⎡ L ⎤
0
⎢ .. ⎥
⎢ . ⎥
Bq j = ⎢ ⎥ ∈ R n i = 1, · · · , N , C = 0 · · · 0 1 ∈ R 1×n .
⎣ 0 ⎦
E j
u
L n
k j
The index of the discrete state is given by: j = 1 + n−1 k=0 2 u k+1 .
Note that at this stage the transition between modes, i.e. the switching signal σ (t),
is completely controlled and independent of the continuous state. In the closed loop
control, where the objective is to regulate the current load charge and the capacitors
voltages at their references, the switching signal σ (t) is designed depending on the
evolution of continuous state to ensure the desired performances. Our objective in
this paper is to detect and isolate the faults cells of the converter. The considered cells
faults are the blocking of the cells. In the next section, we will give the sufficient
conditions which ensure the detection of faults. For that, we introduce the following
assumptions:
Assumptions:
– The sequence of discrete modes are observable and defined by
(t0 , q0 )(t1 , q1 ) . . . (t N , q N ).
– Only finite number of switches occur on any finite time interval (non-zenoness).
– System (2) satisfies the minimal dwell time i.e.:
∃T > 0, tk+1 − tk ≥ T , k = 0, . . . , N − 1.
– There is one and only one discrete fault occurs at a given instant.
3 Discrete Fault Diagnosis of Switched Linear Systems
Figure 2 summarized the principle stapes of the proposed daignosis scheme

Fig. 2 Block diagram representation of proposed FDI scheme
3.1 Detection of Fault Based on Distinguishability Criteria
The discrete mode of (2) contains modes of operation without faults and malfunctions
modes. If we want to study the detectability of the discrete faults based on the
measurement, the discrete modes need to be distinguishable.
We will recall some interesting results on the discrete faults diagnosis of the
switched affine systems.
A. Premilinary and definitions
Consider the switched affine system (2). The exogenous, switching signal σ (t) :
R + → Q is a piecewise constant right-continuous function taking value within the
set Q = {q1 , q2 , . . . , q N}. Moreover, σ (t) = qk , tk−1 ≤ t < tk means that the
subsystem (or the mode) Aqk , Bqk , Cqk are active in the corresponding
interval
and
σ (tk+ ) = qlimplies the system
is switched from the subsystem A qk , Bqk , C qk to the
subsystem Aql , Bql , Cql at the switching instant tk .
We will address the discrete state observation from the continuous input output
measurements. To this end, some preliminary definitions are needed.
Definition 1 (indistinguishability): For system (2), the mode qi ∈ Q is indistin-

guishable from the mode q j ∈ Q in finite time [0, t f ] if for every initial conditions
j j
x0i ∈ R n , x0 ∈ R n , we have y i (t, x0i , u(t)) = y j (t, x0 , u(t)) ∀t ∈ [0, t f ]. Otherwise
the two discrete modes are distinguishable.
Definition 2 (Mode observability): system (2) is said to be mode observable if every

pair of modes (qi , q j ) ∈ Q 2 are distinguishable.
The system (2) is said to be completely mode observable if every pair (qi , q j ) ∈ Q 2
of modes, qi distinguishable from q j (Or system (2) is said to be completely mode
observable if every pair of discrete modes is (2) distinguishable).
From these definitions, it is clear that we need the external differential represen-
tation of (2). This representation takes the form,
n qi −1

(n qi )
ak i y (k) (t) = bqi i = 1, . . . , N
q
y (t) + (3)
k=0
q
where ak i , bqi , i = 1, . . . , N are real numbers. We can put (3) in the following
compact form,
Dqi [s]y(t) = bqi i = 1, . . . , N (4)
Where s = dtd denote the differential operator and Dqi (s) ∈ R[s], i = 1, · · · , N
are polynomials differentials operators of the ring R[s]. We have immediately the
following properties,
T
• rank C T AqTi C T · · · (AqTi )n−1 C T = n qi i = 1, . . . , N
• n qi ≤ Mqi where Mqi , i = 1, . . . , N are the degree of the minimal polynomial of

Aqi .
• For each mode qi , the polynomial Dqi is unique.
Using the above definitions and (4) with the work [14], it is possible to derive an
algebraic condition on the mode observability.
B. Distinguishability of discrete modes
Consider the external differential representations (EDR) of two modes qi and qj ,

Dqi [s]y(t) = bqi /s
(5)
Dq j [s]y(t) = bq j /s
It is clear that the two discrete modes are indiscernible if they exhibit
qj
the same
qi
external differential representation which is equivalent to Dbq [s] = Dbq [s] . This condi-
i j
tion guaranties a strong distinguishability of the two modes which means that two
systems have different output trajectories. But we can find two different modes with
different external differential representations but with common output trajectories.
Indeed, this is the case if we look to the steady state regime of (5). In this regime,
the Eq. (5) became:
q
a0 i −bqi y 0
q = (6)
a0 j −bq j 1 0
bqi bq j
Proposition 1 The discrete qi , q j are distinguishable if and only if q
a0 i
= qj .
a0
i j
Proof From the condition abi = b j , we can focuses on the steady state solution of
0 a0
(6). Then, the Eq. (6) has non trivial solutions if the matrix is singular.˛
Illustratif Example: Consider the following switched system represented by two

modes:

ẋ(t) = −x(t) + 1 ẋ(t) = −2x(t) + 2
q1 : q2 :
y(t) = x(t) y(t) = x(t)
The EDR are given respectively by:

[s + 1]y(t) = 1
(7)
[s + 2]y(t) = 2
These two modes have different input-output differential representation but there
steady state regime are the same. As a result the two modes are indistinguishable.
3.2 Isolation of Fault Based on Algebraic Estimation
The isolation of fault,pursues the following steps:

1. Algebraic continuous state estimation
Consider a real-valued, N-th degree polynomial function of time:
N
y N(i) (0) i
y(t) ≈ y N (t) = t (8)
i=0
i!
Where y N(i) (0), i = 1, . . . , N is the unknown constant coefficients. The goal is to

obtain estimates of the time-derivatives of y(t), up to order N.
The following theorem gives a characterization of this estimate.
Theorem 1 [15] For all t ≥ T , the j order time-derivative estimate ŷ ( j) (t) of the
polynomial signal y(t) as defined in (8) satisfies the convolution:
T
( j)
ŷ (t) = H j (T, τ )y(t − τ )dτ, j = 0, . . . , N (9)
0
Where the convolution kernel:

(N + j + 1)!(N + 1)!
H j (T, τ ) =
T N + j+1
N− j

J
(T − τ )k1 +k2 (−τ ) N −k1 −k2
k1 =0 k2
k1 !k2 !(N − j − k1 )!( j − k2 )!(N − k1 − k2 )!(k1 + k2 )!(N − k1 + 1)
Depends on the order j of the time derivative to be estimated and on an arbitrary

constant time window length.
Assume that the dimension of observable state space is pi than the equations of
the output and their derivatives given by:
⎧
⎪
⎪ y(t) = cqi x(t)
⎪
⎨ ẏ(t) = cqi Aqi x(t) + Bqi u(t)
⎪ ..
⎪
⎪ .
⎩ p −1 p −2 p −3
y ( pi −1) (t) = cqi Aqii x(t) + cqi Aqii u(t) + cqi Aqii u̇(t) + · · · + Bqi u ( pi −2) (t)
(10)
It is clear that to obtain the observable state of the system, we must inverse
these equations by replacing the above derivatives by the derivatives given by (2) by
selecting N ≥ pi − 1.
Remarks The algorithm of this observer cannot be applied in the case when the
output derivatives is noisy, because the derivative operation is increasing noise which
gives a bad estimations of the observable state.
For observe the sequence of discrete modes on performs the same reasoning as
the observation of one mode, by accumulating information on the state to obtain the
complete state.
The principle of this observer shows that this observer is relatively simple because
we need to inverse the observability matrix to find the state. Additionally if there is
no disturbance on the measurement the algebraic observer will be the best choice to
reconstruct the system states.
2. Generation and evaluation of the residual
Residual generators are typically constructed by using a mathematical model of the

system. For instance, a residual can be obtained as the comparison between a value
estimated by a model and the corresponding measured quantity. The residue will
take the following form:
ri = ei , i = 1, . . . , N
A residual is a signal that is zero when the system under diagnosis is fault-free,
and non-zero when particular faults are present in the system.
4 Cells Faults Diagnosis for the Multicellullar Converter
For the multicellular converter, we can distinguish between continuous failures and
discrete failures. The first type is related to the storage elements (capacitors and
inductances) whereas the second one is related to the cells (switches). However, it is
common that the switches are most exposure to failure. This is due to the increasing
internal temperature problem of the power interrupter caused by an overload due to
a power surge causing an avalanche.
In this work, we are interested to analyze of the cell faults diagnosis of the multicell
converter (there are no parametric faults for). More precisely, we are interested on
the detection of the ‘blocking’ in closed or in open state of the cells. Thus, two states
of fault can occur in each cell.
– Application to the three cells converter (Simulation result and discussion)
The desired sequence of discrete modes are determined by the PWM control with
a period T = 0.02s and αr e f = 0.3 to provide the boolean control as in [16]. The
parametre of the converter are:E = 1500V, R = 10, L = 1H, c1 = 720μF, c2 =
720μF, and there initial condition of the proposed observer given by: [ vc1 vc2 i L ] =
[ 300 600 20 ]
Figure 3 show the evolution of the floating voltages and currents of cells without
faults. we can notice the natural balancing around the desired capacitor voltages:
vc1 = E/3 = 500, vc2 = 2E/3 = 1000.
The estimation error converges to zero in finite time which is clear in Fig. 4. These
results confirm the effectiveness of the proposed observer.
We must know the consequences as a preliminary of various blockings of cells on
the variables of state of the system. With this intention, one established signatures of
operation with faults. The following figures show the estimation errors of the state
space variables. Faults are introduced at time t = 0.6 s (Figs. 5 and 6).
We must know the consequences as a preliminary of various blockings of cells
on the state variables of the system. With this intention, one establishes signatures
of operation with faults. The results can be summarized in the signature Table 1 that
describes the consequences of the introduction of fault in order to model the residuals
vector.
Therefore, three residuals associated with each switch are defined, ri with i =
1, 2, 3. These residuals are generated by a threshold test of estimation errors and
can take three states 0, −1 or 1. State 0 to indicate that there is no fault, the state
−1 for fault in switch blocked in opening and state 1 for fault in switch blocked in
closing (Figs. 7 and 8).
550
real V c1
500 estimed V c1
450 510
500
400
490
350
480
0.6 0.605 0.61 0.615 0.62
300
250
0 0.2 0.4 0.6 0.8 1
Time (s)
-a-
1000
1000
990
500
980
0.6 0.605 0.61
real V
c2
estimed V
c2
0
0 0.2 0.4 0.6 0.8 1
Time (s)
-b-
Fig. 3 Evolution of the real states and the estimated states (a and b)
5 Conclusion
In this paper, an algebraic approach for the cells faults diagnosis for the multicell
converter has been proposed. The principle of this approach is based in detecting
the blocking of the cells using the analytic distinguishability condition, and than
know the nature of this fault, whether is locked in opening or closing. For that we
have used an algebraic observer. Simulation results have shown the effectiveness of
the proposed approach. As a future work, a diagnosis methodologie in the tolerant
control (FTC) will be investigated.
400 20
0.03
0
300
E s t im a t io n e rro r V c 2 (V )
Estim ation error V c 1 (V )
0.02
0.05
-20
0.01
200
0.04
0 -40
0.03
100 -0.01
-60
-0.02 0.02
0.2 0.205 0.21 0.215 0.22
0 -80
0.01
0.2 0.21 0.22 0.23 0.24
-100 -100
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)
Fig. 4 Evolution of the error state
0 0 30
-100 -50 25
Estimation error V I
Estimation error V c 2
Estimation error Vc1
-200
-100 20
-300
-150 15
-400
-200 10
-500
-600 -250 5
-700 -300 0
0 0.2 0.4 0.6 0.8 1 0 0.5 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s) Time (s)
Fig. 5 Blocking of u 1 = 0
700 300 40
600 250 30
Estimation error V c 2
Estimation error Vc1
Estimation error V I
500 20
200
400 10
150
300 0
100
200 -10
100 50 -20
0 0 -30
0 0.5 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fig. 6 Blocking of u 1 = 1
Table 1 Table of signature in

Blocked switch r1 r2 r3
the presence of fault
No fault 0 0 0
u 1 blocked in opening −1 0 0
u 1 blocked in closing 1 0 0
u 2 blocked in opening 0 −1 0
u 3 blocked in opening 0 0 −1
0 1 1
-0.2
0.5 0.5
-0.4
r3
r2
r1
0 0
-0.6
-0.5 -0.5
-0.8
-1 -1 -1
0 0.2 0.4 0.6 0.8 1 0 0.5 1 0 0.2 0.4 0.6 0.8 1
Fig. 7 Residuals for blocking of u 1 = 1
1 1 1
0.8
0.5 0.5
0.6
r1
0 0
r3
r2
0.4
-0.5 -0.5
0.2
0 -1 -1
0 0.2 0.4 0.6 0.8 1 0 0.5 1 0 0.2 0.4 0.6 0.8 1
Fig. 8 Residuals for blocking of u 1 = 0
References
1. Meynard T, Foch H (1992) Multi-level Conversion: high voltage choppers and voltages sources
inverters. In: Proceeding of the 23rd annual power electronics specialists conference. IEEE
2. Wilkinson RH, Meynard T, Du H, Mouton T (2006) Natural balance of multicell converter: the
general case. IEE Trans Power Electron 21(6):1658–1666
3. Shahbazi M (2012) Contribution a l’etude des convertisseurs statiques Ac-dc-Ac tolerants aux
defauts. Ph.D. thesis, Lorrine, Fance
4. Benmansour K, Benalia A, Djemai M, Deleon J (2007) Hybrid control of a multicellular
converter, non-linear analysis. Hybrid Syst (1):16–19
5. Ameur I, Gazzam N, Benmiloud M et al (2019) Int J Dyn Control. https://doi.org/10.1007/s40
435-019-00592-z
6. Haddadi K, Gazzam N, Benalia A (2019) Algebraic observer design for switched linear systems
applied to multicellular converters for estimating capacitor voltages. Electroteh Electron Autom
67(1):28–34
7. Gazzam N, Benalia A (2018) Geometric approach for the observability of switched systems:
application to three-cells DC/DC converter. In: 2018 International conference on applied smart
systems (ICASS). IEEE, pp 1–5
8. Cocquempot V, El Mezyani T, Staroswiecki M (2005) Fault detection and isolation for hybrid
systems using structured parity residuals. In: IEEE conference on decision and control
9. Simani S, Bonfe M (2004) Modelling and identification of residual generator functions for fault
detection and isolation of a small aircraft. In: 43rd IEEE conference on decision and control,
pp 4324–4329
10. Christophe C, Cocquempot V, Jiang B (2002) Link between high gain observer-based residual
and parity space one. In: Proceedings of the American control conference anchorage, pp 2100–
2105
11. Maradi S, Benmansour, Herizi KK, Tadijine M, Bouchrit MS (2013) Sliding mode and fault
tolerant for multicellular converter four quadrants. J Electr Power Syst Res 128–139
12. Benzineb O, Taibi F, Laleg-Kirati TM, Boucherit MS, Tadjine M (2013) Control and fault
diagnosis based Sliding mode observer of a multicellular converter: hybrid approach. J Electr
Eng 64(1):1335–3632
13. Gateau G, Fadel M, Maussion P, Bensaid R, Meynard T (2002) Multicell converters: active
control and observation of flying capacitor voltages. IEEE Trans Ind Electron 49:998–1008
14. Fliess M, Join C, Perruquetti W (2008) Real-time estimation for switched linear systems. In:
IEEE conference on decision and control
15. Zehetner J, Reger J, Horn M (2007) A derivative estimation toolbox based on algebraic methods
theory and practice. In: 16th IEEE international conference on control applications
16. Gazzam N, Benalia A (2016) Observability analysis and observer design of multicellular
converters. In: 2016 8th international conference on modelling, identification and control
(ICMIC). IEEE, pp 763–767
Fault Diagnosis of Uncertain Hybrid
Actuators Based Model Predictive
Control
Zahaf Abdelmalek, Bounemeur Abdelhamid, Bououden Sofiane,

Abstract This paper presents an online controlling and fault diagnosis scheme for
hybrid (electric and pneumatic) actuators with nonlinear uncertain discrete time
system. The proposed approach features a modular observer based on model predic-
tive control (FDMPC). The fault detection design can detect faulty states and para-
metric faults based on predictions states of hybrid system that using to guarantees no
false-positives alarms due to systems uncertainties and occurred faults. On the other
hand, based on the fault conditions that have occurred, the proposed scheme is able
to reconstruct the time varying faults directly without using an isolation and identi-
fication task, by calculating approximating the unknown parameter values based on
necessary conditions. The simulation results of a hybrid actuators of an industrial
arm shown the effectiveness of the proposed approach.
Keywords Fault diagnosis · Hybrid actuators system · Model predictive control

(MPC)
Z. Abdelmalek (B) · B. Abdelhamid

Department of Electronics, Faculty of Technology, University Frères Mentouri Constantine 1,
B. Abdelhamid
B. Sofiane
Department of Electronics, University of Khenchela, Khenchela, Algeria
I. Boulkaibet
American University of the Middle East, Egaila, Kuwait

https://doi.org/10.1007/978-981-15-6403-1_66
962 Z. Abdelmalek et al.
1 Introduction
The real applications systems converted to complex performances on account of

faults and external disturbances, which lead to systems weakness. Due to continu-
ally improve controlling process and delivered high performances several approaches
are proposed. The diagnostic techniques have continue increasing interest by the
researchers in the area of applied research and industrial domain. Basically, the
useful of this strategies are keeping a good processing and guarantee the safeness
of this systems by according an appropriate fault diagnosis scheme. Generally, fault
diagnosis includes three aspects: fault detection loop, fault isolation procedure and
fault identification reconstruction performances systems. The analyzing of fault diag-
nosis can be addressed from paper to another, for example studying the problem of
fault in wind turbine energy [1, 2] each paper proposed a model based fault diag-
nosis. The online fault diagnosis model is analyzed for nonlinear power systems [3],
this strategy used also to keep high performances of flexible manufacturing system
[4]. To generate an appropriate fault diagnosis scheme, the designing of input signal
using zonotops based outputs systems is presented in [5]. Combined between fault
tolerant and fault diagnosis against actuators faults for wind turbine is analyzed [6].
Some diagnosis scheme of failures focuses on one task of fault diagnosis as fault
isolation loop for fuzzy systems [7]. This kind of systems are analyzed against input
faults by using a proportional integral observer for fault diagnosis compensation
[8]. Also an observer technique is employed for fault diagnosis and fault tolerant
control for unknown time-varying fault occurred on nonlinear systems [9]. In fact,
several approaches on fault diagnosis is presented, to give an overview of existing
methods, an investigation about fault diagnosis and fault tolerant control is presented
in [10, 11].
On the other hand, model predictive control is considered as efficient technique,
for this reason a combination with fault diagnosis strategy increase the ability to
compensate the effects of faults. In [12] an integration of fault diagnosis with model
predictive control in analyzed. In the same vein, model predictive control based
fault detection and reconstruction algorithm for autonomous driving vehicle [13]. In
this paper, an approach for controlling hybrid actuators is proposed based on fault
diagnosis predictive control.
The contribution of this paper is presenting an online controlling and fault diag-
nosis strategy for hybrid actuators systems. Basically, this approach consists of two-
folders a fault detection design and fault states reconstruction procedure. The first is
based on predicted hybrid states system to ovoid a false-positive alarms due to hybrid
systems uncertainties and occurred faults, according the dynamics error. Secondly,
a reconstructing procedure is applied with necessary conditions, that able to calcu-
late approximately the desirable and unknown parameter values based on predictive
control performances and input control based on dynamic error. The optimal control
problem of hybrid dynamic model is ensuring the stability and robustness based on
sufficient conditions. The results show the effectiveness of the studying control law
by stabilizing the constrained faulty hybrid systems.
Fault Diagnosis of Uncertain Hybrid Actuators … 963
The remain paper is organized as follows: the notations and background elements
is introduced in Sect. 2. Section 3 presents the proposed strategy to monitoring and
controlling the actuators hybrid system. In Sect. 4, we illustrate the electiveness of
the proposed scheme by simulation results.
2 Problem Formulation
The healthy hybrid discrete-time system is represented as:

x(k + 1) = Ai x(k) + Bi u(k)
(1)
y(k + 1) = Ci x(k)
To estimate the state in the feedback channel, an observer is designed for the
faulty uncertain hybrid discrete-time system is:
⎧
⎨ x̂(k + 1) = Ai x̂(k) + Bi u(k) + L y(k) − ŷ(k) + Fa fˆf (k)
ŷ(k + 1) = Ci x̂(k) + Fs fˆf (k) (2)
⎩
fˆf (k) = F y(k) − ŷ f (k)
On the other hand, the dynamics error is defined as:
e(k) = x(k) − x̂(k) (3)
The faulty uncertain hybrid system can be written as:

⎧
⎨ x̂(k + 1) = Ai x̂(k) + Bi u(k) + (LCi + Fa FCi )e(k)
ŷ(k + 1) = Ci x̂(k) + Fs FCi e(k) (4)
⎩
fˆf (k) = F y(k) − ŷ f (k)
where Ai , Bi and Ci are compatible state matrices, K is the controller gain matrix
and L is the observer gain matrix which can be designed by standard methods. Fa ,
and Fs are the actuator (system) and the sensor fault parameters, respectively; ŷ is
the sensors measurement, fˆf is the output faults.
Therefore, the control estimation to be generated on the controller side is:
u(k) = K x̂(k) (5)

3 Fault Diagnosis Predictive Control
From Eq. (4) we consider the next closed loop discrete-time hybrid system:

x̂(k + 1) = (Ai + +Bi )x̂(k) + (LCi + Fa FCi )e(k)
(6)
ŷ(k + 1) = Ci x̂(k) + Fs FCi e(k)
The previous uncertain hybrid system can be written as:

x̃(k + 1) = A p (k)x̃(k) +
B p ũ(k)
(7)
ỹ p (k + 1) = C p x̃(k)

(Ai + Bi K ) 0n Ci (L + Fa F)
Ap = , Bp = ,
Ci 0n Fs FCi

p = I 0n
C
0n I
The ỹ p is the output of the augmented hybrid system, and the control input u(k),
represents the desired energy of the current step k. The matrix A p represents the

system matrix of the preview model while the matrix B p is the input. The C p matrix
is the weighting matrix used to construct performance of optimization quadratic
function J.

h
2
min Jh (k) = ỹ P (k + j/k) − yd (k + j/k)
u(k+ j/k) j>0
j=0

h
2
+ λ ũ(k + j/k) (8)
j=0
ŷh,min ≤ ŷh (k + j/k) ≤ ŷh,max , j ≥ 0, h = 1, 2, . . . , q

u h,min ≤ u h (k + j/k) ≤ u h,max , j ≥ 0, h = 1, 2, . . . , p
where
⎧
⎨ ỹ P (k) = [y(k + 1)y(k + 2) . . . y(k + N )]
T
ũ(k) = [u(k)u(k
+ 1) . . . u(k + N )] T
(9)
⎩
λ = diag λ(k + 1) . . . λ(k + N )
2
where, ỹ P (k + j/k) − yd (k + j/k) , represents de weighting sum of future error
between estimate states and desired trajectory.
2
And λ ũ(k + j/k) represents the energy of input control.
The h and λ represents the prediction step and weighting factor for input,
respectively. And k is the step of present state.
Then, the predictive model ỹ P (k) is expressed as follows
ỹ P (k + 1) = G(k)x̃(k) + H u(k) + M ũ(k) (10)
where,

G= Cp A
p C
p A p A
2p . . . C Np T
⎡ ⎤
0n 0n . . . 0n
⎢ C p Bp 0n . . . 0n ⎥
⎢ ⎥
H =⎢ .. .. . . .. ⎥
⎣ . . . . ⎦
N −1
C p A p Bp 0n . . . 0n 0
T
M= Cp p A
BpC p p A
Bp . . . C Np −1
Bp
In Eq. (10), matrices G, H and F are respectively the augmented state matrix
system, input control and faults occurred on actuators/sensors matrices of the state
space of hybrid system in (8).
On the other hand, in this paper the u term indicates the desired energy for the
position control of h step prediction horizon for industrial hybrid actuator arm. More-
over, on the online optimization for relaxation conditions we hold the (10) in (8),
the cost function can be arranged as (11). In this paper, with the constrains of the
control input, the cost function can have written as mentioned in (11), which the
performance index J defined to compute the optimal solution and predicted states.
1 T
min u P ũ + R T u Subject to W u ≤ Z (11)
u, j>0 2
where W and Z are representing the upper and lower energy limits, respectively.
To deal with actuators and sensors failures the P and R matrices in (11) are defined
as follows:
P = λV T V + H T H (12)
R = H T (G(k)x̃(k) + Mu(k)) (13)
where
T T
W = Wupper , Wlower , Z = Z upper
T
, Z lower
T
⎡ ⎤
v1,1 v1,2 0 . . . 0
⎢ 0 v2,2 v2,1 . . . 0 ⎥
⎢ ⎥
⎢ .. ⎥
V =⎢
⎢ 0 0 0 ... . ⎥
⎥
⎢ . .. .. ⎥
⎣ .. . 0 . vn−1,n ⎦
0 0 0 . . . vn,n
In this stage to test the convexity and feasibility of (11) for the FDMPC, the
definiteness of Matrix P hasbeen evaluated. The quadratic function in (11) is convex
if and only if matrix P is positive semi-definite. Therefore, the quadratic function
defined in (11) is strictly convex if and only if matrix P is positive definite. Then,
the optimal control input can be computed based on the Quadratic Programming
optimization.
In this part, we applied the proposed approach to an industrial arm with hybrid
actuators, for tracking trajectory.
The dynamics model is structured as hybrid actuation, the combination of a pneu-
matic and an electric motor actuator. The both faults (sensors and actuators) and
external disturbances are considered for proving the effectiveness of the proposed
scheme.
The dynamics hybrid actuation robot muscle model is described by the general
equation given by [13, 14], which the friction term, external disturbance, and
environment contact is taking into consideration.
The joint hybrid dynamic model is expressed as follows:
⎧
⎨ I (θ )θ̈ + M θ, θ̇ θ̇ + mgl cos(θ ) = τop v p , vm
τop v p , vm = τ p (P, θ ) + τm (vm , θ ) (14)
⎩
Ṗ = −Fvp P + v p

where, I (θ ) ∈ Rn∗n is the Inertia Matrix, M θ, θ̇ ∈ Rn∗n is external forces, friction
forces and environment contact matrix, m and l are the mass and length of the
arm respectivily. θ and θ̇ denotes the joint angle and velocity. P is the pneumatic
pressure, τop is the effective output torque generated by the electric motor τm and
the pneumatic actuator τ p . v p and vm denote the input voltages for the pneumatic
pressure and electric motor. Thus, we note the transformation coefficient from the
pneumatic pressure to a electric torque Fvp like in [14] .
The weighting matrices are:
Q0 = I, R0 = 1,
Table 1 Industrial arm

Parameters Description Numerical value
parameters
l Length of the arm 0.5 m
m Weight of the arm 0.45 kg
g Gravity 9.8 m/s2
θ Range angle joint −π/5 ~ π/5
1.5
Reference
FDMPC Approach
0.5
-0.5
-1
0 1 2 3 4 5
Time [sec]
Fig. 1 Evolution of position
The results from FDMPC, desired position, input control and position error are
presented in Figs. 1, 2 and 3.
The red line in Figs. 1 and 3 represent the predicted position and position error
of an industrial hybrid actuators arm by the MPC scheme constructed in this study
without faults.
– The case of occurred faults
In this section, the study approach shows the ability and robustness for reconstruction
states system in presence of faults based on fault diagnosis model predictive control.
The result in Figs. 4 and 5 shows that the conception of optimal control law gives
an acceptable convergence time to desired trajectory. Figure 6 shows the energy of
control scheme to operating the occurred faults.
Fig. 2 Evolution of control input
0.4
0.2
Position Error [rad]
-0.2
-0.4
-0.6
-0.8
0 1 2 3 4 5
Time [sec]
Fig. 3 Evolution of position error


Fig. 6 Evolution of position error
5 Conclusion
The paper presents a strategy for diagnosis of faulty hybrid actuators systems based
on predictive control. A scheme of reconstructing the faulty states is proposed based
on two stage; the first is to predict the no false-positive alarm, secondly a compen-
sation procedure using predictions states and input signal based on dynamics error
by calculating the desired values. The systems performances are guarantee in case
of uncertainties or defaults. In spite of occurred faults and disturbances the stability,
robustness and the feasibility are ensured for a class of hybrid systems.
References
1. Simani S, Farsoni S, Castaldi P (2015) Fault diagnosis of a wind turbine benchmark via
identified fuzzy models. IEEE Trans Ind Electron 62(6):3775–3782
2. Sanchez H, Escobet T, Puig V, Odgaard P (2015) Fault diagnosis of an advanced wind turbine
benchmark using interval-based ARRs and observers. IEEE Trans Ind Electron 62(6):3783–
3793
3. Pana W, Yuanb Y, Sandberg H, Gonçalves J, Stan GB (2015) Online fault diagnosis for nonlinear
power systems. Automatica 55:27–36
4. Fakhfakh O, Toguyeni A, Korbaa O (2016) On-line fault diagnosis of FMS based on flows
analysis. J Intell Manuf
5. Scott JK, Findeisen R, Braatz RD, Raimondoc DM (2014) Input design for guaranteed fault
diagnosis using zonotopes. Automatica 50:1580–1589
6. Badihi H, Zhang Y, Hong H (2015) Wind turbine fault diagnosis and fault-tolerant torque load
control against actuator faults. IEEE Trans Control Syst Technol 23(4):1351–1372
7. Schwung A, Beck M, Adamy J (2015) Fault diagnosis of dynamical systems using recurrent
fuzzy systems with application to an electrohydraulic servo axis. Fuzzy Sets Syst 277:138–153
8. Youssef T, Karimi HR, Chadli M (2014) Faults diagnosis based on proportional integral
observer for TS fuzzy model with unmeasurable premise variable. In: IEEE international
conference on fuzzy systems
9. Kabore R, Wang H (2001) Design of fault diagnosis filters and fault-tolerant control for a class
of nonlinear systems. IEEE Trans Autom Control 46(11):1805–1810
10. Gao Z, Cecati C, Ding SX (2015) A survey of fault diagnosis and fault-tolerant techniques—
Part I: fault diagnosis with model-based and signal-based approaches. IEEE Trans Ind Electron
62(6):3757–3767
11. Gao Z, Cecati C, Ding SX (2015) A survey of fault diagnosis and fault-tolerant techniques—
Part II: fault diagnosis with knowledge-based and hybrid/active approaches. IEEE Trans Ind
Electron 62(6):3768–3774
12. Prakash J, Narasimhan S, Patwardhan SC (2005) Integrating model based fault diagnosis with
model predictive control. Ind Eng Chem Res 44(12):4344–4360
13. Shin D, Seitz F, Khatib O, Cutkosky M (2010) Analysis of torque capacities in hybrid actuation
for human-friendly robot design. In: IEEE international conference on robotics and automation,
Anchorage, Alaska, USA, pp 799–804
14. Ishihara K, Morimoto J (2018) An optimal control strategy for hybrid actuator systems:
application to an artificial muscle with electric motor assist. Neural Netw 99:92–100
Fault Tolerant Predictive Control
for Constrained Hybrid Systems
with Sensors Failures
Zahaf Abdelmalek, Mohamed Chemachema, Bououden Sofiane,

Abstract In this paper, a fault tolerant predictive control scheme is developed for
class of hybrid systems in the presence of unknown sensor’s faults and external
disturbances. Based on the error dynamics and estimated states a new control law
scheme is proposed for the reformulation of the optimization quadratic program-
ming problem of a nonlinear faulty of hybrid system sensors based on predictive
control. The proposed design allowed to cope with sensors faults, disturbances and
compensate the faulty state effect of hybrid systems. Stability Conditions expressed
on terms of LMIs is guaranteed based on Lyapunov function. Numerical Simula-
tions on industrial manipulator are presented to show the effectiveness of proposed
scheme.
Keywords Sensors failures · Hybrid system · Model predictive control · Linear

matrix inequality (LMI)
1 Introduction
The hybrid system control aspect has taken attention of developers in last decades,
especially in industrial domain for its tolerance and functional in control process. In
real time a good prediction of future behavior is useful technic to ensure the perfect
Z. Abdelmalek (B) · M. Chemachema

Department of Electronics, Faculty of Technology, University Frères Mentouri Constantine 1,
M. Chemachema
B. Sofiane
Department of Electronics, University of Khenchela, Khenchela, Algeria
I. Boulkaibet
American University of the Middle East, Egaila, Kuwait

https://doi.org/10.1007/978-981-15-6403-1_67
functionally of the hybrid systems, as the employing of MPC approach. Basically,

the model for the prediction is used to predict and estimate the next performance
and always searching for the optimal solution at each sample time; however, many
approaches are presented to get the best solution for improving the control system
and online optimization time. Robustness is very important criteria the guarantee
the high performances of predictive control, which proposed in [1] to controlling
the nonlinear active suspension systems. A photovoltaic/diesel battery hybrid power
system used a model predictive control to switched energy dispatching [2]. In hybrid
electric vehicle, the predictive control is involved for obtaining the cost optimal
energy management [3]. Also, this technic is used to control a robot arm position of
a discrete time of hybrid actuators pneumatic-electric [4]. The advantage of model
predictive control is that can have combined with other strategies as neural network
for hybrid systems [5]. Moreover, derivative free state estimators have using for
combination between state estimation and nonlinear predictive control for improving
the control of autonomous hybrid system [6].
In fact, model predictive control has been widely used in the real time and
industrial application to cope with undesired inputs, therefore, can have occurred
mismatch functional in presence of faults and disturbances. This situation leads
many researchers to propose strategies to keep the highest systems performances.
A fuzzy adaptive controller based on back-stepping is proposed to design a fault-
tolerant tracking control for MIMO nonlinear systems with failures [7]. To compen-
sate the load current sensor failures, a fuzzy logic-based power management algo-
rithm is implemented to develop a fault tolerant control scheme [8]. This scheme
is studied and proposed in [9] by combination of feedback linearization controller,
with a state/fault estimator; to compensate the faults acting in the control Aircraft
system. The FTC scheme based on model predictive control is analyzed in [10]
which based on two hierarchical controllers: the pre-compensators model predictive
and global model predictive to compensate the faults effects of the wind turbine.
In [11] authors present a fault tolerant control based on model predictive control,
that used the estimated state for the optimal solution problem for tracking control.
Moreover, the optimization problem is designing to obtain relaxation conditions for
robustness and stability of hybrid model predictive control problems, this concept is
started with [12]. These performances are analyzed and presented in [13, 14] based
on Lyapunov function for stability of hybrid and switched system.
In this paper, we proposed a fault tolerant predictive control scheme for hybrid
process, based on new control feedback law a relaxation conditions are proposed
for the conception of estimator model predictive control for discrete faulty hybrid
dynamic system. The proposed design copes with sensor’ failures by employing the
optimal solution of predictive control procedure. Then, to improve system perfor-
mances in the case of faults and disturbances signals a relaxed control feedback law
is proposed, which based on the error dynamics and estimated states, the main idea
is to penalize the control signal if there is undesired control value. On the other
hand, this control scheme ensure the stability and robustness of constrained problem
based on Lyapunov function. A simulation results of an industrial arm shows the
Fault Tolerant Predictive Control for Constrained … 975
effectiveness and robustness of the proposed scheme by stabilizing the constrained

faulty systems.
This paper is presented as follows: Sect. 2 introduces the notations and results of
the basic elements. Section 3 presents the fault tolerant predictive control approach
of the closed-loop hybrid system with sensors failures. We conclude by Sect. 4,
simulation results are presented.
We consider the following hybrid discrete-time system represented as:

x(k + 1) = Ai x(k) + Bi u(k)
(1)
y(k + 1) = Ci x(k)
It is assumed that the states of the hybrid system are not measurable in this paper,
so to compensate the sensors fault in the feedback channel, an observer is designed:
⎧
⎨ x̂ f (k + 1) = Ai x̂ f (k) + Bi u f (k) + L y(k) − ŷ f (k)
ŷ f (k) = Ci x̂ f (k) + Fs fˆf (k) (2)
⎩
fˆf (k) = F y(k) − ŷ f (k)
Therefore, the hybrid control prediction to be generated on the controller side is:
u f (k) = u(k) + K f ((x(k) − x̂ f (k)) (3.a)
u(k) = K x̂ f (k) (3.b)
u f (k) = u(k) + K f e(k) (3.c)
where Ai , Bi and Ci are state matrices, K and K f are the controller gains matrices
and L is the observer gain matrix which can be designed by optimization methods.
Let us consider the following problem, which minimizes the given objective function
in an infinite horizon:
min max J∞ (k) (4.a)

u(k+ j/k)=k x̂(k+ j/k) j>0
ŷh,min ≤ ŷh (k + j/k) ≤ ŷh,max , j ≥ 0, h = 1, 2, . . . , q
∞

J∞ (k) = X (k + j) + U (k + j) (4.b)
j=0
Assumption: For any matrices W, V and a symmetric matrices Z > 0 and R > 0,
Consider the problem of finding an appropriate matrix R, the following statements
are equivalent and hold:
1/(W + V )T Z (W + V ) > 0 (5.a)

2/ − W T Z W − V T Z V < W T Z V + V T Z W + R (5.b)
Proof
(W + V )T Z (W + V ) > 0 ⇒

WT ZW + V T ZV + WT ZV + V T ZW > 0

− WT ZW − V T ZV < WT ZV + V T ZW
To be sure that inequality is always verified and the first term is always borne we can
write:

−W T Z W − V T Z V < W T Z V + V T Z W + R
End of Assumption.
The estimation error dynamics is defined as:
e(k) = x(k) − x̂ f (k) (6)
Using (1) and (2) we get the next closed loop discrete-time system:
⎧
⎨ x̂ f (k + 1) = (Ai + B i K )x̂ f (k) + Bi K f + LCi e(k)
e(k + 1) = Ai − Bi K f − LCi e(k) (7)
⎩
e f (k + 1) = −LCi e(k) − F Fs e f (k)
The previous augmented system can be written as:
x̃(k)
x̃(k + 1) = A(k) (8.a)
⎡ ⎤ ⎡ ⎤
1 (k)
A (Ai + Bi K ) Bi K f + LCi 0n 0n 0n
⎢A (k) ⎥ ⎢ (Ai − LCi ) 0n ⎥
⎢ 2 ⎥ ⎢ 0n 0n 0n ⎥
=⎢
A(k) ⎢A (k) ⎥ ⎢
⎥=⎢ 0n LCi F Fs 0n
⎥
0n ⎥
⎢ 3 ⎥ ⎢ ⎥
⎣ A4 (k) ⎦ ⎣ In 0n 0n 0n 0n ⎦
5 (k)
A 0n In 0n 0n 0n
where:
T
x̃(k) = x̂ Tf (k)e T (k)e Tf (k)x̂ Tf (k − 1)e T (k − 1) (8.b)
The new system can describe as follows:
1 (k)x̃(k), e(k + 1) = A
x̂ f (k + 1) = A 2 (k)x̃(k)
e f (k + 1) = A 4 (k)x̃(k),
3 (k)x̃(k), x̂ f (k) = A
5 (k)x̃(k),
e(k) = A
3 Fault Tolerant Predictive Control
Theorem Consider the closed loop estimate of a faulty hybrid system (7). Let the
input feedback controller be defined by (3), which is based on the extended state
observer, which meets the performance (2) for calculate the optimal control problem
of hybrid system; is globally asymptotically stable if there exists a positive define
matrix Q > 0; Y, Y f , L , F, Fs and G satisfying the following convex optimization
problem:
min γ (9.a)
Q,F,G,L

−1 x̂ Tf (k/k)
<0 (9.b)
x̂ f (k/k) Q
⎡ ⎤
GT A3T + A 3 G − P −1 ∗ ∗ ∗ ∗
⎢ 1 (A G + B Y ) A 1 Q ∗ ∗ ∗ ⎥
⎢ 4 i i ⎥
⎢ 1 ⎥
⎢ L A 5 0 n Q ∗ ∗ ⎥<0 (9.c)
⎢ 4
1/2 ⎥
⎣ Q 0 A1 G 0n 0n γ I ∗ ⎦

1/2
R0 Y A 4 + Y f A5 0n 0n 0n γ I

W Ci A4 + F Fs A
5

>0 (9.d)
4 + F Fs A
A 5 T CiT P
Proof Recall the closed-loop system in (7) and consider the following Lyapunov
function candidate:
V (x(k/k)) = x̂ Tf (k/k)P x̂ f (k/k) (10)
To ensure the stability of (2), we have:
V (x̂ f (k + j + 1/k)) − V (x̂ f (k + j/k)) ≤ −[X (k + j) + U (k + j)] (11)
−V (x̂ f (k/k)) ≤ −J∞ (k) (12)

We can write it:
max J∞ (k) ≤ V (x̂ f (k/k)) ≤ γ (13)

Ai ,Bi , j>0
Using schur’s complement to (13) we obtain:

−1 x̂ Tf (k/k)
<0 (14)
x̂ f (k/k) Q
– To ensure the stability of system (2), we have (7) it will be:
V (x̂ f (k + 1/k)) − V (x̂ f (k/k))

≤ − x̂ Tf (k/k)Q 0 x̂ f (k/k)

+ u Tf (k/k)R0 u f (k/k) (15)
That can be writing as:
[x̂ Tf (k + 1/k)P x̂ f (k + 1/k)] − [x̂ Tf (k/k)P x̂ f (k/k)]

< −[(x̂ Tf (k/k)Q 0 x̂ f (k/k))]

+(u Tf (k/k)R0 u f (k/k))
With substitution of x̂ f (k + 1/k) by (2) we obtain:
[((Ai + Bi K )x̂ f (k/k) + LCi e(k))T P((Ai + Bi K )x̂ f (k/k)

+ LCi e(k)) − [x̂ Tf (k/k)P x̂ f (k/k)]]
< −x̂ Tf (k/k)Q 0 x̂ f (k/k)

+ (K x(k) + K f e(k))T R0 K x(k) + K f e(k)
We can write:
1 + LCi A
((Ai + Bi K ) A 1 + LCi A
5 )T P((Ai + Bi K ) A 1T P A
5 ) − A 1

< −A 1T Q 0 A 4 + K f A
1 − K A 5 T R0 K A 4 + K f A
5
We multiple in the left by G T and by G in the right we get:
1 + LCi G A
((Ai G + Bi F) A 5 )T P((Ai G + Bi F) A
1 + LCi G A 5 ) − G T A 1 G
1T P A

1T Q 0 A
< −G T A 1 G − Y A4 + Y f A
5 R0 Y A
T
4 + Y f A
5 (16)
We replace P = γ Q −1 to the precedent inequality, we find:

(G T A 1 G − P −1 ) + ((Ai G + Bi F) A
1T + A 1 + LCi G A 5 )T
P((Ai G + Bi F) A1 + LCi G A5 ) − G T A
1T Q 0 A
1 G
T
− YA 4 + Y f A
5 R0 Y A 4 + Y f A
5 > 0
Using the previous assumption in this inequality we get:
(G T A 1 G − P −1 ) + 1/2((Ai G + Bi F) A
1T + A 1 )T

P((Ai G + Bi F) A 1 ) + 1/2 LA5 T P LA5

+ GT A1T Q 0 A
1 G + Y A 4 + Y f A 4 + Y f A
5 T R0 Y A 5 < 0 (17)
Using generalized schur’s complement to previous equation, we obtain:

⎡ ⎤
3T + A
GT A 3 G − P −1 ∗ ∗ ∗ ∗
⎢ 1 (A G + B Y ) A 1 ∗ ∗ ∗ ⎥
⎢ 4 i i Q ⎥
⎢ 1 ⎥
⎢ L A 5 0n Q ∗ ∗ ⎥<0 (18)
⎢ 4
1/2 ⎥
⎣ Q0 A 1G 0n 0n γI ∗ ⎦
1/2

R0 Y A 4 + Y f A5 0n 0n 0n γI
– Output Constraints:
yh,min ≤ yh (k + j/k) ≤ yh,max , ⇔ |yh (k + j/k)| ≤ yh,max ,
f or : j ≥ 0, h = 1, 2, . . . , q
ymax = W
y(k + j/k)max max y j (k + j/k)

j
With (7), we can write:

−1
5 P x(k)
4 + F Fs A

maxy(k)max ≥ maxCi A
j>0 j>0 max
The relaxed condition for output constraint can written as follows:

W Ci A4 + F Fs A
5

>0 (19)
4 + F Fs A
A 5 T CiT P
End of proof.
A. Example
In this paper, the dynamics model is structured as hybrid actuation, the combination
of a pneumatic and an electric motor actuator. Only the sensors’ faults and external
disturbances are considered for proving the robustness of the proposed scheme.
Generally, taking into consideration the friction term, external disturbance, and
environment contact, a rigid dynamics hybrid actuation robot muscle model is
described by the general equation given by [15, 16].
The joint hybrid dynamic model is expressed as follows:
⎧
⎨ I (θ )θ̈ + M θ, θ̇ θ̇ + mgl cos(θ ) = τop v p , vm
τop v p , vm = τ p (P, θ ) + τm (vm , θ ) (20)
⎩
Ṗ = −Fvp P + v p

where, I (θ ) ∈ Rn∗n is the Inertia Matrix, M θ, θ̇ ∈ Rn∗n is external forces, friction
forces and environment contact matrix, m and l are the mass and length of the arm
respectively. θ and θ̇ denotes the joint angle and velocity. P is the pneumatic pressure,
τop is the effective output torque generated by the electric motor τm and the pneumatic
actuator τ p . v p and vm denote the input voltages for the pneumatic pressure and
electric motor.
Thus, we note the transformation coefficient from the pneumatic pressure to a
electric torque Fvp , like in [16].
In order to achieve this, a unique state-space representation can be established to
describe the hybrid actuations system.
Where x is a state variable defined as:
T
x = θ, θ̇ , P (21)
The input u includes pneumatic and electric torque

T
u = τ p , τm (22)
The weighting matrices are:
Q0 = I, R0 = 1,
Figures 1 and 2 shows up dynamics for controlling position of an industrial arm,

which converges to reference position at some time for different initial states values.
Also, the optimal control law of hybrid system shows a good behavior a good control
with time with different energies in initial time.

To show the ability of the proposed strategy to handle with effects of faulty
hybrid systems, Figs. 3 and 4 depict how the optimal control law of hybrid system
performs to compensate failures of hybrid sensors’ faults and external disturbances
in acceptable range time.
Fig. 3 Evolution of position in presence of sensors faults/disturbances
Fig. 4 Evolution of control input in presence of sensors faults/disturbances

5 Conclusion
In this paper, a fault tolerant predictive control for hybrid systems with sensors fail-
ures and disturbances was presented. We implement the model predictive control to
reconstruct the states sensors for the feedback information that use for the optimiza-
tion of the control law, on the other hand, an observer is used to handle and estimated
the external disturbances. The gains of the optimal controller are based on solving
the relaxation optimization conditions based on new robust control law. Moreover,
the stability and the feasibility of the proposed approach is proved for a class of
nonlinear hybrid systems.
References
1. Bououden S, Chadli M, Karimi HR (2016) A robust predictive control design for nonlinear
active suspension systems. Asian J Control 18(1):122–132
2. Zhu B, Tazvinga H, Xia X (2015) Switched model predictive control for energy dispatching
of a photovoltaic-diesel-battery hybrid power system. IEEE Trans Control Syst Technol
23(3):1229–1236
3. Zhu B, Tazvinga H, Xia X (2020) Switched model predictive control for energy dispatching of
a photovoltaic-diesel-battery hybrid power system. IEEE Trans Power Electron 35(1):382–392
4. Bone GM, Xue M, Flett J (2015) Position control of hybrid pneumatic-electric actuators using
discrete-valued model predictive control. Mechatronics 25:1–10
5. Potocnik B, Mušic G, Škrjanc I, Zupancic B (2008) Model-based predictive control of hybrid
systems: a probabilistic neural-network approach to real-time control. J Intell Robot Syst 51:45–
63
6. Prakash J, Patwardhan SC, Shah SL (2010) State estimation and nonlinear predictive control of
autonomous hybrid system using derivative free state estimators. J Process Control 20(7):787–
799
7. Bounemeur A, Chemachema M, Essounbouli N (2018) Indirect adaptive fuzzy fault-tolerant
tracking control for MIMO nonlinear systems with actuator and sensor failures. ISA Trans
79:45–61
8. Tiar M, Betka A, Drid S, Abdeddaim S, Sellali M, Medjmadj S (2019) Fault tolerant control
of a smart PV-grid hybrid system. IET Renew Power 13(13):2451–2461
9. Tang Y, Patton RJ (2012) Active FTC for non-linear aircraft based on feedback linearization
and robust estimation. IFAC Proc Vol 45(20):210–215
10. Yang X, Maciejowski JM (2012) Fault-tolerant model predictive control of a wind turbine
benchmark. IFAC Proc Vol 45(20):337–342
11. Li J, Zhang D, Wang Z (2020) Novel MPC-based fault tolerant tracking control against sensor
faults. Asian J Control 22(1):1–14
12. Kothare MV, Balakrishnan V, Morari M (1996) Robust constrained model predictive control
using linear matrix inequalities. Automatica 32(10):1361–1379
13. Lazar M (2006) Model predictive control of hybrid systems: stability and robustness. University
of Technology, Eindhoven
14. Goebel R, Sanfelice RG, Teel AR (2012) Hybrid dynamical systems: modeling, stability, and
robustness. Princeton University Press. http://www.jstor.org/stable/j.ctt7s02z
15. Shin D, Seitz F, Khatib O, Cutkosky M (2010) Analysis of torque capacities in hybrid actuation
for human-friendly robot design. In: IEEE international conference on robotics and automation,
Anchorage, Alaska, USA, pp 799–804
16. Ishihara K, Morimoto J (2018) An optimal control strategy for hybrid actuator systems:
application to an artificial muscle with electric motor assist. Neural Netw 99:92–100
Active Adaptive Fuzzy Fault-Tolerant
Control for a Class of Nonlinear Systems
with Actuator Faults
Abdelhamid Bounemeur and Mohamed Chemachema
Abstract This paper deals with the problem of faulty and perturbed class of
nonlinear systems. The universal approximator system is used to cope with the nonlin-
earities and the actuator faults. Two controllers are proposed, the first is an adaptive
controller and the second is a robust controller. The proposed scheme allows to deal
with the external disturbances theoretically instead of approximation. The stability
is proved using Lyapunov theory. Simulation example on an inverted pendulums is
carried out to show the prompt and the accuracy of the proposed strategy.
Keywords Active fault-tolerant · Feedback linearization · Nonlinear systems ·

Lyapunov stability
1 Introduction
Recent years have been witnessed many techniques strategies in the area of control
theory, fault-tolerant control was one of the popular strategies. In fault tolerant
control, the problem is to design a control law such that when a fault occurs from
a sensor and/or actuator, the system keeps its own properties in terms of stability
and tracking performances. The main idea to ensure these properties is to make the
control law immune against these faults by canceling its bad effects.
Fault-tolerant control (FTC) techniques can be classified into passive and active
techniques [1, 2]. In passive fault-tolerant control (PFTC) systems, a single controller
with fixed structure and parameters is utilized to deal with all possible failure
scenarios assumed, a priori, to be known. Furthermore, the passive technique does
not require online fault detection, diagnosis, and control reconfiguration. The imple-
mentation of this technique is easy but it is more conservative [3–5]. Many passive
A. Bounemeur (B) · M. Chemachema

Laboratory of Automatic, Robotic and Control Systems, Department of Electronics, Faculty of
Engineering, University of Constantine 1, Constantine, Algeria
M. Chemachema

https://doi.org/10.1007/978-981-15-6403-1_68
986 A. Bounemeur and M. Chemachema
fault tolerant control approaches have been developed in the literature. For linear
systems, a reliable H∞ guaranteed cost control technique is proposed based on the
linear matrix inequality (LMI) technique [6]. Moreover, a reliable H∞ controller is
developed in [7, 8] based on LMI approach and static output feedback control. In
[9], a reliable non-fragile H∞ compensation filter is investigated. In [10–12], authors
proposed fault tolerant control techniques specifically for non-linear systems. In [10]
a combination of L 2 /H∞ fuzzy static output feedback control is proposed based on
iterative LMI approach. A reliable H∞ filter is designed in [11] for a class of nonlinear
networked systems using Takagi-Sugeno (T-S) fuzzy model and LMI technique. A
static output feedback fuzzy controller is employed in [12] for T-S fuzzy systems
with sensor faults based on LMIS .
In [13], authors combined adaptive disturbance observer with back-stepping
method under time-varying sensor faults for a class of nonlinear systems. In [14] an
active FT tracking controller to deal with sensor faults is proposed for a vehicle system
based on T-S fuzzy representation. In [15], an active fault-tolerant control approach
is presented for vehicle active suspension systems in finite-frequency region. In [16],
an active fault-tolerant controller is developed using neural networks combined with
sliding mode and the H2 performance index for spacecraft system under the loss of
effectiveness actuator fault. In [17], an active FTC based on back stepping is devel-
oped for a class of MIMO uncertain nonlinear systems. In this work, four kinds of
velocity sensor faults including bias, drift, loss of accuracy and loss of effective-
ness are considered. In [18], an active fault tolerant control, based on differential
geometry, and a fault detection and diagnosis are presented for a class of nonlinear
systems subjected to an additive fault. In [19], the authors investigated an adaptive
fault tolerant control for a class of nonlinear systems based on neural networks and
implicit function theorem with unknown bias and loss of effectiveness actuator faults.
Authors in [20], have investigated an indirect adaptive fuzzy fault-tolerant scheme
for a class of nonlinear systems with both actuator and sensor faults. A combina-
tion of fuzzy systems and backstepping technique allowed the online estimation of
all adaptive parameters and ensured the boundedness of all signals involved in the
closed-loop system.
In this present paper, an active fault-tolerant control scheme is presented. With
the help of Fuzzy Logic Systems (FLSs), The uncertain dynamic is approximated
along with the actuator faults, the proposed robust controller allows to deal with
the approximation errors due to the use of the fuzzy systems. Simulation results are
presented to show the effectiveness of the control strategy.
The remainder of the paper is organized as follows: Sect. 2 introduces the
problem formulation. In Sect. 3, an active fuzzy fault-tolerant control is developed
and designed along with the stability analysis derivation. Extensive simulations on
two-inverted pendulum is given in Sect. 4. The conclusion is presented in Sect. 5.
Active Adaptive Fuzzy Fault-Tolerant Control … 987
2 Plant Model and Control Objective
1. Problem formulation and preliminaries
⎧
⎪
⎪ ẋ1 = x2
⎪
⎪
⎪
⎨ ẋ2 = x3
.. (1)
⎪ .
⎪
⎪
⎪
⎪ ẋ = f (x1 , x2 , . . . xn ) + g(x1 , x2 , . . . xn )u
⎩ n
y = x1
We can put in this forms:

x (n) = f x, ẋ, . . . , x (n−1) + g x, ẋ, . . . , x (n−1) u y=x
Or equivalently

y (n) = f (x) + g(x)u
(2)
y = x1
where x = [x1 , . . . , xn ]T ∈ Rn , is the vector of the system;u ∈ R is the scalar

control input; y ∈ R is the scalar system output; f (x) and g(x) are unknown smooth
nonlinear functions.
If we consider the faults model given in the following Table 1.
where t f i denotes the time instant of failure
of the

ith sensor/actuator and
bi
denotes
its accuracy coefficient such that bε −b̄0i , b̄0i , where b̄0i > 0. kε k̄i , 1 , where
k̄i > 0 denotes the minimum sensor and actuator
effectiveness,

in which bi and k
are slowly varying respectively within −b̄0i , b̄0i and k̄i , 1 .
The faulty actuator will be written as
u(t) = k(t)u(t) + b(t)
Table 1 Actuator faults

Actuators Faults kinds Conditions Faults names
u(t) u(t) + b b(t) = 0, Bias (Lock in place)
u(t) + b(t) |b(t)| = λt, Drift
0<λ1
u(t) + b(t) |bi (t)| < b̄0i , Loss of accuracy
ḃi (t) → 0
k(t)u(t) 0 < k̄i ≤ k(t) ≤ 1 Loss of effectiveness
If we introduce the actuator faults, Eq. (8) will take the following form

y (n) = f (x) + f a (x, u) + g(x)u
(3)
y = x1
with f a (x, u) = g((k − 1)u + b(t))

In respect the dynamic of the system (3), the following assumptions will be made
Assumption 1 The order n of the system is know.
Assumption 2 The state vector is available for measurement.
Assumption 3 There exists an unknown continuous positive function D(x) as:
|ġ(x)| ≤ D(x)
The objective is to design an adaptive fuzzy controller for system (3) such that the
system output y(t) follows a desired trajectory yd (t) while all signals in the closed-
loop system remain bounded.
Regarding the development of the control law, the following assumption should
also be made:
Assumption 4 The desired trajectory yd (t) and its time derivatives yd(i) , i = 1, . . . , n
are smooth and bounded.
Assumption 5 The control gain g(x) is different to zero and its sign is known we
suppose that g(x) > g > 0 with g is unknown constant
Assumption 6 The approximation error is bounded as
|ε(x)| ≤ ε̄u
Now, let us define the tracking error vector as

T
e = e, ė, . . . ; e(n−1) ∈ Rn (4)
With
e(t) = yd (t) − y(t) (5)
e(n) = yd(n) − y (n) (6)
e(n) = yd(n) − f (x) − f a (x) − g(x)u (7)
If these functions ( f (x), g(x), f a (x) and D(x)) are knows, the objective of
control, can be made if we choose the ideal control law:
v − f (x) − f a (x) D(x) e T Pe

u = u∗ = + (8)
g(x) 2g(x)2 e T P B
where P is the solution of Lyapunov equation

v − f (x) − f a (x) D(x) e T Pe
e(n) = yd(n) − f (x) − g(x) + (9)
g(x) 2g(x)2 e T P B

D(x) e T Pe
e(n) = −k T e − g(x) (10)
2g(x)2 e T P B
We can write the dynamic error

D(x) e T Pe
ė = Ae + B −g(x) (11)
2g(x)2 e T P B
where
⎡ ⎤ ⎡ ⎤
0 1 ··· 0 0
⎢ . . ⎥ ⎢ .. ⎥
⎢ .. .. ⎥
A=⎢ ··· 0 ⎥ B=⎢
⎢ .⎥⎥
⎢ ⎥ ⎣
⎣ 0 0 ··· 1 ⎦ 0⎦
−kn −kn−1 . . . −kn 1
Define
1
V = e T Pe (12)
2g(x)
where P is a symmetric positive definite matrix satisfying the Lyapunov equation.
A T P + P A = −Q (13)
where Q > 0 using (12) and the error Eq. (10), we have the time derivative of V can
be write in this forms:

1 D(x) e T Pe
V̇ = − e T Qe + e T P Bg(x)−1 −g(x)
2g(x) 2g(x)2 e T P B
1
− e T Pe ġ(x)g(x)−2 (14)
2
1
V̇ = − e T Qe + V̇1 (15)
2g(x)
D(x)e T Pe 1 T
V̇1 = − − e Pe ġ(x)g(x)−2 (16)
2g(x)2 2
Using assumption 3 we obtain

D(x)e T Pe 1 T
−2
V̇1 ≤ − 2
+ e Peg(x) D(x) = 0 (17)
2g(x) 2
Using the Assumption 5
1
V̇ ≤ − e T Qe ≤ 0 (18)
2g(x)
Finally we conclude that the tracking errors and its derivatives converge asymp-
totically to zero e(i) (t) → 0 as t → ∞ for i = 0, 1, . . . , n − 1 [1], and the system
is globally stable.
Since the nonlinear function f (x), f a (x) g(x) and D(x) are unknown, the imple-
mentation of the ideal control law is impossible. In this situation, our goal is to
approach this ideal control law using fuzzy systems.
2. Description of fuzzy systems
It is shown that fuzzy systems (FS) are capable for approximating any real continuous
function on a compact set with arbitrary precision.
Let x = [x1 , . . . , xn ]T be the input of the FS and its output. For each xi is associated
j
m i fuzzy sets Fi in X i its universe of discourse, as for xi ∈ X i there is at least one
degree of membership μ F j (xi ) = 0 where i = 1, 2, . . . , n and = 1, 2, . . . , m i .
i n
The rules base of the FS has .N = i=1 m i . fuzzy rules of the form:
Rk : i f x1 is F̄1k and . . . and xn is F̄nk Then y = f k (x), k = 1, . . . , N

where F̄ik ∈ Fi1 , . . . , Fim i are linguistic values and f k (x) is a numerical function
of the output variable. In general, f k (x) is a polynomial function depending on input
variables. If f k (x) is a polynomial of zero order such as f k (x) = a k , we have the
Takagi-Sugeno zero order fuzzy system which is used in this work. Each rule has
a numerical conclusion, the output of the FS is obtained by calculating a weighted
average as given by the following relationship:
y(x) = w T (x)θ (19)
with

• θ = a 1 . . . a N : values of the conclusions in the fuzzy rules.

• w(x) = [w1 (x) . . . w N (x)]T ,
where
μk (x)
w N (x) = N , k = 1, . . . , N (20)
j=1 μ j (x)
n
with .μk (x) = i=1 μ F̄ik ., F̄ik ∈ Fi1 , . . . , Fim i .
which represents the weight of rule Rk .
3 Controller Design and Stability Analysis
In this section, our interest is to approach the ideal control law to realize tracking
of a given reference trajectory. To achieve these objectives fuzzy system is used
to estimate the control law as a whole (direct approach). According to property of
the universal approximation [21] of fuzzy systems, the ideal control law can be
approached by a fuzzy system of the form (19) as follows
u ∗ = w T (x)θ ∗ + ε(x) (21)
With ε(x) the approximation error, w(x) is a vector of fuzzy basis functions
assumed properly set in advance by the user, and θ ∗ is somehow the vector of optimal
parameters minimizing function |ε(x)|.

θ ∗ = argminθ supx u ∗ − w T (x)θ (22)
We assume that the approximation error is bounded as follows:
|ε(x)| ≤ ε̄u
With ε̄. An unknown positive constant, Note the terminal ε̄u depends on the fuzzy
basis functions selected, the more these functions are suitably chosen, the more
terminal ε̄u is small.
Since the optimal parameters θ ∗ are unknown, it is necessary to estimate for the
synthesis of controller. Whether θ the estimate of θ ∗ and will be calculated from
an adaptation algorithm, besides the fuzzy adaptive approximation of ideal law is
defined by
û = w T θ + u r (23)
Now we consider the following control law:
u = û = w T θ + u r (24)
u r it is a robust control to get around the problem of approximation errors, it is

defined by
σ2
u r = sign e T P B ε̂u − (25)
eT P B
The estimations parameters use the following laws:
θ̇ = γ e T P Bw(x) (26)

ε̂˙ u = n u e T P B (27)
σ̇ = −δ0 σ (28)
when σ is time varying parameter.

with n f > 0, γ > 0, δ0 > 0.
Theorem Consider the system (1). We assume that the assumptions (1–6) Are satis-
fied. The control law defined by (28) and (29) with adaptions law (30–32) ensures
the following properties:
• The tracking error and its derivatives converge to zero, e(i) (t) → 0 when t → ∞
for i = 0, 1, . . . n − 1.
• The output of the system, and its derivatives up to the order (n − 1) and the control
signal are bounded: y(t), ẏ(t), . . . , y n−1 (t), u(t) ∈ L ∞
Proof
e(n) = yd(n) − y (n) = yd(n) − f (x)
− f a (x) − g(x)u + g(x)u ∗ − g(x)u ∗ (29)
with u ∗ it is the ideal control law

e(n) = yd(n) − f (x) − f a (x) + g(x) u ∗ − u − g(x)u ∗ (30)
Replacing the Eqs. (7) and (35) become

D(x) e T Pe
e(n) = −k T e + g(x) u ∗ − u − g(x) (31)
2g(x)2 e T P B
u ∗ = w T θ ∗ + εu (x) (32)
u = û = w T θ + u r (33)
Then
u ∗ − u = w T θ̃ + εu (x) − u r (34)
with
θ̃ = θ ∗ − θ (35)

(n) D(x) e T Pe
e = −k e + g(x) w θ̃ + εu (x) − u r − g(x)
T T
(36)
2g(x)2 e T P B
Then the dynamics of the error can be written as follows:

D(x) e T Pe
ė = Ae + B g(x) w θ̃ + εu (x) − u r − g(x)
T
(37)
2g(x)2 e T P B
With
⎡ ⎤ ⎡ ⎤
0 1 ··· 0 0
⎢ . .. ⎥ ⎢ .. ⎥
⎢ .. ⎥
A=⎢ . ··· 0 ⎥ B=⎢
⎢.⎥
⎥
⎢ ⎥ ⎣
⎣ 0 0 ··· 1 ⎦ 0⎦
−kn −kn−1 . . . −kn 1
Until (|s I − A|) = s (n) + k1 s (n−1) + . . . + kn is stable (A stable), we know that

there exists a symmetric positive definite matrix P(n, n) that satisfies the Lyapunov
equation:
A T P + P A = −Q (38)
where Q is a symmetric positive definite matrix of arbitrary dimensions (nxn).

Whether V , is the Lyapunov function, then
1 1 T 1 2 1 2
V = e T Pe + θ̃ θ̃ + ε̃ + σ (39)
2g(x) 2γ 2n u u 2δ0
1
V̇ = − e T Qe
2g(x)

−1 D(x) e T Pe
+ e P Bg(x)
T
g(x) w θ̃ + εu (x) − u r − g(x)
T
2g(x)2 e T P B
T
1 e Pe 1 1 1
− ġ(x) − θ̃ T θ̇ − ε̃u ε̂˙ u + σ σ̇ (40)
2 g(x) 2 γ nu δ0
1 1 e T Pe
V̇ = − e T Qe − ġ(x) + V̇ 1 + V̇ 2 (41)
2g(x) 2 g(x)2
With
1 T
V̇ 1 = e T P Bw T θ̃ − θ̃ θ̇ (42)
γ
D(x)e T Pe 1 1
V̇ 2 = e T P Bεu (x) − e T P Bu r − − ε̃u ε̂˙ u + σ σ̇ (43)
2g(x) 2 nu δ0
Using Eq. (30)
V̇ 1 = 0 (44)
Using Assumption 6
D(x)e T Pe 1 1
V̇ 2 ≤ e T P B ε̄u − e T P Bu r − − ε̃u ε̂˙ u + σ σ̇ (45)
2g(x) 2 nu δ0
Using Eqs. (29), (31) and (32)
D(x)e T Pe
V̇ 2 ≤ − (46)
2g(x)2
1 1 e T Pe D(x)e T Pe
V̇ ≤ − e T Qe − 2
ġ(x) − (47)
2g(x) 2 g(x) 2g(x)2
Using the Assumption 3
1 1
− e T Qe > − e T Qe ≤ 0 (48)
2g(x) 2g
Using the fact that g > 0, then
V̇ < 0 (49)
Hence V εL ∞ , which implies that the signals e(t), θ̃ (t), ε̃(t) and δ(t) are bounded.
This implies the boundedness of x(t), u(t), and ė(t), by using Babalat’s lemma [1],
we conclure that the tracking errors and its derivatives converge asymptotically to
zero e(i) (t) → 0 as t → ∞ for i = 0, 1, . . . , n − 1.
In this section, to demonstrate the effectiveness of the direct adaptive fuzzy control
algorithms, we consider the tracking control of the inverted pendulum system shown
in Fig. 1. The dynamic equations of such system are given by [22–25].
⎧
⎨ ẋ1 = x2
ẋ = f (x) + g(x)u (50)
⎩ 2
y = x1
with
Fig. 1 The inverted

pendulum system

gsinx1 − mlx22 cosx1 sinx1 / m p + m c
f (x) = (51)
l 4/3 − m p cos2 x1 / m p + m c

cosx1 / m p + m c
g(x) = (52)
l 4/3 − m p cos2 x1 / m p + m c
where x1 is the rotation movement,x2 is the rotation velocity,g = 9.8 m/s2 is the
acceleration due to gravity,m c is the mass of the cart,m p is the mass of the pole,l is
the half-length of the pole, and u is the applied force, values m c = 1 kg,m p = 0.1 kg
and l = 0.5 m are choosen in this example, the control objective is to force the output
system to follow the desired trajectory yd (t) = sin(t).Note that the desired trajectory
allowed a maximum swing of 1 rad,while it is limited to 0.1 rad in [7].
One fuzzy system in the form of (21) is used to approximate the ideal control law.
The fuzzy system have x1 (t) and x2 (t) as inputs.xi , has five Gaussian membership
functions defined as

1 xi + 0.8 2
μ Fi1 (xi ) = exp −
2 0.2

1 xi + 0.4 2
2 0.2

1 xi 2
2 0.2
Fig. 2 y(t) (solid lines) and 1.5

yd (t) (dashed lines)
1
Angular Position (rad)

0.5
-0.5
-1
-1.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)

1 xi − 0.4 2
2 0.2

1 xi − 0.8 2
2 0.2
The initial value of the parameter θ (t) is set randomly. The various parameters
used in this simulation are chosen as follows: k = [1, 2], P = [155; 55], B =
[0; 1]γ = 200, δ0 = 10, n u = 0.001.
The initial value of θ (0) is chosen randomly between (−2 and 2) and ε̂u (0) = 0,
σ (0) = 1.
The simulation results of the position y = x1 and the angular velocity ẏ = x2 are
shown in Figs. 2 and 3, respectively. The control law input signal u(t) is shown in
Fig. 4. The tracking error signal e(t) is shown in Fig. 5, we can find that the actual
trajectories converge to desired trajectories in few time, this simulation show the
tracking capability of the proposed controller.
5 Conclusion
In this paper, a direct adaptive fuzzy control for class of unknown (SISO) nonlinear
systems is developed. In the design, fuzzy logic systems are used for estimating the
adaptive control law and robustifying control term to compensate for approximation
errors, the developed adaptive controller does not require the mathematical model
of the plant, the boundedness of all signals in the closed -loop system and ensure
the convergence of the tracking error to zero. The main contribution of this paper is
the development of an adaptive controller with control gain is a nonlinear function,
Fig. 3 ẏ(t) (solid lines) and 1.5

ẏd (t) (dashed lines)
Angular Velocity (rad/s)

0.5
-0.5
-1
-1.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)
Fig. 4 Control input signal 20
15
10
-5
-10
-15
-20
-25
0 2 4 6 8 10 12 14 16 18 20
Time (s)
in addition this approach relaxes the a priori knowledge of the lower bound of the
control gain and the upper bounds of the approximation errors.
In the last, one example of simulation show the tracking performances of the
proposed method. In future works, we propose to improve adaption law with
optimization of various parameters using (PSO).
Fig. 5 Tracking error signal 0.2
0.1
-0.1
-0.2
-0.3
-0.4
-0.5
-0.6
0 2 4 6 8 10 12 14 16 18 20
Time (s)
References
1. Zhang YM, Jiang J (2008) Bibliographical review on reconfigurable fault-tolerant control

systems. Ann Rev Control 32(2):229–252
2. Li S, Sauter D, Aubrun C, Yamé J (2008) Stability guaranteed active fault- tolerant control of
networked control systems. J Control Sci Eng. https://doi.org/10.1155/2008/189064
3. Ye SJ, Zhang YM, Wang XM, Rabbath CA (2009) Robust fault tolerant control using on-line
control re-allocation with application to aircraft. In: Proceedings of the 2009 American control
conference St Louis Riverfront USA IEEE, pp 5534–5539
4. Fu Y, Chai T (2012) Nonlinear adaptive decoupling control based on neural networks and
multiple models. Int J Innov Comput Inf Control 8(3A):1867–1878
5. Nguang SK, Shi P. H (2003) Fuzzy output feedback control design for nonlinear systems: an
LMI approach. IEEE Trans Fuzzy Syst 11(3):331–340
6. Zhang D, Su H, Chu J, Wang Z (2008) Satisfactory reliable H1 guaranteed cost control with
D-stability and control input constraints. IET Control Theory Appl 2(8):643–653
7. Fujun C, Yu C, Wu W, Fuwen Y (2008) Robust reliable H1 control in the presence of
sensor failure. In: Proceedings of the 7th world congress on intelligent control and automation
(WCICA), vol 8, pp 143–8
8. Bo Y, Fuzhong W Jianhua W (2010) Reliable output feedback for linear systems with sensor
mixed faults. In: Proceedings of the 8th world congress on intelligent control and automation
(WCICA), pp 509–13
9. Jin X-Z, Yang G-H, Ye D (2013) Insensitive reliable H1 filtering against sensor failures. Inf
Sci 224:188–99
10. Wu H-N, Zhang H-Y (2005) Reliable mixed L2/H1 fuzzy static output feedback control for
nonlinear systems with sensor faults. Automatica 41(11):1925–1932
11. Zhao X, Tian E, Wei J, Yuan Y (2011) Reliable H1 filter design for nonlinear networked
control systems with probabilistic sensor failure. In: Proceedings of the 30th Chinese control
conference (CCC), pp 4170–4175
12. Wang C, Dong J, Yang G, Kang H (2013) Fuzzy fault tolerant control for nonlinear systems with
sensor faults. In: Proceedings of the 32nd Chinese control conference (CCC), pp 6214–6219
13. Sun H, Guo L (2014) Composite adaptive disturbance observer based control and back-
stepping method for nonlinear system with multiple mismatched disturbances. J Franklin Inst
351(2):1027–1041
14. Aouaouda S, Chadli M, Boukhnifer M, Karimi HR (2014) Robust fault tolerant tracking
controller design for vehicle dynamics: a descriptor approach. Mechatronics. https://doi.org/
10.1016/j.mechatronics.2014.09.011
15. Qiu J, Ren M, Zhao Y, Guo Y (2011) Active fault-tolerant control for vehicle active suspension
systems infinite-frequency domain. IET Control Theory Appl 5(13):1544–1550
16. Liu C-S, Jiang B (2013) H2 fault tolerant controller design for a class of nonlinear systems
with a spacecraft control application. Acta Autom Sinica 39(2):188–196
17. Khebbache H, Tadjine M, Labiod S, Boulkroune A (2015) Adaptive sensor-fault tolerant control
for a class of multivariable uncertain nonlinear systems. ISA Trans. https://doi.org/10.1016/j.
isatra.2014.10.001
18. Castaldi P, Mimmo N, Simani S (2014) Differential geometry based active fault tolerant control
for aircraft. Control Eng Pract 32:227–235
19. Shen Q, Jiang B, Shi P, Lim C-C (2014) Novel NNS-based fault tolerant control scheme with
fault alarm. IEEE Trans Cybern 44(11):2190–2201
20. Bounemeur H, Chemachema M, Essounbouli N (2018) Indirect adaptive fuzzy fault-tolerant
tracking control for MIMO nonlinear systems with actuator and sensor failures. ISA Trans.
https://doi.org/10.1016/j.isatra.2018.04.014
21. Wang LX, Mendel JM (1992) Fuzzy basis functions, universal approximation and orthogonal
least-squares learning. IEEE Trans Neural Netw 3(5):807–814
22. Wang LX (1994) Adaptive fuzzy systems and control. In: Englewood cliffs. Prentice-Hall, New
Jersey
23. Abdelhamid B, Mouhamed C, Najib E (2017) Optimal indirect robust adaptive fuzzy control
using PSO for MIMO nonlinear systems. In: international conference on electrical engineering
and control applications, pp 208–224, Springer, Cham
24. Abdelhamid B, Mouhamed C, Najib E (2016) Indirect robust adaptive fuzzy control of uncertain
two link robot manipulator. In: international conference on electrical engineering and control
applications, pp 125–139, Springer, Cham
25. Bounemeur A, Chemachema M, Essounbouli N (2014) New approach of robust direct adaptive
control of a class of SISO nonlinear systems. In: 2014 15th international conference on sciences
and techniques of automatic control and computer engineering (STA), pp 725–730, IEEE
Signal and Communications (SC)
Simultaneous Localization and Mapping
Algorithm based on 3D Laser
for Unmanned Aerial Vehicle
Fethi Demim, Abdelkrim Nemra, Oumima Mouali, Mehdia Hedir,

Abdenebi Rouigueb, Mustapha Hamerlain, Mohamed Amine Bendoumi,
and Abdelouahab Bazoula
Abstract This paper focuses on developing a robust solution for the Simultaneous
Localization and Mapping (SLAM) problem to increase the autonomy of Unmanned
Aerial Vehicles (UAV). The investigated topics are related to data fusion, localization,
features extraction and matching, map building, 3D pose estimation and SLAM. One
important aspect of the autonomous navigation which should be investigated as well
as the fusion of data from the different sensor. The data fusion algorithms are very
important and their performances are closely dependent on both performances of the
F. Demim (B) · A. Nemra · M. A. Bendoumi · A. Bazoula

Laboratoire des Véhicules Autonomes Intelligents, Ecole Militaire Polytechnique, EMP, Bordj El
Bahri, Algiers, Algeria
A. Nemra
M. A. Bendoumi
A. Bazoula
O. Mouali
Faculté des Sciences Mathématiques, Physiques et Naturelles, Université d’el Manar, Tunis,
Tunisia
M. Hedir
Image Processing and Radiation Laboratory, Faculty of Electronic and Informatics, USTHB, Bab
A. Rouigueb
Data Fusion and Analysis Laboratory, Ecole Militaire Polytechnique, EMP, Bordj El Bahri,
Algiers, Algeria
M. Hamerlain
Centre de Développement des, Technologie Avancées, CDTA, Baba Hassen, Algeria

https://doi.org/10.1007/978-981-15-6403-1_69
1004 F. Demim et al.
constructed map and the accuracy of the UAV position within this map. Optimal and
robust filter using centred Gaussian noises is implemented. The original contribution
of this work is the proposition and the adaptation of the Extended Kalman Filter (EKF)
to solve the Inertial Navigation Systems INS/3D laser UAV navigation problem.
Simulation results for 3D flight scenario are presented to demonstrate the advantages
of the hybrid localization based on EKF compared with results of the INS relative
localization based technique. Good results were obtained with the EKF using noise
process and/or measurement noise characteristics, particularly in the case of centred
Gaussian noises.
Keywords Autonomous navigation · Unmanned aerial vehicle · Localization, map

building
1 Introduction
Autonomous vehicles must operate in their environment without any human inter-
vention. Today, Unmanned Aerial Vehicles (UAVs) are rapidly becoming a challenge
for military forces and civil airspace due to their various applications such as surveil-
lance, reconnaissance, communication relay, target designation and payload delivery.
They include a wide variety of robotic aircraft of varying size, shape and flight, the
characteristics and level of operational autonomy. In recent years, the production
and development of drones has exploded in military arenas. The increase of UAV
autonomy can be achieved by developing tools able to provide an accurate automatic
localization in an accurate environment map. SLAM problem looks at the ability of
an autonomous vehicle, starting in a partially known or unknown environment, to
incrementally build an environment map and simultaneously localize itself within this
map. New challenges in SLAM are delivering methods enabling large-scale imple-
mentations in increasingly unstructured environments and especially in situations
where Global Positioning System (GPS) like solutions are unavailable or unreliable
such as urban canyons and urban environments. In principle, implementing SLAM in
three dimensions (3D) is an extension of the two-dimensional (2D) case. However 3D
SLAM involves significant added complexity due to the more general vehicle motion
model and most importantly increased sensing and feature modelling complexity. The
challenge now is to demonstrate SLAM solutions for a UAV navigating in large scale
natural environment without recourse to GPS.
The genesis of the SLAM probabilistic problem occurred at the IEEE Robotics
and Automation conference held in San Francisco, California in 1986. At that time,
probabilistic methods were just beginning to be introduced into robotics and artificial
intelligence (AI) problems. A number of researchers have focused on the applica-
tion of theoretical estimation methods to mapping and localization problems. These
include Peter Cheeseman, Jim Crowley and Hugh Durrant-Whyte. A lot of research
has been done on coherent mapping. Subsequently, Raja Chatila, Oliver Faugeras,
Randal Smith and others also made useful contributions to the conversation. They
Simultaneous Localization and Mapping Algorithm … 1005
conclude that the coherent probabilistic mapping was a fundamental element in

robotics with major conceptual and computer problems that needed to be solved.
Over the next few years, a number of key papers were produced.
The work of Smith et al. [1] and Durrant-Whyte [2] established a statistical basis
for the analysis and description of the relationship between landmarks and geometric
uncertainty. A key element of this work was to show that there must be a high degree
of correlation between estimates of the location of different landmark points on a
map and that, indeed, these correlations would increase with successive observations.
In the same year, Ayache and Faugeras [3] undertook preliminary work on visual
work based on sonar navigation systems for the chercehrus Crowley [4], Chatila
and Laumond [5, 6]. Navigation of mobile robots using stochastic algorithms using
Kalman filter are introduced. Interest in SLAM has grown exponentially in recent
years, and workshops continue to be held at both ICRA and IROS. The SLAM
summer school ran in 2004 in Toulouse then in Oxford in 2006, where many papers
were published on the inertial SLAM [7, 8].
Currently, the Visual SLAM becomes a very challenging axe of research and many
papers were published in [9, 10]. Moreover, with the development of Aerial Vehicles,
the Airborne SLAM is seen as an essential area of research and many research works
on Real time, Airborne and Visual SLAM are investigated in [11, 12]. The work
presented by Chan et al. talks about UAV based bridge inspection application systems
[13]. After that, Gheisari et al. discuss the unmanned aerial Systems (UAS) for
construction safety applications in proceedings of Construction Research Congress
[14]. A visual monitoring infrastructure systems via camera-equipped with UAVs is
proposed [15]. In 2017, Khaloo et al. present AUV inspection of the placer river trail
bridge through image-based 3D modelling [16]. The survey on vision-based UAV
navigation is presented in 2018 by [17]. The main motivation behind this research,
as the introduction has suggested, is the autonomous navigation of aerial vehicles.
Today, the navigation of aerial vehicles becomes a challenging axe of research, firstly,
because the airborne have six degrees of freedom, which make it more complicated
than the mobile robot; secondly, the massive number of applications of aerial vehicle
in military and civil applications. The position determination of the object is then
done by combining the relative positions and their times. Since measurement systems
are most often flawed, errors accumulate and quickly become significant. An inertial
navigation system makes use of an Inertial Measurement Unit (IMU) to sense the
vehicle’s rotation rates and accelerations. This data is then used to obtain vehicle
states such as position, velocity and attitude at high sampling rates which are crucial
for airborne guidance and control. However, using only INS as proprioceptive sensor
can gives diverging errors. These errors of localization can cause damage that is
sometimes unacceptable. INS diverging errors due to the integration process, requires
absolute sensors in order to constrain these drifts.
In this paper, we have presented some notions to solve the problem of UAV loca-
tization and mapping simultaneously, with a stereoscopic sensor based on laser and
through point of interest mapping. However, an experimental evaluation is necessary
to determine the most appropriate method for such an application under specific
conditions. To do this, we studied our system in the different scenarios. The use of
an estimation method based on Extended Kalman filter is necessary to improve the

accuracy of the system (UAV position, and the representation of the environment).
For the INS models the internal Extended Kalman Filter fuses in real-time inertial
and 3D laser data for enhanced position and orientation measurements in urban and
canyons environments.
The paper is organized as follows. Section 2 depicts the process model of the
UAV and the observation model. The Sect. 3 describes the SLAM of UAV and also
proposes our version of the implementation of the EKF-SLAM algorithm in detail.
Simulation results are presented in Sect. 4. Finally, conclusion and future works are
provided in Sect. 5.
2 Process Model of UAV
The quadrotor (UAV) has six degrees of freedom, three rotational movements and
three translational movements. It must be able to perform unexpected manoeuvres
in its environment. The construction of such a map should be incremental, merging
the successive perceptions acquired by the UAV’s sensors during its movement (see
Fig. 1a). Modeling is the most important step in mobile robotics. In our case, it
consists in representing the UAV and its sensors in the environment in which it
operates. In this section, we will first present the UAV kinematic model used to
obtain its position, then we determine the geometry of our sensor which is the laser,
then we present the direct and inverse observation models for the UAV (Fig. 3 ).
Let us present the UAV with the following nonlinear model

Ẋ (t) = f (X (t), u(t), t)
(1)
Ż (t) = h(X (t), u(t), t)
(a) UAV presentation. (b) Hokuyo 3D laser.
Fig. 1 Hokuyo laser integrated in UAV

20
15
10
5
0
-5
-10
-15
-20
20
15
10 20
5 15
0 10
5
-5 0
-10 -5
-15 -10
-15
-20 -20
Fig. 2 Simulation of the qudrotor model
T
where X = X auv , ag is the state vector, which contains position, velocity and
Euler angles, ag = [a1 , . . . , an ]T represents the coordinates in the global frame of
the ith landmark, and U represents the inertial measurement outputs.
The vehicle’s state vector X auv is given by:
X auv = [x, y, z, u, v, w, r, φ, θ, ψ]T (2)
UAV localization has been formulated based on the INS sensor. This latter consists
of 3-axis gyroscopes, that provide to the roll, pitch and yaw of the body frame. It also
has 3-axis accelerometers, which provide accelerations
along the three body axes.
INS sensor provides the accelerations ax , a y , az and the rotation rates ( p, q, r ) of
the UAV with high update rates, which can then be transformed and processed to
provide its position (x, y, z), velocity (u, v, w) and attitude (φ, θ, ψ) resulting in
an inertial navigation system, where φ is pitch, θ is roll and ψ is yaw. Let U =
T
u, v, w, ax , a y , az , be the control vector, which is assigned with the UAV (angular
rates and accelerations).
The navigation equations require defining at least two frames. One for the
body/inertial representation (UAV), and one for the navigation frame representa-
tion. Then, the equations of motion can be given by a simple integrations and frame
transformations.
The Euler angle rates φ̇, θ̇ and ψ̇ can be calculated using:
⎛ ⎞ ⎛ ⎞⎛ ⎞
φ̇ 1 sin(φ) tan(θ ) cos(φ) tan(θ ) p
⎝ θ̇ ⎠ = ⎝ 0 cos(φ) − sin(φ) ⎠⎝ q ⎠ (3)
ψ̇ 0 sin(φ) sec(θ ) cos(φ) sec(θ ) r
Assuming that the inertial measurement unit is at the vehicle center of gravity,
the true vehicle acceleration in the body frame is given by u̇, v̇ and ẇ:
⎛ ⎞ ⎛ ⎞
u̇ ax + vr − wq + g sin(θ )
⎝ v̇ ⎠ = ⎝ a y − u r + w p − g cos(θ ) sin(φ) ⎠ (4)
ẇ az + u q − v p − g cos(θ ) cos(φ)
The resulting acceleration vector is integrated with respect to time to obtain the
velocity of the vehicle in the body frame:
⎛ ⎞ ⎛ ⎞
u u̇
⎝ v ⎠ = ⎝ v̇ ⎠dt (5)
w ẇ
The velocity vector is then integrated to read the position of the vehicle in the
body frame. If the velocity is transformed down to navigation frame and integrated,
we get the position (x, y, z)T vector in the navigation frame.
⎛ ⎞ ⎛ ⎞
x u
⎝ y ⎠ = Cdct
T
(φ, θ, ψ)⎝ v ⎠dt (6)
z w
where Cdct is the Direct Cosine Transform matrix that rotates a vector form body
frame to the navigation frame.
⎛ ⎞
c(θ )c(ψ) c(θ )s(ψ) −s(θ )
Cdct = ⎝ s(θ )s(φ)c(ψ) − s(ψ)c(φ) s(ψ)s(θ )s(φ) + c(ψ)c(φ) s(φ)c(θ ) ⎠ (7)
s(θ )c(φ)c(ψ) + s(ψ)s(φ) s(φ)s(θ )c(φ) − c(ψ)s(θ ) c(φ)c(θ )
where c and s represent con and sin of angles respectively, then, the nonlinear INS
process state model is given by the following equation:
⎛ ⎞
T
Cdct (u, v, w)T
⎜ ⎟
f (X, U ) = ⎝ (u̇, v̇, ẇ)T ⎠ (8)
T
φ̇, θ̇ , ψ̇
A. Observation based Model
All robot apprentices have built a robot that avoids obstacles with contact sensors
or infrared distance sensors. But for more advanced robots, which have to move in
complex environments, higher-level sensors are needed. Different choices are then
available to the robot designer, depending on the type of mobile flying robot and its
environment. Hokuyo laser rangefinders allow us to measure and map environments.
They are compact 50 × 50 × 70 mm, lightweight 160 g and powered by 5 V, which
allows them to be easily integrated into unmanned aerial vehicle, even small ones.
The different models offer spans from 4 to 30 m (see Fig. 1b).
Hokuyo 3D laser scanners use the same 2D laser technique, but with a mirror
that pivots on two axes. During each scan, the Hokuyo URG - 04 - LX laser range
finder returns 683 distance information ρ corresponding to the 683 angles at which
it has measured via its USB interface. They can be retrieved in a Comma-Separated
Values (CSV) file or directly in C ++ software using the Application Programming
Interface (API) provided by Hokuyo. It is easier to transform this information into x
and y coordinates. Considering the angle β varies from −120◦ to 120◦ and the angle
0 corresponds to the y axis we obtain:

x = ρ sin(β)
(9)
y = ρ cos(β)
T
Let ag = x g , yg , z g be the landmark coordinates in the global reference frame,
the direct observation model is used to express the coordinates of an estimated
landmark with respect to the UAV coordinates as (Fig. 3):
Validation of models
5.5
4.5
4
15
10 15
5 10
0 5
-5 0
-5
-10 -10
Observation inverse model Observation model

6 6
5 5
4 4
20 20
10 15 10 15
10 10
0 5 0 5
0 0
-10 -10 -5 -10 -10 -5
Fig. 3 Validation of observation models


Z = h X, ag (10)
⎛ ⎞ ⎛ ⎞
xl T xg − x
Z = ⎝ yl ⎠ = Cdct ⎝ yg − y ⎠ (11)
zl zg − z
T
where x g , yg , z g represents the coordinates of landmark in the global frame;
(xl , yl , zl )T represents the coordinates of landmark in the local frame; (x, y, z)T
represents the position of the UAV in the global frame.
The coordinate of a new landmark anew with respect to the global frame that
defines the inverse observation model ag,new is initialized by solving the following
equation:
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
xg xl x
ag,new T ⎝
= h −1 (X, Z ) = ⎝ yg ⎠ = Cdct yl ⎠ + ⎝ y ⎠ (12)
zg zl z
3 SLAM-UAV
SLAM is the problem of constructing a model of the environment being traversed

with on board sensors [18], while at the same time maintaining an estimate of the
vehicle location within the model [19, 20]. As an alternative approach of UAV, there
is a robust filter with zero-mean Gaussian noise, known as the nonlinear extended
Kalman filter (EKF). This filter provides a robust and stable estimate to modeling
uncertainties and errors [21, 22]. This can be interpreted as follows: the system and
observation models are accurate besides when the process and observation noises
are uncorrelated zero-mean Gaussian with known covariance then the EKF gives a
good accuracy for position estimation [23, 24].
X (k + 1) = f (X (k), U (k + 1), W (k + 1)) (13)
Z (k + 1) = h(X (k + 1), V (k + 1)) (14)

4 EKF-SLAM Algorithm
A. Initialization
The state vector associated with the submarine robot and the initial covariance matrix
is presented as follows [25]:
T
X (0) = X auv (0) = 0 0 0 0 0 0 0 0 0 (15)

09×9 09×3N
P(0) = Puav (0) = (16)
03N ×9 σa2N
where σa2N is initialized accordingly with the precision of the laser sensor.
B. Prediction Time Update
The prediction stage is a process, which deals with vehicle motion increase the
uncertainty of the vehicle pose estimate. First, the state vector is augmented with a
T
control input U = u, v, w, ax , a y , az .
C. Data association and update

∂f ∂f ∂f ∂f ∂f ∂f ∂f ∂f ∂f
Fuav = ∂ x ∂ y ∂z ∂u ∂v ∂w ∂φ ∂θ ∂ψ (17)

where Fuav represents the Jacobian of f X uav , U (k) with respect to X uav .
From the kinematic model, Fuav is:
∂f
Fuav = X uav (k), 0 (18)
∂ X uav
The prediction equations are presented as follows:

X (k + 1) = f X (k), U (k)
(19)
P (k + 1/k) = ∇ Fuav P(k/k)∇ Fuav
T
+ ∇ Fu Q∇ FuT
where Q represents a noise with covariance matrix of input control U.

⎡ ⎤
σu2 0 0 0 0 0
⎢ 0 σv2 0 ⎥
⎢ 0 0 0 ⎥
⎢ ⎥
⎢ 0 0 σw2 0 0 0 ⎥
Q=⎢ ⎥ (20)
⎢ 0 0 0 σax
2
0 0 ⎥
⎢ ⎥
⎣ 0 0 0 0 σay
2
0 ⎦
0 0 0 0 0 σaz
2

∇ Fuav and ∇ Fu are the Jacobian matrices of f X (k), U (k) with respect to X (k)
and U (k) respectively such as :

Fuav (k) 09×3N
∇ Fuav (k) = (21)
03N ×9 I3N ×3N
and

Fu (k)
∇ Fu = (22)
03N ×6
The Jacobian necessary for this approximation is denoted by Fu , it is the derivative

motion model f with respect to the input control U.
⎡ ⎤
∂ f X (k), U (k)
∂f ∂f ∂f
Fu = ⎣ ⎦ = ∂f
∂u
∂f ∂f
∂v ∂w ∂ax ∂a y ∂az (23)
∂U (k)
The covariance matrix can be written as follows:
P(k + 1/k) = ∇ Fuav P(k/k)∇ Fuav

T
+ ∇ Fu Q∇ FuT (24)
To calculate the prediction error between the measure and its prediction, the
innovation term S j,i is written as follows :

v j,i (k + 1) = Z uav
j (k + 1) − h i X (k + 1/k) (25)
S j,i = Hi (k + 1)P(k + 1/k)HiT (k + 1) + R
where Hi (k + 1) sets the Jacobian matrix of h i with respect to the augmented state

vector X (k + 1/k).
∂h i (k)
Hi (k + 1) = X (k + 1/k), 0 (26)
∂X

Hi (k + 1) = Huav (k) . . . Hai (k) . . . 0 0 (27)
such as:
∂h i (k) ∂h i (k)
Huav (k) = , Hai (k) = (28)
∂ X uav (k) ∂ai (k)
The correspondence for each data association with an unknown association is done
by techniques of association, such as the minimization of the Euclidean distance or the
distance of Mahalanobis. To verify the validity of the corresponding, the Mahalanobis

distance is implemented using [21, 23]:
D j,i = v j,i (k + 1)S −1

j,i (k + 1)v j,i (k + 1)
T
(29)

d = argmin D j,i < δα (30)
where d defines the minimization of the distance from the Mahalanobis and δα
represents the data association threshold. The term ν represents the innovation matrix
between the actual sensor measurement Z iuav (k + 1) and the prediction of this same
measurement, S j,i being its corresponding covariance matrix. This is necessary to
calculate K j,i (k + 1), the Kalman gain, which is chosen to correct the estimate
and minimize the error covariance P(k + 1) after the update. For all the measures
observed at time (k + 1), we get a real measure: Z iuav (k + 1). The predicted measure
at time (k + 1) is obtained as follows:

Z iuav (k + 1) = h i X (k + 1) (31)
Since h i (k) depends on the position of the UAV X uav (k) and the position of the
ith feature ai in the associated state vector with the extraction feature. For this, we
rewrite it in the following form:
Hiuav (k + 1) = Hi (k + 1) × ∇Wiuav (32)
where:
⎡ ⎤
1 0 0 0 ··· 0 0 0 0 0 0 0 0 ··· 0
⎢0 ··· ··· 0⎥
⎢ 1 0 0 0 0 0 0 0 0 0 0 ⎥
⎢ ⎥
⎢0 0 1 0 ··· 0 0 0 0 0 0 0 0 ··· 0⎥
⎢ ⎥
⎢0 0 0 0 ··· 0 1 0 0 0 0 0 0 ··· 0⎥
⎢ ⎥
⎢0 0 0 0 ··· 0 0 1 0 0 0 0 0 ··· 0⎥
⎢ ⎥
∇Wiuav =⎢
⎢0 0 0 0 ··· 0 0 0 1 0 0 0 0 ··· 0⎥⎥ (33)
⎢0 ··· ··· 0⎥
⎢ 0 0 0 0 0 0 0 1 0 0 0 ⎥
⎢0 ··· ··· 0⎥
⎢ 0 0 0 0 0 0 0 0 1 0 0 ⎥
⎢ ⎥
⎢0 0 0 0 ··· 0 0 0 0 0 0 0 0 ··· 0⎥
⎢ ⎥
⎣0 0 0 0 ··· 0 0 0 0 0 0 0 0 ··· 0⎦
0 0 0 0 ··· 0 0 0 0 0 0 0 0 ··· 0
For all the associated landmarks between the actual measure and the predicted
measure, at time (k + 1), we have the following correction equations:
T
S j,i (k + 1) = Hiuav (k + 1)P(k + 1/k) Hiuav (k + 1) + R j (34)
T T
K j,i (k + 1) = P(k + 1/k) Hiuav (k + 1) S j,i (k + 1) (35)

X (k + 1/k + 1) = X (k + 1) + K j,i (k + 1)Z iuav (k + 1) − h i X (k + 1) (36)

P(k + 1/k + 1) = I11×11 − K j,i (k + 1)Hiuav (k + 1) P(k + 1/k) (37)
where the covariance matrix on the measurement noise is represented as:

⎡ ⎤
σρi 0 0
R j = ⎣ 0 σβi 0 ⎦ (38)
0 0 σz i
Even as the environment is investigated, new features are detected and should
be put on the accumulation map. In this case, the state vector and the output error
estimate matrix are calculated from the new observation.
D. Map management
It is a step of increasing the state vector by the new detector feature. A critical issue
in a SLAM algorithm is to decide how to represent the joint distribution over vehicle
poses and feature map [22, 23]. In particular, the map management is very important
due to the fact that maps are usually represented by a high number of parameters.
5 Implementation, Simulations and Discussion
In this section, we discuss the functionality of the Inertial Navigation Systems (INS)
by specifying the impact of the accumulation of errors due to the succession of
integrations given by the INS model, and then we specify the major role of data
fusion between the laser and INS which gave us reliable and accurate estimates
supported by Matlab software simulations.
We will evaluate its estimation accuracy and robustness in the face of centered
Gaussian white noise.
The simulation environment is in three-dimensional xyz space 20 m × 20 m × 20 m
containing 65 static landmarks (see Fig. 2). The quadrotor performs a well-defined
movement from the origin as an initial position (0, 0, 0) in an against clockwise
direction, with T = 0.05s represents a sampling rate.
The simulation results provided in the following figures represent the esti-
mated UAV position obtained using the extended Kalman filter respectively with
the kind of uncertainties. We will propose a simulation with a validation of an
INS model. The different INS commands are presented as follows: accelerations
(ax = 0.1m/s2 , a y = 0.2m/s2 , az = −9.81m/s2 ) as well as angular velocities
( p = sin(i), q = cos(i), r = 0) where i represents the number of iterations varied
between 0 and 255 with a pitch of 1. As the quadrotor moves, it detects the 3D land-
marks that are included in its field of view. The drone is modeled as shown in the
Fig. 2. The external sensor used in our work is a HOKAYU laser rangefinder type
with a 5 m of the beam, the detected points are distributed over a range from 0 to
180◦ , which are presented as green semi-circle.
In order to test the reliability of the proposed algorithm, it can be noticed that
the map of EKF provides the best result, in this case, as can be seen from Figs. 4
and 5, the EKF-SLAM algorithm gives the best results position and mapping than
INS localization algorithm. As it can be seen, the trajectory obtained by running the
EKF with performing the position updates with laser measurements is more accurate
compared to the trajectory acquired by INS.
On the other hand, without using laser data, the trajectory obtained only by inte-
grating the velocity measurements and does not update the estimate information,
the process is inherently affected by drift. In order to test the robustness, accuracy
and efficiency of our EKF-SLAM algorithm, several simulations were performed.
According to these simulation results using INS model, we have generated a noise
input into its control, in order to specify the errors on accelerations and angular veloc-
ities lead to errors in constant growth position and speed of the quadrotor components
due to integration.
We have shown the navigation errors, and there are three position errors (x, y, z),
three speed errors (u, v, w), two attitude errors (φ, θ ) and the heading error (ψ). We
notice that over time, noise increases due to the accumulation of integration errors.
The comparison results in terms of noise is presented in Fig. 6. Trajectory estimated
INS(XYZ)-without noise
0.8
INS(X)-without noise
0.6
INS(Y)-without noise
Position (m)
0.4 INS(Z)-without noise
0.2
-0.2
0 50 100 150 200 250
Algorithm iterations
INS(UVW)-without noise INS(Phi-Theta-Psi)-without noise

0.8 0.02
INS(U)-without noise INS(Phi)-without noise
0.6
Velocity (m/s)
INS(V)-without noise INS(Theta)-without noise

Angle (rad)
0.01
INS(W)-without noise INS(Psi)-without noise
0.4
0
0.2
-0.01
0
-0.2 -0.02
0 50 100 150 200 250 0 50 100 150 200 250
Algorithm iterations Algorithm iterations
Fig. 4 INS simulation result without noise

INS(XYZ)-with noise
0.2
0
Positions (m)
-0.2
-0.4 INS(X)-with noise

INS(Y)-with noise
-0.6
INS(Z)-with noise
-0.8
0 50 100 150 200 250
INS(UVW)-with noise INS(Phi-Theta-Psi)-with noise

0.5 0.15
INS(Phi)-with noise
Velocity (m/s)
0 0.1 INS(Theta)-with noise
Angle (rad)
0.05 INS(Psi)-with noise
-0.5
INS(U)-with noise 0
-1 INS(V)-with noise
INS(W)-with noise -0.05
-1.5 -0.1
0 50 100 150 200 250 0 50 100 150 200 250
Fig. 5 INS simulation result with noise
INS(X)-without noise
0.5
Positions (m)
INS(X)-with noise
INS(Y)-without noise
0 INS(Y)-with noise
INS(Z)-without noise
INS(Z)-with noise
-0.5
-1
0 50 100 150 200 250
INS(U)-without noise INS(Phi)-without noise

INS(U)-with noise INS(Phi)-with noise
INS(V)-without noise INS(Theta)-without noise

1 INS(V)-with noise
0.15 INS(Theta)-with noise
INS(W)-without noise INS(Psi)-without noise
0.5 INS(W)-with noise 0.1 INS(Psi)-with noise
Velocity(m/s)
Angle (rad)
0 0.05
-0.5 0
-1 -0.05
-1.5 -0.1
0 50 100 150 200 250 0 50 100 150 200 250
Fig. 6 Comparison results in terms of noise

position by the EKF-SLAM algorithm, as shown in Fig. 7, is more accurate in

comparison to the INS relative localization algorithm. As can be seen in Fig. 8, the
evolution of the UAV velocities (u, v, w) following the three axes respectively are
presented. As can be shown from these figures the EKF-SLAM performs better than
Position X Position Y
11.1 2.3
True path True path
11.08
Estimated EKF-SLAM 2.2 Estimated EKF-SLAM
11.06 INS path INS path
Y-uav(m)
X-uav(m)
11.04 2.1
11.02
2
11
10.98 1.9
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
Position Z
5.03
5.025
True path
5.02 Estimated EKF-SLAM
Z-uav(m)
5.015 INS path
5.01
5.005
5
4.995
0 10 20 30 40 50 60 70 80
Fig. 7 UAV pose estimation following x, y and z axes
Velocity U Velocity V
0.35 0.8
0.3 True path True path
Estimated EKF-SLAM 0.6 Estimated EKF-SLAM

0.25
U-uav(m/s)
V-uav(m/s)
INS path INS path

0.2
0.4
0.15
0.1 0.2
0.05
0 0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
Velocity W
0.04
W-uav(rad/s)
0.02
-0.02 True path

Estimated EKF-SLAM
-0.04 INS path
-0.06
0 10 20 30 40 50 60 70 80
Fig. 8 UAV velocities estimation following u, v and w axes

INS velocities. The estimated trajectories following φ, θ and ψ angles is presented in

Fig. 9. We see clearly from Fig. 10 that the estimated covariance’s following x, y, z
axes and φ, θ , ψ angles are presented. Moreover, from Fig. 11, the pose errors of
the EKF-SLAM algorithm for different axes and angles increase significantly and
the UAV.
Angle Phi Angle Theta
0.2 0.1
True path
Theta-uav(rad)
0.15 Estimated EKF-SLAM
0.05
Phi-uav(rad)
INS path
0.1
0
0.05 True path
Estimated EKF-SLAM -0.05
0
INS path
-0.05 -0.1
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
Angle Psi
0.04
0.03
Psi-uav(rad)
0.02
0.01 True path

Estimated EKF-SLAM
0
INS path
-0.01
0 10 20 30 40 50 60 70 80
Fig. 9 Estimated trajectories following φ, θ and ψ angles
X-Covariance Y-Covariance Z-Covariance

1.6 2 1
X-Cov Y-Cov Z-Cov
Covariance(m 2 )
Covariance(m2 )
1.4
Covariance(m2 )
1.5 0.8
1.2
1 0.6
1
0.5 0.4
0.8
0 0.2
0 20 40 60 80 0 20 40 60 80 0 20 40 60 80
Algorithm iterations Algorithm iterations Algorithm iterations
x 10
-3 Psi-Covariance -4
x 10
Theta-Covariance -3
x 10
Phi-Covariance
1.5 4 2
Psi-Cov Theta-Cov Phi-Cov
Covariance(m 2 )
Covariance(m 2 )
Covariance(m2 )
1 2 1
0.5 0 0
0 -2 -1
-0.5 -4 -2
-1 -6 -3
0 20 40 60 80 0 20 40 60 80 0 20 40 60 80
Fig. 10 Estimated covariances for different axes and angles

x 10
-3 X-Error Error-Y x 10
-3 Error-Z
3 0.01 4
Z-Error
Y-Position (m)
2.5 X-Error Y-Error
Z-Position (m)
0.008
X-Position(m)
3
2
0.006
1.5 2
0.004
1
0.002 1
0.5
0 0 0
0 20 40 60 80 0 20 40 60 80 0 20 40 60 80
Error-Psi Error-Theta Error-Phi

0.02 Theta-Position (m) 0.015 0.02
Phi-Position (m)
Psi-Position (m)
0.015 0.015
0.01
0.01 0.01
0.005
0.005 0.005
Psi-Errors Theta-Error Phi-Error
0 0 0
0 20 40 60 80 0 20 40 60 80 0 20 40 60 80
Fig. 11 Estimated position errors for different axes and angles
6 Conclusion
The aim of this paper is the application of SLAM techniques to the field of aerial
vehicle navigation using a Hokuyo type 3D laser that are necessary to solve in
order to increase UAV autonomy. UAV localization has been investigated. Then,
the implementation of simultaneous localization and mapping solution using laser
observation was presented. INS/3D Laser UAV localization technique using EKF has
been proposed and implemented. Simulation results have demonstrated the efficiency
of the proposed method to construct a map of the environment and to estimate well
positions of autonomous vehicle within this map.
The area of research that can be proposed for future investigation is the applica-
tion and validation of the robust visual SLAM real time application. Furthermore,
one important aspect to be explored in the future is the real time constraints with
complexity algorithms and computation time. This latter is very important to select
the suitable algorithm for real time implementation.
References
1. Smith R, Cheesman P (1987) On the representation of spatial uncertainty. Int J Robot Res
5:56–68
2. Durrant-Whyte HF (1988) Uncertain geometry in robotics. IEEE Trans Robot Automat 4:23–31
3. Ayache N, Faugeras O (1988) Building, registrating, and fusing noisy visual maps. Int J Robot
Res 7:5–45
4. Crowley J (1989) World modeling and position estimation for a mobile robot using ultra-sonic
ranging. In: Proceedings of IEEE international conference on robotics and automation, pp
674–681
5. Chatila R, Laumond JP (1985) Position referencing and consistent world modeling for mobile
robots. In: Proceedings of IEEE international conference robot and automation, pp 138–143
6. Smith R, Self M, Cheeseman P (1990) Estimating uncertain spatial relationships in robotics.
In: Cox IJ, Wilfon GT (eds) Autonomous robot vehicles. Springer, New York, pp 167–193
7. Kim JH, Sukkarieh S (2007) Real-time implementation of airborne inertial-SLAM. Robot
Auton Syst 55:62–71
8. Kim J, Yoon KJ, Kim JS, Kweon I (2006) Visual SLAM by single-camera catadioptric stereo
SICE-ICASE. In: Proceedings of international joint conference. Bexco, Busan, Korea, pp 18–21
9. Jeong WY, Lee KM (2006) Visual SLAM with line and corner features. In Proceedings of the
IEEE/RSJ international conference on intelligent robots and systems. Beijing, China
10. Davison AJ, Reid I, Molton N (2007) Stasse MonoSLAM: Real-time single camera SLAM.
IEEE Trans, PAMI
11. Kim JH, Sukkarieh S (2006) Robust multi-loop airborne SLAM in unknown wind envi-
ronments. In Proceedings of the IEEE international conference on robotics and automation.
Orlando, Florida
12. Caruso D, Engel J, Cremers D (2015) Large-scale direct SLAM for omnidirectional cameras.
In: Proceedings of international conference on intelligent robots and systems
13. Chan B, Guan H, Jo J, Blumenstein M (2015) Towards UAV-based bridge inspection systems:
a review and an application perspective. Struct Monit Maintenance 2:283–300
14. Gheisari M, Esmaeili B (2016) Unmanned Aerial Systems (UAS) for construction safety
applications. In: Proceedings of international conference on intelligent robots and syste in
proceedings of construction research congress, pp 2642–2650
15. Ham Y, An KK, Lin JJ, Golparvar-Fard M (2016) Visual monitoring of civil infrastructure
systems via camera-equipped Unmanned Aerial Vehicles (UAVs): a review of related works. J
Vis Eng 4:1–8
16. Khaloo A, Lattanzi D, Cunningham K, Dell-andrea R, Riley M (2017) Unmanned aerial vehicle
inspection of the placer river trail bridge through image-based 3D Modelling. J Struct Infrastruct
Eng 1–13
17. Yuncheng LU, Zhucun X, Gui-Song X, Orcid I, Liangpei Z (2018) Survey on vision-based
UAV navigation. J Geo-spatial Inf Sci 21:21–32
18. Demim F, Nemra A, Louadj K (2016) Robust SVSF-SLAM for unmanned vehicle in unknown
environment. J IFAC-Papers OnLine Sci Direct 49:386–394
19. Demim F. et al (2016) Simultaneous localization and mapping algorithm for unmanned ground
vehicle with SVSF Filter. In: IEEE 8th international conference on modelling, identification
and control, pp 155–162
20. Demim F. et al (2017) A new adaptive smooth variable structure filter SLAM algorithm for
unmanned vehicle. In: IEEE 6th international conference on systems and control, pp 6–13
21. Nemra A, Aouf N (2009) Robust airborne 3D visual simultaneous localization and mapping
with observability and consistency analysis. J Intell Robot Syst 55:345–376
22. Demim F, Nemra A, Louadj K, Hamerlain M, Bazoula A (2017) An adaptive SVSF-SLAM
algorithm to improve the success and solving the UGVs cooperation problem. J Exp Theor
Artif Intell
23. Demim F et al (2018) Simultaneous localization, mapping, and path planning for unmanned
vehicle using optimal control. J Adv Mech Eng 10:1–25
24. Demim F. et a (2018) Visual SVSF-SLAM algorithm based on adaptive boundary layer width.
In ICEECA2017, ed., Lecture notes in electrical engineering book series. Springer Nature
Switzerland AG, pp 97–112
25. Demim F, Nemra A, Boucheloukh A, Hamerlain M, Bazoula A (2018) SLAM problem for
autonomous underwater vehicle using SVSF filter. In: IEEE 25th international conference on
signal, image, vision and their applications. Maribor, Slovenia, pp 1–5
Comparison Between Gabor Filters
and Wavelets Transform
for Classification of Textured Images
Abdelkader Zitouni, Fatiha Benkouider, Fatima Chouireb,

and Mohammed Belkheiri
Abstract In the present research we have used wavelet transform and Gabor filters
to extract texture features in order to classify textured-images. Gabor features are
efficient in finding class boundaries, whereas the wavelets can represent textures at
different scales and offer great discriminatory power between textures with strong
resemblances. So, in this paper, we attempt to make a comparison between the feature
extraction techniques based on Gabor filters and the wavelet transform with the
purpose of classifying textured images. In the first step, we applied those two feature
extraction strategies on textured images in order to get more information. After
that in the second step, estimated feature vector of each pixel is sent to the neural
networks classifier for labelling. The performance of the segmentation algorithms
was evaluated on synthetic images from Brodatz and DTD datasets. The obtained
classification results confirm the power of the wavelet transform features compared
to Gabor filters features in classification of textured images.
Keywords Image segmentation · Classification · Texture features · Gabor filters ·

Wavelets
1 Introduction
Image segmentation is an important step in any image analysis process. It is a low-

level processing that precedes the measurement, comprehension and decision stage.
A. Zitouni (B) · F. Benkouider · F. Chouireb · M. Belkheiri

Department of Electronics, University Amar Telidji, Laghouat, Algeria
F. Benkouider
F. Chouireb
M. Belkheiri

https://doi.org/10.1007/978-981-15-6403-1_70
1022 A. Zitouni et al.
Its aim is to divide the image into related and homogeneous regions in the sense of
a homogeneity criterion that is difficult to define, particularly in the case of textured
regions [1].
Texture, which is almost ubiquitous in images, is one of the most important
features used to identify different regions of an image. Its study is difficult because
there is no universal definition to apprehend the notion of texture. Texture segmenta-
tion and classification approaches have been used in literature and have shown their
importance in various applications such as computer vision [2], image synthesis [3],
biomedical image processing [4] and remote sensing [5–7]. Several methods have
been proposed to address the problem of texture analysis. They can be divided into
three main approaches: statistical, structural and spectral [8].
The purpose of texture analysis is to extract the characteristic properties of the
object in an image and express them in a form of feature vector. The obtained repre-
sentation will serve as a basis for later steps in image analysis and interpretation [9].
However, the search for discriminating parameters characterizing the texture and
the use of these parameters for the segmentation still remains a delicate problem,
without a universal solution. It is in this context that we suggest in this paper struc-
tural segmentation methods of textured images based on Gabor filters and the wavelet
transform coefficients.
Two phases are usually involved in the texture classification process, namely the
learning phase and the recognition phase. The learning phase aims to build a texture
content model of each texture class present in the training data, which usually consists
of images with known class labels. In the recognition phase, the texture content of the
unknown sample is initially described using the same texture analysis method. Then,
the textural features are compared to those of the training images with a classification
algorithm, and the sample is assigned to the class with the best match [10].
Texture segmentation may be block wise or pixel wise. Texture features in the
vicinity of each pixel within the image are evaluated in the pixel wise segmenta-
tion schemes. The benefit of pixel wise segmentation over block wise lies in the
elimination of blockiness. Hence, we chose pixel wise segmentation for features
extraction.
In the present research, we attempt to use two types of feature extraction tech-
niques with the purpose of classifying textured images. These features are structural
features obtained using Gabor filters and the wavelet transform coefficients.
In the first step of our work, we applied these two features extraction strategies
on textured image in order to get more information that enables the classifiers to
discriminate the textures. In the second step, an appropriate classification algorithm
is applied using the set of extracted features in the previous step. The neural networks
classifier is chosen among the most well-known classifiers. It was initiated in [11, 12]
and the estimated feature vector of each pixel is sent to the neural networks classifier
for labeling. For the training phase of the neural networks, images of 320 × 320
size were used, and 6400 patterns are selected randomly from each image, which
represents 6.25% of the total number of patterns.
The proposed segmentation algorithms performance was verified on synthetic
images from Brodatz and DTD datasets. The obtained classification results confirm
Comparison Between Gabor Filters and Wavelets Transform … 1023
the power of the wavelet transform features compared to Gabor filters features in
classification of textured images.
The rest of this paper is organized as follows: Sect. 2 gathers the background
theory of various techniques use in this paper. Then, Sect. 3 is dedicated to give
in details the segmentation process and the achieved performance of the proposed
methods. Finally, in Sect. 4, we conclude and recommend possibilities for future
work.
2 Feature Extraction Algorithms
A. The wavelet analysis
Since the work of Grossman and Morlet [13], the wavelets transform has appeared
as a powerful tool to solve problems in different application. The wavelet transform
decomposes the input signal into a series of wavelet functions Ψa,b (t) derived from
a mother function Ψ (t) given by dilatation (factor a) and translation (factor b) oper-
ations. Figure 1 shows some examples of wavelets that are usually used in image
processing [1].
Wavelet analysis transforms a finite energy signal in the spatial domain into
another finite energy signal in the spatio-frequency domain. The components Ca,b
of this new signal, described in Eq. (1), are called wavelet coefficients. In an image,
these coefficients offer information on the local variation of the gray levels around a
given pixel. The more significant is this variation, the higher they are [8].
Fig. 1 Classical examples of wavelets. a Wavelet of Morlet, b Wavelet of Mexican hat, c Wavelet
of Meyer
Fig. 2 Use of 1-stage

discrete wavelet transform LL HL
Original DWT
(DWT)
image
LH HH
+∞
Ca,b = x(t)ψa,b (t)dt, (1)
−∞
where

1 t −b
ψa,b (t) = √ ψ with a = 0 (2)
a a
The major advantage of wavelets as compared to other frequency methods, as

Fourier transform, is that it offers both frequency and spatial locality [10]. In 1989,
Mallat [14] proposed a multi-resolution decomposition algorithm based on wavelet
transform. The algorithm decomposes an input image into a set of detail images and
an approximation image using a filter bank comprising a high pass filter (HP) and a
low pass filter (LP). At each decomposition level the size of the transformed images
is reduced by a factor of two [8]. The discrete wavelet transform of a 2-D image can
be obtained by performing the filtering successively along horizontal and vertical
directions (separable filter bank) [15]. Four images are then generated at each level.
Figure 2 illustrates an example of decomposition of the image on one level.
In Fig. 2, LL contains vertical and horizontal low frequencies, it represents approx-
imation coefficients; LH contains vertical high and horizontal low frequencies, it
represents vertical details; HL contains vertical low and horizontal high frequencies,
it represents horizontal details; HH contains vertical and horizontal high frequencies,
it represents diagonal details.
The values of transformed coefficients in detail and approximation images (sub-
band images) represent the necessary features that capture useful discrimination
information for texture segmentation [16].
(1) Choice of the wavelet: In our work, we have used a second order biorthogonal
spline wavelet. This wavelet is often used for texture analysis due to its excel-
lent location in the frequency and spatial domains and its sensitivity to local
singularity and correlation of the image [8].
(2) Calculation of the texture indices: One of the most used indices for character-
izing the texture in the spatio-frequency plane is the measurement of energy.
Because the transformed images have different frequencies, orientations and
scales, the energy index is a local measure of the wavelet coefficient distribu-
tion according to the frequency, the orientation and the scale. It has been used
successfully for classification and segmentation of textures [8].
The expression of energy expressed in standardized form is given by [8]:
Fig. 3 Wavelet features extraction process
1
E= C(i, j)2 (3)
N R
The second index used in conjunction with energy is the measurement of the local
mean of the wavelet coefficients given by [8]:
1
M= |C(i, j)|, (4)
N R
where N denotes the number of pixels, named by the indices (i, j), and enclosed in
the neighborhood R on which the texture is estimated.
The evaluation of these indices was done on a sliding window W. The energy
and the local mean on the sliding window are calculated from the resultant sub-band
images. So, the feature vector of each window is made of eight parameters V = [ELL ,
ELH , EHL , EHH , MLL , MLH , MHL , MHH ], as seen in Fig. 3.
Various tests were carried out on a series of window sizes going from 5 × 5 to 25
× 25. The uppermost good classification rate was attained for a window of 11 × 11
dimension. The estimated features vector of every window is used as an input to the
neural networks classifier for labeling, and the score for the window provided by the
neural networks is allocated to its central pixel.
B. Gabor Filters
Gabor wavelets have been widely used in diverse fields of image processing like
texture classification, fingerprint identification, image coding, and edge detection
[17]. A Gabor wavelet is a linear filter whose impulse response is defined as a
harmonic function multiplied by a Gaussian function. Gabor filter orientation and
frequency representations are alike to those of human visual system [18].
A Gabor filter in the spatial domain is given by [19]:
h(x, y) = g(x, y) cos(2π(u 0 x)), (5)
where g(x, y) is a 2D Gaussian defined as:

x2 2
− 21 + y2
g(x, y) = e σx2 σy
, (6)
and u 0 is the spatial frequency of a cosine along the x axis and orientation of 0°, and
(σ x and σ y ) are the spatial variances of the Gaussian envelope along the x and y axes,
successively [19]. The remaining filters, with orientations of 0°, 45°, 90° and 135°
are obtained by rotating the co-ordinate system, or by using the rotation equation for
the (x, y) plane, where θ0 represents the rotation angle [19]:

x2 2
− 21 + y2
h(x, y) = e σx2 σy
cos(2π u 0 (x cos θ0 + y sin θ0 )) (7)
The orientation selectivity properties of the Gabor Filter are far clearer in the
frequency domain. So, the Fourier Transform of Eq. (7) is given by the subsequent
expression [19]:

1 (u − u 0 )2 v2 1 (u + u 0 )2 v2
H (x, y) = A exp − + + exp − + ,
2 σu2 σv2 2 σu2 σv2
(8)
where σu = 1/2π σx , σv = 1/2π σ y and A = 2π σx σ y .

Figure 4 shows an even-symmetric Gabor filter and its frequency representation
in a 64 × 64 array, and orientation of 0°.
In this research, an ensemble of wavelets is created using a mother wavelet like
in Eq. (7). Then the entire image is used as the input to the wavelet set, as shown in
Fig. 4 Symmetric Gabor

filter and its frequency
representation. a Spatial
representation of the Gabor
Filter, b Frequency
representation of the Gabor
Filter
Fig. 5 Gabor features extraction process
Fig. 5. By doing so, the input image is transformed into a set of N filtered images.
As the filters output are complex valued, we exploit the magnitude of these values.
Two important parameters within Gabor filters are the number of frequencies
(N f ) and directions (N d ) [17], for the √having√size of√320 × 320√used in this
√images
work, five radial frequencies (N f = 2/26 , 2/25 , 2/24 , 2/23 , and 2/22 ) with
4 orientations (N d = 0°, 45°, 90°, 135°) were adopted according to [20]. So, the
features vector is composed of twenty parameters (5 frequencies × 4 orientations).
Finally, the estimated features vector is sent to the neural networks classifier for
labeling.
3 Experiment Results
The performance of the proposed algorithms by segmenting textured images is

assessed using many images with different textures and different shapes. They are
collected in the wild and in cluttered conditions, from Brodatz and DTD datasets
[21, 22], as shown in the first column of Table 1. Each image (a), (b), (c), (d), contain
the same textures that are present in different forms and shapes (square, rectangle,
triangle, circle, and curved shape) to check the performance that could be achieved
by the proposed segmentation techniques. For all experiments, the percentage of
right classification is used as an indicator to assess the classification precision.
The results are illustrated in the 3th and 4th columns of the Table 1. These last
results clearly confirm the power of the wavelet transform features in textured-images
segmentation and the ability of Gabor features to find class boundaries.
Table 2 demonstrates the recognition rates that were achieved by utilizing wavelet
transform and Gabor features.
According to the 2nd and 3rd columns of Table 2, the results obtained by utilizing
the wavelet transform shows that the quality of the segmented images is enhanced
and that the recognition rates are improved compared with the rates obtained by
utilizing Gabor filters. Wavelets offer great discriminatory power between textures
with strong resemblances, while Gabor features are preferable in the areas close to
the class borders.
We note that in each images (a), (b), (c), (d) we have tested the same textures in
different shapes and we observe that our segmentation algorithms give almost the
same classification results whatever the form of the used textures.
Table 1 Experimental results. The first column contains textured images in different shapes. The
second column contains the ground truth map of different images. The third and fourth columns
are the classification results using respectively Gabor filters, wavelet transform
Original image Ground truth map Gabor Filters Wavelet Transform
(a1)
(a2)
(a3)
(b1)
(b2)
(b3)
(c1)
(c2)
(continued)
Table 1 (continued)
(c3)
(d1)
(d2)
(d3)
Table 2 Classification rates

Image Gabor filters (%) Wavelet (%)
of textured images
(a1) 90.32 94.03
(a2) 93.35 94.42
(a3) 92.53 92.99
(b1) 92.66 94.93
(b2) 94.02 94.57
(b3) 93.76 94.67
(c1) 92.24 94.71
(c2) 93.88 95.41
(c3) 94.14 95.74
(d1) 93.08 94.71
(d2) 92.87 94.74
(d3) 92.98 93.55
Consequently, by our feature extraction methods, the segmentation results of the

textured images are acceptable, and, as future work, it seems that by combining these
two types of features, we may be able to use the advantages of them both.
4 Conclusion
In this article, we have suggested two approaches for the segmentation of textured
images. We started by extracting the features using Gabor filters and the wavelet
transform. After that, each estimated vector of features for every pixel was sent
to the neural networks classifier for labeling. The proposed algorithms were tested
on synthetic images from Brodatz and DTD datasets. The obtained classification
results confirm the power of the wavelet transform features compared to Gabor
filters features in classification of textured images.
References
1. Sahbani MH, Hamrouni K (2005) Segmentation d’images texturées par transformée en

ondelettes et classification C-moyenne floue. In Proceedings of 3rd international conference
sciences of electronic, technologies of information and telecommunications, Tunisia, 27–31
March
2. Wild P, Hofbauer H, Ferryman J et al (2017) Robust iris image segmentation. In: Rathgeb
C, Busch C (eds) Iris and periocular biometric recognition. IET Book Series on Advances in
Biometrics. The Institution of Engineering and Technology, London, pp 57–82
3. Peña-Barragán JM, Ngugi MK, Plant RE (2011) Object-based crop identification using multiple
vegetation indices, textural features and crop phenology. Remote Sens Environ 115(6):1301–
1316
4. Sasikala M, Kumaravel N, Subhashini L (2006) Automatic tumor segmentation using optimal
texture features. In IET 3rd international conference on advances in medical, signal and
information processing, MEDSIP, pp 1–4
5. Filiberto P, Gema G, Pedro GS et al (2009) Multi-spectral texture characterisation for remote
sensing image segmentation. In Proceedings of 4th Iberian pattern recognition and image
analysis conference, pp 257–264
6. Pu H, Sun D-W, Ma J et al (2014) Using wavelet textural features of visible and near infrared
hyperspectral image to differentiate between fresh and frozen-thawed pork. Food Bioprocess
Tech 7(11):3088–3099
7. Su H, Sheng Y, Du P et al (2015) Hyperspectral image classification based on volumetric
texture and dimensionality reduction. Front Earth Sci 9(2):225–236
8. Iftene T, Safia A (2004) Comparaison Entre La Matrice De Cooccurrence Et La Transformation
En Ondelettes Pour La Classification Texturale Des Images Hrv (Xs) De Spot. Télédétection
4(1):39–49
9. Iggane M, Mammass D, Khenchaf A et al (2009) Segmentation d’images sonar par matrice
de co-occurrence. In Proceedings of 5th international conference sciences of electronic,
technologies of information and telecommunications, Tunisia, 22–26 March
10. Anibou Ch, Saidi MN, Aboutajdine D (2015) Classification of textured images based on discrete
wavelet transform and information fusion. J Inf Process Syst 11(3):421–437
11. Hérault J, Jutten, Ch. Réseaux neuronaux et traitement du signal (Hermès, 1994)
12. Jodouin JF. Les réseaux de neurones. Principes et définitions (Hermès, 1994)
13. Grossman A, Morlet J (1984) Decomposition of hardy functions into square integrable wavelets
of constant shape. SIAM J Math Anal 15(4):723–736
14. Mallat S (1989) A theory of multiresolution signal decomposition: the wavelet representation.
IEEE Trans Pattern Anal Machine Intell 11(7):674–693
15. Scheunders P, Livens S, Van de Wouwer G et al (1997) Wavelet-based texture analysis. Int J
Comput Sci Inf Manag, December
16. Arivazhagan S, Ganesan L (2003) Texture segmentation using wavelet transform. Pattern
Recogn Lett 24(16):3197–3203
17. Mirzapour F, Ghassemian H (2015) Fast GLCM and Gabor filters for texture classification of
very high resolution remote sensing images. Int J Inf Commun Technol Res 7(3):21–30
18. Sivalingamaiah M, Reddy BDV (2012) Texture segmentation using multichannel Gabor
filtering. IOSR J Electron Commun Eng 2:22–26
19. Bosnjak A, Montilla G, Torrealba V (1998) Medical images segmentation using Gabor filters
applied to echocardiographic images. Comput Cardiol IEEE 25:457–460
20. Jain AK, Farrokhnia F (1991) Unsupervised texture segmentation using Gabor filters. Pattern
Recogn 24:1167–1186
21. Cimpoi M, Maji S, Kokkinos I et al (2014) Describing textures in the wild. In Proceedings of
IEEE conference computer vision and pattern recognition, Columbus, OH, pp 3606–3613
22. Cimpoi M, Maji S, Vedaldi A (2015) Deep filter banks for texture recognition and segmentation.
In Proceedings of IEEE conference computer vision and pattern recognition (CVPR), Boston,
MA, pp 3828–3836
High Level Modeling and Hardware
Implementation of Image Processing
Algorithms Using XSG
Kamel Messaoudi, Hakim Doghmane, Hocine Bourouba, E. B. Bourennane,

and S. Toumi
Abstract Design of Systems-on-Chip has become very common especially with

the remarkable advances in the field of high-level system modeling. In recent years,
Matlab also offers a Simulink interface for the design of hardware systems. From a
high-level specification, Matlab provides self-generation of HDL codes and/or FPGA
configuration codes while providing other benefits of easy simulation. In addition, a
large part of the Systems-on-Chip use at least one image processing algorithm and at
the same time border detection is one of the most used algorithms. This paper presents
a study and a hardware implementation of various algorithms of borders detection
realized under Xilinx System-Generator. The various algorithms are implemented
using Xilinx FPGA device, the simulation and synthesis results are also compared.
We use the Xilinx Zed-Board for physical implementations in-the-loop.
Keywords Edge detection · Hardware implementation · Image processing ·

VHDL · Xilinx system generator · Xilinx Z-Board
1 Introduction
In recent years and with the remarkable advances of powerful and low-cost processing
chips, applications of signal processing (and in particular image processing) have
grown dramatically in importance [1]. In-fact, media information such as audio,
images, and video have come to be necessary for various computer applications [2]
such as medical, video surveillance, target recognition, robotics application, etc…
K. Messaoudi (B) · E. B. Bourennane · S. Toumi

Electrical Engineering Department, Mohemed Cherif Mssaadia University, Souk-Ahras, Algeria
H. Doghmane · H. Bourouba
Laboratory of Inverse Problems, Modeling, Information and Systems (PIMIS), 8 Mai 1945
University, Guelma, Algeria
H. Bourouba

https://doi.org/10.1007/978-981-15-6403-1_71
1034 K. Messaoudi et al.
Image processing has a strong mathematical basis where the quality is enhanced by
using various processing techniques. The main objective of image processing is to
improve the quality of the images for human interpretation and analysis [3].
Several basic processes are applied to a digital image defined in a digital computer
as being a matrix of pixels. Acquisition, Enhancement, Restoration, Segmentation
and Analysis are the steps needed by just about every application which involves
image processing [4]. Edges are one of the most significant features for image
processing and many computer vision applications. In various image processing
algorithms, the filtering operation (and in particular the edge detection) is gener-
ally the first step used in order to eliminate noises and to reduce the amount of
data. In digital image, the edge defines the object boundaries within the image and
occurs when discontinuities present in the pixel values. Edge detection is a pre-
processing step for many image processing algorithms such as image enhancement,
image segmentation, object tracking, object recognition and image/video coding [3].
Currently, Matlab-Simulink is enriched with features and templates that simul-
taneously enable data processing (using fixed-point precision) and ensure self-
generation of synthesizable VHDL codes for FPGA targets. In addition to the
easy simulation provided by Matlab-Simulink and using the new Xilinx System-
Generator (XSG tool), we can directly implement the proposed Simulink models
on FPGA through the various simulation and co-simulation steps, the Input/Output
(I/O) assignment, the temporary and spatial constraints and at the synthesis and
implementation.
This paper presents a study with hardware implementations of edge detection algo-
rithms using the XSG tool. The idea is to compare Matlab-Simulink image processing
results (using floating point precision) with XSG image processing results (using
fixed-point precision). The idea is also to implement advanced image processing
algorithms and in particular edge detection methods. Using the new “in the loop”
technique, a first part of the proposed model is simulated under Simulink using
the main processor and a second part is implemented in real time on FPGA device.
Matlab-Simulink will provide real-time communication and links between both parts.
The rest of this paper is organized as follows: we first define image processing
and edge detection in Sect. 2. The XSG design flow is presented and discussed in
Sect. 3. In Sect. 4, we present and detailed the proposed hardware implementation
of edge detection algorithms. Simulation and synthesizes results are discussed in
Sect. 5. Finally, we present the conclusions and outline future work in Sect. 6.
2 Image Processing and Edge Detection
Image processing is a very vast field that has known, and still knows, an important
development for a few decades. By digital image processing is meant all the tech-
niques enabling a digital image to be modified in order to improve the visual quality
or to extract the relevant information from it.
High Level Modeling and Hardware Implementation of Image … 1035
Image processing is a discipline widely used in several fields namely: medicine,

military field, target detection, robotics, etc. This discipline is highly sought also in
the field of automation and robotics where important decisions are often taken from
the images processed by various algorithms.
There are number of well defined processes which go to make up a typical image
application. Acquisition, Enhancement, Restoration, Segmentation and Analysis are
the steps needed by just about every application which involves image processing
[2]. Edges are one of the most significant features for image processing and many
computer vision applications. Edge detection is one of the classical studies since it
is perquisite thing prior to perform any image processing operation. It has provoked
interest among research fraternity since several years [3].
A. Edge detection
Edge detection is used as a pre-processing step for many image processing algorithms
such as image enhancement, image segmentation, tracking and image/video coding.
This is a basic operation in image processing, and refers to the process identifying
and locating sharp discontinuities in an image; the discontinuities are abrupt changes
in pixel intensity which characterize boundaries of objects in a scene [5].
In the literature, various gradient based edge detection algorithms are proposed
such as Robert, Prewitt, Sobel and Canny (as example). These operators are firstly
linear and secondly use simple convolution masks. The elements of these masks are
often in the set {−2, −1, 0, 1, 2} which makes easier the convolution operations
in the mathematical expressions of the filters. Indeed, in a mathematical operation
the multiplication by (−1, 0 and 1) can be suppressed and the multiplications by
2 and by −2 can be realized using additions or more simply by left offsets. This
is an advantage in a software and/or hardware implementation generally to avoid
the expensive multiplication operations (hardware resources and computing time).
The ultimate goal is to speed up calculations of mathematical expressions in filters
especially with larger amounts of data in the images.
B. Hardware implementation of edge detection algorithms
Most implementations of image processing algorithms are software and run on

personnel computer and generally do not respect real-time. The purpose of these
software implementations is to verify the effectiveness of the proposed algorithms
by simulations on synthetic images or real images. These images are extracted from
known databases to compare the simulation results.
In the literature, various hardware implementations are proposed for edge detec-
tion algorithms in order to ensure real time processing of different tools and applica-
tions. In [1, 3, 6], various hardware implementations of gradient based edge detec-
tion algorithms (such as Robert, Prewitt, Sobel and Canny) using Xilinx System
Generator tool have been proposed. These edge detection algorithms are firstly
designed and simulated in Matlab-Simulink and then reproduced using only Xilinx
modules (these modules are recently added in Matlab-Simulink). Traditional simula-
tion modules are used to read and display images in Matlab-Simulink (co-simulation).
Designed algorithms are efficiently compiled, synthesized and implemented using

FPGA devices.
In this paper, we propose a real-time hardware implementation of some image
processing algorithms. In fact, the basic idea of the proposed implementations
consists in the realization of a platform of acquisition and display of images while
adding software and/or hardware modules. This platform can be used in multiple and
varied applications of automatic and industrial computing.
3 XSG Design Flot
Currently, FPGA devices are widely used either to realize prototypes or hardware
and real-time implementations of processing algorithms. In the industrial and espe-
cially in the real-time control of systems field FPGA devices are also used [7].
Moreover, a new tool called Xilinx System Generator (XSG) is added recently under
Matlab-Simulink. XSG is a new design tool that enables the model-based Simulink
design environment for hardware implementation (VHDL and RTL levels) and FPGA
design. Using the XSG tool, previous experience on hardware implementations and
RTL design methodologies are not required. XSG maintains a very harmonious
level with Simulink Blockset, but at the same time allows automatic translation
of models to synthesizable and efficient hardware implementations using a Xilinx
specific Blockset.
The use of Matlab-Simulink allows the easy creation of testbench for simula-
tion and co-simulation of Simulink models and Xilinx models at the same time.
In a hardware implementation under Simulink and System Generator various steps
are necessary from Description of the algorithm in mathematical terms to the auto
generation of HDL code and the synthesis of results.
Figure 1 shows the System Generator design flow. System Generator works
within the Simulink model-based design methodology [7]. Firstly, the algorithm
of the application can be developed and implemented using the standard Simulink
Blockset. Matlab-Simulink uses floating-point numerical precision and without hard-
ware detail. This software implementation can be verified using Simulink simulation
results. System Generator can be used to specify the hardware implementation details
for the FPGA devices using Xilinx DSP Blockset. The hardware implementation can
be verified both using Simulink simulation and System Generator simulation. Simu-
lation results can be compared to propose necessary modifications. System Gener-
ator invokes Xilinx Core Generator to generate VHDL code, the NGC file or the
Bitstream for the specified FPGA device. Several steps are required before and after
the generation of synthesizable codes.
Fig. 1 System-Generator design flow [7]
4 Hardware Implementation of Edge Detection Algorithms
In this section, we will present some hardware implementations of some image

processing algorithms using the XSG tool. First, we change the image format (bi-
dimensional matrix of pixels) into a one-dimensional signal manipulated by the
processing modules provided by Xilinx. Then, we implement the image processing
algorithms using the Matlab-Simulink modules. After that, we cross-refer to the
implementation of the same algorithms using Xilinx processing modules.
In the second implementation, we copied the same Simulink model of the image
processing algorithm using only the Xilinx Blockset. Several tests are carried out for:
(i) the good resizing of various variables, (ii) the synchronization of the operating
frequencies of the Simulink blocks and (iii) the Xilinx blocks. The purpose of this part
is to ensure the transition from floating-point precision (used by Simulink blocks)
to fixed-point precision (used by Xilinx blocks) to make the results comparable.
Figure 2 shows the flowchart of the proposed main idea.
Fig. 2 In-the-loop implementation of the proposed idea
A co-simulation is necessary in order to compare simulation results and to correct

errors in the proposed models. Therefore, several adaptations and configurations are
necessary for each Xilinx processing module. The advantage of these Xilinx-based
models is the automatic transition to synthesizable HDL codes, which are used. We
use these HDL codes for FPGA configuration. In addition, for rapid implementation
and physical verification [7], these HDL codes are combined with the “in-the-loop”
technique.
A. The processing platform
Matlab® is software mainly adapted to matrix calculation. A simple scalar number

is considered under Matlab as a matrix of size 1 × 1. An image in Matlab is saved as
an N × M matrix (N × M pixels, where N and M represent the number of rows and
columns, respectively). All operations used in Matlab are applicable on the image
matrix. In hardware implementations using an HDL, the details of the operations
applied to the simple data are necessary for optimizing the implementations. In order,
to manipulate an image, it is necessary to convert the matrix into a vector of pixels. In
this section, we will take advantage of the reading and display tools available in the
Matlab-Simulink libraries while taking advantage of the Xilinx System Generator
tool for the realization of hardware accelerators in HDL.
The proposed models combined with the image reading and display modules are
used for checking the software implementations of the processing algorithms. Then,
we propose models based only on Xilinx modules (the XSG library). For the reading
and the display of the images for these last models we use the same Matlab modules
and we compare the results.
a. Pre-processing Matlab model
An image in Matlab is defined as a two-dimensional (2D) arrangement. To meet
the requirement of hardware implementations, the image should be pretreated and
given as a one-dimensional (1D) vector. The Matlab-Simulink implementation
based on a model used for image preprocessing is shown in Fig. 3. In order, to
process a 2D image, it is first necessary to calculate the transposed image and
then, it is necessary to convert the 2D matrix into a 1D vector using the “convert
2D to 1D” module. Next, we use the “to frame” and “Unbuffer” modules to read
the vector with one pixel every time with a predefined scalar data format and at
a higher sampling rate.
b. Post-processing model
As shown in Fig. 4, the post-processing module is used to convert the output of
the image result to the original format. This module is used before the display
of the processed image and to compare simulation results. The element modules
are practically the inverse modules of the first chain, we start with the conversion
of the floating point data using “Data type conversion” then, the “Buffer” and
the “Convert 1-D to 2-D” modules for the conversion of pixels in the form of a
matrix. Finally, we calculate the transposed matrix with the display of the image
(Fig. 4).
B. The proposed hardware implementations
Fig. 3 The used pre-processing model in the proposed implementation
Fig. 4 The used post-processing model in the proposed implementation

In this section, we propose hardware implementations for several filters using Xilinx
Blockset. We compare simulation results (co-simulation) and we discuss the synthesis
results. Then, we show the “in the loop” execution of the proposed implementations.
In order, to optimize these implementations, two large parts are proposed:
• The first part concerning the memory management: this part is utilized for the
preparation of pixels concerned by the treatment at each time. In fact, the filtering
operation is adapted to a window of size 3 × 3 or 5 × 5 pixels over the entire
image matrix.
• The second part is used to define the architecture of the used filter in order to
calculate resulting pixel according to the coefficients of the filter window.
a. Memory management for the used filters
To ensure the continuity of the pixels subject of processing, we used three (or
five) rows of flip-flops according to the number of pixels in the image. If we
process an image of size N × M, we use M flip-flops in each line [8]. We have
configured these three rows of flip-flops as a FIFO or a shift register.
At the output of this block, we can recover 3 × 3 pixels being processed and
each rising edge of the clock a pixel shift is made to recover the 3 × 3 pixels
subject of the next processing. By shifting the pixels, we can recover the 3 × 3
pixels processing subjects in the same position, which implies a significant gain
in hardware implementations. Indeed, instead of shifting the window of the 3 ×
3 pixels in the NxM size image, it is sufficient to shift the pixels of the image and
recover the 3 × 3 pixels in the same place (Fig. 5).
b. The processing module architecture

This part starts with the point-by-point multiplication of the two windows (the
window of 3 × 3 pixels and the window of coefficients). This multiplication is
followed by a sum and in some cases by a division. The mathematical formula
is given by:
Fig. 5 Block diagram of the memory management strategy used for the implemented filters

3
3
The pixel result = coefficient(i, j) ∗ pixel(i, j) (1)
i=1 j=1
In order to avoid multiplication and division operations (which are costly in terms
of FPGA resources consumption) and since most filters have simple coefficients
(−1, 0, and 1), this formula is replaced by another formula based on addition,
subtraction and shifting operations. At the end of this part, we use a block for the
adjustment of data (the value of the pixel to be in the domain 0–255).
C. The proposed implementation for the 3 × 3 Sobel filter
Figure 6 shows the proposed hardware implementation under System-Generator for

the 3 × 3 Sobel filter. We also notice the presence of the main module (System-
Generator) to ensure the automatic conversion of models in HDL level. Several
adaptations and data processing’s are also necessary to ensure simulation results
comparable with the initial filter and to ensure the self-generation of a correct and
synthesizable HDL codes.
A. Simulation results
The proposed implementation based on Xilinx Blockset operates using fixed-point

precision. The processing results are different from the results of the Matlab filter
that operates using floating-point precision. Figures 7 and 8 show the simulation
results of the proposed implementations for the Sobel and Prewitt filters applied on
two images extracted from Matlab library.
B. Synthesis results
After self-generation operation of the HDL code using System-Generator from a

Matlab-Simulink model, a synthesis step is required using one of the Xilinx tools.
In this work, we use the Xilinx-ISE tool. This software allows editing, verification,
behavioral and functional simulation as well as the synthesis and generation of FPGA
configuration files.
In this work, several filters were made under Matlab-Simulink we have subse-
quently converted these models into VHDL. The synthesis results of these codes are
shown in Tables 1. Two options are possible for the synthesis operation: synthesis
using RAM blocks or without use of the RAM blocks. These results show the use
around 300 slices registers and around 270 slice LUTs as well as the use of 2 RAM
block. The number of resources consumed depends on the filter used and the size of
the processed image.
1042
Fig. 6 Hardware implementation of 3 × 3 Sobel filter using system-generator

K. Messaoudi et al.
Original image (ball.gif)

320x250pixels
Matlab function The proposed filter

Sobel 3x3 Sobel 3x3
Matlab function The proposed filter

Prewitt 3x3 Prewitt 3x3
Fig. 7 Simulation results of filters using Matlab-Simulink (the image mri.tif)
6 Conclusion
In this paper we have proposed hardware implementations of image processing algo-

rithms using Xilinx System Generator. It is in the context of rapid prototyping of
real-time implementation of image processing applications using FPGA devices with
ever greater integration. As processing algorithms, we have chosen the border detec-
tion filters (High-Pass Filters) because of theirs utilizations in several domains and
applications. For each filter, we start with a high-level implementation under Matlab-
Simulink. Subsequently, we implement the same filter always under Matlab-Simulink
but using the Xilinx Blockset. The advantage of this second implementation is that it
is automatically convertible into HDL level and synthesizable for an FPGA device.
Original image (mri.tif) Matlab function

128x128pixels Sobel 3x3
The proposed filter The proposed filter

Sobel 3x3 Prewitt 3x3
Fig. 8 Simulation results of filters using Matlab-Simulink (the image mri.tif)
Table 1 Synthesis results

Filter & size Image & size Slice registers Slice luts Block RAM/FIFO
Available on used FPGA device 106400 53200 140
Sobel 3 × 3 football.jpg 320 × 343 264 2
256pixels
Sobel 3 × 3 mri.jpg 128 × 355 275 2
128pixels
Prewitt 3 × 3 football.jpg 320 × 249 227 2
256pixels
Prewitt 3 × 3 Mri.jpg 128 × 343 264 2
128pixels
The use of both implementations allows the co-simulation as well as the compar-
ison of the simulation results. The co-simulation results show the effectiveness of
the proposed implementations despite the operation with fixed-point precision. The
synthesis results show the consumption of less than 1% of FPGA resources using
the Xilinx Zed-Board. As perspectives of this work, we propose to implement other
filters that require more computing time as well as the transition to other types of
image processing.
References
1. Elamaran V, Rajkumar G (2012) FPGA implementation of point processes using Xilinx system
generator. J Theor Appl Inf Technol 41(2):201–206
2. Alareqi M, Elgouri R, Hlou L (2014) High level FPGA modeling for image processing algorithms
using Xilinx system generator. Int J Comput Sci Telecom 5(6), June
3. Avinash GM, Shah AM (2017) FPGA implementation of gradient based edge detection
algorithms. Int J Innov Res Comput Commun Eng 5(5), May
4. Bin Othman MF, Abdullah N, Bin Ahmad Rusli NA (2010) An overview of MRI brain classifica-
tion using FPGA implementation. In IEEE symposium on industrial electronics & Applications
(ISIEA), pp 623–628, Oct
5. Rakesh MR (2013) Design and simulation of Matlab/Simulink model for edge detection
techniques in image segmentation. Int J Adv Res Electr 2(12):5828–5834
6. Gupta A, Vaishnav H, Garg H (2015) Image processing using Xilinx system generator (XSG)
in FPGA 2(6):119–125, Sept
7. Messaoudi K, Yahi A, Messaoudi N, Bourennane EB, Toumi S (2016) Adaptive hardware imple-
mentation for the deblocking filter used in H.264/AVC using system generator. In ICESTI’16,
International Conference on Embedded Systems in Telecommunications and Instrumentations,
Annaba, Algeria, 24–26 Oct 2016
8. Messaoudi K, Bourennane EB, Toumi S, Kerkouche E, Labbani O (2010) Memory require-
ments and simulation platform for the implementation of the H.264 encoder modules. In IEEE
international conference on image processing theory, tools and applications IPTA’10, Paris, pp
133–137, Juillet 2010
Proposal for a Radar Detection
Architecture Based on the Knowledge
Based Systems Exploitation
Abdellatif Rouabah, Hamza Zeraoula, M’hamed Hamadouche,

and Kamal Tourche
Abstract In this paper, we propose a radar detection technique, based on the Knowl-
edge Based Systems (KBS) exploitation. The proposed detector called Knowledge
Based Systems-Adaptive Linear Combined (KBS-ALC). This technique combines
the prior knowledge of the environment and the ability of the Adaptive Linear
Combined-Constant False Alarm Rate (ALC-CFAR) detector to automatically adapt
the detection threshold to the changing environment, particularly in the presence of
interfering targets. The simulations results show that the KBS-ALC detector performs
better compared with the KBS-Dynamic detector in terms of detection capability in
different situations, especially in the presence of interfering targets. The proposed
architecture has been implemented on the EZ-KIT-LITE ADSP TS 201 processing
board and the real-time processing constraints have been evaluated and discussed.
Keywords ALC-CFAR · Detection · ADSP TS 201 · GIS · Implementation · KBS
1 Introduction
Several types of CFAR detectors have been proposed in the literature, each one
of these types is adapted to a given operating environment. The classic forms of
adaptive threshold detectors are CA-CFAR (Cell Averaging) [1] for a homogeneous
A. Rouabah (B)
Département: Signal et Communication, Ecole Militaire Polytechnique, Bordj El Bahri, Alger,
Algérie
H. Zeraoula · K. Tourche
Département Radar, Ecole Supérieure -Ali Chabati, Reghaia Alger, Algérie
K. Tourche
M. Hamadouche
Faculté de Technologie, Univerisité de Boumerdes, Boumerdes, Algérie

https://doi.org/10.1007/978-981-15-6403-1_72
1048 A. Rouabah et al.
environment, SO (Smallest Of) [2] in presence of interfering targets only in one half
reference window, OS (Ordered Statistics) [3] in the presence of interfering targets
and the GO (Greatest Of) [4] in the presence of clutter edges.
The scale factor used for the detection threshold calculation depends on the clutter
distribution (Rayleigh, Weibull, exponential, log-normal…). In practice, the dynamic
change of the environment (sea, land, forests, sand….) in the same observation direc-
tion causes the degradation of detection performances in terms of target detection
losses and getting additional false alarms.
2 Problem Position
The major problem of CFAR detectors is the way of calculating the cell’s detection
threshold using the clutter power of neighboring cells, and since this calculation in
CFAR detectors does not have prior knowledge of the environment, then each CFAR
detector may be suitable for a special clutter environment, and not for all [5]. The
exploitation of the KBS offers a priori knowledge on the clutter for each cell in
order to solve the problems of non homogeneity of the environment [6] (sea, land,
forests, sand….) by taking into account only reference cells of the same type of GIS
(same type of clutter) as that of the CUT (Cell Under Test) for the estimation of
the detection threshold level in order to improve the detection quality of targets in
different operating situations of a radar.
3 KBS Establishment
In this work, we have considered two sources for the KBS establishment: GIS
(Geographic Information System) and Kolmogorov-Smirnov test (K-S test).
A. Geographic Information System (GIS)
The Geographic Information Systems is an expression related to information manage-

ment with a geographic component stored as a vector with associated attributes [7]
which are used to characterize clutter distribution in each cell range.
B. Kolmogorov-Smirnov test (K-S test)
Kolmogorov-Smirnov test is based on the comparison between two distributions [8],

we can use it as an identification test to recognize the distribution of the clutter in
each cell using signals from a high-resolution radar to get a sufficient number of
samples for a good identification.
The used data in this test are [9]:
• The set of random samples to be tested X1, X2,…..Xn of a certain function of
unknown distribution, denoted F(x);
Proposal for a Radar Detection Architecture … 1049
• S (x): empirical distribution function based on the random samples to be tested;

• F*(x): indicates a known distribution function. The test value T is the largest
vertical distance between S(x) and F*(x) (denoted by the “Sup”).

T = Supx F∗ (x) − S(x) (1)
According to the value of T, we deduce the concordance between F(x) and F*(x).
4 Detection Based on KBS
In this part, we present two detectors that use KBS in the detection threshold
estimation, called: KBS-Static and KBS-Dynamic.
A. KBS-Static
The exploitation of the information provided by the GIS makes it possible to assign
to the CUT and K reference cells the corresponding attributes Xcut and Xi , i =
1, 2, . . . K [10]. Moreover, for all i ∈ {1, …, K} the evaluation of the homogeneity
indicator D (i) as follows:

0 si Xi = Xcut
D(i) = (2)
1 si Xi = Xcut
The set containing the indices of the selected data is obtained as follows:

– If ki=1 D(i) ≥ Ks such that D(i) = 1 and Ks is a fixed value, then is the set of Ks
smallest indices, it is equivalent to choosing as formation data the homogeneous
Ks in the closest space to the CUT.
cells
– If ki=1 D(i) < Ks such as D (i) = 1, so consists of homogeneous cells indices

and Ks − ki=1 D(i) indices of non homogeneous cells closest to the CUT.
After the cell selection procedure, a CA-CFAR detector is used to make the final
decision about target presence.
B. KBS-Dynamic
The main difference between KBS-Static and KBS-Dynamic relies on the fact that
the first requires a fixed number of cells while the second chooses it dynamically
[10]. In other words, the prior information is still used to choose the cells number to
be considered in the threshold estimation.
Kmax and Kmin are the maximum and the minimum numbers of data exploited
for the threshold calculation. The set containing the indices of the selected data is
obtained as follows:

– If ki=1 D(i) ≥ Kmax , such as D(i) = 1, then is the set Kmax of the smallest
indices. It is equivalent to choose as selected data the homogeneous Kmax cells
in the closest space to the CUT.
k
– If Kmin ≤ i=1 D(i) ≤ Kmax , such as D(i) = 1, then is the set of all the
indices.
All homogeneous cells closest to the CUT are exploited to calculate the threshold.

– If ki=1 D(i) < Kmin , such as D(i) = 1, then is the set of homogeneous cells

indices and the Kmin − ki=1 D(i) indices of non homogeneous cells closest to
the CUT.
5 Detection Based on KBS Performances Analysis
The performances are analyzed in two parts; in the first one, the Kolmogorov-Smirnov
identification test is used as a source of knowledge for analyzing the detection
based on KBS exploitation performances in terms of effective false alarm rate in
a non homogeneous environment, in the second one we use the GIS as a source of
knowledge in order to analyze the performances of this technique in targets presence
case.
A. Effective False Alarm Rate Performances
We consider a clutter with three distributions (Rayleigh, Normal, and Exponential)

that PDF (Probability Density Function) expressions are:
Rayleigh distribution:

x x2
fX (x) = 2 exp − 2 (3)
σ 2σ
With σ is the scale parameter.

Normal distribution:

1 (x − μ)2
fX (x) = exp − (4)
2ψ2 2σ2
With μ is the mean and σ 2 is the variance.

Exponential distribution:
fX (x) = λ exp(−λx) (5)
With λ > 0 is the rate or inverse scale.

We use the Kolmogorov-Smirnov test to identify the distribution in each cell that
contains 30 samples to create the source of knowledge. After, we apply the detectors
(CA, GO, SO, OS and KBS-Dynamic) using a reference window of 24 cells to show
the influence of the non homogeneity of the clutter on the detection.
Figures 1, 2, 3, 4 and 5 show the false alarms for the CA, GO, SO, OS and
KBS-Dynamic detectors respectively.
Table 1 shows the number of false alarms for each detector.
CA-CFAR: the number of false alarms is 97 which is high, because the environ-
ment is non homogeneous, composed of three different distributions and the scale
factor T is calculated for a homogeneous Rayleigh distribution.
90
100
120 60
80
60
150 30
40
20
180 0
r = (cos(t)2+sin(t)2) 12
210 330
240 300
270
r = (cos(t)2+sin(t)2) 88
Fig. 1 False alarms for CA-CFAR
90
100
120 60
80
60
150 30
40
20
180 0
r = (cos(t)2+sin(t)2) 12
210 330
240 300
270
r = (cos(t)2+sin(t)2) 88
Fig. 2 False alarms for GO-CFAR

90
100
120 60
80
60
150 30
40
20
180 0
r = (cos(t)2+sin(t)2) 12
210 330
240 300
270
r = (cos(t)2+sin(t)2) 88
Fig. 3 False alarms for SO-CFAR
90
100
120 60
80
60
150 30
40
20
180 0
r = (cos(t)2+sin(t)2) 12
210 330
240 300
270
r = (cos(t)2+sin(t)2) 88
Fig. 4 False alarms for OS-CFAR
GO-CFAR: the number of false alarms is 68, it is relatively small compared to

CA-CFAR, it returns to the principle of the detector which chooses the largest value
of the sums of the samples between the two half-windows, which generates a high
detection threshold which reduces false alarms number.
SO-CFAR: the number of false alarms is 143 which is very high because of
detector principle which chooses the half window that have smallest samples sum,
this lowers the threshold below the level of the echo especially in the transition zones
from one distribution to another, then the number of false alarms increases.
90
100
120 60
80
60
150 30
40
20
180 0
r = (cos(t)2+sin(t)2) 12
210 330
240 300
270
r = (cos(t)2+sin(t)2) 88
Fig. 5 False alarms for KBS-Dynamic
Table 1 Number of false

CFAR detectors Number of false alarms
alarms for each detector
CA-CFAR 97
GO-CFAR 68
SO-CFAR 143
OS-CFAR 82
KBS-Dynamic 26
OS-CFAR: the number of false alarms is 82 which is reduced compared to CA-

CFAR because at the beginning of the transition from a higher level distribution to
another, the first samples are treated as interfering targets that are not considered in
the threshold estimation in OS-CFAR, so the false alarms are reduced.
KBS-Dynamic: the number of false alarms is reduced to 26 for the reason of the
detector principle which is based on a priori knowledge to choose the cells of the
reference window which have the same type of clutter as the cell under test for the
threshold estimation. Moreover, in the transition zones where the number of cells
with the same clutter type is reduced, the detector also reduces the number of cells
that enters in the threshold estimation, thus a better estimation (the threshold adapts
with the transitions of clutter).
B. Detection probability performances
To evaluate the detection probability performances for the KBS exploitation in targets
presence case, we use the geographic information system as a source of KBS, We
consider a vector of distance cells in a fixed azimuthal direction in different situations.
Fig. 6 Thresholds of CA-CFAR and KBS-Dynamic detectors in the absence of interfering targets
(1) Absence of interfering targets
We consider the following conditions:

– Size of the used reference window is 16.
– No interfering targets in the reference window.
Figure 6 shows thresholds of CA-CFAR and KBS-Dynamic Detectors in the
absence of interfering targets.
In this figure, targets of cells 3 and 43 are missed by CA-CFAR and detected by
KBS-Dynamic. This is because they are located in weak clutter cells that correspond
to a GIS of 0 and are surrounded by cells of a higher clutter which influences on the
threshold estimation of the CA-CFAR, while GIS information helps KBS-Dynamic
to keep a low threshold, which allows it to detect the targets.
(2) Presence of interfering targets in cells with different GIS
Figure 7 shows the received signal in the presence of interfering targets in cells
with different GIS as well as the evolution of the thresholds of the CA-CFAR and
KBS-Dynamic detectors.
We note that the target located in the cell distance 13 is missed by CA-CFAR
detector because of the presence of an interfering target that increases its threshold
level. But for the KBS-Dynamic, as the GIS level of the interfering target cell is 2,
is different from that of the cell under test which is 0, then it does not participate in
the threshold estimation so no influence.
Fig. 7 CA-CFAR and KBS-Dynamic thresholds in case of interfering targets presence in cells with
different GIS
(3) Presence of interfering targets in cells with Same GIS
Figure 8 shows the presence of multiple targets in the reference window and with
the same level of GIS.
Fig. 8 Presence of interfering targets in cells with same GIS

In case of target presence in the reference window and with the same GIS of
the cell under test (Fig. 8), the detection threshold estimation takes into account the
interfering target, this causes the loss of detection of the target in question.
KBS-Dynamic detector has good performances in the presence of interfering
targets without influence on the GIS. However, if there is influence on the GIS the
detector loses its performances.
To take advantage of the performance of this detector and resolve the problem of
interfering targets, we propose another detector called KBS-ALC that combines the
advantages of KBS-Dynamic to ensure better performances in heterogeneous areas
and also advantages of ALC-CFAR detector in case of interfering targets presence.
6 The Proposed Detector (KBS-ALC)
The ALC-CFAR detector combines the performance of CA-CFAR in a homogeneous

environment and OS-CFAR in the presence of interfering targets [11]. We have also
seen that KBS-Dynamic detector loses its performance in case of interfering targets
presence. To benefit from the advantages of the ALC-CFAR detector and at the same
time from those of KBS-Dynamic, we combine them by first applying the KBS-
Dynamic and use the prior knowledge on the environment for the choice of cells
with the same characteristics as the cell under test, and then apply the ALC-CFAR
detector on the chosen samples for the threshold estimation.
A. KBS-ALC with presence of interfering targets in cells with same GIS
Figure 9 presents a received signal with the presence of interfering targets, between
targets in cells of the same GIS as the cell under test and other that have different
GIS.
We find that despite the existence of interfering targets in the reference window
and even in cells with same GIS, the KBS-ALC detector detects the targets because
of its principle which is based on the application of KBS-OS detector in presence
of interfering targets case (after the calculation of the degree of homogeneity of
the reference window), so the interfering targets don’t affect the estimation of the
detection threshold.
7 DSP Implementation of KBS Detectors
In this part, we present the implementation results of KBS-dynamic and the proposed
detector (KBS-ALC) on EZ KIT LITE ADSP TS201 processing board.
We consider a situation with the presence of 06 targets, targets number 2 and 3
are located in two cells with different GIS while the cells of targets 5 and 6 have the
same GIS. Implementation results are shown in Figs. 10 and 11.
Fig. 9 KBS-ALC and KBS-Dynamic thresholds in presence of interfering targets in cells with
same GIS
Fig. 10 KBS-Dynamic implementation results
From the two Figs. 10 and 11, we can see that the presence of interfering target
in cell with the same GIS as the cell under test causes target detection loss for
KBS-Dynamic, while KBS-ALC solves this problem.
A. Evaluation of Execution Time
Real-time processing constraints are managed to meet the requirements of the ASR-
12 radar, whose characteristics are [12]:
– A repetition frequency of 1200 Hz corresponding to a period of 0.833 ms;
– A rotation speed of v = 12 rpm;
– Pulse width: 1 μs.
Fig. 11 KBS-ALC implementation results
Table 2 Execution time of KBS-Dynamic and KBS-ALC (in clock cycles)

N 8 16 24 32
KBS-Dynamic 503 890 631 043 791 920 884 000
KBS-ALC 1 746 980 1 833 840 1 856 320 1 908 396
Using an ADSP TS 201 processor with a cycle time of 1.67 ns [13] requires that
the total number of cycles does not exceed 500,000 cycles.
We calculated the execution time of the KBS Dynamic and KBS-ALC algorithms
(cycles) for 800 distance cells and for different values of N (size of the reference
window). The obtained results are shown in the Table 2.
From Table 2, we notice that the execution time of KBS-ALC is much higher than
that of KBS-Dynamic. This comes down to the nature of the operations performed
by the two detectors to estimate the threshold. The ordering operation of the samples,
the homogeneity factor calculation and the threshold combination of CA-CFAR and
OS-CFAR are additional calculations of KBS-ALC compared with KBS-Dynamics.
We also note that for both detectors, the real-time processing constraint imposed
by the radar ASR-12 is not respected because, for all the values of N, the processing
time of all the samples exceeds the critical time of a real-time application for the
radar ASR-12 which is 500 000 cycles.
Table 3 shows how many cells can be processed for each detector and for each
reference window.
To respect the real-time constraint required by the ASR-12 radar, one of the
proposed solutions is to proceed a coherent integration of Nc pulses.
Table 3 Number of treated cells in real time

N 8 16 24 32
Number of treated cells with KBS-Dynamic 793 633 505 452
Number of treated cells with KBS-ALC 228 218 215 209
The ASR-12 radar used parameters are:

Rotation speed v = 12 rpm, the opening angle at −3db is θ = 1.4◦ , PRF =
1200 Hz.
θ ∗ PRF
Nc = (6)
6∗v
For the ASR-12 radar, Nc = 23.
For an integration period = 23, the maximum tolerable time will be multiplied
by 23.
Tint−max = 23 ∗ Tmax = 23 ∗ 500 000 = 11 500 000 cycles.
With: Tint−max is the maximum tolerable time in case of integration of 23

pulses. After the integration of 23 pulses, the tolerable maximum processing time is
sufficiently large for the application of the proposed detector.
8 Conclusion
The comparative study of some detectors performances shows the superiority of

the KBS-Dynamic detector in terms of reducing the number of false alarms in a
heterogeneous environment. However, in the case of interfering targets presence
located in cells that have the same type of clutter (same GIS) as the cell under test,
the KBS-Dynamic detector presents performance degradation. In order to improve
the performances of this existing detector (KBS-Dynamic), we have proposed a
new detection architecture called KBS-ALC, based on the exploitation of a priori
knowledge on the radar coverage area on the one hand and ALC-CFAR detector
capabilities to automatically adapt to environmental changes on the other hand.
The simulation results show that the KBS-ALC detector performs better compared
with the KBS-Dynamic detector in terms of detection capability in different situa-
tions, especially in the presence of interfering targets. The proposed detection algo-
rithms are implemented on the DSP board type EZ-KIT LITE ADSP TS 201. For eval-
uation of the calculation times, we have considered the KBS-Dynamic and KBS-ALC
detectors.
References
1. Finn HM, Johnson RS (1968) Adaptive detection mode with threshold control as a function of
spatially sampled clutter level estimate. RCA Rev 29(3):414–464
2. Weiss M (1982) Analysis of some modified cell-averaging CFAR processors in multiple target
situations. IEEE Trans Aerosp Electron Syst, AE, Jan 1982
3. Rohling JH (1983) Radar CFAR thresholding in clutter and multiple target situations. IEEE
Trans Aerosp Electron Syst AES-19:608–621
4. Hansen GV, Sawyers JH (1980) Detestability loss due to greatest-of selection in a cell averaging
CFAR. IEEE Trans Aerosp Electron Syst 16:115–118
5. Conte E, Maio AD, Galdi C (2004) Statistical analysis of real clutter at different range
resolutions. IEEE Trans AES 40(3):903–918
6. Maio AD, Farina A, Foglia G (2010) Knowledge-aided radar detector & their application to
live data. IEEE Trans AES 46(1):170–183
7. Merri PS Ph.D. Introduction to GIS, 6 May 2005 ©Micro Images, Inc. 2000–2005
8. Conover WJ (1999) Practical nonparametric statistical, 3rd edn. New York
9. Lawless JF (2003) Statistical models and methods for lifetime data. New York
10. Foglia G (2006) Knowledge-based adaptive radar detection, Doctor Thesis, Universita Degli
Studi Di Cassino Scuola Di Dottorato in Ingegneria, Italy
11. Magaz B, Belouchrani A, Hamadouche M (2011) A new adaptive linear combined CFAR
detection in presence of interfering targets. Prog Electromagn Res 34:P167–P187
12. Cole EL, De Cesare PA, Martineau MJ, Baker RS, Buswell SM (1998) A next generation solid
state air traffic control Radar. In IEEE Radar conference, Dallas, Tx, 12–13 May 1998
13. Tigersharc embedded processor data sheet. analog devices, 2003
Comparative Analysis and Simulation
of Propulsion and Energy Systems
for Satellite Communication
Jalal Eddine Benmansour, Boulanouar Khouane, and Tarek Badis
Abstract The most common problem in the operation of satellite communication

especially in geostationary equatorial orbit (GEO) is to guide it and position it at the
desired orbit with the minimum quantity of propellant, wherein the fuel consump-
tion is highly important parameter that limits the lifetime of the GEO satellite. The
present paper is concerned with the orbital transfer of a satellite to its final orbit. The
comparative study has been made by using electric propulsion compared to conven-
tional chemical propulsion. The numerical simulations are then drawn to study the
performance of two propulsion devices for transfer orbit, which vary in their speed
of transfer and energy efficiency.
Keywords Satellite communication · Orbit transfer · Low and high thrust
1 Introduction
Geosynchronous orbit satellite having a special role in field of communication, broad-

casting and data applications. Since the beginning of the space era, there are hundreds
of satellites orbiting in geosynchronous orbit. The strategy to deliver a satellite in
GEO is relied in propulsion system design. There are many propulsions systems,
which can be classified in two basic categories: Chemical and Electrical propulsion
Systems. However, the difference between these two is a high power propulsion
J. E. Benmansour (B) · T. Badis

Département de recherche en mécanique spatiale, Centre de développement des satellites, Oran,
Algeria
T. Badis
B. Khouane
Département Mission et Systèmes Spatiaux, Centre de développement des satellites, Oran, Algeria

https://doi.org/10.1007/978-981-15-6403-1_73
1062 J. E. Benmansour et al.
system in shorter periods [1]. Whither the transfer time to reach the GEO with chem-
ical propulsion belt ranges from few hours to few days while the electrical propulsion
can reach months [2].
This paper focus on the orbit transfer strategy of geosynchronous satellite. First,
some differences between thrust satellites that is limited by the power available will
be discussed. Then, a multiple Hohmann transfers for mission planning in terms
of propellant consumption to design a mission scenario of GEO satellite will be
presented.
Finally, Numerical simulations demonstrate the validity of transfer methods
designed for a GEO satellite eventually launched from Hammaghir “Algeria”.
2 Propulsion System Design
In order to accelerate spacecraft and artificial satellites in space, there exits two
solutions system design Chemical Propulsion (CP) and Electric Propulsion (EP).
A. Chemical Propulsion System Design

(1) Propellant Consumption
The energy needed in the chemical propulsion design for accelerating the satellite is
provide by the propellant itself in which the change of the velocity of the satellite is
then [3–5].

md + m p
V = Isp · g ln (1)
md
where Isp is the specific impulsion, g is the acceleration by gravity, m d is the delivered
mass at any time (the total mass of the satellite), and m p is the propellant mass on
the satellite.
Using Eq. (1), the required propellant mass can be written as:

m p = m d eV /(Vex ) − 1 = m d eV /( Isp ∗g) − 1 (2)
where Vex is the propellant exhaust velocity.

Therefore, the propellant mass consumption equation within each V maneuver
can be presented as:

−V
m = m 1 − exp (3)
g Isp
The velocity change V needed to obtain an inclination change i is obtained

from as:
Comparative Analysis and Simulation of Propulsion and Energy … 1063
Fig. 1 Hohmann transfer

orbit scheme
V = (V12 + V22 − 2V1 V2 cos i)1/ 2 (4)
(2) Hohmann Transfer Orbit
The Hohmann transfer is the two-pulse maneuver that consumes the least possible
energy for transfer between two coplanar circular orbits that have a common focus.
As shown in Fig. 1, the Hohmann transfer is an elliptical orbit tangent to the two
circles on his apse line, the apoapsis and the perispace of the transfer ellipse are
the radii of the outer and inner circles, respectively [6]. Evidently, only half of the
ellipse is flown during the maneuver, which can occur in both directions, from the
inner circle to the outer circle, or vice versa.
Leaving from A on the inner circle, an increasing V A in the direction of the
flight is necessary for the satellite to reach the elliptical trajectory at a higher energy.
After grounding from A to B, another speed increment before the VB places the
satellite on the outer circular orbit, an even higher energy. Obviously without this
last delta-v, the spacecraft would remain on the Hohmann transfer ellipse and return
to A. Therefore The total energy expenditure is reflected in the total requirement in
delta-V, such as Vtotal = V A + VB .
The eccentricity of an elliptical orbit can be calculated from its radius to periapsis
r p and its radius to apoapsis ra by [4, 6]:
ra − r p
e= (5)
ra + r p
The semi-major axis can be obtained as:
1
a= (ra + rb ) (6)
2
The radius to periapsis is given by:
h2 1
rp = (7)
μ 1+e
where h is the angular momentum and μ is the Gravitational parameter.

Combining these last two equations yields:
h2 1
rp =
μ 1 + rra −r
+r
p
a p
h ra + r p
2
= (8)
μ 2ra
Hence, the angular momentum h can be obtained as

ra r p
h= 2μ (9)
ra + r p
Thus, according to (9) it is useful formula for analysing Hohmann transfers, Note
that for circular orbits with a radius r, the angular momentum can be simplified as
√
h= μr (10)
Consequently, the velocity can be computed in the following form:

1 1
V = 2μ − (11)
r 2a
B. Electrical propulsion system design
Compared to chemical propulsion, it is clear that electric propulsion can carry larger
payloads for space missions. However, this device must thrust for a longer period T to
produce a desired change velocity (V ). The transfer time for low thrust propulsion
is on the order of days and sometimes months.
The transfer time t f of the satellite is computed from the propellant mass and the
constant mass-flow rate
mp
tbur n = (12)
ṁ
It is clear that the spacecraft mass increases, so does the amount of propellant
needed for the same transfer.
We can compute the necessary input power as [7].
2
ṁ Isp g0
P= (13)
2η
where η is the thruster efficiency

As defined in [8], the EP is the gas acceleration to generate thrust via electric
heating, electric body forces, or both electric and magnetic body forces.
The greatest advantage of the electric propulsion is the very high exhaust veloc-
ities, resulting in a reduction of the fuel consumption. Although electric propulsion
has low thrust, the high specific impulse guarantees a reduced burned fuel along the
maneuver.
To compute the lifetime of the satellite (the payload mass), a satellite mass budget
is compute similar to the one used in [4, 9].
The orbit period of two satellites orbiting each other in a circular or elliptic orbit
can be compute from the Kepler’s third law as the following formula [10]:

a3
T = 2π (14)
μ
3 Numerical Simulation
The following results help explain the proposed GTO design procedure which is
used to the transfer orbit for a geostationary satellite launched from Hammaghir.
“Bechar”, the simulations were carried out by Matlab-Simulink.
• Injection Orbit:
Perigee Altitude: 200 km
Apogee Altitude: 36000 km
Inclination 30.9°
• GEO Orbit:
Radius 42164 km
Inclination 0°
• Number of Maneuvers: 3 (2 intermediate orbits)
For our propellant mass calculations, assume that the thrusters have a specific
impulse Isp of 290 s and the gravity of earth is g = 9.81 m s2 .
• The satellite mass is Ms = 1000 kg and the propellant mass is m p = 600 kg.
According to the equations presented above, the Tables 1 and 2 is presented as
follow:
According to the results obtained from Table 2, the propellant mass consumption
depends on the satellite mass and the Isp of the thrusters used.
It can be seen that the total mass consumed is 457.97 kg which represent 76% of
the whole propellant mass used in satellite.
Table 1 Orbital design

Orbit r p (Km) ra (Km) I (Deg) a (Km) e Time (h)
Transfer orbit 6578 42378 30.9 24478 0.73 5.29
1st intermediate orbit 17307 42373 12.87 29840 0.42 14.24
2nd intermediate orbit 30061 42375 6.43 36218 0.17 19.05
Final orbit 42378 42378 0.00 42378 0.00 24.11
Table 2 Maneuvers
V i (Km/s) m i (Kg) V i (m/s)
parameters
Transfer orbit 1.59 / 1477
1st intermediate orbit 2.33 238.32 773.7
2nd intermediate orbit 2.79 132.00 540.9
Final orbit 3.07 87.65 426
Figures 2 and 3 show the orbit maneuver during GTO mission where after the
transfer orbit only two intermediate orbit have been needed to reach the target orbit
with a total time of 62.69 h.
The time responses of various parameters during the electric transfer phase is
shown in Figs. 4, 5 and 6. Figure 4 shows the evolution of inclination from a start
value 30.9° to the target value of the 0°. As can be seen the target inclination can be
reached in 66 days.
In Fig. 5, it is clear that the velocities of geostationary orbit has been achieved
which is 3.07 × 103 ms−1 that is always over the same point on the earth. As shown
in Fig. 6 the semi-major axis has been successfully increased to the desired value
(42164 km).
Hohmann Transfer: Orbital Maneuver

1.5
1
Z coordinate (ER)
0.5
-0.5
-1
-6 -4 -2 0 2 4 6
Y coordinate (ER)
Fig. 2 Transfer orbit simulation results “Plan Y-Z

Hohmann Transfer: Orbital Maneuver

6
4
Y coordinate (ER)
-2
-4
-6
-8 -6 -4 -2 0 2 4 6 8
X coordinate (ER)
Fig. 3 Transfer orbit simulation results “Plan X-Y”
35
30
Inclination (degrees)
25
20
15
10
0
0 10 20 30 40 50 60 70
Time (days)
Fig. 4 Time response of inclination
According to the simulation result in Fig. 7, it can be deduced that for the electrical
propulsion many revolutions of intermediate are needed in order to reach the desired
orbit. with computed time of 74 days.
7.5
6.5
Velocity (km/s)
5.5
4.5
3.5
3
0 10 20 30 40 50 60 70
Time (days)
Fig. 5 Time response of velocity
10 4
4.5
4
Semi-major axis (km)
3.5
2.5
1.5
0.5
0 10 20 30 40 50 60 70
Time (days)
Fig. 6 Time response of semi-major axis
4 Conclusion
Both chemical propulsion described by Hohmann transfer and electric propulsion

have advantages and disadvantages. For the low thrust trajectory that use electric
propulsion generally require a larger value of delta V than Hohmann trajectories,
but require less propellant mass owing to the much greater specific impulse of the
engine. Therefore, the low-thrust propulsion generally will be the best solution for a
mass constrained mission.
Further research is required to consider design of a hybrid chemical/electric
propulsion orbital transfer satellite.
Fig. 7 Spacecraft trajectory from LEO to GEO
References
1. Bate RR, Mueller DD, White JE (1971) Fundamentals of astrodynamics. Courier Corporation
2. Wertz JR, Everett DF, Puschell JJ (2011) Space Mission Engineering the New SMAD.
Microcosm Press
3. Koppel CR (1999) Advantages of a continuous thrust strategy from a geosynchronous transfer
orbit, using high specific impulse thrusters. In 14th international symposium on space flight
dynamics–ISSFD XIV, February
4. Sidi MJ (1997) Spacecraft dynamics and control. Cambridge University Press, New York, NY,
USA
5. Hernandez E, Bolivar J, Wang Q (2011) Geosynchronous transfer orbit design: a practical
approach. In Modeling, simulation and applied optimization (ICMSAO), 2011 4th international
conference on, IEEE, pp 1–4
6. Curtis HD (2013) Orbital mechanics for engineering students. Butterworth-Heinemann
7. Alan J (2004) Representative mission trade studies for low-thrust transfers to geosynchronous
orbit. In AIAA/AAS astrodynamics specialist conference and exhibit, p 5086
8. Starinova OL, Gorbunova IV, Fain MK, Khabibullin RM, Shornikov AY. Application of electric
propulsion for motion control of spacecraft which function on non-keplerian orbits. Procedia
Eng 185:291–298
9. Thomas U, Sellmaier F, Schmidhuber M (2015) Spacecraft Operations. Springer Vienna
10. Berlin P (1988) The geostationary applications satellite. No. 2. Cambridge University Press
Narrowband Sensor-Antenna Based
on an Interdigital Capacitor
Rafik Khelladi, Mustapha Djeddou, and Farid Ghanem
Abstract This paper proposes a design of narrowband sensor-antenna based on an

Interdigital Capacitor (IDC) for the dielectric measurement of planar materials. An
interdigital capacitor was etched along the radiating edge of a narrowband rectangular
patch antenna fed by a microstrip transmission line. The sensitivity of the proposed
sensor-antenna was performed by the shift in the operating frequency when planar
dielectric materials were placed above the patch. The prototype has been validated
by experiments.
Keywords Sensor-antenna · Narrowband microstrip antenna · Dielectric constant

(permittivity) measurement
1 Introduction
Accurate measurement of the relative permittivity of different materials plays an

important role for antennas and microwave circuit design [1]. In addition, character-
ization of dielectric materials has lots of applications such as quality control in the
food industry, bio-sensing, subsurface detection, water impurity, gaz sensing… [2,
3]. Various techniques for measuring the dielectric properties have been proposed.
Generally, dielectric measurement methods can be categorized into non-resonant
methods [4] and resonant methods [5]. In non-resonant methods, the relative permit-
tivity of dielectric materials are fundamentally deduced from their characteristic
R. Khelladi (B) · M. Djeddou

Telecommunication Laboratory, Ecole Militaire Polytechnique, Algiers, Algeria
M. Djeddou
F. Ghanem
Telecom Product Direction, R&D&I Brandt Group, Cevital Industry Pole, Algiers, Algeria

https://doi.org/10.1007/978-981-15-6403-1_74
1072 R. Khelladi et al.
impedance and wave velocities of EM waves, measured by the reflection and trans-
mission characteristics [1, 6]. In the resonate methods, the permittivity can be derived
from the shift in resonance frequency and factor quality change of resonators. Reso-
nant methods generally have higher accuracies than non-resonant methods and are
good for low-loss materials [7].
This paper studies the design and implementation of a new prototype for
measuring the relative permittivity of planar materials. The design consists of a
narrowband rectangular antenna based on an IDC etched along the radiating edge
of the patch used for permittivity measurement. Through measuring the resonant
frequency of the input reflection coefficient (S 11 ), the permittivity can be determined
by the proposed sensor-antenna when a dielectric Material Under Test (MUT) is
placed above the IDC area.
2 Sensor-Antenna Design
2.1 Narrowband Microstrip Patch Antenna
In order to have an accurate data measurement, a microstrip patch antenna with very
low bandwidth is adopted in our design.
Figure 1 shows the geometry of the proposed narrowband antenna. It consists of
a simple inset-fed rectangular microstrip patch antenna. The patch is designed on a
Fig. 1 Layout of the designed narrowband patch antenna

Narrowbad Sensor-Antenna Based on an Interdigital Capacitor 1073
Table 1 Parameters of the narrowband patch antenna

Ws Wp Wg Wf Win
49.4 49.4 2.4 2 2
Ls Lp Lf Lin
39.2 38.4 5.5 8.8
Fig. 2 S11 parameter of the

designed narrowband
antenna
Teflon substrate with relative dielectric constant of E = 2.57 and tan δ = 0.0018. The
thickness of the substrate is h = 0.8 mm.
The patch parameters shown in Table 1 are determined by using the transmission
line model [8]. The units of all parameters are given in millimeter.
Lp and Wp are the length and width of the patch, respectively. Lf and Wf are
the length and width of the microstrip feed line. An inset feed with the dimensions
of Lin x Win has been created between the transmission feed line and the patch in
order to improve the impedance match. The design has been simulated by using the
transiet solver in CST Microwave Studio. The boundary condition has been set to
open add space and the port used is waveguide port. The simulated return loss S 11
of the designed patch antenna is illustrated in Fig. 2.
From Fig. 2, the patch antenna is operating at the frequency of 2.4 GHz. It
provides at 10 dB return loss a very narrow bandwidth of 11.36 MHz with a fractional
bandwidth of 0.47%. The realized gain of the antenna is 5.18 dB.
2.2 Interdigital-Capacitor-Loaded Microstrip Patch Antenna
Figure 3 presents the layout of the designed sensor-antenna based on an IDC.

Fig. 3 The layout of the

proposed sensor antenna
a IDC-loaded patch antenna,
b parameters of the IDC
The geometric parameters of the proposed sensor-antenna were optimized for

higher sensitivity through a parametric study. Note that the lengths Lp and Ls are set
to be 37 mm and 42.6 mm, respectiveley. The parameters of the IDC are illustrated
in Table 2.
Figure 4 shows the electric field distribution of the proposed IDC-loaded
microstrip sensor-antenna. As illustrated in Fig. 4, the electric field distributions
are concentrated in the IDC area which is mainly used for permittivity sensing.
According to the MUT permittivity, the S 11 characteristic of the sensor-antenna
is compared as shown in Fig. 5. The MUT was placed above the IDC, and its relative
Table 2 Parameters of the IDC

W’ W L G S
14.7 0.7 9.1 0.1 0.3
Fig. 4 Electric field

distribution at 2.4 GHz
Fig. 5 Simulated S11

parameter of the
sensor-antenna according to
the relatuve permittivity of
the MTU
permittivity (E) was varied from 2 to 10. The dimensions of the MUT are the same
of those of the IDC and its thickness is chosen to be 1.6 mm.
As depicted in Fig. 5, when the relative permittivity (E) of the MUT changes
from 2 to 10, a shift in the operating frequency of the sensor-antenna is occurred
from 2.4368 GHz to 2.3675 GHz. This is can be explained as follows; the change
of relative permittivity of the MUT which is loaded above the IDC, affects the total
capacitance and effective relative permittivity related to the microstrip patch antenna,
and hence the operating frequency of the sensor-antenna.
The sensitivity performance (S) of the designed sensor-antenna is presented in
Fig. 6. It is defined as the shift of the resonant frequency compared to the unloaded
case, over respective permittivity variation of the MUT [9].
Fig. 6 Sensitivity of the

designed sensor-antenna

fr fr,unloaded − fr,loaded
S= =
εr εr,unloaded − εr,loaded
It can be seen from Fig. 6 that S of the three first resonant frequencies is about
0.0096 GHz. Then it decreases up to S = 0.008 GHz for the relative permittivity
ranging from 5 to 10.
3 Experiment Results and Discussion
A prototype of the designed sensor-antenna described in the previous section has been
fabricated, as shown in Fig. 7. The return loss measurements were performed using
an Agilent HP8719ES vector network analyzer. Two dielectric MUTs have been
tested. The MUTs are: the FR-4 substrate (E = 4.3, tan δ = 0.0018, h = 1.6 mm)
and the Teflon substrate (E = 2.57, tan δ = 0.0018, h = 0.8 mm).
The measured S11 characteristics of the fabricated sensor-antenna when the two
MUTs are placed on the patch and for the unloaded case, are illustrated in Fig. 8.
The behavior of the proposed sensor-antenna could be observed from the
frequency response. From the experiment results of Fig. 8, it is shown that the reso-
nant frequency of the sensor-antenna is orderly modified with different loaded MUTs.
Indeed, the resonant frequency of the proposed sensor-antenna for the unloaded case
is 2.427 GHz. This resonant frequency moves from 2.411 GHz for the Teflon MUT
(E = 2.57) to 2.397 GHz for the FR-4 MUT (E = 4.3).
In order to compare the simulated and measured responses of the proposed sensor-
antenna, the return loss S11 parameters have been simulated with the two used MUTs
(Teflon and FR-4) placed above the IDC area. The results of the comparison are given
in Figs. 9, 10 and 11.
Fig. 7 Photograph of the

fabricated sensor-antenna
based on an IDC
Fig. 8 Measured S11

parameter for the MUTs
A good agreement can be noted with a small shift between the simulated and
measured results. This difference is mainly due to the implementation imperfection,
the cutting precision of the fabricated antenna which makes difference in dimensions
compared to the simulated one. Finally, the relative permittivity related to both the
designed antenna and the used MUTs could be slightly different from the one reported
in the substrates data sheet.
Fig. 9 Simulated/Measured
S11 parameter for the
unloaded case
S11 parameter for the Teflon
MUT
S11 parameter for the FR-4
MUT
4 Conclusion
In this paper, we have proposed a design of sensor-antenna base on an IDC capable of

measuring the relative permittivity of planar dielectric materials. To demonstrate the
concept, a prototype of the sensor-antenna has been fabricated, and measurements
have been performed by using two different MUTs placed above the patch. Reason-
able agreement between the simulated and measured results could be noticed, which
validates the proposed concept.
References
1. Saeed K, Shafique MF, Byrne MB, Hunter C (2012) Planar microwave sensors for complex
permittivity characterization of materials and their applications. Appl Measur Syst:319–350
2. Bobay MS, Ramahi OM (2012) Material characterization using complementary split-ring
resonators. IEEE Trans Instrum Measur 61(11):3039–3046
3. Saghlatoon H, Mirzavand R, Honari MM, Pedram M (2018) Sensor antenna transmitter system
for material detection in wireless-sensor-node applications. IEEE Sens J 18(21):8812–8819
4. Caijun Z, Quanxing J, Shenhui J (2011) Calibration-independent and position-insensitive trans-
mission/reflection method for permittivity measurement with one sample in coaxial line. IEEE
Trans Electromagn Compat 53(3):684–689
5. Santra M, Limaye KU (2005) Estimation of complex permittivity of arbitrary shape and size
dielectric samples using cavity measurement technique at microwave frequencies. IEEE Trans
Microw Theory Tech 53(2):718–722
6. Yeo J, Lee J (2019) High-sensitivity microwave sensor based on an interdigital-capacitor-shaped
defected ground structure for permittivity characterization. Sensors 19(3):498
7. Jiang J, Geyi W (2014) Development of a new prototype system for measuring the permittivity
of dielectric materials. J Eng 2014(6):302–304
8. Balanis CA (2005) Antenna theory: analysis and design. 3rd ed. Wiley and Sons Inc.
9. Yeo J, Lee J (2019) Slot-loaded microstrip patch sensor antenna for high-sensitivity permittivity
characterization. Electronics 8(5):502
Secure Color Image Transmission Based
on the Impulsive Synchronization
of Fractional-Order Chaotic Maps Over
a Single Channel
Ouerdia Megherbi, Hamid Hamiche, Saïd Djennoune,

and Maamar Bettayeb
Abstract In this paper, we propose a novel single-channel secure transmission

scheme based on the impulsive synchronization of fractional-order discrete chaotic
systems. A discrete fractional-order impulsive observer has been developed at the
receiver level in order to reconstruct the states of a fractional-order Hénon map
used at the transmitter level. Robust chaos-based encryption and decryption algo-
rithms have been used to cipher an original color image and retrieve it as soon as the
synchronization is established between the transmitter and the receiver.
Keywords Chaos-based transmission · Synchronization of chaotic systems ·

Fractional-order maps · Impulsive synchronization
1 Introduction
Over the last decades, image encryption has attracted more attention due to the various
applications of image data over social networks and via internet. Indeed, images are
the most frequently shared, sent and downloaded media files. This is why, their
confidentiality against undesirable intrusions becomes the crucial objective when
they are sent.
O. Megherbi (B) · H. Hamiche · S. Djennoune

Laboratoire de Conception et Conduite des Systèmes de Production (L2CSP), Université de
Mouloud Mammeri (UMMTO), Tizi-Ouzou, Algeria
H. Hamiche
S. Djennoune
M. Bettayeb
Department of Electrical and Computer Engineering, University of Sharjah, Sharjah, UAE

https://doi.org/10.1007/978-981-15-6403-1_75
1082 O. Megherbi et al.
In the last years, in order to secure images transmission, chaotic systems have
been widely included in encryption algorithms [1]. In fact, the sensitivity to initial
conditions and control parameters as well as the random like behavior of the chaotic
systems are so interesting to use them in cipher algorithms [2, 3]. Another attractive
idea that has fast gained ground more recently is the use of fractional-order chaotic
systems instead of the integer-order ones into communication schemes. It has been
noticed, indeed, that fractional-order chaotic systems have brought better complexity
and robustness to encryption algorithms [4]. Furthermore, fractional derivative orders
can be considered as additional secret keys and thus make the key space larger [5].
For the purpose of using chaotic systems, either with integer or fractional orders,
for data transmission, the chaotic generator used for masking data at the transmitter
level, should be available at the receiver side. Synchronization between the trans-
mitter and receiver should then be established. Since the first and pioneering works
of Pecora and Carroll [6], many synchronization methods have been presented in
the literature and extended later to synchronize fractional-order chaotic systems [7].
Among them, observer-based synchronization approaches are increasingly studied
and developed. In fact, according to Nijmeijer and Mareels research works [8],
synchronization problem can be considered as an observer design one. As results,
more and more observers have been proposed to synchronize fractional order chaotic
systems. For instance, we can cite the fractional-order delayed step by step observers
[9], the fractional-order sliding-mode observers [10] and also the fractional-order
impulsive observers we choose to use in this paper. In fact, the design of impulsive
observers to estimate the states of a chaotic system at discrete instants is one of the
methods that offer many advantages especially for secure transmission applications
[11].
In this paper, a novel transmission scheme has been proposed on the basis of
synchronization of fractional-order discrete-time chaotic systems using fractional-
order impulsive observers. The rest of the paper is organized as follows. In Sect. 2, we
present some preliminaries on fractional-order discrete chaotic systems. Section 3
is devoted to describe in more details the proposed image transmission scheme. The
simulation results will be provided in Sect. 4. In order to verify the robustness of
our scheme, some tests are defined and applied in Sect. 5. This paper will be ended
by a summarizing conclusion and some ideas to enhance the communication system
presented in this paper.
2 Preliminaries on Fractional-Order Discrete-Time

Systems
In this section, we give some basic definitions and concepts on fractional-order

discrete systems. In order to model fractional-order chaotic systems, many definitions
can be used. In our case, we use the following Grünwald-Letnikov definition:
Secure Color Image Transmission Based on the Impulsive … 1083
Définition The α-order Grünwald -Letnikov difference for a given state vector x is
as follows [12]:

k
α
α x(t) = (−1) j x(k − j) (1)
j
j=0
where α > 0 denotes the fractional order, x(k) = [x1 (k)x2 (k) · · · xn (k)]T ∈ Rn is
the state vector
atthe current time k and x (k-j) is the delayed state vector.
α
The term can be computed from the following relation:
j

1 f or j = 0
α(α−1)···(α− j+1) (2)
j!
f or j > 0
After some calculus and substitutions, we obtain the following conventional

representation of fractional-order discrete-time systems

L
x(k + 1) = f (x(k)) + (α − 1)x(k) − C p x(k − p) (3)
p=1

α
with C p = (−1) p+1
.
p+1
The constant L is considered as the memory size of the system. It is chosen
according to the application. Furthermore, note that (3) is obtained in the commen-
surate case since the differentiation order α is supposed the same for all the state
variables. Otherwise, in the non commensurate case, we should consider all the
derivative orders α1 , α2 , …, αn of the system states [13]. Equation (3) becomes then

L
x(k + 1) = Ax(k) + ϕ(x(k)) − Ad ( p)x(k − p) (4)
p=1
where n × n matrix A describes the linear non delayed part of the system and the
vector ϕ(x(k)) denotes the nonlinear terms. Ad ( p) are diagonal n ×n matrices which
contain the coefficients C p presented previously with respect to the derivative orders,
namely:
⎡ ⎤
C p1 0 ··· 0
⎢ ... 0 ⎥
⎢ 0 C p2 ⎥
Ad ( p) = ⎢
⎢ .. .. .. .. ⎥
⎣ . . . . ⎥ ⎦
0 0 · · · C pn
3 New Image Transmission Based on the Impulsive

Synchronization of Fractional-Order Chaotic Maps
Before developing our transmission scheme, we recall at first some results on the
impulsive synchronization between fractional-order chaotic systems by mean of
impulsive observers. As mentioned in the introduction, the impulsive synchroniza-
tion consists of the use of samples of states of a master system to force the states
of another system called slave system to have the same evolution as those of the
master system. In our case, an impulsive observer is proposed to estimate the states
of System (4).
The proposed fractional-order observer is defined by the following equations:
⎧
⎪
⎪ x̂(k + 1) = A x̂(k) + ϕ(x̂(k)
⎪
⎪
⎨ − A ( p)x̂(k − p)
⎪ f or k = ki
L
d
(5)
⎪
⎪
p=1
⎪
⎪ + f or k = ki ,
⎪
⎩ x̂ ki = x̂(ki ) − B(ki )e(ki ),
i = 1, 2, · · · i = 1, 2, · · ·
where x̂(k) ∈ Rn is the estimated state vector and e(k) = x(k) − x̂(k) is the error
vector.
ki is a subset of discrete-time {k} chosen so that 0 < 1 = k1 < k2 < · · · < ki <
· · · , with ki → ∞ as i → ∞. It denotes, in fact, the instants at which the samples of
the master system’s states are sent to the response system. In this paper, we assume
that the distance between two successive instants ki is constant and we denote it by
d. Thus, d = ki+1 − ki , i = 1, 2, · · · , ki+ denotes the instants just after ki . Matrices
B(ki ) are symmetric and contain control factors applied to the observer sates at the
instant ki [14].
The synchronization error between systems (4) and (5) is defined as follows:
⎧
⎪
⎪ e(k + 1) = Ae(k) + ϕ̃ x(k), x̂(k)
⎪
⎪
⎪
⎨ L
f or k = ki
− Ad ( p)e(k − p)
(6)
⎪
⎪
p=1
⎪
⎪ f or k = ki ,
⎪
⎩ e(k) = B(ki )e(ki ),
i = 1, 2, · · ·

where e(k) = [e1 (k)e2 (k) . . . en (k)]T is the vector error and ϕ̃ x(k), x̂(k) =
ϕ(x(k)) − ϕ x̂(k) .
To establish impulsive synchronization, it is necessary to study the stability of
system (6). This is why the theorem given in [13] has been used.
Now, we present the transmission scheme using the fractional-order chaotic map
and its impulsive observer to transmit color image. The communication system is
composed of two main blocks: a transmitter and a receiver. In the following, we
describe in more details the content of each block.
Fig. 1 Block diagram of the transmitter
A. Transmitter structure
The transmitter contains three basic elements: The fractional-order chaotic map, an
encryption algorithm to encrypt the original image and a multiplexer. Figure 1 shows
the block diagram of the transmitter side. In the following, we further develop the
role of each element.
• The fractional-order chaotic system
The fractional-order chaotic map used in this paper is the fractional-order discrete
Hénon map. It is used to generate more complex chaotic sequences for an encryption
purpose [15].
The fractional-order Hénon map is obtained using the Grünwald-Letnikov
difference, and is given by the following equations:
⎧
⎪
⎪ x (k + 1) = 1 − ax12 (k) + x2 (k) + (α1 − 1)x1 (k)
⎪ 1
⎪
⎪
⎪ L
⎪
⎨ − C p1 x1 (k − p)
p=1
(7)
⎪
⎪ x2 (k + 1) = bx1 (k) + (α2 − 1)x2 (k)
⎪
⎪
⎪
⎪ L
⎪
⎩ − C p2 x2 (k − p)
p=1
where [x1 x2 ]T ∈ R2 is the state vector, y is the output and a, b are two constant
parameters. 0 ≤ α1 ≤ 1 and 0 ≤ α2 ≤ 1 are the derivative orders and coefficients
C p are defined as follows:

α1 α2
C p1 = (−1) p+1
, C p2 = (−1) p+1
.
p+1 p+1
For the parameters values a = 1.4 and b = 0.3 and by choosing the derivative
orders α1 = 0.997, α2 = 0.998 and the memory size L = 250, the fractional-order
Fig. 2 Chaotic attractor of the fractional-order Hénon map
Hénon map presents a chaotic behavior. The strange attractor obtained on the phase
plan of states x2 versus x1 is illustrated on Fig. 2.
System (7) can be rewritten by using the state representation as follows

L
x(k + 1) = Ax(k) + ϕ(x(k)) − Ad ( p)x(k − p) (8)
p=1
where matrices A, Ad and the nonlinear vector ϕ(x(k)) are defined as

α1 − 1 1 C p1 0
A= , Ad ( p) =
b α2 − 1 0 C p2

1 − ax12 (k)
ϕ(x(k)) =
0
The system output denoted y(k) can be defined in the following way

x1 (ki ) i f k = ki
y(k) = (9)
0 i f k = ki
• The encryption algorithm
Many operations have been performed to encrypt the color image before sending it
on the channel. They are described in the following steps.
Step 1: The plain image noted I is represented by a M × N × P matrix where
M is the number of rows, N the number of columns and P the number of color
channels and is equal to 3. The obtained matrix is reshaped to get a set of decimal
values

V = V1,1,1 , · · · , VM,N ,1 , V1,1,2 , · · · VM,N ,2 , V1,1,3 , · · · , VM,N ,3 .
Step 2: Two chaotic sequences Z 1 and Z 2 are generated from the chaotic states x1
and x2 of the fractional-order Hénon map and using two nonlinear functions f 1
and f 2 .
Step 3: The chaotic sequences Z 1 and Z 2 are processed using the ceil and mod

operations, to generate two new sets of positive integer elements Z 1 and Z 2 .

Step 4: The positive integer sequences Z 1 , Z 2 and the sequence V corresponding
to the plain image are converted to binary sequences S B1 and S B2 and VB .
Step 5: XOR operation is performed between the bits of the chaotic sequences
and the bits of the plain image as follows

C B = S B1 ⊕ S B2 ⊕ VB
Step 6: The bits of the binary matrix C B are flipped.

Step 7: The previously flipped binary matrix is converted to a decimal sequence
Dc whose elements, once arranged appropriately, represent the pixels of the cipher
image IC .
• The multiplexer
As we use a single channel to transmit both the synchronization signal y and the
cipher image sequence Dc , a multiplexer (Mux on the bloc diagram) should be used
to select at any instant, one of the two signals and forward it into the public channel.
Thus, when the address input of the multiplexer is A0 (k) = 1, the output y(k) of
the fractional-order Hénon map will be sent to achieve impulsive synchronization,
otherwise, i.e. if A0 (k) = 0, we send the sequence Dc corresponding to the encrypted
image.
B. Receiver structure
As shown on Fig. 3, the receiver bloc is also composed of three elements: the
demultiplexer noted Demux, the fractional order impulsive observer and a decryption
algorithm. The role of each bloc will be described below.
• The demultiplexer
The demultiplexer is used to send the input signal of the receiver to the decryption
bloc and the fractional-order impulsive observer according to the value of the address
input A0 which is defined in the same way as for the multiplexer.
• The discrete fractional-order impulsive observer
In order, to recover the original image I, states x1 and x2 should be available at the
receiver side. This is why, we use the following fractional-order impulsive observer
to estimate the two states of the fractional-order Hénon map.
Fig. 3 Block diagram of the receiver
⎧
⎪ x̂(k + 1) = A x̂(k) + ϕ x̂(k)
⎪
⎨ L
i f k = ki
− Ad ( p)x̂(k − p) (10)
⎪
⎪
⎩ + p=1
x̂ ki = x̂(ki ) − B(ki )e(ki ) i f k = ki
T
where x̂ = x̂1 x̂2 ∈ R2 is the estimated states vector. Matrices A and Ad are the
same as for the master system (8).
B(ki ) ∈ R2×2 are chosen equal in our case. Since
we send only samples of the
b1 0
state x1 , they are defined as follows: B = .
0 0
• The decryption algorithm
In order to retrieve the plain image, we should perform the inverse operations
mentioned in the encryption algorithm from the latter step to the first one and use
the estimated chaotic states x̂1 and x̂2 .
We provide, in this section, some results of the synchronization and image transmis-
sion process obtained by using the fractional impulsive observer and the proposed
encryption algorithm.
We provide at first, some impulsive synchronization results obtained by simulating
the proposed synchronization scheme on MATLAB. The initial conditions are chosen
equal to (x1 (0), x2 (0)) = (0.8, −0.5) for the master fractional-order Hénon map and
(x̂1 (0), x̂2 (0)) = (1, −3) for the impulsive observer. The control gain b1 is fixed to
b1 = −99.98 × 10−2 . The memory size chosen while modelling the fractional-order
Hénon map and its observer is L = 250. Note that, these values are chosen so as
to ensure the stability and the convergence of the error system (6). Furthermore, we
choose the period d = 3. Figures 4 and 5 show respectively the reconstruction of
states x1 and x2 by the impulsive observer. We can see that after a short time, the
states x1 and x2 have been estimated successfully. In addition, the synchronization
errors illustrated on Fig. 6 converge after a short number of iterations and confirm
the synchronization establishment.
Fig. 4 Synchronized states x1 and x̂1
Fig. 5 Synchronized states x2 and x̂2

Fig. 6 Synchronization errors e1 and e2
Fig. 7 Transmission of Lena image: a Lena plain image, b The encrypted image, c The recovered
image
Lena color image of size 220 × 220 × 3 has been transmitted using the proposed
scheme. Figure 7 illustrates both the plain image, the encrypted image and the
recovered image at the receiver level using the proposed algorithm. We can see
from the obtained results that the whole image has been recovered thanks to the
fractional-order impulsive observer.
However, these results are not sufficient, and a security analysis should be
achieved. The robustness analysis will be described in the following section.
5 Robustness Analysis
To check the robustness of the proposed transmission scheme, some security tests
have been applied to the proposed encryption algorithm. The analysis principle and
the obtained results will be given below.
Fig. 8 Lena plain image and the corresponding color channels histograms
A. Histogram Analysis
The image histogram is a powerful tool for analyzing the distribution of pixels for
each intensity level. From Fig. 9, we can notice that the histograms of the encrypted
image corresponding to the three color channels are uniform and different from those
of the plain image illustrated on Fig. 8. Hence, the encryption algorithm is robust
against statistical attacks based on histogram analysis [9].
B. Correlation analysis
The correlation can be defined as the relation between adjacent pixels in a given
image. In the plain image, the correlation is high between a pixel and its neigh-
bors in horizontal, vertical and diagonal directions [16]. On the other hand, an effi-
cient encryption algorithm should reduce the correlation between adjacent pixels. In
order to test the pixels correlation in the cipher image, the correlation coefficient is
computed for each color channel among a number of adjacent pixels. The correlation
Fig. 9 Lena cipher image and the corresponding color channels histograms
Table 1 Correlation coefficients of Lena image before and after encryption

Direction Plain image Encrypted image
R G B R G B
Horizontal 0.9373 0.9349 0.8778 0.0331 0.0140 -0.0109
Vertical 0.9666 0.9661 0.9389 0.0179 -0.0536 0.0184
Diagonal 0.9098 0.9065 0.8285 0.0200 0.0045 0.0022
coefficient is determined by the following equations:
cov(x, y)
ρx y = √ √ (11)
D(x) D(y)
1 N
where cov(x, y) = i=1 ([x i − E(x)])([yi − E(y)]), D(x) =
1 N 1 N
N
N i=1 (x i − E(x)) 2
and E(x) = N i=1 (x i ). x i and yi are the values of
two adjacent pixels in the plain or encrypted image. The obtained coefficient value
belongs to the interval [0,1]. For high correlated pixels, the coefficient value is close
or equal to 1. Otherwise, the coefficient is close or equal to 0. Table 1 gives the
obtained correlation coefficients before and after encryption of Lena image using
our algorithm. We can see that the pixels correlation in the encrypted image is weak.
Thus, the proposed algorithm has successfully destroyed the correlation between
adjacent pixels.
C. Information entropy
Information entropy is a basic criterion to show the randomness of the encrypted

data. Information entropy of a cipher image can be calculated as follows:
Q−1 1
H (IC ) = Pr o(ci ) × log2 (12)
i=0 pr o(ci )
Pr o(ci ) denotes the probability of each intensity level. Incase of equal probability,
255 1
the information entropy is maximal and equals H (IC ) = i=0 256
× log2 256 = 8.
In our case, information entropy of each color component in the plain and cipher
images are given in Table 2. We can see that the information entropy of the encrypted
image is very close to the ideal value 8. This implies that the proposed algorithm is
robust upon entropy attacks.
D. Key sensitivity
Table 2 Entropy values for Lena image

Image Plain image Encrypted image
R G B R G B
Lena 7.2898 7.5835 7.0720 7.9935 7.9952 7.9947
The key sensitivity analysis defines how a small change δ on the value of one of the
encryption keys can affect the transmission process. A robust encryption algorithm
should be highly sensitive to keys variation in order to resist to differential attacks
[16]. Four secret keys have been used in our algorithm namely the control parameters
a and b of the fractional-order Hénon map and the derivative orders α1 and α2 . The
sensitivity is measured by the NPCR (Number of Pixels Change Rate) and UACI
(Unified Average Changing Intensity) rates defined as follows.

D(i, j, p)
ij
N PC R = × 100% (13)
M×N×p
⎡ ⎤
1 IC1 (i, j, p) − IC2 (i, j, p)
U AC I = ⎣ ⎦ × 100% (14)
M × N × P i, j, p 255
where M, N and P represent respectively the height, the width and the depth of the
cipher image IC . Matrix D elements are computed as follows:

1 i f IC1 (i, j, p) = IC2 (i, j, p)
D(i, j) = (15)
0 i f IC1 (i, j, p) = IC2 (i, j, p)
The optimal values of NPCR and UACI rates are NPCR = 99.6094 et UACI
= 33.4635. Table 3 illustrates the sensitivity of our encryption scheme to the keys
variations. From the obtained values of NPCR and UACI rates, we can conclude that
our encryption algorithm is very sensitive to keys values changes and can successfully
withstand differential attacks.
Table 3 Sensitivity to secrete

Modified key Variation NPCR (%) UACI (%)
keys variations
δ
a 10−15 99.5544 33.3491
b 10−16 99.5241 33.5348
α1 10−16 99.5572 33.5083
α2 10−16 99.5224 33.4514
E. Key space analysis
The key space size is the total number of the different keys values that can be used
in the encryption process.
The key space should be large enough to resist to brute force attacks. Using the
results of the previous analysis, the key space size is computed as follows:

TK = 1016×3+15 ≈ 2209
The obtained value is greater than 2100 which is the minimal size required for
chaos-based encryption schemes to resist against attacks based on the key space
analysis.
6 Conclusion
In this paper, we have proposed a new color image transmission scheme through
a single channel and based on the impulsive synchronization of fractional-order
chaotic Hénon map. The impulsive synchronization has been carried out by the use
of a fractional-order impulsive observer. Once the synchronization conditions filled
and the stability of the synchronization errors reached, the synchronized systems
have been exploited to transmit a color image. The simulation results have shown
that the image has been recovered with success thanks to the synchronization process.
Furthermore, and after several security and robustness tests, we have noticed that our
scheme is secure and robust against many attacks such as the differential attacks, the
statistical attacks and the brute-force attacks.
Several prospects can be considered to improve the present work, the first one is
to enhance the robustness against channel noises. Another idea is to implement the
proposed communication scheme on programmable devices such as Digital Signal
Processors (DSP) or FPGA circuits.
References
1. Hamiche H, Lahdir M, Tahanout M, Djennoune S (2012) Masking digital image using a novel
technique based on a transmission chaotic system and SPIHT coding algorithm. Int J Adv
Comput Sci Appl 3(12):228–234
2. Megherbi O, Guermah S, Hamiche H, Djennoune S, Ghanes M (2013) A novel transmission
scheme based on impulsive synchronization of two Colpitts chaotic systems. In 3rd international
conference on systems and control, pp 117–122
3. Hamiche H (2011) Inversion à gauche des systèmes dynamiques hybrides chaotiques: Appli-
cation à la transmission sécurisée de données , PHD Thesis, University Mouloud Mammeri of
Tizi-Ouzou, Algeria
4. Zouad F, Kemih K, Hamiche H (2019) A new secure communication scheme using fractional
order delayed chaotic system: design and electronics circuit simulation. Analog Integr Circ Sig
Process 99(3):619–632
5. Megherbi O, Kassim S, Hamiche H, Djennoune S, Bettayeb M, Barbot J.-P (2017) Robust
image transmission based on coupled fractional-order chaotic maps. In Siam conference on
control and its applications, USA
6. Pecora LM, Carroll TL (1990) Synchronization in chaotic systems. Phys Rev Lett 64(8):821–
824
7. Ouannas A, Khennaoui A-A, Grassi G, Bendoukha S (2018) On the Q-S chaos synchronization
of fractional-order discrete-time systems: general method and examples. Discret Dyn Nat Soc
2018: 8 pages. Article ID 2950357
8. Nijmeijer H, Mareels IMY (1997) An observer looks at synchronization. IEEE Trans Circuits
Syst I: Fundam Theor Appl 44(10):882–890
9. Kassim S, Hamiche H, Djennoune S, Bettayeb M (2017) A novel secure image transmis-
sion scheme based on synchronization of fractional-order discrete-time hyperchaotic systems.
Nonlinear Dyn 88(4):2473–2489
10. Li Y, Hou B (2018) Observer-based sliding mode synchronization for a class of fractional-
order chaotic neural networks. Adv Differ Equ 2018:146. https://doi.org/10.1186/s13662-018-
1588-9
11. Yang T, Chua LO (1997) Impulsive stabilization for control and synchronization of chaotic
systems: theory and application to secure communication. Trans Circuits Syst I: Fundam Theor
Appl 44(10):976–988
12. Kilbas AA, Srivastava HM, Trujillo JJ (2006) Theory and application of fractional differential
equations. In North Holland Mathematics Studies. Editor Jan van Mill, Elsevier, Amsterdam
13. Megherbi O, Hamiche H, Djennoune S, Bettayeb M (2017) A new contribution for the impulsive
synchronization of fractional-order discrete-time chaotic systems. Nonlinear Dyn 90(3):1519–
1533
14. Hamiche H, Megherbi O, Kara R, Saddaoui R, Laghrouche M, Djennoune S (2017) A new
implementation of an impulsive synchronization of two discrete-time hyperchaotic systems
using Arduino-Uno boards. Int J Modell, Identif Control 28(2):177–186
15. Huang L, Wang L, Shi D (2016) Discrete Fractional order chaotic systems synchronization
based on the variable structure control with a new discrete reaching-law. IEEE/CAA J Autom
Sinica
16. Hamiche H, Kassim S, Megherbi O, Djennoune S, Bettayeb M (2018) Secure digital data
communication based on fractional-order chaotic maps. In Advanced synchronization control
and bifurcation of chaotic fractional-order systems, pp 438–467
Performance of FBMC for Backhauling
5G
Abdelhamid Bounegab and Mohammed Boulesbaa
Abstract FBMCs (Filter Bank Multi-Carrier) are potential candidates for future
generations of 5G communication systems because of their advantages over OFDM
(Orthogonal Frequency-Division Multiplexing). One of the main advantages is the
improvement of the spectral efficiency thanks to the absence of the cyclic prefix (CP)
and to the waveforms well localized in time and in frequency and as well, thanks to
the use of some specific filters (PHYDYAS: PHYsical layer for Dynamic Spectrum
Access and cognitive radio project filter). The simulation programs carried out in this
work made it possible to highlight the interest of an FBMC/OQAM (Offset Quadra-
ture Amplitude Modulation) technique implemented in SMT (Staggered Multi tone)
with a FS-FBMC structure (Frequency Spreading). The BER (Bit Error Rate) and
PSD (Power Spectral Density) curves plotted in LOS (Line Of Sight) and Non-LOS
environments and with Rayleigh and Rice channels, have shown interests FBMC
connections for backhaul between small cells future mobile networks 5G.
Keywords Multicarrier modulations · Wireless cannel · Filter bank multi-carrier ·

5G · OQAM · Filter PHYDYAS
1 Introduction
In recent years, multi-carrier modulations have shown their effectiveness in wired

and wireless broadband communications. OFDM is the most adopted by several
standards: DAB, DVB-T, ADSL, IEEE 802.11a/g (Wi-Fi), IEEE 802.16 (WiMAX)
and LTE.
Today a debate is being lively conducted on the choice of the multi carrier tech-
nology to be deployed for modern 5G wireless communication systems, which use
A. Bounegab (B) · M. Boulesbaa

L.R.P.P.S. Laboratory, University Kasdi Merbah of Ouargla, Ouargla, Algérie
Electronic and Telecommunications Department, University Kasdi Merbah of Ouargla, 30000
Ouargla, Algérie

https://doi.org/10.1007/978-981-15-6403-1_76
1098 A. Bounegab and M. Boulesbaa
small cells (typically in urban areas) federate by a set of relays. To meet the challenge
of capacity, end to get benefit from new value-added services such as the V2x, the
D2D, IoT, smart home etc. …, we must find other types of modulation that repel
the OFDM modulation limits. In order to examine and analyze the performance of
a multi-carrier modulation alternative to the OFDM Which has some weaknesses
such as: the CP which reduces the nominal transmission rate, the very high level of
side lobes causing a loss spectral efficiency, the high rate of PAPR (peak-to-average
power ratio); as well as temporal de-synchronization.
An interesting multi-carrier modulation technique is proposed in this work. This
technique, called FBMC, uses a FIR (finite impulse response) filter bank. Further-
more, it offers better spectral efficiency since it requires neither guard interval nor
cyclic prefix.
The objective of this study is to make a theoretical and practical examination
(by simulation) of the candidate FBMC/OQAM modulation for the future wireless
communication system modern 5G that uses small cells (typically in urban areas)
to federate by one relay set. We propose hereby to conduct a study of modulation
performance FBMC/OQAM for a backhaul link between the different relays, these
in different LOS scenarios (existence of visibility between transmitter and receiver)
and NLOS (no visibility).
2 Filter Bank Multi-Carrier (FBMC)
A. Basic principles FBMC
FBMC technique (Filter Bank Multi-Carrier) is a multi-carrier modulation based

filter bank. It is thought to be a main transmission technology for future modern
wireless 5G. FBMC technology uses a filter bank FIR well localized in frequency
and time with less side lobes, which makes it less susceptible to errors. Its successful
contribution is in combining FBMC/OQAM and be an alternative solution OFDM
she retains the advantages and improves weaknesses.
FBMC based techniques can be divided into three main categories; FMT modu-
lation (filtered multi tone) [1, 2]. The techniques called CMT (Cosine Modulated
Multi tone) and SMT [3] Maintaining the Integrity of the Specifications.
B. Refers filter PHYDYAS prototype [4]
The realization of this filter is based on the frequency sampling technique (Technical
frequency sampling) of determining the frequency coefficients respect the Nyquist
criteria and orthogonally and build from these coefficients. The frequency response
is sampled at a frequency points K*M, uniformly spaced of Wk = 2k/KM by inter-
polation and then the impulse response by inversing Fourier transform. Determining
frequency coefficients depends on the overlap factor K = L*M, where L is the number
of coefficients of the impulse response of the filter and M is the number of subcarriers.
Performance of FBMC for Backhauling 5G 1099
Table 1 Frequency coefficients of the prototype filter PHYDYAS for K = 2; 3 and 4

K H0 H1 H2 H3 σ2
√
2 1 2/2 – – −35
3 1 0.911438 0.411438 – −44
√
4 1 0.971960 2/2 0.235147 −65
(a) (b)
Fig. 1 Frequency response of (b) and temporal (a) of the prototype filter PHYDYAS [5]
Half Nyquist filter frequency coefficients HK are obtained for K = 2.3 and 4 are
shown in Table 1.
Figure 1b shows the frequency response, this frequency is obtained from the
frequency coefficients above using the following interpolation formula [1, 4]:
−1

K
sin π f − Mk
K
MK
H( f ) = HK (1)
K =−(K −1)
M K sin π f − MKK
The FIR low-pass prototype filter, shown in Fig. 1a below, was obtained through
using the inverse transform quartermaster is governing by the following expression:

K −1
2.π.K
h[n] = H0 + 2. (−1) HK . cos
K
.(n + 1) (2)
K =1
K .N

K −1
Kt
h[t] = 1 + 2. HK . cos 2π (3)
K =1
KT
Fig. 2 Section of a filter bank based on the prototype with K = 4 [ 6]
From this representation, we can clearly see that the undulations observed are
almost negligible. Once the prototype filter is designed, the filter bank is obtained
by frequency offsets k/M, as in the case of FFT.
Figure 2 shows the section of a filter bank based on the prototype with K = 4.
The index k is obtained by multiplying the coefficients of filter prototype e j2ki/M
the main point is that the subcarriers with the same index type (odd index) do not
overlap, which is a very interesting result for the system design. Indeed, a particular
sub-carrier overlaps in the frequency domain with its contiguous only as represented
in Fig. 2.
C. General structure PPN of the technical FBMC/ OQAM
In the general structure PPN (Poly phase Network) of the technical FBMC/ OQAM
represented in Figs. 3 and 4 has issue was the bloc OQAM pre-processing and IFFT
bank of synthesis filters. At the reception, there is the post-processing block OQAM
FFT and the bank anal filters.
The complexity is taking the disadvantage of FBMC/OQAM system considera-
tion. However this latter gets reduced with multi-phase structure PPN.
The FS-FBMC technique is the alternative structure FBMC-p that assembled the
principles of OFDM and benefits of FBMC. It can also provide a perfect equalization
of the transmission channel.
Fig. 3 Block diagram of a bank filters synthesis [7]
Fig. 4 Block diagram of a bank filters analysis [7]
3 Results, Analysis and Discussion
In this part, a multi-carrier modulation OQAM/FBMC is processed and implemented

in order to show its performance in wireless communication. We developed a program
using Matlab for simulation of OQAM/ FBMC technique with FS-FBMC implemen-
tation structure. We are interested in taking it as a prototype filter, the filter PHYDYAS
which according to the literature, has the best performance in terms of mitigation
lobes off strips. To show the performance of the modulation OQAM/FBMC, we will
compare it with the conventional OFDM modulation in four comparative parameters.
The first criterion is the spectral efficiency conditioning and dimensioning multi user
of a mobile transmission system (Fig. 5).
The second is the bit error rate (BER) which measures the number of errors
induced by the transmission system (Figs. 6, 7, 8 and 9). The third parameter is the
Fig. 5 Principle frequency spreading FS FBMC [8]

(a) (b)
Fig. 6 Spectral efficiency OFDM waveform (a) and FBMC (b)
Fig. 7 BER performance based on SNR variation order QAM
rate of the peak at the mean value (PAPR), which is a very important parameter for
the choice of amplifiers in a transmission chain (Fig. 10). The fourth criterion is the
complexity of the system as measured by the number of the operations carried out
by the relevant transmission technology (Fig. 11).
The simulation is carried out by varying the intrinsic parameters modulations
Multi-carrier, namely the modulation order (QPSK, QAM- 16, QAM- 64…).
The transmission channel is an essential element of all wireless transmission
systems; we then conducted our simulation considering several types of wireless
channels, The AWGN channel, Rayleigh channel and Rice channel and model ITU
Fig. 8 BER according to ITU SNR Rayleigh distribution model
Fig. 9 BER based on the ITU model for SNR distribution RICE
Fig. 10 BER according to the SNR through a 5G channel (Model TDL)
Fig. 11 Block diagram of a bank filters analysis

will be in consideration. In addition, the recently standardized by the 3GPP chan-

nels for the next generation of mobile wireless systems 5G TDL (Tapped Delay
Line) will be simulated. These channels will be simulated under the two scenarios
LOS and NLOS PDP (Power Density Profile) with different standardized by 3GPP
recommendations (TR 25 890 V1.0.0, TR and TR 38.901 36.803v1.1.0 v 3.2.0) [9].
A. Results
(1) Spectral Performance:
The power spectral density (PSD) is one of the basic criteria in performance
evaluation spectral of a given system.
Figure 6 shows, the spectral efficiency of FBMC is very high with very small side
lobes compared to OFDM, which suffer strong side lobes.
The advantage the FBMC is that the same filter bank can be used for processing
the reception signal and spectrum capacity benches analysis of high-resolution filters
that can be exploited for systems of 5G.
(2) BER performance based on SNR:
Figure 7 shows the system performance in terms of BER in a canal AWGN depending
on the variation of the modulation order. Note that the increase in number of the
constellations of the QAM (4QAM, 16-QAM, 64 QAM and 256 QAM) requires
an increase in SNR and therefore better equalization. The SNR equal to 10 dB for
4QAM against an SNR of 30 dB for 256 QAM for a BER equal to 10−5 .
Figure 8 shows the SNR has values between 22 dB and 32 dB for a BER 10− 5
Using general settings and PDP model ITU with a Rayleigh distribution with TS
(sample time) equal to 2.6001e−7 , the Doppler spectrum as Jakes Doppler.
Figure 9 shows that the SNR takes values between 24 dB and 33 dB with a BER
of 10− 5 Using general settings and PDP model ITU with Rice distribution defined
by the following properties.
TS equal to 2.6001e−7 , the Doppler spectrum as Doppler Jakes, K Rician factor
equal to 10 and the Doppler effect of the direct LOS path fd (Doppler shift) equal to
100.
Figure 10 shows the SNR has values between 21dB and 27dB for general scenarios
of the future 5G channel characterized by small cells and millimeter waves.
(3) Performance in PAPR:
Figure 11 shows the Constraints with higher PAPR in FBMC and OFDM.
(4) Complexity of FBMC/ OQAM system
To trace the curves showing the complexity of the system based on the number of
subcarriers, we need to implement the formulas of the complexity with different
possible structure [8].
B. Analysis and Discussion
Figure 6 presents the performance of FBMC system, in terms of spectral efficiency,

with respect to OFDM. In this figure, clearly observed that the spectral efficiency
is FBMC very high with respect to the OFDM. This significant gain is because all
the subcarriers OFDM suffer strong side lobes and a CP interval spectral efficiency
reduced is obtained.
Figures 7, 8, 9 and 10, we present the BER curves of the system based on the SNR
(Signal to Noise Rate), that was made after the signal was transmitted across different
channel type such as AWGN, RICE and RAYLEIGH (ITU model) using equaliza-
tion ZF (zero forcing) which translates the robustness of this FBMC waveform for
channels multipath in many scenarios.
From the results above performance is shown in terms that the quality of trans-
mission is improved by using the FBMC/4OQAM system in an FS-FBMC structure
via a 5G channel or via a channel in the ITU model (previous generations of mobile
radio systems). The order increase modulation to be care rate increase in requires an
estimation step channel and a high performance equalizer to keep the same quality
of transmission.
Figure 11 shows the frequency responses of PHYDYAS filter used in FBMC and
that of the rectangular window used in OFDM. In this figure, we can clearly see that
the side lobes of PHYDYAS filter (FBMC) are reduced compared to the main lobe
and attenuate rapidly until they become virtually zero, unlike those of the rectangular
waveform (OFDM) that has a magnitude comparable to that of their main lobe that
barely decrease. It is a problem of PAPR, which will have an influence on energy
consumption, and the rapid saturation of amplifier and as well a kind of benefit that
reduces side lobes or makes it negligible.
The last Fig. 12 shows that the complexity of the system is a factor of evaluation
of a system; the complexity of FBMC over OFDM varies depending on the structure
of the system with or without a polyphase structure.
4 Conclusion
The work carried out in this paper deals with the multi-carrier modulations based
on benches FBMC filters and these three variants FMT, CMT and SMT. All these
modulations are based on a prototype filter of non-rectangular shape (case of the
OFDM) chosen, in our work, in accordance with PHYDYAS filter. FBMC modulation
is associated with orthogonal mapping OQAM for better efficiency.
Fig. 12 Complexity of FBMC technical and OFDM
It was necessary to introduce the theoretical foundations of wireless communica-

tion channels, mainly the 5G channels, as well as the principles of OQAM/FBMC
modulation and all the differences and similarities with OFDM. In addition, we
presented the two structures of implementation of the OQAM/FBMC system PPN
and FS-FBMC. We also presented a comparison between the prototype filter
PHYDYAS and the rectangular filter OFDM, where the chosen filter has a better
location in time and in frequency, as well as an excellent spectral efficiency where
the out-of-band lobes are negligible.
The performance simulation results of the OQAM/FBMC system, with the struc-
ture FSFBMC, in different types of wireless channels are acceptable. For Rice chan-
nels and Rayleigh, and mainly those designed for the 5G, a ZF forcing equalization
is necessary.
The complexity has been reduced with the help of FFT/IFFT in the OFDM;
we find a reasonable additional complexity for our OQAM/FBMC system, using a
poly-phase structure.
Finally, the FBMC technique, in combination with the OQAM modulation, in a
structure poly-phase, implemented in FS-FBMC solves the constraints encountered
with the OFDM and gives us a better results with the 5G channel.
References
1. Bellanger M (2006) Digital signal processing, theory and practice, dunod, 8th edn
2. Cherubini G, Eleftheriou E, Olcer S (1999) Filtered multitone modulation for vdsl. In: Global
telecommunications conference (GLOBECOM 99), vol 2, pp 1139–1144
3. Farhang-Boroujeny B (2003) Multicarrier modulation with blind detection capability using
cosine modulated filter banks. IEEE Trans Commun 51(12):2057–2070
4. Bellanger M (2001) Specification and design of a prototype filter for filter bank based multicarrier
transmission. In: IEEE international conference on acoustics, speech, and signal processing.
Proceedings, (ICASSP ‘01), vol 4
5. Anritsu (2016) Understanding 5G REV1
6. Bellanger (2010) FBMC physical layer: a primer, PHYDYAS
7. Hidalgo Stitz T (2010) Filter bank techniques for the physical layer in wireless. Tampere
University of Technology
8. Berg V, Doré J-B, Noguet D (2014) Microprocessors and Microsystems. homepage diary www.
elsevier.com/locate/micpro, CEA-LETI, Minatec in Grenoble, France
9. ETSI (2017) 5G;Study on channel model for frequencies from 0.5 to 100 GHz, 3GPP TR 38.901
version 14.0.0 Release 14), in may 2017
Flexible FPGA-Based RFID Emulation
Platform for Experimental Purpose
Ibrahim Mezzah, Omar Kermia, and Hamimi Chemali
Abstract Rapid advances in radio frequency identification (RFID) applications

are the driving force behind the development of new tools for embracing
frequently optimal solutions. In this work, an FPGA-based RFID emulation plat-
form conforming to EPC UHF Gen2 protocol has been developed to be used as a
development and evaluation tool to promote and extend RFID solutions. This plat-
form is mainly composed of a highly flexible architecture of a tag and a reader.
The VHDL designed platform around an IP core microcontroller serves as an effi-
cient and a low cost research development environment. The platform utility and its
features are demonstrated through analysis and evaluation of fault impact on a UHF
tag architecture.
Keywords RFID · Platform · Tag · Transponder · Reader · Emulation · EPC UHF

Gen2 · FPGA · Wireless technologies
1 Introduction
Radio frequency identification (RFID) is widespread in several currently application

fields such as supply chain management, transportation, wireless sensor networks
and internet of things. This technology is in continuous progress where new appli-
cations, new solutions and new protocol improvement appear frequently. However,
prototyping and evaluation tools as simulation and emulation platforms are essential
and play a significant role in this progress [1]. These platforms are very useful and can
I. Mezzah (B) · O. Kermia

Centre for Development of Advanced Technologies (CDTA), Algiers, Algeria
O. Kermia
H. Chemali
Department of Electronics, University of Setif, Setif, Algeria

https://doi.org/10.1007/978-981-15-6403-1_77
1112 I. Mezzah et al.
be employed in diverse aspects such as prototyping, sensing application, protocols

evaluation, performance assessment, attacks emulation and security analysis.
In the last decade, several RFID simulation environments have been reported.
They focus generally on the evaluation of a single RFID component whether tags
or readers [2], while other environments consider the physical and logical layer
simulation [3]. Some of them use hardware support for the implementation and
the validation. As RFID is highly up-to-date and a very interesting area, RFID tag
and reader prototypes realisation are then frequently reported. In these reported
prototypes, developed tag circuits are generally used in sensing applications [4]
or emulation means [5]. These circuits have been developed around microcontrollers
[5] or FPGA circuits [6]. For reader reported platforms, several works such as in
[7] rely on microcontrollers, on FPGA [8] and others on combination of DSP and
FPGA circuits [3]. In this work, the two RFID components, a tag and a reader,
are considered simultaneously in a designed flexible and high performance FPGA
based RFID platform. A tag and a reader are built conforming to EPC UHF Gen2
protocol [9]. The MCIP microcontroller IP core [10] is used in the development of
both components. Some helpful modules are added as peripherals to MCIP to provide
simple and powerful architectures developed in VHDL. The developed tag and reader
constitute therefore an experimental platform whose flexibility is doubly guaranteed
by FPGA design circuit flexibility and monitoring software. The developed tag and
reader are completely independent and, thus, can be used together or separately.
The paper is organised as follows: Sects. 2 and 3 present details on the design of the
tag and the reader of the developed platform. In Sect. 4 the physical implementation
of the system tag-reader is presented, while an application example is given in Sect. 5.
The last section concludes this paper.
2 Design of the Tag
A variety of known RFID tag (transponder) architectures deployed in real appli-

cations or proposed in research works rely on architectures processor based, inte-
grating hardware and software parts, as well as on architectures solely hardware built.
Each category has its advantages over the other; for example architectures based on
processors are more flexible due to the software flexibility and efficiently support
complexity and RFID protocols features. On the other hand, hardware-based archi-
tectures support basic RFID protocols functionalities in small area, low power and
low cost solutions.
Since tag architecture flexibility is targeted in an emulation platform, we have
developed, in this work, a new architecture based on MCIP microcontroller IP core
[10]. The configurability of MCIP is thoroughly explored to build the tag architecture
where new modules, such as “Transmitter” and “Receiver”, are designed and added
to the MCIP as peripherals (Fig. 1). Several tag functionalities are therefore hardware
implemented in these peripherals, which in turn considerably reduced the software
part and made it easier to adopt. The resulting tag is then so flexible, that it is possible
Flexible FPGA-Based RFID Emulation Platform … 1113
Input
Receiver
Output
MCIP Transmitter
(Configurable Microcontroller)
Backs. clock
Data Bus
Timer
PRNG
Program Data
Memory Memory Tag Memory
Fig. 1 RFID tag block diagram
to add any optional or costumed functionality conforming to the EPC UHF Gen2
version 2 protocol [11]. In addition, attention is directed to developing and applying
new RFID solutions such as those related to the authentication and the security.
As shown in Fig. 1, six peripherals are developed and added to MCIP to complete
the tag hardware design. The “Receiver” and the “Transmitter” constitute the main
modules since concerned by the generation and the reception of frames. For the others
peripherals, the “Backscatter clk” generates the used clock in the replay frames gener-
ation, the “Timer” counts the time periods, the “PRNG” is a pseudo random number
generator and finally the “Tag Memory” which includes the “Reserved memory”,
“EPC memory”, “TID memory” and “User memory” according to the protocol
[9]. The “Receiver” block contents is illustrated in Fig. 2 where commands are
acquired through monitoring the two registers “RXCON” and “RXSTAT” under the
supervision of “PIE decoder” commands.
TRcalH TRcalL
RTcal
Received
Data Bus
data Tag
16-bit shift register input
Interrupt Load PIE Decoder
counte
RXdataH RXdataL
RXCON CRC5 CRC5/CRC16

CRC16 Decoder
RXSTAT
Fig. 2 Tag receiver block diagram

Software part has been developed using C language in conjunction with assembly
code in order to obtain an optimal program. Several functions are defined and organ-
ised in twenty functions in addition to the main program and the interrupt subroutine.
Figure 3 illustrates the program functions tree and the size of each function. As shown
in this figure, “Frame reception”, “Open state” and “Secured state” functions occupy
a large volume in the program since they integrate the majority of functionalities.
As “Open state” and “Secured state” functions include nearly the same functional-
ities, they could be regrouped. In our case we have voluntary kept them separated
to distinguish the different states of the tag activity. The total size of the obtained
program is 5944 bytes while the used RAM is only 28 bytes. For optimising the
power consumption, the sleep mode is extensively employed. During the tag-reader
communication, the tag remains in the sleep mode about 90% of the time.
Implementation results of the developed tag on Virtex-5 FPGA (XC5VLX50T)
are given in Table 1. The number of used slice registers is 674 slices, while the number
of used slice LUTs is 3475 slices which represent 12% of the FPGA resources. These
main 698 bytes

frame_reception 1117 bytes
New_Round 242 bytes
inventoried_transition 66 bytes
loading_Slot_counter 88 bytes
Slot_counter_checking 56 bytes
Arbitrate_state 100 bytes
Slot_counter_checking
loading_Slot_counter
Reply_state 138 bytes
loading_Slot_counter
Slot_counter_checking
Acknowledged_state 182 bytes
inventoried_transition
Open_state 994 bytes
Write_process 42 bytes
Secured_state 1038 bytes
Write_process
Open_Kill_seq_state 266 bytes
Open_Access_seq_state 268 bytes
reply_generation 106 bytes
RN16_tag_reply 24 bytes
handle_tag_reply 24 bytes
EPC_tag_reply 250 bytes
Data_tag_reply 102 bytes
Write_success_tag_reply 36 bytes
Error_tag_reply 44 bytes
interrupt 4 bytes
Fig. 3 Functions tree of the tag software with the program size of each function
Table 1 Tag area occupation on Virtex-5 (XC5VLX50T)

Logic Used Available Utilization (%)
Number of slice registers 674 28800 2
Number of slice LUTs 3475 28800 12
Number of DSP 48Es 1 48 2
results concern the implementation of both hardware and software parts. Note that
the tag program memory (software part) occupies about 7% of FPGA resources.
3 Design of the Reader
Following the same methodology of the tag development, the designing of the reader
is completed by the development of two additional peripherals: the “Transmitter”
and The “Receiver” (Fig. 4). Three other MCIP peripherals are then used: the “I/O
Port” for LCD display, the “Timer0” for counting time periods and the “UART”
for enable serial communication. Several functions such as PIE encoding, delim-
iter/RTcal/TRcal generation and CRC5/CRC16 calculation [9] are hardware imple-
mented on the “Transmitter” module, in such a way as to facilitate the commands
generation and to easily add new commands through software. The design effort is
focused on the receiving part since the decoding process is always more complicated
than the encoding process. Indeed, only FM0 decoding is integrated in this current
version of the reader. The reader implementation reported area occupation of 6% on
Virtex-5 XCVLX50T FPGA (Table 2).
Output
Transmitter
MCIP Input
(Configurable Microcontroller) Receiver
I/O Port
Data Bus
Timer0
Program Data
Memory Memory UART
Fig. 4 RFID reader block diagram

Table 2 Reader area occupation on Virtex-5 (XC5VLX50T)

Logic Used Available Utilization (%)
Number of slice registers 708 28800 2
Number of slice LUTs 1900 28800 6
Number of DSP 48Es 1 48 2
4 Physical Implementation
Both developed circuits of the tag and the reader and all the defined modules are
completely validated in the simulation process. A physical implementation test is
then achieved by implementing the tag and the reader in a single FPGA using wired
communication between the two, this therefore constitutes an emulation of a perfect
RFID communication. For this implementation, we have used Digilent Genesys
board which includes a Virtex-5 FPGA (Fig. 5).
An oscilloscope is also used to visualize the communication between the tag and
the reader. Figure 6 illustrates the captured signals during the communication process
where the reader sends some commands that allow the tag to enter in the open state.
Fig. 5 Physical system implementation on Digilent Genesys board

RN16 EPC handle
Tag
Select Query ACK Req_RN

Reader
RN16 (FM0)
Tag
Query
Reader
Fig. 6 Captured signals during the communication process reader-tag
5 Application Example: RFID Tag Robustness

and Security Evaluation
A fault emulation system for RFID tag robustness and security evaluation is derived
from the developed platform and from the approach presented in [12]. This system
(Fig. 7) contains a couple of tag-reader implemented on a single FPGA with fault
emulation mechanism which permits the emulation of transient faults including single
or multiple bit upset (SBU, MBU) and single event transient (SET). In fact, faults
are injected into one tag, while the other is kept fault free. In addition, the present
solution has excluded the radio frequency communication by implementing in a
single FPGA both readers and tags. This choice greatly simplifies the emulation
process upon keeping valid analysis outcomes since the tag behaviour is willingly
maintained unaffected through this exclusion [13].
As illustrated in Fig. 7, the fault emulation system includes specific monitoring
circuits which control readers, tags and the communication. This defined structure
delivers information about injected faults and effects. Fault injection operation is
carried out through a module based on a VHDL fault injection technique [14]. In
addition, an emulation controller manages every system component and synchronises
the emulation process [15].
FPGA Fault free

tag
Reader1
monitoring
circuit
FPGA board
Fault Faulty host

injection tag
Reader2 computer
module
Emulation controller
Fig. 7 RFID fault emulation platform
An exhaustive single SEU analysis is then performed on the targeted tag, following
the communication scenario illustrated in Fig. 6. As shown in this scenario, the reader
starts by sending a ‘Select command’. Going through necessary steps of inventories
and accesses, the tag replies at the end with ‘Handle’ and enters in ‘Open state’.
For every flip-flop (FF) within the tag, a single fault is injected at every clock cycle,
except cycles for which the tag remains idle since injected faults will produce the
same result. Every tag FF, in total of 314 FFs, has been the subject of 6598 SEUs.
Consequently, in this experimentation, 2071772 SEUs are autonomously introduced
and analysed.
According to their effects, injected faults are classified in five defined categories
as follows: 1. Silent (S): communication scenario achieved with no error detection.
2. Frame Alteration (FA): communication scenario achieved in presence of frame
alteration. 3. Latent (L): communication scenario achieved but accompanied with
memory content altered which in turn may produce failure on forthcoming cycles.
4. Latent with Frame Alteration (L&FA): similar to latent but in presence of frame
alteration. 5. Failure (F): no achievement of communication scenario.
Table 3 summarises the deduced errors, they are divided into two categories,
major errors and minor errors, according to their impact on the tag security and
reliability. Note that one fault may produce several errors at the same time. Minor
errors have limited impact on the tag. For example, with ‘valid altered frame’ error,
frames either include non-destructive encoding faults or have delays without any
scenario corruption or breaking the communication. The remaining minor errors
may be recovered at new interrogations or when initiating a new round. In addition,
the presence of major errors, although in a minority (eight in total), is damaging
since the tag security and data contents are affected. ‘Enter into the Secured state’
error is a serious tag security threat since it can lead to secure data delivering or
data altering unwillingly. In addition, ‘enter into the Killed state’ and ‘change of tag
memory content’, considered as major errors, have not been detected through the
performed exhaustive injection.
Table 3 Summary of produced errors for the SEU experimentation

Errors Nb. Prop Fault category
Major Enter into the secured state 8 1/258,971 F
Enter into the killed state 0 0% F
Change of tag memory content 0 0% F
Minor Valid altered frame 202,176 9.76% L&FA/FA
Frame corruption 68,143 3.29% F
Invalid frame 40,875 1.97% F
Identification failed 157,795 7.62% F
Exiting the current round 61,368 2.96% F
Change of inventoried parameters 34,926 1.69% F/L/L&FA
Change of query parameters 107,085 5.17% F/L/L&FA
6 Conclusion
In this work, a UHF RFID emulation platform based on FPGA is developed. It

constitutes a low cost and a useful tool to carry out multiple tasks in prototyping,
protocol evaluation, attacks emulation and security analysis. The usefulness of this
platform is demonstrated through an exhaustive single SEU fault emulation on an
RFID tag. This experimentation has allowed the evaluation of the tag robustness and
the identification of the most vulnerable tag parts.
References
1. Steger C, Janek A, Weiß R, Derbek V, Jantscher M, Preishuber-Pfluegl J, Pistauer M (2010)

Simulators and emulators for different abstraction layers of UHF RFID Systems. RFID
SYSTEMS
2. Derbek V, Steger C, Weiß R, Wischounig D, Preishuber-Pfluegl J, Pistauer M (2007) Simulation
platform for UHF RFID. In: Proceedings of the conference on Design, automation and test in
Europe. EDA Consortium, pp 918–923
3. Angerer C, Langwieser R, Rupp M (2012) Evaluation and exploration of RFID systems by
rapid prototyping. Pers Ubiquit Comput 16(3):309–321
4. Sample AP, Yeager DJ, Powledge PS, Mamishev AV, Smith JR (2008) Design of an RFID-based
battery-free programmable sensing platform. IEEE Trans Instrum Meas 57(11):2608–2615
5. Plos T, Aigner M, Baier T, Feldhofer M, Hutter M, Korak T, Wenger E (2012) Semi-passive
RFID development platform for implementing and attacking security tags. Int J RFID Secur
Crypt 1(16–24):39
6. Abdelmalek O, Hely D, Beroulle V, Mezzah I (2013) An UHF RFID emulation platform with
fault injection and real time monitoring capabilities. In: 2013 8th international on design and
test symposium (IDT). IEEE, pp 1–2
7. Nikitin PV, Ramamurthy S, Martinez R (2013) Simple low cost UHF RFID reader. In: IEEE
international conference on RFID. pp 126–127
8. Li X, Wang C (2011) Design and FPGA verification of UHF RFID reader digital baseband. In:
2011 International conference on IEEE electrical and control engineering (ICECE)
9. EPC Radio-Frequency Identity Protocols Class-1 Generation-2 UHF RFID, Protocol for
Communications at 860—960 MHz”, Version 1.2.0, October 2008, EPCglobal Inc., http://
gs1.org
10. Mezzah I, Chemali H, Mezzah S, Kermia O, Abdelmalek O (2015) MCIP: High configurable
8-bit microcontroller IP-core. In: 2015 science and information conference (SAI). London, pp
1387–1390
11. EPC Radio-Frequency Identity Protocols Class-1 Generation-2 Version 2 UHF RFID, Protocol
for Communications at 860—960 MHz”, Version 2.0.1, April 2015, EPCglobal Inc., http://
gs1.org
12. López-Ongil C, García-Valderas M, Portela-García M, Entrena L (2007) Autonomous fault
emulation: a new FPGA-based acceleration system for hardness evaluation. IEEE Trans Nuclear
Sci 54(1):252–261
13. Mezzah I, Chemali H, Kermia O (2017) Emulation-based fault analysis on RFID tags for
robustness and security evaluation. Microelectron Reliab 69:115–125
14. Lala PK (2012) Transient and permanent fault injection in VHDL description of digital circuits.
Circ Syst 3(2):192–199
15. Mezzah I, Kermia O, Chemali H, Abdelmalek O, Beroulle V, Hély D (2013) Assertion based
on-line fault detection applied on UHF RFID tag. In: 2013 8th IEEE design and test symposium.
Marrakesh, pp 1–5
Human Activity Recognition Based
on Ensemble Classifier Model
Souhila Kahlouche and Mahmoud Belhocine
Abstract In this work, we address the problem of Human Activity Recognition

(HAR), applied to service robot. In addition, a real time HRI system able to under-
stand some common interactive human activities is designed. To classify activities
into different classes, a combination of three supervised machine-learning algo-
rithms: Support Vector Machine (SVM), Decision Tree (DT) and Artificial Neural
Network (ANN) is used, based on the idea that a set of classifiers improve machine
learning results. Our approach uses as input a view invariant 3D data of skeleton
joints, which are rich body movement information recorded from a single Microsoft
Kinect camera to create specific dataset of six interactive activities. The algorithm
was able to successfully classify and recognize activities being performed in front
of the camera. The system framework is realized on the Robot Operating System
(ROS), and real-life activity interaction between our service robot and the user was
conducted to demonstrate the effectiveness of the developed HRI system.
Keywords Human Robot Interaction (HRI) · Human Activities Recognition

(HAR) · Ensemble learning · Robot Operating System (ROS) · Service robot ·
RGB-D camera
1 Introduction
Human Activity Recognition (HAR) is a very important aspect for social robot inter-
action. This aspect has to be considered in every Human Robot Interaction (HRI)
system, which generally uses a combination of sensors and modalities, including
vision, audio, and touch to recognize human actions.
S. Kahlouche (B)
Ecole Nationale Supérieure de l’Informatique ESI, Oued Smar, Algiers, Algeria
M. Belhocine
Centre de Développement des Technologies Avancées CDTA, Baba Hassen, Algiers, Algeria

https://doi.org/10.1007/978-981-15-6403-1_78
1122 S. Kahlouche and M. Belhocine
Machine learning algorithms have been widely applied for this purpose. Using
Neural Network, Fiaz and Ijaz [1] proposed a method to design an intelligent human
activity monitoring system to detect and track suspicious activity in a surveillance
environment. A three-layer perceptron was used for classifying human activities
from information of distance and motion vector. Foroughi et al. [2] applied four-
layer MLPs to learn Eigen-motions, which are the movement patterns extracted by
an Eigen space technique, for motion classification and falling detection.
Using Support Vector algorithm, Schuldt et al. [3] aimed to recognize human
activities by extracting local space-time features in a video. In addition, Decision
tree algorithm also constitutes one of the most used machine learning methods in
HAR problem. Stauffer and Grimson [4] built a hierarchical classification binary
tree by accumulated joint co-occurrence statistics of the representations in a video
sequence. Moreover, Ribeiro and Santos-Victor [5] investigated the use of hierar-
chical binary tree classifiers in human activity recognition. For each node of the
binary tree classifier, a Bayesian classifier was used and the likelihood functions
were modelled and systematically learned as Gaussian mixtures.
Recently, the availability of low-cost and high-accuracy depth sensors, using
the out-of-the-box skeletal tracker framework [6] leads to a large community of
researcher to investigate in motion and pose tracking fields. Using this sensor,
Chrungoo et al. [7] released an RGBD human-robot interaction dataset consisting of
18 activities including ten stylized gestures and eight conventional activities of daily
living. The dataset includes both communicative and non-communicative actions;
they used basic SVM classifier that was achieved near perfect accuracy on a closed
set of eight activities.
Chen et al. [8] proposed an intuitive fact that not all skeletal joints are involved
in a particular activity, and only a few active joints are significant and informative
for a certain activity. By concentrating on the active joints and abandoning the other
inactive parts, they generate more discriminative and robust features, which can deal
better with intra-class variation problem.
In this work, we take into consideration that a set of classifiers is designed to
obtain better performance than any of individual one. For this purpose, we combine
three single classifiers: Support Vector Machine, Decision Tree and Neural Network
to classify six interactive activities represented by 3D coordinate of skeleton data.
The major contributions of this work are:
– The proposed viewpoint invariant transformation in the preprocessing step, which
can eliminate observation variations caused by viewpoint differences.
– Providing ROS package and a specific dataset containing interactive activities for
the HRI system. This dataset is a challenging one, once there is significant intra-
class variation in the realizations of the same activity. For example, sometimes
‘Call’ activity is done with the left hand while sometimes with the right hand.
Human Activity Recognition Based on Ensemble Classifier Model 1123
2 Human-Robot Interaction System Overview
Figure 1 shows the framework of human-robot interaction system. It mainly includes

two modules:
The off line process where the dataset is created using RGBD camera to obtain
the human skeleton data stream, from which features are extracted and pretreated.
Three supervised machine-learning algorithms (SVM, DT, ANN) are then trained.
processing
Off line
Dataset creation
Features extraction/preprocessing
ANN training SVM training DTree training

processing
Pretained models Detected User

Online
Voting process No
Start interaction
Yes
Speech Recognition
Feeding Features
Action Recognition
Stop Going Call Coming Hello Pointing
Robot Reaction
Motion interaction Voice interaction
Fig. 1 The framework of human-robot interaction system

Table 1 Robot reactions

Activity Robot reaction
Hello Voice reaction: Hello welcome to CDTA
Call Voice reaction: Please wait, I am coming
Motion reaction: Approaching user and start interaction
Stop Voice reaction: ok, I will stop here
Motion: Stop moving
Pointing Voice reaction: Ok, I will go there
Motion reaction: Go to pointed position
Coming Voice reaction: How can I help you?
Motion reaction: step back and prepare to begin interaction
Going Voice reaction: Good-bye, thank you for your visit
Motion: turn back and stop interaction
The online process where the activity being performed is inferred using voting
process of the above classifier and the robot runs the action according to rules that
are described in Table 1.
3 The Off-Line Process
A. Dataset Creation
Our dataset has been created using OpenNI tracker framework, which allows the
skeleton tracking at 30 fps, providing the three-dimensional Euclidean coordinates
and three Euler angles of rotation in the 3D space for each joint with respect to the
sensor. The dataset contains six interactive activities performed by different indi-
viduals, among the activities. Four are static, where the user does not move from
camera view field like (hello, Stop, call), and two dynamic activities where the user
can leave the camera view field like (going, coming), Fig. 2 shows some examples
of our dataset.
B. Features Extraction
In order to detect and recognize activities related to interactive service robot, we have
selected nine joints in the upper body human skeleton (see Fig. 3), since joints on
low limbs introduce more noise than useful information to distinguish activities.
(1) Construction of Feature Vectors
Using the above information, we have computed dataset of features as follows:

– 3D coordinates of the 9 selected joints: (x, y, z).
Fig. 2 Examples of our dataset A Hello, B Call, C Stop
1- Head
2- Neck
Y 3- Torso
4- Left Shoulder
5- Left Elbow
6- Right shoulder
X 7- Right Elbow
8- Left Hand
Z 9- Right Hand
Fig. 3 The selected human skeleton joints
– 3D rotation angles in the space described by a sequence of three rotations

according to the three axes X, Y and Z. Let be the rotation angle about the
x-axis which can be computed as:
⎡ ⎤
1 0 0
Rx (Ψ ) = ⎣ 0 cos − sin ⎦ (1)
0 sin cos
Similarly, let θ and be the rotation angles about the y-axis and z axis respectively:
⎡ ⎤
cos θ 0 sin θ
R y (θ ) = ⎣ 0 1 0 ⎦ (2)
− sin θ 0 cos θ
⎡ ⎤
cos Φ − sin Φ 0
Rz (Φ) = ⎣ sin Φ sin Φ 0 ⎦ (3)
0 0 1
Hence, the feature vector has the following form:
Time X Y Z θ φ ψ
(2) Feature Preprocessing
A preprocessing step is applied to the 3D skeleton data, it consists of the following:

Translation: to guarantee the same origin of the coordinates system for all
acquired frames; the reference was set to the torso of the human skeleton;
Smoothing: in order to attenuate noise introduced by the sensor, we have applied
a moving average filter with five (5) neighbor’s data points, to filter the noise and
smooth the data.
Normalization: a normalization step is applied to accommodate and reduce the
influence of different user’s height and limb lengths; therefore, all skeleton 3D
coordinates are normalized as follows;
Di − min(Di )
Di−normal =
max(Di ) − min(Di )
where max(Di ) and min(Di ) are the maximum and minimum values of a feature set.
Symetrization: It is done to disambiguate between mirrored versions of the same
activity (e.g. gestures performed by right and left-handed people); it is required for
activities like Hello, Call and pointing. It is just necessary to consider a new sample
based on a mirrored version of the original 3D skeleton data.
C. Training
An ensemble-learning model is applied in order to obtain better prediction result than

any individual classifiers [9]. This is done by using multiple learning algorithms; in
this work, we have used a fusion of three single classifiers, which are nonlinear
support vector machine, Neural Network and Decision Tree (see Fig. 4).
In order to validate the activities recognition models, 70% of our dataset has been
used for training while 30% for testing.
(1) Support Vector Machine: is discriminative classifier, which outputs the optimum
hyperplane that separates the training data. Margin is defined as the distance
Train dataset SVM

Final
Voting
Dataset ANN Prediction
Test dataset Dtree
Fig. 4 Ensemble learning
Stop Call Hello Coming Going Pointing

Stop 0.91 0.09 0.0. 0.00 0.00 0.00
Call 0.00 1.00 0.00 0.00 0.00 0.00
Hello 0.00 0.20 0.80 0.00 0.00 0.00
Coming 0.50 0.00 0.00 0.33 0.17 0.00
Going 0.00 0.00 0.00 0.00 1.00 0.00
Pointing 0.00 0.00 0.00 0.00 0.00 1.00
Fig. 5 Confusion matrix of SVM
of data from the separating hyperplane. The classification results of SVM that
are applied to our dataset are presented in a confusion matrix, in which the
performance measures of accuracy is 84% (Fig. 5).
(2) Decision Trees: The goal of Decision Trees is to generate a model that predicts
the classification labels based on decision rules according to the training data.
At each level of the tree, the model tries to maximize information gain by
considering all possible splits based on the used features.
Binary tree classification is simple and fast, but the separation rules in each
node are difficult to be general for other cases, making it difficult for complex
scenarios.
The classification results are shown in Fig. 6 where the accuracy result is 94%.
(3) Artificial Neural Networks (ANNs) use the concept of biological neuron
networks to supply nonlinear capability via more discriminative feature trans-
formation through hidden layers. This provides more effective classification
through multilayer perceptrons (MLPs), which apply the back propagation (BP)
supervised learning algorithms to compute suitable connection weights and

Stop 0.95 0.00 0.00 0.00 0.00 0.05
Call 0.24 0.53 0.12 0.00 0.00 0.12
Hello 0.04 0.08 0.88 0.00 0.00 0.00
Coming 0.00 0.12 0.12 0.76 0.00 0.00
Going 0.00 0.00 0.00 0.23 0.75 0.02
Pointing 0.00 0.41 0.00 0.09 0.02 0.48
Fig. 6 Confusion matrix of decision tree


Stop 0.95 0.00 0.0. 0.05 0.00 0.00
Call 0.00 0.97 0.00 0.03 0.00 0.00
Hello 0.00 0.00 1.00 0.00 0.00 0.00
Coming 0.00 0.00 0.00 1.00 0.00 0.00
Going 0.00 0.00 0.00 0.19 0.76 0.05
Pointing 0.00 0.04 0.00 0.00 0.00 0.96
Fig. 7 Confusion matrix of ANN
biases of the network. ANN is applied on our dataset with ‘relu’ as activation
function and achieved 72.5% of accuracy (see Fig. 7).
4 The Online Process
Once the Human/robot interaction is started, the pertained models are loaded and the
prediction of new features is done by voting.
A. Voting Process
The ensemble learning involves training more than one model in the same dataset,
then using each of the trained models to make a prediction before combining the
predictions in some way to make a final outcome or prediction.
To combine multiple classifiers; we have used majority-voting process as follows:
Let Mi = {m1, m2 ,…., mn }, be the ensemble of model trained and Cj = {c1, c2 ,….,
cl }, be the classes.
The vote for each class j and each model i is computed by:

1 i f m i votes for c j
vi, j =
0 else
n
The total vote for each class j is then calculated by V j = i=0 vi, j , j = 1, . . . , l
The final decision V final is then the majority vote.
V f inal = arg max(V j)
Figure 8 shows the confusion matrix of activity recognition using voting process,
which achieves accuracy of 97.1%.
B. Robot Reaction
A multimodal interaction is designed for human robot interaction scenarios to provide

natural communication. Once an activity is recognized by the classification step, the
proper reaction is taken from the look-up table to be executed by the mobile robot.
Stop 1.00 0.00 0.0. 0.00 0.00 0.00

Call 0.00 1.00 0.00 0.00 0.00 0.00
Hello 0.00 0.00 0.97 0.00 0.00 0.03
Coming 0.00 0.00 0.00 0.96 0.04 0.00
Going 0.00 0.00 0.00 0.00 1.00 0.00
Pointing 0.00 0.10 0.00 0.00 0.00 0.90
Fig. 8 Confusion matrix of voting process
The developed HRI system has been implemented on Robot Operating System (ROS)
and tested using the B21r mobile robot in real scenarios.
Our dataset contains several intra-class variations among different realizations of
the same activity. For example, hello wave is done with the left or right hand. Another
challenging feature is that the activity sequences are registered from different views.
The validation technique adopted for activity recognition is based on the strategy
of learning from different persons and testing with an unseen person.
From Table 2, we can see that activity recognition accuracy of the fusion approach
is much better than the base classifiers one. Therefore it has been used for the real
time application.
The experimental tests using the voting process approach for a real time appli-
cation acquire 4 s of RGBD sensor data for features extraction and classification.
After that, a final decision is made for activity recognition and a robot reaction is
performed.
The proposed framework was capable of recognizing different interactive activ-
ities that happens sequentially in case of a person transition from one activity to
another one.
Figure 9 shows a scenario where a person attempts to interact with a service robot.
(a) First, the person salutes the robot, ‘Hello’ action is recognized and the robot
reacts with voice mode and says: ‘Hello, welcome to CDTA’.
Table 2 Comparison of fusion accuracy with base classifiers

Method Accuracy (%)
SVM 84
ANN 72.5
DTree 94
Fusion 97.1
Fig. 9 Human robot interaction scenario

(b) The person is calling the robot, and the activity is recognized as ‘Call’ and then
the robot interacts with voice mode, and says: “Please wait, I am coming”, at
the same time approaching the user.
(c) The user is pointing position; the activity is recognized as “pointing”, and the
voice reaction of the robot is: “Ok, I will go there”, and the robot moves to the
pointed position.
(d) The user decides to stop the robot; the activity is recognized as ‘Stop’, and the
robot stops moving.
(e) The user decides to go away, the activity “Going” is recognized and the robot
reacts with voice: “Good-bye, thank you for your visit”, and turns back and
stops interaction.
6 Conclusion
This work demonstrates progress towards a HRI system for a service robot, capable
of understanding interactive human activities in real time.
In the preprocessing step, we have proposed view invariant transformation applied
to the data, captured by the Microsoft Kinect camera to guarantee view invariant
features.
Three machine-learning algorithms which are SVM, Decision Tree, and ANN,
have been used in this work and their classification results are compared. Based on
those classifiers, a fusion approach has been proposed using voting process, and the
experimentation shows that they are enough for obtaining good results.
Our HRI system was able to correctly recognize the performed activity and it
successfully reacts according to the situations.
However, the same activity may have different behavior interpretations depending
on the context in which it is performed. More specifically, the time (when), the place
(where), the intention (why), the manner (how). In the future, we will focus on the
study of that additional contextual information, which will improve the interactivity
of our service robot.
References
1. Fiaz MK, Ijaz B (2010) Vision based human activity tracking using artificial neural networks.
In: Proceedings of IEEE international conference on intelligent and advanced systems (ICIAS).
Malaysia, pp 1–5
2. Foroughi H, Naseri A, Saberi A, Yazdi HS (2008) An eigenspace-based approach for human
fall detection using integrated time motion image and neural network. In: Proceeding of IEEE
9th international conference on signal processing (ICSP). Beijing, China, pp 1499–1503
3. Schuldt C, Laptev I, Caputo B (2004) Recognizing human actions: a local SVM approach.
In Proceedings of the 17th ieee international conference on pattern recognition (ICPR), vol 3.
Cambridge, UK, pp 32–36
4. Stauffer C, Grimson WEL (2000) Learning patterns of activity using real-time tracking. IEEE
Trans Pattern Anal Mach Intell 22:747–757
5. Ribeiro PC, Santos-Victor J (2005) Human activity recognition from video: modeling, feature
selection and classification architecture. In: Proceedings of the international workshop on human
activity recognition and modelling (HAREM), vol 1. Oxford, UK, pp 61–70
6. Shotton J, Fitzgibbon A, Cook M, Sharp T, Finocchio M, Moore R, Kipman A, Blake A (2011)
Real-time human pose recognition in parts from single depth images. In: IEEE conference on
computer vision and pattern recognition
7. Chrungoo A, Manimaran SS, Ravindran B (2014) Activity recognition for natural human robot
interaction. In: Social robotics. Springer, Berlin, pp 84–94
8. Chen H, Wang G, Xue JH, He L (2016) A novel hierarchical framework for human action
recognition. Pattern Recogn 55:148–159
9. Zhou Z-H (2012) Ensemble methods: foundations and algorithms. CRC Press
Security in the Internet of Things: Recent
Challenges and Solutions
Hamza Belkhiri, Abderraouf Messai, Mohamed Belaoued, and Farhi Haider
Abstract Internet of things (IoT) is one of the technological revolutions of the

new computing technology era, and is considered as the Web 3.0, in which billions
of smart devices are interconnected. Although these connected objects offer many
advantages, they are a critical risk in terms of security and privacy. In this paper,
we present a review of the most recent threats and attacks targeting the IoT as well
as the most relevant existing solutions that deal with these threats. Moreover, we
provide an evaluation of these solutions with regards to IoT’s future trends. Finally,
we provide our recommendations to insure a high security level against the different
threats targeting IoT devices.
Keywords Internet of Thing · IoT · IoT security · DoS/DDoS attack · WSN · IoT
malware
1 Introduction
The connection of a group of objects and devices of various kinds through the Internet
network constitutes the so-called Internet of things [1]. This term was introduced in
1999 Kevin Ashton [2]. The company Cisco systems estimated that in April 2014,
there were about 12.1 billion of devices connected to the Internet, indicating that the
H. Belkhiri · A. Messai · F. Haider (B)

Department of Electronics, University Freres Mentouri Constantine 1, Constantine, Algeria
H. Belkhiri
A. Messai
M. Belaoued
Lire Laboratory Software Technologies and Information Systems Department, University of
Constantine 2, Constantine, Algeria

https://doi.org/10.1007/978-981-15-6403-1_79
1134 H. Belkhiri et al.
number in 2020 will reach 50 billion devices [3]. It is true that the number is huge, and
it is due to the fact that there is more and more devices that have become connected
to the Internet so that phones and computers are no longer the only ones connected
to Internet networks. Indeed, this connectivity went beyond several things such as
environmental sensors, traffic cameras and household electronics… etc. [4]. Thus,
this connectivity generates large amounts of data traffic and to deal with that we use
cloud computing technology that is characterized by its ability to store, analyze and
process data [5].
The Internet of things has a wide range of applications, most notably smart cities,
smart homes, and health care, which have facilitated our daily lives, but they have
collided with one of the most important security and privacy challenges that are often
linked to the lives of users [6]. To deal with billions of objects or devices interacting
with one another or with humans security mechanisms must be established in order
to face the variety of attacks that threatens the security and privacy of IoT, such as
denial of service attacks (DoS), malware (ransomware, spyware attacks, etc.), and
routing attacks [7].
In order to plan the path of data in the internet, objects must be built on a structure
in which the physical world is connected to the virtual world via wireless networks
[8]. This structure must also be able to meet challenges in improving its performance
and continuing to improve quality [7]. Garcia et al. [9] proposed an architecture for
IoT that consists of 5 layers: Perception Layer, Network Layer, Middleware Layer,
Application Layer and Business Layer.
In this paper we provide a brief review of the most common attacks targeting
internet of things (IoT) at three levels of the aforementioned architecture, namely
the perception layer, the network layer and the application layer, which are in charge
of data collection, data transmission and task accomplishment respectively [10, 11].
Moreover, and after reviewing the literature works, we will present what we believe
are the most notable existing solutions published in the last five years that deal with
the aforementioned attacks.
This paper is organized as follows: Section 2 describes the most common attacks
in IoT and existing solutions. In Sect. 3, we will discuss the existing solutions and
provide our recommendations. Finally, Sect. 4 concludes our paper and underlines
our perspectives.
2 Most Common Attacks in IoT and Existing Solutions
Any person depends on information in all aspects of his/her life and seeks to main-
tain it in a proper way. In the Internet of things context, data are exchanged between
devices trough the Internet, and due to its nature, the data path is subject to several
types of attacks. Despite the physical attacks that target devices or components [12],
data-driven ones aim to illegitimately access and/or damage users’ data [13]. Further-
more, some attacks target some specific communication protocols (wireless proto-
cols) that connect billions of devices together, and/or a specific service in order to
Security in the Internet of Things … 1135
make them unavailable [12]. Thus, the collected, stored, processed, and exchanged
data is not entirely secure. In this section, we will highlight the most common attacks
that target IoT with regards to the previously discussed IoT layered architecture. We
provide a detailed taxonomy of these attacks (See Fig. 1), as well as the most recent
and the notable proposed solutions against these attacks. Note that we focused on
what we believe are the most common attacks namely: DoS/DDos, Ransomware,
and SinkHole Attacks.
A. Dos and DDoS Attacks
In denial of service (DoS) attack, a huge amount of invalid packets is brought to

the target system (server dumping, dropping web sites) in a short time, causing it to
be slowed down or damaged and therefore the service is denied even to legitimate
users [14]. In the case of the Distributed Denial of Service Attacks (DDoS), a large
Fig. 1 Taxonomy of attacks

in IoT Attacks in IoT
Perception Network Application

layer layer layer
Physical Jamming DHCP

attacks attacks Starvation
DDOS
MiMT attacks Rootkit
attacks attacks
Routing
DOS attacks APTs
attacks attacks
Sinkhole
attacks Malwares
Eavesdropping
attacks attack
Wormhole
attacks
Ransomware
RFID attacks
Hello flood
spoofing
attacks
Spyware
Selective attacks
forwarding
attacks
number of computers controlled by cybercriminals are exploited constituting the so-

called Botnets [16]. The purpose of Botnets is to increase data traffic on a single web
server in order to dump it and making it unavailable [17]. Protection against DDoS
should be done at early stages, since it is difficult distinguish between legitimate and
malicious traffic once the attack is occurring [15].
In 2017, Nayak et al. [18] introduced a mechanism for detecting DOS attack in
Wireless Sensor Networks (WSN) using knowledge of network behavior based on
both anomaly-based and signature-based identifiers. The authors focused on these
methods to identify infiltration through the installation of IDS unit according to the
structure of the WSN. The analysis is performed on the data collected by the sensor
node and in case the data is collection is interrupted, an alarm is triggered. The
proposed mechanism is able successfully detect DoS attack, however, it needs to be
evaluated based on more performance measures.
Kasinathan et al. [19] proposed a mechanism to detect Denial-of-Service (DoS)
attack in 6LoWPAN (IPv6 LoW Power wireless Area Networks) networks after
monitoring the traffic and observing the flow and consistency of network data reported
by IDS. The authors suggested the use of IDS-SURICATA interface associated with
6LoWPAN in order to provide an accurate packets analysis received from IDS Probe
(IDS-P) to pass through a decoder down to the detection engine. The mechanism has
been subjected to evaluation through a penetration test where the network is injected
by tool Metasploit that have the capability to sniff and inject forged packets into the
6LoWPAN network (UDP flood attack) to be detected by giving alerts. The proposed
mechanism achieved a high true positive rate for DoS flooding attack detection.
In their work of 2015, Lee et al. [20] proposed a lightweight system to detect
intrusions in the 6LoWPANs. This system is based on the exhaustion of energy from
the harmful node (affected by its massive reception of packets) under the influence
of jamming attack. The Authors used the energy consumption as the main feature to
detect the attack. Indeed, when a certain threshold is exceeded, the IDS classifies the
node as harmful. The simulation results show that the system is efficient and accurate,
although it lacks some performance indicators (true positive and true negative).
Pu, in [21] proposed a scheme for DoS attacks detection that cause power depletion
in Low power and Lossy Networks (LLNs), which is based on misconduct threshold.
The scheme consists in recording the number of packets received in each adjacent
node in the observation table during the control window and recording also the
number of redirection behaviors detected for each adjacent node in the detection
table. The calculated threshold was then compared with a dynamic threshold for
separation and isolation of the harmful node of the network. The experimental results
show that the proposed scheme is capable of detecting the energy depletion attack
and reducing its power consumption (isolation). However, authors did not explain
how they obtained the dynamic threshold.
Xia et al. in [22] proposed a mechanism to detect DDoS attack in IoT devices.
This mechanism is based on the analysis of wireless traffic, and the authors used
Arduino as a sensor and data collector. This data is handled by the central server
after being transferred by the sensor. They also proposed a deep learning model
based on Auto-encoder network for separating and detecting malicious behavior.
Their experimental results, namely the true positive rate indicate that the classifier
gives acceptable results. However, the overall accuracy has not been mentioned by
the authors.
Yan et al. in [23] introduced a mechanism to detect DDoS attacks at all levels
of the IoT architecture. The proposed work consists of edge computing, and it
is mainly based on software defined networking (SDN) based Industrial Internet
of Things (IIoT) gateways (SDNIGW) which in turn uses several mechanisms to
protect the perception devices such as data encryption and fog computing. It also
includes controllers and methods to mitigate attacks by dealing with bandwidth and
detecting malicious programs after the IIoT management control unit (IMCU) scans
the network flow traffic data. Finally, cloud computing rely on deep learning model
for attack detection. Experimental results show that the mechanism is able to respond
to DDoS. However, the proposed approach has not been evaluated with regards to
the effectiveness (ex. Recall, Precision, etc.).
Lyu et al. in [24] proposed a mechanism that aims to deliver data reliably, by
detecting DoS attacks in WSN. The proposed work consists of trust-based geographic
opportunistic routing, selective authentication algorithm, and cooperative verification
scheme. Where a selective algorithm isolates the invalid data before sending the
packets and that based on a public and private key encryption at each sensor node
level in order to verify and sign the packets. The trust-based geographic opportunistic
routing is used to characterize the trust of wireless links based on the SSI-based
trust model. The opportunistic routing and candidate selection forwarder are used
to ensure high and reliable delivery of data. Finally, they used the notification and
verification mechanism to isolate illegal signatures quickly through the exchange of
verification information between candidates by relying on cooperative verification
scheme. Experimental results show that the protocol SelGOR works on high and
reliable delivery of data at a Low computational cost and also resists to DoS attacks.
However, the end-to-end delay could become quite long when a high node verification
probability is decided.
B. Ransomware Attack
Ransomware are one of the most dammaging type of malware. A malware is a

computer program that is designed to accomplish unauthorized actions without
the user’s consent [25]. A Ransomware encrypts the data or prevent users from
accessing their system [13]. Sajjan and Ghorpade [26], classfied ransomware attacks
in two categories namely Crypto Ransomware and Locker Ransomware. Crypto
Ransomware encrypts the users’ data and asks for a ransom in exchange of the
decryption key [26]. Locker Ransomware takes the control of a device by disabling
it or changing its behavior, which makes the attack more destructive [27].
Azmoodeh et al. in [28] proposed a technique for detecting crypto-ransomware
in internet of things (IoT) networks based on energy consumption fingerprint using
algorithms or workbooks to identify the class of each energy-consuming series of the
samples used for them. The K-Nearest Neighbor (KNN) learning algorithm is used
on a data set for predictive work that is based on a dynamic time warping (DTW) for
its ability to deal with the distortions of time between two series in the time adopted.
Support Vector Machine (SVM), Neural Network (NN), and Random Forest (RF)
algorithms which give accurate prediction constructing on a set of decision trees.
In their work, the authors focused on how to improve the performance of these
algorithms by assuming the size of a window (time value) and add sub-samples after
splitting samples using energy into sub-samples and using database of sub-sample
the sub-sample is trained to determine the sample class, after observing the results of
the experiment we see that the proposed work has a high detection rate with a good
accuracy rate. After observing the experimental results, we see that the proposed
work contains a high detection rate and a good accuracy rate, but it has not been
applied in a real environment. It is clear to us that when it is deploy in a real-world
IoT network.
Also Dash et al. in [29] proposed model to deal with the ransomware attack
using machine learning in internet of things by analyzing the flow of network traffic
where when relying on accurate algorithms and dealing with the huge amount of
malicious files. Authors used K-Nearest Neighbors (KNN), Support Vector Machines
(SVM), Random Forest, Neural Network that contribute to decision-making, to better
predict, and to solve problems of classification and regression. When algorithms
are subjected to any evaluations when analyzing network packets or dealing with
traffic, ransomware is detected by the K-Nearest Neighbor algorithm based on the
value of k which have an important role in predict, ransomware are highlighted by
building decision trees by Random Forest. The results of the evaluation showed that
the proposed approach is effective and accurate in detecting ransomware. However,
relying on ML may generate false positives when the data sets that have been used
are small.
C. Sinkhole Attack
Sinkhole attack is intended to make the sensors send imaginary signals to higher
layers (the processing layer) in the network by creating a malicious node to attract
traffic [34]. This node is centered in the place where it is easy to withdraw the packets
destined to the base station and then participate in the routing process for the attack
to be imposed [30, 31].
In 2016, Singh and Kaur [32] proposed an approach for detecting a sinkhole
attack deployed in area containing a limited number of wireless sensors. Using the
AODV routing protocol, the best path to send packets between the source and the
destination was determined when calculating the resulting delay and knowing the
distance of the adjacent node from the source node. Based on that, they could find out
the location of the compromised node. The algorithm evaluation showed its ability
to improve rate of packets which are successfully delivered (Throughput) and packet
loss while providing higher energy efficiency. However, we see that relying on the
delay indicator in detecting the malicious node, especially if it is long may lead to
waste of network resources before isolating the malicious node.
In 2013, Shahraki et al. [33] proposed a mechanism to detect sinkhole attack in
the wireless sensor networks based on two stages: The firs one consists in finding
suspicious nodes and then locate the intruder (malicious node). The same algorithm
was also used to collect and analyze the consistency of the network data in the
target area by calculating the attack zone and obtaining the data to be renewed in
the sensors. The results of the evaluation showed that the algorithm is accurate in
detecting the harmful node however; they are particularly weak when there is an
increase in colluding nodes.
In 2018, Chowdhury and Kiswani [34] proposed a mechanism to detect routing
attacks by applying two algorithms namely Key Match Algorithm (KMA) and
Cluster Based Algorithm (CBA). The latter algorithms were designed to address
attacks targeting DODAG (A Destination Oriented Directed Acyclic Graph), which
is considered as the core of RPL (Routing Protocol for Low Power and Lossy
Networks). These algorithms allow obtaining coverage, detection and prevention
of the suspect nodes in the network. The algorithms were evaluated on two types of
routing attacks namely: Selective Forwarding attack and sinkhole attack. The exper-
imental results showed the ability of the proposed solution to detect the location
and block of the harmful node. However, the proposed work lacks of performance
evaluation measures such as accuracy, precision, etc. (Table 1).
3 Discussion
After reviewing the literature works produced in the last five years. It turns out that
IoT technology still suffers from many problems and threats. In this paper, we have
presented a study of the most common attacks targeting Internet of things.
We also analyzed the proposed solutions in these literature namely Ransomware
attacks [28, 29], Sinkhole attack [32–34], and DoS attack [18–24], and thus can
provide some suggestions in order to improve their efficiency.
For instance, for the Ransomware case, we suggested to rely on the behavior
model of the malware instead of the file’s characteristics. For sinkhole attacks, we
suggest the use opportunistic routing and authentication algorithm to address the
issue related to the increase of the colluding nodes. As for DoS attack, we suggest
imposing on each node a public key and private key to verify and sign data packets
to isolate counterfeit packages and establish the behavior model of the attack that
target several layers, including the perception layer. For example, Wireless Sensor
Networks (WSN), which are used in various applications such as smart health and
military applications and traffic management, which collect data, and often these data
are sensitive, and their loss could lead to material and even human losses especially
in light of the integrate of these sensors with smart devices and connect them to
the internet (IoT), which makes it easier for the attacker to penetrate these networks
(WSN) and access to information.
Table 1 Summary of the existing works
1140
Attack Work Year Description Advantage Disadvantage Suggestions

DoS Attack Lyu et al. [24] 2019 Uses the SelGOR protocol, High (data) and reliable The end-to-end delay could Modification of the value of
which is based on trust-based delivery of data at a low become quite long when a the node verification
geographic opportunistic computational cost high node verification probability using the linear
routing and selective probability is decided function
authentication algorithm to
separate invalid from
legitimate data. Cooperative
verification scheme based on
notification and verification is
used to isolate illegal
signatures through the
exchange of verification
information between
candidates
Xia et al. [22] 2019 Based on a deep learning The AEN have a high The system performances we see using machine
model (Auto-encoder detection rate and a low false should be evaluated using the learning approach for the
network) for detecting positives rate accuracy metric IEEE 802.11 wireless
malicious behavior based on network, can detect more
data collected from Arduino IoT devices anomalies
senso
Pu [21] 2019 Records the number of Low energy consumption It is not clear how the Imposing on each node a
received and redirected dynamic misconduct threshold public and a private key to
packets in each adjacent node is obtained verify and sign data packets
in the OT and DT during the to isolate illegitimate ones
control window.
Based on the misconduct
threshold, the network’s
harmful node is detected and
isolated
(continued)
H. Belkhiri et al.
Table 1 (continued)
Yan et al. [23] 2018 Based on control units and Is able to respond to DDOS Lacks of evaluating metrics Use blockchain technologies
addresses attacks at all levels attacks at all levels of the IoT such as Recall, Accuracy, etc. in
of IoT’s architecture, such as architecture the proposed framework
data encryption, malware
detection, and uses a deep
learning model for DDoS
attack detection
Nayak et al. 2017 Uses both anomaly detection IDS unit collects the data Lacks detection indicators on Install IDS in all sensor
[18] and signature-based detection accurately, and analyzes it to the malware(TPR,FPR…)and nodes to avoid sinks received
on WSN network behavior find out the behavior of each performance evaluation a lot of invalid data from
data sensor node measures(Accuracy, Recall…) DoS attackers
Security in the Internet of Things …
Lee et al. 2015 Depending on a lightweight The system is efficient and Lacks detection indicators on Using more evaluation
[20] system to detect intrusions in accurate the malware (True positive metrics (TPR, FPR, etc.) can
the 6LoWPAN network. and negative) allow a better assessment of
Based on the exhaustion of the effectiveness
energy from the compromised performance of the proposed
node to detect attacks method can be obtained
Kasinathan et al. [19] 2013 Monitoring the traffic and High true positive rate The mechanism deals with Using centralized monitoring
observing the flow and only one type of attacks (DoS) and intrusion prevention
consistency of network data systems, to deal with the
reported by IDS in 6LoWPAN different types of attacks
networks through a
penetration test (UDP flood
attack)
Ransomw-are attacks Dash et al. in [29] 2018 Based on well-known machine ML offers high detection Uses a relatively small dataset Using a larger dataset and try
learning algorithms such as accuracy to establish the behavior
SVM, KNN, Random Forest, model of ransomware attacks
and Neural Network applied
on network data traffic flow
generated by the ransomware
(continued)
1141
Table 1 (continued)
1142

Azmoodeh et al. [28] 2017 crypto-ransomware was High detection and accuracy It is not applied in a real we see deploying proposed
detected based on energy rates environment approach in a real-world IoT
consumption to distinguish network contributes to
ransomware from consequently better ratings
non-malicious applications we have the ability to
fingerprint using algorithms or improve the detection rate,
workbooks.The sequence of especially the accuracy rate
applications’ energy
consumption is splitted into
several sequences of power
usage subsamples, which are
then classified to build
aggregated subsample’s class
labels
Sinkhole attack Chowdhury and Nis Kiswani 2018 Applying KMA and CBA Detection and isolation of The proposed work lacks Combine both detection
[34] algorithms to address the the malicious node performance evaluation techniques(KMA and CBA)
attack targeting the measures such as accuracy, To increase
Destination Oriented Directed precision, etc Increase the true positive rate
Acyclic Graph (DODAG) of
Routing Protocol for
Low-Power and Lossy
Networks (RPL). These
algorithms work to obtain
coverage, detection and
prevention of the suspect node
in the network
Singh and Kaur [32] 2016 Using the Ad Hoc On-demand Provides higher power Relying on the delay indicator Routing contributes to
Distance Vector Protocol efficiency in detecting the malicious increased productivity and
(AODV) to determine the best node may lead to waste the thwarting the attack
path to send packets between network resources before
the source and the destination isolating the malicious node
based on the calculated delay
and the distance of the
adjacent node from the source
node. This allowed locating
the compromised node
(continued)
H. Belkhiri et al.
Table 1 (continued)
Shahraki et al. [33] 2013 Finding Suspicious nodes and Locating the malicious node Weak success rate in intruder Use of opportunistic routing
then locate the intruder accurately identification particularly
(malicious node) This low and when increasing colluding
effective algorithm was also nodes
used to collect and analyze the
consistency of the network
data in the target area by
calculating the attack zone
and obtain the data to be
renewed in the sensors
Security in the Internet of Things …
1143
4 Conclusion
The rapid development of Device and smart objects linked to each other and related
to the Internet (IOT) has led to an increase in the data rate; this data may affect the
privacy of users in the absence of security. This paper provides an overview of the
most common attacks targeting the internet of things, the existing solutions, and also
our recommendations to improve these solutions. When reviewing the literature, it
became clear to us how dangerous and divers are these attacks and therefore it is
necessary to have proper mechanisms that contribute in addressing these security
issues to ensure the protection of these systems and ensure the reliability of data.
As a perspective, we are working on providing our own solution for IoT devices
security, and due to the importance of WSNs, this solution will focus on DoS attacks
detection in WSNs.
References
1. https://itsocial.fr/innovation/objets-connectes/quest-liot-definition-deliot-linternetobjets/
2. Madakam S, Ramaswamy R, Tripathi S (2015) Internet of Things(IoT): a literature review. J
Comput Commun 3:164
3. Cisco Internet Business Solutions Group (IBSG) (2011) The Internet of Things—how the next
evolution of the internet is evolution of the internet is changing everything. CISCO white Paper,
pp 1–11
4. Greengard S (2015) The Internet of Things. MIT Press
5. Čolaković A, Hadžialić M (2018) Internet of Things (IoT): a review of enabling technologies,
challenges, and open research issues. Comput Netw 144:17–39
6. Kouicem DE, Bouabdallah A, Lakhlef H (2018) Internet of things security: a top-down survey.
Comput Netw 141:199–221
7. Alansari Z et al (2018) Internet of Things: infrastructure, architecture, security and privacy.
In: 2018 international conference electronics and communications enginneering (ICCECE).
Southend, United Kingdom, pp 150–155
8. https://transformation-digitale.pwc.fr/data/la-bonne-architecture-iot
9. Marques G, Garcia N, Pombo N (2017) A survey on IoT: architect ures elements, applications,
QoS, platforms and security concepts. In: Mavromoustakis CX, Mastorakis G, Dobre C (eds)
Advances in mobile cloud computing and big data in the 5G era, vol. 22. Springer International
Publishing, Cham, pp 115–130
10. Kumar V, Sharma DN (2017) Internet of Things: security challenges and there proposed
solutions. Int J Eng Technol Comput Res 7
11. Sharma SK, Wang X (2017) Live data analytics with collaborative edge and cloud processing
in wireless IoT networks. IEEE Access 5:4621–4635
12. Atzori L, Iera A, Morabito G (2010) The Internet of Things: a survey. Comput Netw
54(15):2787–2805
13. https://www.trendmicro.com/vinfo/us/security/definition/ransomware
14. https://searchsecurity.techtarget.com/definition/distributed-denial-of-service-attack
15. Chhabra M, Gupta B, Almomani A (2013) A novel solution to handle DDOS attack in MANET.
J Inf Secur 04(03):165–179
16. Venkatesan R, Kumar GA, Nandhan MR (2018) A novel approach to detect Ddos attack
through virtual. In: 2018 IEEE international conference on system, computation, automation
and networking, pp. 1–6
17. Venkatesan R, Devi DR, Keerthana R, Kumar AA (2018) A nove approach for detecting
DDOSattack in hids using. In: 2018 IEEE international conference on system, computation,
automation and networking, pp 1–5
18. Nayak P (2018) A review on DoS attack for WSN : defense and detection mechanisms. In:
2017 international conference on energy, communication, data analytics and soft computing,
pp 453–461
19. Kasinathan P, Pastrone C, Spirito MA, Vinkovits M (2013) Denial-of service detection in
6LoWPAN based Internet of Things. In: International conference on wireless and mobile
computing, networking and communications, pp 600–607
20. Lee T, Wen C, Chang L (2014) A lightweight intrusion detection scheme based on energy
consumption analysis in 6LowPAN,” p. file: ///C:/Users/yy/Desktop/VERSION 2 PDF/IEEE-
Che
21. Pu C (2019) Energy depletion attack against routing protocol in the Internet of Things, pp 1–4
22. Xia F, Song H, Xu C (2019) Securing the wireless environment of IoT. In: 2018 IEEE
international conference of safety produce informatization, pp 315–318
23. Yan Q, Huang W, Luo X, Gong Q, Yu FR (2018) A multi-level DDoS mitigation framework
for the industrial Internet of Things. IEEE Commun Mag 56(2):30–36
24. Lyu C, Zhang X, Liu Z, Chi CH (2019) Selective authentication based geographic opportunistic
routing in wireless sensor Networks for Internet of Things against DoS attacks. IEEE Access
7:31068–31082
25. Belaoued M, Mazouzi S (2016) A chi-square-based decision for real time malware detection
using PE-file features. J Inform Process Syst 12(4):644–660
26. Sajjan RS, Ghorpade VR (2017) Ransomware attacks: radical menace for cloud computing,
pp 1640–1646
27. Yaqoob I et al (2017) The rise of ransomware and emerging security Challenge in the Internet
of Things. Comput Netw 129:444–458
28. Azmoodeh A, Dehghantanha A, Conti M, Choo KKR (2018) Detec-ting crypto-ransomware
in IoT networks based on energy consumption footprint. J Ambient Intell Humaniz Comput
9(4):1141–1152
29. Dash A (2018) Ransomware auto-detection in IoT devices using machine learning, pp 0–10
30. Krontiris I, Giannetsos T, Dimitriou T (2008) Launching a sinkhole attack in wireless sensor
networks; The intruder side, pp 526– 531
31. Abduvaliyev A, Pathan AK, Zhou J, Roman R, Wong W, On the vital areas of intrusion detection
systems in wireless senso networks, pp 1–16
32. Kaur M (2016) Detection and mitigation of sinkhole attack in wireless sensor network, pp
218–222
33. Shahraki S, Razzaque ASSMA, Naraei P, Farrokhtala A (2013) Detection of sinkhole attack
in wireless sensor networks
34. Choudhary S (2018) Detection and prevention of routing attacks in Internet of Things. In:
2018 17th IEEE international conference on trust, security and privacy in computing and
communications/ 12th IEEE international conference on big data science engineering, vol 201,
pp 1537–1540
Characterization of Melanoma Using
Convolutional Neural Networks
and Dermoscopic Images
Abdelghani Tafsast, Mohamed Laid Hadjili, Ayache Bouakaz,

and Nabil Benoudjit
Abstract Characterization of melanoma from moles in skin lesions using dermo-

scopic images, guides to an efficient treatment. In this study a new approach to
analyze dermoscopic images and characterize melanoma from other benign skin
lesions, using Convolutional Neural Networks (CNN), is implemented. The database
employed in this experimental study consists of dermoscopic skin lesion images. This
database contains malignant lesions called melanomas and benign skin lesions called
moles. In order to optimize the characterization of melanoma, a data augmentation
process is used. The best accuracy of melanoma classification obtained in this study
is equal to 81.6% comparing to 75.6% obtained using a known CNN model called
LeNet-5. These results show that the suggested CNN model works well for melanoma
characterization using dermoscopic images.
Keywords Melanoma characterization · Dermoscopic images · Benign skin

lesion · Malignant skin lesion · Convolutional neural networks · CNN
1 Introduction
Melanoma cancer develops from the pigment-containing cells of the skin called
melanocytes [1]. Melanocytes can also form non-cancerous growths called moles.
Melanomas typically take place in the skin, but may rarely take place in the
intestines, eye, or mouth [1, 2]. For women, they most commonly tack place on the
legs, while for men they usually occur on the back [2]. Sometimes, melanomas grow
A. Tafsast (B) · N. Benoudjit

Laboratoire d’Automatique Avancée et d’Analyse des Systèmes, Université Batna 2, Batna,
Algeria
M. L. Hadjili
Ecole Supérieure d’Informatique, (HE2B-ESI), Brussels, Belgium
A. Bouakaz
UMR Inserm U1253 Imagerie et Cerveau, Université de Tours, Tours, France

https://doi.org/10.1007/978-981-15-6403-1_80
1148 A. Tafsast et al.
from a mole with changes like irregular boundaries, change in color, increase in
shape, or skin breakdown [1].
The main cause of melanoma is Ultraviolet light (UV) coverage in those have low
levels of skin pigment [2–5]. The UV light may be from the sun or either from other
sources like tanning devices [2]. About 25% of melanomas develop from moles [2].
Those who have poor immune function, with many moles, are at greater risk [1].
For the treatment, the melanoma is usually removal by surgery [1]. In those with
a little larger cancer, nearby lymph nodes may be analyzed for propagation [1]. Most
people are cured if the spread did not occur [1]. When melanoma has spread, biolog-
ical therapy, immunotherapy, chemotherapy or radiotherapy may improve survival
[1, 6].
For the diagnosis of melanoma, the visual examination is the most frequent diag-
nostic technique [7]. Analysis of any skin lesion that has symptoms of being poten-
tially malignant is important [1]. Moles that are irregular in shape or color are usually
treated as candidates. To identify melanomas, it is suggested to recognize them, to
frequently examine moles for changes such as size, shape, or color, and to consult a
qualified physician when a candidate appears [8].
Early symptoms of malignant lesion are changes to the color or shape of existing
moles, or the appearance of a new lump anywhere on the skin in the case of
nodular melanoma. Early signs of malignant lesion are summarized by the mnemonic
ABCDEF for Asymmetry, Borders, Color, Diameter, Evolving over time, and Funny
looking [9, 10]. This classification is not applicable to nodular melanoma, which has
its own classifications summarized by the mnemonic EFG for Elevated above the
skin surface, Firm to the touch, and Growing [11].
Characterization of melanoma from benign moles presents a big challenge. For
instance, many melanomas present as lesions smaller than the recommended diameter
length using ABCDEF criteria, and all melanomas are malignant when they first come
out as a small dot. In Addition, physicians usually examine all moles, including those
less than the recommended diameter length. For example, Seborrheic keratosis may
assemble some or all of the ABCD criteria, and can lead to false diagnosis. Doctors
can usually differentiate seborrheic keratosis from melanoma with dermatoscopy or
upon examination.
Moreover, as pigmented lesions occurring on the skin, melanoma is willing to
early discovery by expert visual inspection. It is also willing to automated discovery
with image analysis. Given the widespread accessibility of high-resolution cameras,
algorithms that can improve the ability to display and discover troublesome lesions
can be of great value.
For this reason, the use of Artificial Intelligence (AI) for the pattern recognition
in medical engineering is widely preferred, especially for the diagnosis and the
characterization of diseases [12].
Furthermore, the use of deep neural networks such as Convolutional Neural
Networks (CNNs) for the diagnosis using medical images or signals is remarkable
due to the good results obtained [13–15].
In this study we suggest an alternative tool to identify melanomas from benign
moles using dermoscopic images. This system can lead to reduce errors made by
Characterization of Melanoma … 1149
physicians in some cases. The remaining of this manuscript is ordered as follows:

Sect. 2 describes the methods and materials used in this study. Section 3 provides
the results and discussions. The last section presents the conclusion of this study.
2 Materials and Methods
A. Database description
Dermoscopy is an imaging technique that eliminates the surface reflection of skin.

Enhancing the visualization of deeper levels of skin can be made by removing surface
reflection. Prior research has shown that dermoscopy provides improved diagnostic
accuracy in comparison to standard photography, when used by expert dermatologists
[16].
As mentioned in the database description [16, 17], 900 of benign and malignant
skin moles are included in this study. We take 450 samples for the training of the
CNN model and we leave 450 samples for the test. The training set is also divided
into two subsets. 300 samples are taken for the training and 150 samples are left for
the validation and the evaluation of the trained CNN model. The next figure shows
random samples images of both benign and malignant lesions (Fig. 1).
B. Convolutional neural network (CNN)
A CNN consists of an input, an output, and multiple hidden layers. The hidden layers
of a CNN usually consist of convolution layers, activation function layers, subsam-
pling (pooling) layers and fully connected layers. The Fig. 2 shows a well known
CNN called LeNet-5 suggested by LeCun et al. [18]. This CNN model is based on
the convolution process. It is recognized after winning the ImageNet Large Scale
Visual Recognition Challenge (ILSVRC) in 2012 [19, 20]. Since this year, the CNN
appears more than other Artificial Neural Networks (ANN) for the classification of
images. LeNet-5 consists of 4 hidden layers; two pairs of convolution and subsam-
pling layers and two fully connected layers. In LeNet-5, the authors suggest using 6
masks with a size of 5 × 5 for the first pair of layers, and in the second pair, they
have chosen 16 masks with the same dimensions.
C. Suggested CNN model
To suggest new CNN model, a heuristic study is important. For this reason, we test
different topologies on the validation subset to choose the best one according to the
high accuracy (see Eq. 1) obtained. The validation subset is picked randomly from the
training dataset using the leave P out (LPO) technique. From this heuristic study we
suggest a new CNN model with an augmentation of the number of layers comparing
to LeNet-5 [18]. This process leads to a better features extraction. The suggested
CNN consists of 7 hidden layers comparing to 4 hidden layers for the LeNet-5 CNN
model. The first 6 hidden layers are sets of a convolution, an activation function
Fig. 1 Skin lesion types; A benign and B malignant
Fig. 2 CNN suggested by LeCun et al. [18] (LeNet-5)

Fig. 3 Suggested CNN model
using Rectified Linear Unit (ReLU) and a subsamlping (pooling) layer, while the
last hidden layer is a fully connected layer. The number of masks is equal to 5, 10
and 20 with a size of 7 × 7 for the first three hidden layers respectively, while the
number of masks is equal to 30, 40 and 50 with a size of 5 × 5 for the next three
hidden layers, as mentioned in the following figure. We use the fully connected layer
just after the last hidden layer to pile up the features extracted using the convolution
layers, ReLU layers and the pooling layers. Finally, a softmax layer is used to classify
the input image as benign or malignant mole (Fig. 3).
D. Data augmentation
Due to the small size of the database [16], and the advantage of using big databases
for the training of CNNs, we suggest augmenting the database size using different
adjustments on the original images. Hence, for each image we create 8 variants that
present the same lesion but with some adjustments for the reflections, translations,
rotations and scaling. So we reach 2400 images for the training set and we let the
validation set as it is with 150 images.
To evaluate the performance of the suggested CNN model, we use the classification
accuracy defined as:
TP +TN
A= , (1)
T P + T N + FP + FN
where, TP (True Positive) is the correct predicted of a label, TN (True Negative) is

the correct predicted of the other label, FP (False Positive) is the wrong predicted of
a label, and FN (False Negative) is the incorrect predicted of the other label.
Table 1 summarizes the percentage of the classification accuracy of melanoma on
the test set using a CNN trained on the training set without data augmentation, while
Table 1 Results obtained without data augmentation

Models Accuracy (%)
LeNet-5 73.78
Suggested CNN 79.33
Table 2 Results obtained with data augmentation

Models Accuracy (%)
LeNet-5 75.56
Suggested CNN 81.56
Table 2 illustrates the results obtained using a CNN trained on the training images
after applying the data augmentation process on the training set.
It is shown that the suggested CNN model gives the best results comparing to the
LeNet-5 in terms of classifying melanomas from benign moles.
The figures (Figs. 4 and 5) show the confusion matrices of the results obtained
on the images with data augmentation. In the Fig. 4, we find the overall accuracy
Confusion Matrix
328 35 90.4%
Benign
72.9% 7.8% 9.6%
Output Class
75 12 13.8%
Malignant
16.7% 2.7% 86.2%
81.4% 25.5% 75.6%

18.6% 74.5% 24.4%
nt
gn
na
ni
Be
ig
al
M
Target Class
Fig. 4 Confusion matrix for the LeNet-5 [18]

Confusion Matrix
351 12 96.7%
Benign
78.0% 2.7% 3.3%
Output Class
71 16 18.4%
Malignant
15.8% 3.6% 81.6%
83.2% 57.1% 81.6%

16.8% 42.9% 18.4%
t
n
n
ig
na
n
Be
ig
al
M
Target Class
Fig. 5 Confusion matrix for the suggested CNN
obtained using LeNet-5 CNN model which is equal to 75.6% (see the right down
square), as we find also the interclass accuracy which means the accuracy per class.
For the first row, the green square presents the number of images classified correctly
for the first label (benign), with the corresponding rate, while the red square shows
the number of images classified incorrectly, and the grey square presents the accuracy
per class (benign lesion). For the second row, the red square shows the misclassified
images for the second label (malignant), the green square presents the number of
images classified correctly and the grey square presents the accuracy per class for
the second label (malignant). The overall accuracy obtained using LeNet-5 is equal
to the sum of the diagonal (green) rates in the corresponding matrix.
In the Fig. 5, we find the results obtained using the suggested CNN model. As
explained for the previous figure, the overall accuracy obtained is equal to 81.6%.
From these confusion matrices, we find that the suggested CNN model gives the
best results even for the classification interclass.
Results from this study illustrate that the suggested CNN classifier system works
well for melanoma classification comparing to the LeNet-5 CNN model.
4 Conclusion
Melanoma classification presents a big challenge to enhance the treatment protocol.

In this study an automated system to characterize and classify melanoma was
investigated. The technique uses a convolutional neural network (CNN) to classify
melanoma in dermoscopic images.
Results obtained from this study confirm the superiority in terms of classification
accuracy of our suggested CNN classification model over the LeNet-5 model [18].
The obtained classification accuracy is about 81%. The technique, as mentioned in
this study, proves to be effective in improving melanoma characterization and iden-
tification from benign moles in skin lesions, using dermoscopic images. However,
further improvements, in terms of using transfer learning techniques on trained deep
models, are required to ameliorate the results obtained.
References
1. Melanoma treatment-for health professionals (PDQ®). National Cancer Institute. http://

www.cancer.gov/cancertopics/pdq/treatment/melanoma/HealthProfessional. Accessed Jully
20, 2019
2. World Health Organization (WHO). World cancer report, 2014. Chap. 5.14. ISBN 978-
9283204299
3. Goldstein BG, Goldstein AO (2001) Diagnosis and management of malignant melanoma. Am
Family Phys 63(7):1359–1374
4. Howlader N. SEER stat fact sheets: melanoma of the skin. National Cancer Institute website.
http://seer.cancer.gov/statfacts/html/melan.html. Accessed Jully 20, 2019
5. Kanavy HE, Gerstenblith MR (2011) Ultraviolet radiation and melanoma. 30, 4, 30(4):222–228
6. Syn NL, Teng MW, Mok TS, Soo RA (2017) De-novo and acquired resistance to immune
checkpoint targeting. Lancet Oncol 18(12):e731–e741
7. Wurm EM, Soyer HP (2010) Scanning for melanoma. Aust Prescriber 33(5):150–155
8. Friedman RJ, Rigel DS, Kopf AW (1985) Early detection of malignant melanoma: the role of
physician examination and self-examination of the skin. CA: Cancer J Clinic 35(3):130–151
9. CDC—What are the symptoms of skin cancer?. www.cdc.gov. Accessed Jully 21, 2019
10. Jensen JD, Elewski BE (2015) The ABCDEF rule: combining the “ABCDE Rule” and the
“Ugly Duckling Sign” in an effort to improve patient self-screening examinations. J Clinic
Aesthetic Dermatol 8(2):15
11. The EFG of Nodular Melanomas| MoleMap New Zealand. The EFG of Nodular Melanomas|
MoleMap New Zealand. Accessed Jully 21, 2019
12. Douak F, Tafsast A, Fouan D, Ferroudji K, Bouakaz A, Benoudjit N (2016) A wavelet opti-
mization approach for microemboli classification using RF signals. In 2016 IEEE international
on ultrasonics symposium (IUS). IEEE, pp 1–4
13. Tafsast A, Ferroudji K, Hadjili ML, Bouakaz A, Benoudjit N (2018) Automatic microemboli
characterization using convolutional neural networks and radio frequency signals. In: 2018
international conference on communications and electrical engineering (ICCEE). IEEE, pp
1–4
14. Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian M, Sánchez CI (2017) A
survey on deep learning in medical image analysis. Med Image Anal 42:60–88
15. Spanhol FA, Oliveira LS, Petitjean C, Heutte L (2016) Breast cancer histopathological image
classification using convolutional neural networks. In: 2016 international joint conference on
neural networks (IJCNN). IEEE, pp 2560–2567
16. ISIC 2016: Skin lesion analysis towards Melanoma detection. https://challenge.kitware.com/#
challenge/560d7856cad3a57cfde481ba. Accessed June 18, 2019
17. Gutman D, Codella NC, Celebi E, Helba B, Marchetti M, Mishra N, Halpern A (2016) Skin
lesion analysis toward melanoma detection: a challenge at the international symposium on
biomedical imaging (ISBI) 2016, hosted by the international skin imaging collaboration (ISIC).
arXiv:1605.01397
18. LeCun Y, Bottou L, Bengio Y, Haffner P (1998) Gradient-based learning applied to document
recognition. Proc IEEE 86(11):2278–2324
19. Ciresan DC, Meier U, Masci J, Gambardella LM, Schmidhuber J (2011) Flexible, high perfor-
mance convolutional neural networks for image classification. In: Twenty-second international
joint conference on artificial intelligence
20. Ciregan D, Meier U, Schmidhuber J (2012) Multi-column deep neural networks for image
classification. In: 2012 IEEE conference on computer vision and pattern recognition. Institute
of Electrical and Electronics Engineers (IEEE), New York, pp 3642–3649
Advanced Feedforward-and-Feedback
Decorrelation Algorithms for Speech
Quality Enhancement
Rédha Bendoumia
Abstract Recently, efficient speech enhancement and noise reduction techniques

based on two-channel adaptive filtering algorithms have been proposed. Feedforward
(Forward) and Feedback (Backward) blind-source-separation (BSS) structures have
been combined with symmetric adaptive decorrelation (SAD) algorithm using the
fixed-step-size. The main drawback of the SAD algorithms is theirs poor performance
in steady state regime when the fixed step-sizes are selected large. In this study, we
propose two improved versions of the recently proposed Forward-and-Backward
normalized SAD algorithms based on variable step-size approach. In these new
versions, we propose an efficient estimation of optimal step-sizes values, basing on
the minimization of the cross-correlation function of error signal. To confirm the
good performance of proposed algorithms, we have done several simulations in very
noisy situations using system mismatch and output segmental signal-to-noise ratio
criteria.
Keywords Feedforward-and-Feedback · Speech enhancement · Decorrelation

algorithm · Variable step-size
1 Introduction
Adaptive filtering algorithms are frequently employed in digital signal processing,

telecommunications and many other applications because of its robustness [1–3].
Recently, a very important amount of works have investigated in acoustic noise reduc-
tion (ANR) and speech quality enhancement (SQE) applications by using different
adaptive filtering algorithms combined with the two-channel blind-source-separation
structures [4–8].
R. Bendoumia (B)
Department of Electronics, Laboratory LATSI, University of Blida 1, Blida, Algeria

https://doi.org/10.1007/978-981-15-6403-1_81
1158 R. Bendoumia
These approaches are proposed to improve the behavior of ANR and SQE systems
in terms of speed convergence, steady state, misadjustment values and complexity.
These smart combinations between these two structures and these adaptive filtering
algorithms families have given a new insight in the field of acoustic noise reduc-
tion and speech intelligibility enhancement. For example in [5, 7], they are used the
Symmetric Adaptive Decorrelation (SAD) algorithm for signal separation. These
two-channel adaptive filtering algorithms are used to identify the two impulse
responses (IRs) of two-channel convolutive mixture by using the fixed step-size
parameters [9–12].
In this paper, firstly we present a modified forward-and-backward decorrela-
tion algorithm (FND and BND) by inserting normalized step-sizes parameters [13].
Secondly, we propose an implementation of efficient variable step-sizes on (i) modi-
fied FND algorithm (denoted: VS-FND), and (ii) modified BND algorithm (denoted:
VS-BND). These algorithms are proposed for acoustic noise reduction and speech
quality to improve the convergence rate and to enhance the quality of the estimated
output signal.
This paper is presented as follows: in Sect. 2, the two-channel mixing model is
presented. In Sect. 3, we describe the forward SAD algorithm and its normalized
version. Section 4 is reserved to present the classical and modified backward SAD
algorithms. The proposed variable step-sizes algorithms are presented in Sect. 5. The
simulation results are presented in Sect. 6 and finally the conclusion is presented in
Sect. 7.
2 Two-Channel Mixing Systems Problem
In this paper, we consider a convolutive mixing model of two uncorrelated sources

that is shown in Fig. 1 [4–8]. We use a original speech signal s(n) and a punctual
noise signal b(n) in the input of the two-channel mixing model, where h12 (n) and
h21 (n) are the reel impulse responses (IR). In the output of this mixing model, the
observed noisy signals are given as:
Fig. 1 Two-channel convolutive mixing system

Advanced Feedforward-and-Feedback … 1159
p1 (n) = s(n) + b(n) ∗ h21 (n) (1)
p2 (n) = b(n) + s(n) ∗ h12 (n) (2)
3 Forward Normalized Decorrelation Algorithm
The basic symmetric adaptive decorrelation (SAD) algorithm combined with forward
BSS structure, is firstly proposed in [5, 7].
We assume the generating signals s(n) and b(n) to be zero mean and statistically
independent. This implies that they are uncorrelated, i.e. E[s(n)b(n − l)] = 0,∀l.
Only this latter is required for forward SAD algorithm (FSAD) to work. The
description of FSAD algorithm is presented in Fig. 2.
In [5, 7]; it has been proven that the quadratic error minimization is completely
equivalent with the decorrelation between estimated output signals u1 (n) and u2 (n),
with the noise reference presents on the observation p2 (n) and p1 (n) respectively
over the span of adaptive filters:
Cu1 u2 (l) = E[u1 (n)u2 (n − l)] and;

Cu2 u1 (l) = E[u2 (n)u1 (n − l)]
The exact update relations of adaptive filters w12 (n) and w21 (n) by the FSAD
algorithm are given as follows [7]:
w21 (n) = w21 (n − 1) + λ21 u1 (n)u2 (n) (3)
w12 (n) = w12 (n − 1) + λ12 u2 (n)u1 (n) (4)
Fig. 2 A description of basic forward SAD algorithm [7]

1160 R. Bendoumia
Fig. 3 A description of basic BSAD algorithm [13]
λ12 and
λ21 are the step-sizes. They are chosen according to the relations 0 <
λ12 < 2 σ12 and 0 < λ21 < 2 σ22 , where σ12 and σ22 represents the variance of the
two input signals u1 (n) and u2 (n) [7]. In the output of FSAD algorithm presented
in Fig. 3, we can define respectively the estimated speech signal u1 (n), and it a
posteriori error signal e1 (n) in the forward structure as
u1 (n) = p1 (n) − w21

T
(n − 1)p2 (n) (5)
e1 (n) = p1 (n) − w21

T
(n)p2 (n) (6)
When the update formulas of the adaptive filter w21 (n) by the FSAD algorithm is
given as follow:
w21 (n) = w21 (n − 1) + λ21 u1 (n)u2 (n) (7)
Using (5), and inserting (7) into (6), we obtain, e1 (n) = u1 (n)+w21
T
(n − 1)p2 (n)−
− 1) + λ21 u1 (n)u2 (n)] Only, in noise-present segments, e1 (n) = 0, we
pT2 (n)[w21 (n
obtain

u1 (n) 1 − λ21 pT2 (n)p2 (n) = 0 (8)

Using (8) and assuming that u1 (n) = 0 → 1 − λ21 pT2 (n)p2 (n) = 0 , the step-size
λ21 is given by
1
λ21 = (9)
pT2 (n)p2 (n)
Using the same development steps but in the symmetric filter w12 (n) of forward
structure, the step-size λ12 is defined as
1
λ12 = (10)
pT1 (n)p1 (n)
By inserting [(9) and (10), into (3) and (4)] and incorporating respectively the new
normalized step-sizes, denoted: µ12 and µ21 . The two update formulas of w12 (n) and
w21 (n) of the forward normalized decorrelation algorithm are given by
u2 (n)u1 (n)
w12 (n) = w12 (n − 1) + μ12 (11)
p1 (n)p1 (n) +
T
ε
u1 (n)u2 (n)
w21 (n) = w21 (n − 1) + μ21 (12)
p2 (n)p2 (n) +
T
ε
Where ε represents a very small constant to avoid the division by zeros in the
absence speech signal. When μ12 and μ21 take their values between 0 and 2 to
guarantee convergence of w12 (n) and w21 (n) [4].
4 Backward Normalized Decorrelation Algorithm
In second part of this paper, we will focus our interest on the two-channel backward
SAD algorithm with normalized step-sizes parameters that is developed in the next
section [13].
The update equations of w12 (n) and w21 (n) by the backward decorrelation
algorithm are given as follows [13]:
w21 (n) = w21 (n − 1) + λ21 v1 (n)v2 (n) (13)
w12 (n) = w12 (n − 1) + λ12 v2 (n)v1 (n) (14)
In the output of BSAD algorithm presented in Fig. 3, we can define respectively the
estimated speech signal v1 (n), and it a posteriori error signal e1 (n) in the backward
structure as
v1 (n) = p1 (n) − w21

T
(n − 1)v2 (n) (15)
e1 (n) = p1 (n) − w21

T
(n)v2 (n) (16)
When the update formulas of the adaptive filter w21 (n) by the BSAD algorithm
is given as follow:
w21 (n) = w21 (n − 1) + λ21 v1 (n)v2 (n) (17)

1162 R. Bendoumia
Using (15), and inserting (17) into (16), we obtain,
e1 (n) = v1 (n) + w21

T
(n − 1)v2 (n) − v2T (n)[w21 (n − 1) + λ21 v1 (n)v2 (n)
Only, in noise-present segments (silence periods of speech), e1 (n) = 0, we obtain

v1 (n) 1 − λ21 v2T (n)v2 (n) = 0 (18)

The estimated speech v1 (n) = 0 → 1 − λ21 v2T (n)v2 (n) = 0 , the step-size λ21
is given by
1
λ21 = (19)
v2T (n)v2 (n)
By using the same development steps but in symmetric filter w12 (n) of backward
structure, the step-size λ12 is defined as
1
λ12 = (20)
v1T (n)v1 (n)
By inserting [(19) into (13), (20) into (14)] and incorporating respectively the new
normalized step-sizes, denoted: μ12 and μ21 .
The two formulas of w12 (n) and w21 (n) in the modified normalized decorrelation
algorithm are updating by the following relations:
v2 (n)v1 (n)
w12 (n) = w12 (n − 1) + μ12 (21)
v1 (n)v1 (n) + ξBND
T
v1 (n)v2 (n)
w21 (n) = w21 (n − 1) + μ21 (22)
v2 (n)v2 (n) + ξBND
T
µ12 and µ21 are normalize respectively by the

In BND algorithm, the step-sizes
input signals energy v1T (n)v1 (n) and v2T (n)v2 (n) of adaptive filters w12 (n) and
w21 (n). µ12 and µ21 take their values between 0 and 2 to guarantee convergence of
two adaptive filters w12 (n) and w21 (n) [4].
5 Proposed Variable Step-Sizes Normalized Decorrelation

Algorithms
In this section, we describe all formulations of the proposed variable step-size normal-
ized decorrelation algorithms. In these proposed algorithms, we update the two
variable-step-sizes. A great deal of the proposed variable step-size decorrelation
algorithms is to improve the convergence rate of the adaptive filters and reducing
the distortion of the estimated output signals (minimizing the steady-state estimation
error).
In the following, we presents the proposed variable step-size algorithms which
are mainely based on minimization of μ12 (n) and μ21 (n) using recursive estimations.
Noting that the values of these step-sizes are bounded by (μ12,min and μ12,max ,
μ21,min and μ21,max ) to guarantee the adaptive filters convergence of w12 (n) and
w21 (n) respectively, i.e.
0 < μ12,min ≤ μ12 (n) ≤ μ12,max < 2 (23)
0 < μ21,min ≤ μ21 (n) ≤ μ21,max < 2 (24)
A. Proposed variable step-size FND algorithm (VS-FND)
The detailed scheme of the VS-FND algorithm is presented by Fig. 4. To improve

the output estimated speech signal intelligibility of the FND algorithm, we propose
a simple modification by adapting then to apply the variable step-sizes technique.
The update equations of w12 (n) and w21 (n) of proposed variable step-size FND
algorithms are given by:
u2 (n)u1 (n)
w12 (n) = w12 (n − 1) + μ12 (n) (25)
p1 (n)p1 (n) +
T
ε
Fig. 4 A description of VSS forward normalized SAD algorithm

1164 R. Bendoumia
u1 (n)u2 (n)
w21 (n) = w21 (n − 1) + μ21 (n) (26)
pT2 (n)p2 (n) + ε
The proposed and extended VS-FND algorithm is now controlled by the new
variable-step-sizes μ12 (n) and μ21 (n) which are controlled by following relations:
⎧
⎨ μ12,max if γ12 (n) > μ12,max
μ12 (n) = μ12,min if γ12 (n) < μ12,min (27)
⎩
γ12 (n) otherwise
⎧
⎨ μ21,max if γ21 (n) > μ21,max
μ21 (n) = μ21,min if γ21 (n) < μ21,min (28)
⎩
μ21 (n) otherwise
where γ12 (n) and γ21 (n) represent the controlled step-sizes:
γ12 (n) = α γ12 (n − 1) + βc12 (n) (29)
γ21 (n) = αγ21 (n − 1) + βc21 (n) (30)
where 0 < α < 1 and β > 0. c12 (n) and c21 (n) repre-
sent respectively the scalar estimations of the cross-correlation
functions between [u1 (n) and u1 (n − 1), . . . , u1 (n − M + 1)] and
[u2 (n) and u2 (n − 1), . . . , u2 (n − M + 1)]. They are given by:

M1
[u2 (n)u2 (n − k)]2
c12 (n) = λ c12 (n − 1) + (1 − λ) (31)
k=0
p1 (n)2 + ε

M−1
[u1 (n)u1 (n − k)]2
c21 (n) = λ c21 (n − 1) + (1 − λ) (32)
k=0
p2 (n)2 + ε
where λ is a control parameter that is given by 0 < λ < 1.

B. Proposed variable step-size BND algorithm (VS-BND)
The second proposed VS-BND algorithm is shown by Fig. 5. We now propose to

apply the new variable-step-size formulations on the two-channel backward decorre-
lation structure. The update equations of the two adaptive filters of proposed variable
step-size BND algorithms are given by:
v2 (n)v1 (n)
w12 (n) = w12 (n − 1) + μ12 (n) (33)
v1T (n)v1 (n) + ε
Fig. 5 Description of proposed variable step-size BND algorithms
v1 (n)v2 (n)
w21 (n) = w21 (n − 1) + μ21 (n) (34)
v2 (n)v2 (n) +
T
ε
where μ12 (n) and μ21 (n) present respectively the variable-step-sizes of adaptive
filters w12 (n) and w21 (n) that are estimated respectively by
⎧
⎨ μ12,max if ψ12 (n) > μ12,max
μ12 (n) = μ12,min if ψ12 (n) < μ12,min (35)
⎩
ψ12 (n) otherwise
⎧
⎨ μ21,max if ψ21 (n) > μ21,max
μ21 (n) = μ21,min if ψ21 (n) < μ21,min (36)
⎩
ψ12 (n) otherwise
where ψ12 (n) and ψ21 (n) represent the controlled step-sizes in two-channel VSS
backward structure. These controlled step-sizes are estimated by
ψ12 (n) = ϕ ψ12 (n − 1) + σ d12 (n) (37)
ψ21 (n) = ϕ ψ21 (n − 1) + σ d21 (n) (38)

1166 R. Bendoumia
where 0 < ϕ < 1 and σ > 0. d12 (n) and d21 (n) repre-
sent respectively the scalar estimations of the cross-correlation
functions between [v1 (n) and v1 (n − 1), . . . , v1 (n − M + 1)] and
[v2 (n) and v2 (n − 1), . . . , v2 (n − M + 1)]. They are given by:

M1
[v2 (n)v2 (n − k)]2
d12 (n) = ζ d12 (n − 1) + (1 − ζ) (39)
k=0
v1 (n)2 + ε

M−1
[v1 (n)v1 (n − k)]2
d21 (n) = ζ d21 (n − 1) + (1 − ζ) (40)
k=0
v2 (n)2 + ε
where ζ is a control parameter that is given by 0 < ζ < 1.
In this simulations part, we consider the two-channel convolutive mixing model

presented in Fig. 1, and the two statistically independent source signals to generate
the two noisy signals at the output. These two point sources are specified as:
i. s(n) is a speech signal phonetically equilibrated, and
ii. b(n) is a punctual noise that can be white or colored.
All the obtained simulations results are performed at a sampling frequency, Fs =
8 kHz. In the two-channel mixing model [6, 15, 16], we have used impulse responses
h12 (n) and h21 (n) that are convoluted with the original signals (speech and noise) to
generate the two noisy signals p1 (n) and p2 (n). The length of impulse responses is
L = 512. The noisy signals are generated by this model, with input SNR equal –6 dB.
A. System mismatch evaluation
In order to evaluate the convergence time and SM values properties of proposed

algorithms, we have reported in Figs. 6 and 7, the SM evolution in the same
simulation.
From Figs. 6 and 7, we note the good convergence speed performance of the
proposed algorithms compared with modified versions. According to the obtained
results, we observe a fast convergence time of proposed algorithms. However, the
modified algorithms suffer from slow convergence time especially in the very noisy
observations.
B. Segmental signal-to-noise ratio evaluation
In this section, we show the obtained SegSNR results applied to the output speech
signal obtained by modified and proposed algorithms. We have done several exper-
iments according to the input SNR (SNR = –6 dB, 0 dB and 6 dB) by using three
20
FND with μ21 =0.2
10
FND with μ21 =1.2
0 Proposed VS-FND
-10
SM in dB
-20
-30
-40
-50
-60
0 100 200 300 400 500 600 700
Blocks of 128 samples
Fig. 6 SM evaluation of w21 (n) obtained by modified FND and proposed VS-FND algorithm
10
BND with μ 21 = 0.2
0 BND with μ 21 = 1.2
Proposed VS-BND
-10
-20
SM in dB
-30
-40
-50
-60
0 100 200 300 400 500 600 700
Blocks de 128 samples
Fig. 7 SM evaluation of w21 (n) obtained by modified BND and proposed VS-BND algorithm
types of noise (white, USASI, babble and street) with the length of the real and
adaptive filters is large, i.e. L = 512. We note that the obtained results by modified
and proposed algorithms are reported on Table 1 and Fig. 8.
From the obtained results presented in Table 1 and Fig. 8, we note the good
behavior of proposed algorithms. This behavior is noted with different situations,
noisy types and different input SNR. We conclude that the proposed algorithms have
1168
Table 1 Overall CD evaluation obtained by modified FND, BND, proposed VS-FND and proposed VS-BND algorithms. The bold values represent the best
ones
Noise type White USASI Street Babble
Input SNR in dB –6 0 6 – 6 0 6 – 6 0 6 – 6 0 6
SegSNR in dB FND 46,5 47,2 48,9 46,3 47,0 48,6 46,3 46,9 48,2 46,1 46,8 47,8
BND 47,2 48,7 49,8 47,1 48,5 49,7 47,0 47,3 49,4 47,8 48,2 48,7
Proposed VS-FND 48,6 49,2 49,9 48,3 49,1 50,6 48,4 49,1 50,3 48,1 49,8 50,8
Proposed VS-BND 49,6 50,3 50,9 49,3 50,2 51,6 49,4 50,0 51,3 49,2 50,9 51,9
R. Bendoumia
Fig. 8 Segmental SNR evaluation obtained by modified FND, modified BND, proposed VS-FND
and proposed VS-BND algorithms
given the higher SNR values, which means less distortion on the estimated speech
signal. As we note that the proposed VS-BND algorithm shows a best performance
compared with proposed VS-FND algorithm.
7 Conclusion
In this paper, we have proposed two new two-channel variable step-size forward-
and-backward normalized symmetric adaptive decorrelation algorithms (VS-FND
and VS-BND) that improve the behavior of its original fixed-step-size algorithms.
These two proposed algorithms are based on the estimation of optimal step-sizes by
using the decorrelation criterion between the two estimated outputs of each algorithm.
We have carried out several experiments to show the performance of each proposed
algorithm. The SNR criterion has proved the efficiency of these proposed algorithms
in different noisy situations. Basing on SM and SegSNR criteria, when we use the
proposed VS-FND and VS-BND algorithms, we obtain two optimal properties at
same time, faster convergence speed and best speech quality. Finally, we note that
the VS-BND algorithm has shown a slight superiority over VS-FND.
References
1. Al-Kindi MJ, Dunlop J (1989) Improved adaptive noise cancellation in the presence of signal
leakage on the noise reference channel. Signal Process 17(3):241–250
2. Buchner H et al (2005) A generalization of blind source separation algorithms for convolutive
mixtures based on second order statistics. IEEE Trans Speech Audio Process 13(1):120–134
3. Ozeki K, Umeda T (1984) An adaptive filtering algorithm using an orthogonal projection to
an affine subspace and its properties. Electron Commun Jpn 67-A,n(5):19–27
1170 R. Bendoumia
4. Bendoumia R, Djendi M (2015) Two-channel variable-step-size forward-and-backward adap-

tive algorithms for acoustic noise reduction and speech enhancement. Sig Process 108:226–244
5. Gervan SV, Compernolle DV (1995) Signal separation by symmetric adaptive decorrelation:
stability, convergence, and uniqueness. IEEE Trans Signal Process 43(7):1602–1612
6. Bendoumia R, Djendi M (2014) Variable step-sizes new efficient two-channel backward
algorithm for speech intelligibility enhancement: a subband approach. Appl Acoust 76:209–222
7. Gervan SV (1996) Adaptive noise cancellation and signal separation with applications to speech
enhancement. PhD dissertation, University catholic of Leuven, Mars
8. Araki S, Makino S, Aichner R, Nishikawa T, Saruwatari H (2003) Subband based blind source
separation with appropriate processing for each frequency band. In: IEEE. ICA 2003. Japan,
pp 499–504
9. Weinstein E, Feder M, Oppenheim AV (1993) Multi-channel signal separation by decorrelation.
IEEE Trans Speech Audio Process 1(4):405–413
10. Haykin S (1996) Adaptive filter theory. Prentice Hall, NJ
11. Sayed AH (2003) Fundamentals of adaptive filtering. Wiley, New York
12. Bendoumia R, Kerkar M, Bouzakar SA (2019) Acoustic noise reduction by new sub-band
forward symmetric adaptive decorrelation algorithms. Appl Acoust 152:118–126, 69–81
13. Bendoumia R, Djendi M, Guessoum A (2019) Two-Channel Acoustic Noise Reduction by New
Backward Normalized Decorrelation Algorithm. In: Springer Nature Switzerland AG 2019,
ICEECA 2017, Constantine, LNEE 522, pp 464–478
14. José GF, Orlando Z, Tobias J, Seara R (2008) A VSSLMS algorithm based on error
autocorrelation. In: 16th European signal processing conference (EUSIPCO). Lausanne,
Switzerland
15. Djendi M, Bendoumia R (2013) A new adaptive filtering subband algorithm for two-channel
acoustic noise reduction and speech enhancement. Comput Elect Eng 39(8):2531–2550
16. Bendoumia R, Djendi M (2018) Acoustic noise reduction by new two-channel proportionate
forward symmetric adaptive decorrelating algorithms in sparse systems. Appl Acoust 137:69–
81
17. Djendi M, Bendoumia R (2014) A new efficient two-channel backward algorithm for speech
intelligibility enhancement: a subband approach. Appl Acoust 76:209–222
New Log CFAR Radar for Compound
Sea Clutter
Nouh Guidoum, Faouzi Soltani, and Khadidja Belhi
Abstract This paper introduces a new Constant False Alarm Rate (CFAR) detector
designed for compound Gaussian sea clutter. The performance of the proposed
detector is numerically investigated through Monte Carlo simulations assuming sea
clutter following the K distribution, the Generalized Pareto (GP) distribution and
the Compound Inverse Gaussian (CIG) distribution, with the consideration of the
presence of thermal noise and interfering targets in the reference cells. The obtained
results illustrate an excellent false alarm rate regulation especially for spiky clutter.
Keywords CFAR detector · Sea clutter · Compound gaussian
1 Introduction
One of the important missions in radar detection in a maritime environment is to

maintain the concept of CFAR against spiky clutter. Maritime radar performance is
limited due to the severity of the sea clutter where the spikes are processed by the
radar detector as targets. Furthermore, the statistics of sea clutter have been observed
to deviate from Gaussianity [1]. A number of non-Gaussian distributions have been
proposed to model sea clutter. In addition to the log-normal distribution [2], the K
distribution developed by Ward and Watts provides a compound representation that
includes the speckle, which is the fast fluctuating component and the texture that
represents the slowly fluctuating [3]. The Rayleigh distribution is a particular case of
the K distribution. The Generalized Pareto (GP) model with an inverse gamma texture
[4]. Also, the Compound Inverse Gaussian (CIG) accurately describes the amplitude
statistics of sea clutter [5]. Non-coherent radar systems have been the subject of
intensive research for many years. Goldstein proposed the log-t CFAR detector for
the log-normal and Weibull clutters [6]. Weber and Haykin proposed an extension
of the ordered statistic OS-CFAR for two parameter distributions with unknown
average power and unknown skewness. This detector proved to insure the CFAR
N. Guidoum (B) · F. Soltani · K. Belhi

Département D’Electronique, Laboratoire Signaux et Systèmes de Communication, Université
Des Frères Mentouri Constantine 1, Constantine, Algeria

https://doi.org/10.1007/978-981-15-6403-1_82
1172 N. Guidoum et al.
property for the lognormal, Weibull, and Gamma distributions [7]. Armstrong and
Griffiths studied the detection performance in the K distribution for the cell averaging
(CA), greatest of (GO) and order statistics (OS) CFAR, the three detectors suffer a
severe loss in detection especially for spiky clutter [8]. For heavy-tailed Weibull
and lognormal, Norouzi et al. proposed a new CFAR detector based on goodness-
of-fit (GoF). This detector proved to be superior to the CFAR detector based on
binary integration [9]. Watts et al. studied the performance of CA-CFAR in the K
distribution. This detector suffers a significant loss when the clutter is spiky or when
the texture is uncorrelated [10]. Weinberg and Alexopoulos suggested a dual order
statistic detector for the Pareto distribution that achieves the CFAR property [11].
Also, for Pareto model, Weinberg and Glenny explored the log-t detector which
attains the CFAR property [12]. In this work, we propose a new detector for the
non-coherent radar system and we focus on the concept of CFAR property for the
CG distributions plus thermal noise, especially for spiky clutter situations. Then, we
modify the suggested detector to take into account the presence of thermal noise
and interfering targets. This paper is organized as follows: In Sect. 2, we describe
the suggested detector then we briefly recall the expressions of the K plus thermal
noise, the CIG plus thermal noise and GP plus thermal noise distributions. Then, we
present some results of the variations of the probability of false alarm (PF A ) versus
the threshold factor. In Sect. 3, we study the robustness of the suggested detector
in the case of the presence of interfering targets in the reference cells and we show
some results in terms of PF A . The final section concludes this study.
2 The Proposed Detector
The suggested detector, shown in Fig. 1, is designed to be a CFAR in the case of

the presence of thermal noise and interfering targets. The proposed threshold was
found using the trial and error method. The samples of the sea clutter are stored in
a shift register. The detection test in each cell uses a reference window of N cells
adjacent to the cell under test (CUT) which gives a local estimate of the clutter level.
Fig. 1 The block diagram of the proposed detector

New Log CFAR Radar for Compound Sea Clutter 1173
The selection of reference cell number was found after intensive testing of several
values from 2 to 32 cells. A number of cells (N = 2) was found to provide the best
performance for the proposed detector for reasons related to the computing time and
local stationarity of the clutter. The principle of the proposed detector is based on
the comparison the content of the CUT with an adaptive threshold, denoted Th to
choose between two statistical assumptions: the absence of the target H0 and the
presence of the target H1 . Then, Monte Carlo simulations are performed to study the
suggested threshold for: The K, CIG and GP distributions by taking into account the
presence of thermal noise. The threshold corresponding to the suggested detector is
based on four statistics: The logarithm, the standard deviation, the maximum and the
minimum value of the reference cells X :
1
Th = ∗ (((max(log(X )) − min(log(X ))) ∗ τ )
2
+ std(log(X ))) + min(log(X )) (1)
2.1 The K Distribution Plus Thermal Noise
The K distribution contains: the fast-varying component, called speckle and the
slowly fluctuating component, called the texture. The probability density function
( pd f ) of each component is expressed respectively by an Exponential distribution
and a Gamma distribution [3]:

1 x
P(x\y) = exp − (2)
y y
bν ν−1
P(y) = y exp(−by) (3)
(ν)
where ν is the shape parameter, b is the scale parameter. The intensity of the K
distribution is given by [3]:
∞ ν+1 ν−1 √
2b 2x 2
P(x) = P(y)P(x\y)dy = K ν−1 2 bx (4)
(ν)
0
K ν−1 (.) represents the modified Bessel function of the second kind, and (.) is
the Gamma function. In the case of the presence of thermal noise, the pdf of the
speckle component becomes [3]:

1 x
P(x\y) = ex p − (5)
pn + y pn + y
pn = 2σ 2 is the power of the thermal noise. The intensity of the K distribution

plus thermal noise becomes:

bν ∞ y ν−1 x
P(x) = ∫ ex p − − by dy (6)
(ν) 0 pn + y pn + y
2.2 The GP Distribution Plus Thermal Noise
The GP distribution is given by:
αbα
P(x) = (7)
(x + b)α+1
where α is the shape parameter, b is the scale parameter, in the case of the presence
of thermal noise. The resulting pd f becomes [4]:

bα ∞ y −α−1 x b
P(x) = ∫ ex p − ex p − dy (8)
(α) 0 pn + y pn + y y
2.3 The CIG Distribution Plus Thermal Noise
The texture component of the CIG distribution is expressed by inverse Gaussian [5]:

x x 2
P(x\y) = 2 exp (9)
2y 4y 2

λ1/2 (y − μ)2
P(y) = √ ex p −λ (10)
2y 3/2 2μ2 y
It is characterized by the shape parameter λ and the mean μ. In the case of presence
of thermal noise, the speckle can be rewritten as [5]:

x x2
P(x\y) = ex p − (11)
pn+2y 2 / pn + 4y 2 /
And the resulting pdf for 0 ≤ x ≤ ∞ becomes

∞
x x2
P(x) = ex p −
pn + 2y 2 / pn + 4y 2 /
0

λ1/2 (y − μ)2
×√ ex p −λ (12)
2y 3/2 2μ2 y
2.4 False Alarm Regulation in the CG Clutter Plus Thermal

Noise
To illustrate the variations of PF A against the threshold factor τ for the three clutter
distributions mentioned above, we use the MATLAB tool and the following assump-
tions: Non-Coherent detector, variations of PF A from 0 to 10−4 , the thermal noise
power Pn = 2σ 2 was set to 2 dB and different values of scale and shape parameters.
The number of independent Monte Carlo runs is set equal to 106 . Figure 2 shows
the variation of PF A versus the threshold factor τ for the K distribution clutter with
different values of shape parameters ν ≥ 0:01 and different values of β. We observe
that all the curves almost overlap with each other even when the clutter becomes
spikier. Note the same pattern for the GP distribution with α > 1 in Fig. 3. Also, the
same variations for the CIG with λ ≥ 0,05 in Fig. 4. All the results prove that the
suggested detector achieve the CFAR property independently of the shape and scale
parameters. In addition, the presence of thermal noise in the reference cells does not
affect the robustness of this detector.
Fig. 2 Variation of PFA versus the scale factor τ for the K Distribution plus thermal noise
Fig. 3 Variation of PFA versus the scale factor τ for the GP distributionplus thermal noise
Fig. 4 Variation of PFA versus the scale factor τ for the CIG distribution plus thermal noise
3 False Alarm Regulation in the CG Clutter Plus Two

Interfering Targets
To study the effect of the presence of two interfering targets in the reference window,
we modified the first version of the proposed detector. We use four reference cells
instead of two and then after the sorting operation, the two largest are censored. This
means that this regulator can handle up to two interfering targets. The threshold is
calculated using the remaining samples in the same way as in (1). To realize this test,
we use the same hypotheses mentioned above, considering that the interfering targets
are supposed to fluctuate according to the model of Swerling I and the Interference-
to-Clutter Ratio (ICR) is assumed to be equal to 20 dB. For the K distribution Fig. 5
illustrates the changes of PF A versus the threshold factor τ. It is clear that all the curves
almost interfere with each other which mean that the CFAR property is conserved.
For the GP distribution, all the curves shown in Fig. 6 almost overlap. Finally, the
Fig. 5 Variation of PFA versus the scale factor τ for the K distributionplus thermal noise and two
interfering targets
Fig. 6 Variation of PFA versus the scale factor τ for the GP distribution plus thermal noise and two
interfering targets
CIG distribution, the same pattern is observed, that is this detector also keeps the
CFAR characteristic as shown Fig. 7. From the three results, the property CFAR has
been achieved, which means that the presence of interference targets in the reference
window do not affect the robustness of this detector.
Fig. 7 Variation of PFA versus the scale factor τ for the CIG distribution plus thermal noise and
two interfering targets
4 Conclusion
In this paper, we suggested a new non-coherent CFAR detector operating in sea

clutter modeled by the K, generalized Pareto and compound inverse Gaussian distri-
butions. This detector uses two reference cells, which is advantageous for reasons
of computing time and local stationarity of the clutter. In the case of the presence
of the thermal noise this detector preserves a good performance in terms of false
alarm regulation especially in spiky clutter and proved that the PF A variations do not
depend on parameter distributions. Moreover, this detector proved to be robust when
up to two interference targets are present in reference windows.
References
1. Jayaprakash A, Reddy GR, Prasad N (2016) Small target detection within sea clutter based on
fractal analysis. Procedia Technol 24:988–995
2. Trunk GV, George SF (1970) Detection of targets in non-gaussian sea clutter. IEEE Trans
Aerosp Electron Syst AES-6(5):620–628
3. Ward K, Tough R, Watts S (2013) Sea clutter: scattering, the K distribution and radar
performance, 2nd edn. Inst., London, UK
4. Mezache A, Soltani F, Sahed M, Chalabi I (2016) Estimating the Pareto plus noise distribution
parameters using non-integer order moments and [zlog(z)] approaches. IET Radar Sonar Navig
10(1):192–204
5. Mezache A, Soltani F, Sahed M, Chalabi I (2013) A model for non Rayleigh sea clutter
amplitudes using compound Inverse Gaussian distribution. In: 2013 IEEE radar conference
(RadarCon13)
6. Goldstein G (1973) False-alarm regulation in log-normal and weibull clutter. IEEE Trans
Aerosp Electron Syst AES-9(1):84–92
7. Weber P, Haykin S (1985) Ordered statistic CFAR processing for two-parameter distributions
with Variable skewness. IEEE Trans Aerosp Electron Syst AES-21(6):819–821
8. Armstrong B, Griffiths H (1991) CFAR detection of fluctuating targets in spatially correlated

K-distributed clutter. In: IEE Proc F Radar Signal Process 138(2):139; Young M (1989) The
technical writer’s handbook. University Science, Mill Valley, CA
9. Norouzi Y, Gini F, Nayebi M, Greco M (2007) Non-coherent radar CFAR detection based on
goodness-of-fit tests. IET Radar Sonar Navig 1(2):98
10. Watts S, Ward K, Tough R (2007) CFAR loss and gain in K-distributed sea-clutter and thermal
noise. In: IET international conference on radar systems
11. Weinberg G, Alexopoulos A (2016) Analysis of a dual order statistic constant false alarm rate
detector. IEEE Trans Aerosp Electron Syst 52(5):2567–2574
12. Weinberg GV, Glenny VG (2017) Enhancing goldsteins Log-t detector in pareto-distributed
clutter. IEEE Trans Aerosp Electron Syst 53(2):1035–1044
Stochastic Analysis of the Crosstalk
Between Transmission Lines
Hemza Gheddar, M. Melit, and B. Nekhoul
Abstract The main purpose of our work is to give an overall analysis of crosstalk
in transmission lines. Classically, this problem is largely dealt with using a deter-
ministic approach. For this reason, we proceed on a probabilistic approach, the
geometric parameters of the transmission line (height above the ground plane, sepa-
ration between conductors) are considered as random variables. In this work we use
the chaos polynomial technique to calculate the probability density function of the
induced voltage, efficiency and speed these technique with respect to the method of
Mont carol which uses a lot of time.
Keywords Crosstalk · Transmission line · Random · Chaos polynomial ·

Monte-Carlo
1 Introduction
The advancement of process technologies produces very fast and compacted systems,
in this situation, the problem of coupling and interference is automatically foreign.
As a result, crosstalk noise between the signal line increasingly severe by capacitive
inductive coupling and which create an EMC (electromagnetic compatibility). The
crosstalk problem is conventionally treated using a deterministic approach and the
near-end induced voltage Near End (NE) and Far End (FE) is expressed both in the
temporal and frequency domains [1–3]. Unfortunately, these studies demonstrate that
the induced voltage is highly dependent on geometric parameters such as conductor
length, height above ground plane and conductor separation from such a problem.
H. Gheddar (B) · M. Melit · B. Nekhoul

Electrical Engineering, Jijel University, Jijel, Algeria
M. Melit
B. Nekhoul

https://doi.org/10.1007/978-981-15-6403-1_83
1182 H. Gheddar et al.
For this reason, we only assume a probabilistic approach of crosstalk very interesting.
The aim of this study is to construct a methodology for the analysis of the coupling
between the conductors of a transmission line, which takes into account the random
nature of the environment in which the line is located. Probabilistic tools and statis-
tics are good candidates to help them achieve this task. They are used to predict the
level of crosstalk for any given. To illustrate this approach, we consider the case of a
transmission line, composed of two neighboring conductors. The law of probability
of the geometric parameter (height, length and separation) of the conductors is deter-
mined numerically by the uniform law. The parameters of the Probability Density
Function (PDF) [7], mean, variance, are functions of the transmission line, length,
separation and height above a conductive plane.. For the estimated PDF crosstalk,
there are several numerical methods in this study and based on the approach of the
chaos polynomial, effective end fast this approach for crosstalk PDF [8, 9] evaluation
compared to the classic Monte Carlo method, it is Difficult to calculate the polyno-
mial coefficient In this article, we propose the regression method chosen to calculate
the polynomial coefficient.
2 Study Crosstalk by Deterministic Approach
In this part uses the deterministic approach which based on the theory of transmission
line for calculating the induced voltage at near end and fear end, we consider the
three-conductor transmission lines as shown below in Fig. 1a. Where the Z-axis are
representing the line’s axis, is the total length of the line and (Z = 0, Z = )
are the left and right ends respectively. Also we assume that these conductors are
perfect ones and the medium surrounding them is lossless and homogeneous where
is characterized by the permeability μ and the permittivity ε
We find that the transmission-line equations are given by [1]:
d[U (z)]
+ jw[L] [I (z)] = [0] (1)
dz
d[I (z)]
+ jw[C] [U (z)] = [0] (2)
dz
where L is the per-unit length inductance matrix, C is the capacitance matrix and ω
is the radian frequency of excitation and is given by ω = 2π f
From the chain parameter matrix of the line φ which relates the voltages and
current [1], we obtain the following solution:

U (z) U (z 0 )
= [(z − z 0 )] (3)
I (z) I (z 0 )
where
Stochastic Analysis of the Crosstalk Between Transmission Lines 1183
Generator wire
a
Shielded receptor wire
50 Ω 50 Ω
1 V+ + VNE + VFE 50 Ω
50 Ω
Soil
Fig. 1 a Shielded wire, b the single wire receptor configuration

[φ11 (z − z 0 )] [φ12 (z − z 0 )]
[φ(z − z 0 )] = (4)
[φ21 (z − z 0 )] [φ22 (z − z 0 )]
we choose z 0 = 0 and z = , Eq. (3) as become:

[U ()] [U (0)]
= [φ()] (5)
[I ()] [I (0)]

[φ11 ()] [φ12 ()] [U (0)]
= (6)
[φ21 ()] [φ22 ()] [I (0)]
And to have a perfect conductors without dielectric insulation, the sub matrices
φi j () of the chain-parameter matrix φ() must be a 6 × 6 matrices so the entries
will be given as [2]:
φ11 () = cos(β)I3 (7a)

φ12 () = − jvsin(β)L (7b)
φ21 () = − jvsin(β)C (7c)
φ22 () = cos(β)I3 (7e)
where β is the phase constant given by β = ων = 2π λ

,ω is the radian frequency
of excitation, λ is a wavelength at this frequency, λ = νf and ν is the velocity of
propagation in the surrounding medium,v = √μ1v εv .the permeability and permittivity
of free space are denoted by μv and εv respectively the per-unit-length inductance
and capacitance matrices L And C respectively, are given by [3]
Once the overall chain-parameter matrix of the line has been determined, the
terminal condition will be modeled and added. The terminal loads, voltage, and
currents are as shown in Fig. 1a. From this diagram, the termination equation can be
written as
U (0) = V − R0 I (0) (8)
U () = R I () (9)
where V is a 3 × 1 vector and R0 and R are 3 × 3 matrices, given by

⎡ ⎤
Rs 0 0
R0 = ⎣ 0 R N E 0 ⎦ (10)
0 0 RN E
⎡ ⎤
RG 0 0
R = ⎣ 0 R F E 0 ⎦ (11)
0 0 RF E
⎡ ⎤
Vs
V =⎣ 0 ⎦ (12)
0
After the resolution of the system given in formula (3) using the conditions in
formulas (8) and (9). We find the report of crosstalk as following

V0R
N ext = (13)
Vs

VR
Fext = (14)
Vs
3 Estimation of the Crosstalk by Stochastic Approach
Polynomial Chaos technique is applied to the modeling of random cross-section vari-

ations in multiconductor transmission lines [4]. These Variations are due to uncer-
tainties in the geometry and material parameters which reflect into a variability of
the per-unit-length parameters (p.u.l). The frequency domain of the transmission-line
equation is written as:

d
dz
[U (x, ξ )] = −[L(ξ )][I (z, ξ )]
(15)
dz (x, )]
d
[I ξ = −[C(ξ )][U (z, ξ )]
where ξ = [ξ1 , . . . .ξn ] is a multivariate set of n independent random variables that

affect the cross-sectional parameters. So the voltages U and the currents I are random
and dependent on ξ . In short, the idea here is to approximate the dependence on
the random variables ξ which means the expansion in a truncated series of proper
multivariate orthogonal polynomials as following [5].

p−1
X (.., ξ ) ≈ X j (. . .)φ j (ξ ) (16)
j=0
where X (…) stands for Ũ (x, ξ ), I˜(x, ξ ) and the X j are the respective coefficients.
While φ j defines an orthogonal polynomial basis, whose choice relies on the distri-
bution of the random variables ξi . so using Hermit polynomials is the optimal choice
for the Gaussian [7, 6].
By substitution of (15) into (16) and the application of the regression approach
which is written as
⎧
⎪
⎪
p−1 M
⎪
⎨ Ũ (z, ξ ) ≈ α j φ j ξk(i)
k=1
j=0
(17)
⎪
⎪
p−1 M
˜
⎪ I (z, ξ ) ≈
⎩ β j φ j ξk(i)
j=0 k=1
where the expansion coefficients α j and β j can be computed with regression

approach, using the least squares method, the quadratic deviation U between the
random [8, 9].
M
Ũ : Solution approach by PCE, ξk(i) : Constituent plan of experiences
k=1
Variables U and Ũ is minimized with respect to the coefficients
{αk , k = 0, . . . , p − 1}
n
2
U = U (i) − Ũ (i) (18)
i=1
This is equivalent to minimizing the coefficients {αk , k = 0, . . . , p − 1} the

expression:
⎛ ⎞2

n
p−1 M
U = ⎝U (i) − αjφj ξk(i) ⎠ (19)
k=1
i=1 j=0
The minimum is obtained for verifying values of U y :
U
= 0 ∀y ∈ {0, . . . , p − 1} (20)
U y
where
⎛ ⎞

n
p−1
M M
φ y ξk(i) ⎝U (i) − α j φ j ξk(i) ⎠=0 (21)
k=1 k=1
i=1 j=0
∀y =∈ {0, . . . , p − 1}
Finally the system to be solved is the following one:

p−1

n
M M
n
M
αj φ y ξk(i) φ j ξk(i) = U (i) φ y ξk(i) (22)
k=1 k=1 k=1
j=0 i=1 i=1
∀y =∈ {0, . . . , p − 1}
This system can also be written as:

⎡ ⎤

n M M
n M M
⎢ i=1 φ0 ξk(i) φ0 ξk=1 ... φ0 ξk(i) φ P−1 ξk(i) ⎥
⎢ k=1 i=1 k=1 k=1
⎥
⎢ .. .. ⎥
⎢. . ⎥×
⎢ ⎥
⎣ n
(i)
M
(i)
M
(i)
M n M ⎦
φ P−1 ξk φ0 ξ k ··· φ P−1 ξk φ P−1 ξk(i)
i=1 k=1 k=1 i=1 k=1 k=1
⎛ n ⎞
(i) (i)
M
⎛ ⎞
α0 ⎜ i=0 U φ0 ξk k=1 ⎟
⎜ ⎟
⎜ .. ⎟ ⎜ .. ⎟
⎝ . ⎠=⎜ . ⎟ (23)
⎜ M ⎟
α P−1 ⎝ n ⎠
U (i) φ P−1 ξk(i)
i=0 k=1
# $
By solving the system (23), we find the coefficients α0 , . . . α P−1 , so can solve
Eq. (17).
4 Simulation
Case 1: Deterministic approach

We calculated Near end with the parameters values: Vs = 1 V, Rs = 0, RNE =
RFE = R = 50, h = 1 cm, d = 2 cm and = 3.567 m. using a beach of frequency
going from 1 kHz to 100 MHz. the answers are represented in Fig. 1 (Figs. 2, 3 and
4)
Case 2: Stochastic approach
In the second set of simulations, we consider tow geometric parameters defined as
uniform variables heights h and distance d, we compute the induced voltage near-end
and far-end for a set of values of the configuration defined below: The radius of the
wires: 0.4 mm with reference to Fig. 1, the loads are set to Rs = 0, RNE = RFE = RL =
R, we assumed R = 50 , L = 3.567 m. Wire heights and distance defined as uniform
variables in the range [0.9 1.1] cm [0.19 0.21] cm.
We calculate the Probability Density Function (PDF) of induced voltage at the
near end and far-end of shielded wire, using two different ways: Monte Carlo (MC)
and Chaos Polynomial (PC) methods where the results are represented in Figs. 5 and
6 as below (Tables 1 and 2):
The result in Figs. 5 and 6 shows that Monte-Carlo method and PC gives very
close results of PDF, we notice that MC are based on repeated simulations 700*700
times which need more than 15 min were PC uses 2*2 collocation points and less
than 1 s.
Near-end crosstalk
150
MTL
Magnitude [mV]
100
50
0
2 3 4 5 6 7 8
10 10 10 10 10 10 10
Frequency [Hz]
100
Phase [deg]
0
MTL
-100
-200
2 3 4 5 6 7 8
10 10 10 10 10 10 10
Frequency [Hz]
Fig. 2 Near-and crosstalk for the geometry shown in Fig. 1a: Our computation results
Fig. 3 Near-and crosstalk for the geometry shown in Fig. 1a: Published Results [2]
5 Conclusion
This work, discuss the problem of crosstalk in shielded wire using both deterministic
and stochastic method, the challenge point of this method is to model the Electro-
Magnetic (EM) coupling in the high complexity cable. The crosstalk response of the
system is expanded on a multivariate polynomial basis of input random variables in
order to reduce the computational cost compared to MC simulation. The efficiency of
this PC representation consists in providing a very good approximation of the output
Far-end crosstalk
150
MTL
Magnitude [mV]
100
50
0
2 3 4 5 6 7 8
10 10 10 10 10 10 10
Frequency [Hz]
200
MTL
Phase [deg]
-200
2 3 4 5 6 7 8
10 10 10 10 10 10 10
Frequency [Hz]
Fig. 4 Far-and crosstalk for the geometry shown in Fig. 1a: Our Computation results
0.3 MC PDF
F= 10 MHz
PC PDF
0.25
0.2
PDF
0.15
0.1
0.05
0
77 78 79 80 81 82 83 84 85 86 87 88
Magnitude [mV]
Fig. 5 PDF of the frequency-domain near-end crosstalk for 10 MHz
by means of a small number of simulations, despite a large number of uncertain input

parameters.
0.2
MC PDF
F= 10 MHz
PC PDF
0.15
PDF
0.1
0.05
114 116 118 120 122 124 126 128 130

Magnitude [mV]
Fig. 6 PDF of the frequency-domain far-end crosstalk for 10 MHz
Table 1 The statistic

Number of 490 000 MC 4 PC simulations (P =
parameter for near-end the
simulations simulations 2)
frequency 10 MHz
The statistic
parameter
Mean (μ)[mV ] 82.398 82.4711
Variance (σ 2 )[mV ] 1.73022 1.67485
Standard deviation 1.31538 1.29416
(σ )[mV ]
Simulation time 15 mint 1s
Table 2 The statistic

Number of 490 000MC 4 PC simulations (P =
parameter for far-end the
simulations simulations 2)
frequency 10 MHz
The statistic
parameter
Mean (μ)[mV ] 122.235 122.347
Variance (σ 2 )[mV ] 3.92151 3.79935
Standard deviation 1.98028 1.94919
(σ )[mV ]
Simulation time 15 mint 1s
References
1. Paul CR (2002) Solution of the transmission-line equations under the weak-coupling assumption.
IEEE Electromagn Compat 44(3):413–423
2. Manfredi P (2009) Models of complex cables for SPICE simulation at system leve. these
3. Smith WT, Paul CR, Savage JS, Das SK, Cooperider AD, Frazier RK (1994) Crosstalk modeling
for automotive harnesses. IEEE Electromagn Compat 447–4524
4. Ghanem RG, Spanos PD (1991) Stochastic finite elements. A spectral approach. Springer, New
York
5. Spina D, De Jonghe D, Deschrijver D, Gielen G, Knockaert L, Dhaene T (2014) tochastic
macromodeling of nonlinear systems via polynomial chaos expansion and transfer function
trajectories. IEEE Trans Microw Theory Techn 62(7):1454–1460
6. Larbi M, Stievano IS, Canavero FG, Besnier P (2018) Variability impact of many design
parameters: the case of a realistic electronic link. IEEE Trans Electromagn Compat 60(1):34–41
7. Stievano S, Manfredi P, Canavero FG (2011) Stochastic analysis of multiconductor cables and
interconnects. IEEE Trans Electromagn Compat 53(2):501–507
8. Papoulis A (1991) Probability, random variables and stochastic processes. McGraw-Hill, New
York
9. Shiran S, Reiser B, Cory H (1993) A probabilistic model for the evaluation of coupling between
transmission lines. IEEE Trans EMC 35(3)
Moving Objects Detection and Tracking
with Camera Motion Compensation
Bousta Mohamed Akli, Nemra Abdelkrim, Hamami Fatima,

and Demim Fethi
Abstract The background subtraction is a widely used approach for Detecting

Moving Objects (DMO) by a static camera using a simple algorithm; however, it
is very sensitive to the local gradual changes of illumination, shadows, non-rigid
moving objects and partial or full target occlusion. In order to overcome these issues,
and to bring more potential to the solution, we propose an effective method for object
detection based on active contour with camera motion compensation. The principle
of active contours is to evolve an initial curve towards the object of interest that corre-
sponds to the boundaries of the moving objects. Once the object has been detected,
then it can be tracked by the Kalman filter. This latter requires many preassumptions
about models and noise characteristics. As an alternative, a new method based on the
Smooth Variable Structure Filter (SVSF) is implemented. The SVSF was introduced
in an effort to provide a more robust estimation strategy. Detection and tracking
algorithms require stable images to recognize a real moving target position. There-
fore, the captured image from cameras which are placed on the moving platform,
have undesired jitters, shakes and blurs. Video stabilization becomes an indispens-
able technique which is focusing on removing unnecessary camera vibrations from
image sequences using a homographic matrix by extracting features using FAST
corner detector and FREAK feature descriptor. The proposed algorithm is validated
in real-world and the obtained results confirm the efficiency and robustness of our
approach.
B. M. Akli · N. Abdelkrim · D. Fethi (B)

École Militaire Polythechnique EMP, BP17 Bordj El Bahri, Algeria
B. M. Akli
N. Abdelkrim
H. Fatima
Ecole Nationale Polytechnique ENP, El Harrach, Algiers, Algeria

https://doi.org/10.1007/978-981-15-6403-1_84
1194 B. M. Akli et al.
Keywords Background subtraction · Tracking · Geodesic active contour · Smooth

variable structure filter · Kalman filter · Video stabilization
1 Introduction
Detection of moving objects from a video sequence, which is based on representing

moving objects by a binary mask in each frame, is an important issue and interested in
many vision-based application such as traffic control, autonomous robot navigation
and intelligent video surveillance. There are mainly two modes of moving object
detection: static camera and moving camera. In the case of static camera, the only
changes between consecutive frames are caused by moving objects. However, these
changes are not due to the target’s motion only, but also to the camera moves too.
In an indoor scene, even under a controlled environment, shadow regions and illu-
mination source changes may occur and are undesired for moving object detection
[1]. For an outdoor environment, typically the environment is not controllable and
not static, and many undesired changes such as branch movement, cloud movement
and illumination variations can cause serious problems for moving object detection
[2]. Many previous works [3, 4] have addressed to the moving object detection in
video sequences captured by a fixed camera in the presence or not of undesired
changes in the scene. These algorithms involve background subtraction that can be
further divided into several other algorithms: Frame difference method, Approxi-
mate median, Running Gaussian average and Mixture of Gaussian [5, 6], temporal
difference, optical flow. Nowadays, various techniques have been extensively studied
with different solutions for object detections. But, we can not find a solution to give
an optimal solution in different situations. In the work of [7, 18], the author has
introduced the algorithm of automatic objects detection using histograms of oriented
gradients (Dalal 2005). Also, some object detection algorithms based-on colour infor-
mation [8, 9] are proposed, but usually we can not use a colour image in the case of
moving camera, because it may be caused a limited computing resources. There is
an approach to detect objects using a machine learning which is proposed by Papa-
georgiou [10]. In this case, it is necessary to create training samples and find some
solution to the interest wide variety of objects and the observation conditions.
When a camera undergoes vibrations or unwanted motions as in the case of a
camera mounted on a moving vehicle or a hand-held camera, it records a jerky video
created by small, random movements. The poor quality of such recording may cause
the video not to be useful (surveillance task or pattern recognition). Therefore, video
stabilization has become an essential process and it is considered among the first step
for a lot of video processing works.
Digital video stabilization is an essential video improvement technology which is
focusing on removed unnecessary camera vibrations from image sequences. In the
proposed method, firstly, feature points are detected with the Features from Acceler-
ated Segment Test (FAST) corner detection algorithm and then these feature points
are extracted using efficient Fast Retina Keypoint (FREAK) descriptor [28, 29]. Then,
Moving Objects Detection and Tracking … 1195
feature points are matched based-on Hamming distance between consecutive frames.
Next, the matched point pairs are fitted to the affine transformation model using an M-
estimator SAmple Consensus (MSAC) algorithm, which is a variant of the RANSAC
algorithm based-approach to estimate inter-frame motion strongly [31]. After that,
the calculated results are used to find out the cumulative motion parameters between
the conjugative frames, and the translational components. The principal idea of video
stabilization is to create a stable background and then to apply a background subtrac-
tion technique, namely to subtract current frame from background to detect moving
objects [30]. Object tracking strategy can also be considered as moving object detec-
tion although its objective is different. Indeed, in object tracking, typically we mark
an object as our desired object (target) and then try to localize it in the next frames
of the video sequence. In the literature survey [11], it has been introduced various
segmentation methods relevant to tracking objects in video and categorized object
tracking into a point - tracking, kernel tracking and silhouette tracking [15, 16]. For
this purpose, two tracking algorithms based-on Kalman filter and SVSF filter are
proposed using a real data.
The paper is organized as follows: Sect. 2 depicts the different technique of moving
object detection and tracking in video sequences captured with a static background.
Section 3 describes two target detection approaches, the former is based-on colour
segmentation, and the latter is based-on the geodesic active contour [19, 20]. Tracking
moving object has been done in Sect. 4 using Kalman filter and the smooth variable
structure filter. Comparison between these filters will be presented in terms of accu-
racy, robustness and smoothness in Sect. 5. Finally, Sect. 6 concludes the paper and
gives some future works.
2 Moving Object Detection Methods with Static

Background
In recent years, many algorithms for moving object detection have been proposed.
The increase of surveillance detection can be achieved by developing tools able to
provide accurate moving object detection. Most of the works carried out today are
still about background subtraction, optical flow, temporal difference and many other
algorithms for detecting moving objects. Today, the most widely algorithms are still
about finding solution to the fundamental problems using background subtraction.
In this case, a frame difference, approximate median, and a mixture of Gaussian are
proposed.
2.1 Background Substraction Method
It is particularly a commonly used technique for object detection in static images

[12–14]. It detects moving regions by subtracting the current image pixel-by-pixel
from a reference background image that is created by averaging images over time
in an initialization period. The basic idea of the background subtraction method is
to initialize a background firstly, then by subtracting the current frame in which the
moving object presents that current frame is subtracted with the background frame to
detect moving object. This method is simple and easy to use and accurately extracts
characteristics of target data, but it is sensitive to the change of external environment.
So, it is applicable to the condition that the background is known.
2.2 Frame Difference Method
The frame difference method is the simplest form of background subtraction [13,
14]. In this method, the current frame is simply subtracted from the background
frame. If the difference in pixel values for every pixel is greater than the threshold,
then the pixel is considered as a part of the foreground otherwise, by the way, it
is considered as background. The frame difference method (FDM) for background
subtraction which used for the static camera is presented.
2.3 Approximate Median Method
In this method, the median filtering buffers the previous N frames of the video [13].
Then, the background frame is calculated from the median of the buffered frame and
the background is subtracted from the current frame to give the foreground pixel.
The Approximate Median method checks whether the pixel in the current frame has
a value that is larger than the corresponding background pixel. If that is the case, the
background pixel is incremented by one. However, if the pixel in the current frame
has a value that is smaller than the corresponding background pixel, the background
pixel is decremented by the last one (Fig. 1).
2.4 Temporal Differencing
This method uses two adjacent frames based on the time series image to subtract and
gets difference images; it is very similar to background subtraction method [12]. After
the subtraction of image, it gives moving target information through the threshold
value. This method is simple and easy to implement.
Fig. 1 Overview of the

System
This is very adaptive to dynamic scene changes however it generally fails in

detecting whole relevant pixels of some types of moving objects. Additionally,
methods need to be adopted in order to detect stopped objects for the success of
higher levels are computationally complex and cannot be used in real-time without
specialized hardware.
3 Moving Object Detection Based on Color Segmentation
The colour threshold technique was carried out based-on the colour information of
the object to extract its images from the background and other objects [17]. This
technique specifies the range of RGB intensities for threshold. The specific steps for
the proposed approach focus on the following:
• Original image;
• Calculate range of RGB intensities for threshold;
• Apply thresholding;
• Morphological operation;
• Binary image (Fig. 2).
Fig. 2 Motion detection by thresholding
4 Moving Object Detection by Geodesic Active Contour
The difficulty of objects detection depends on various factors related to the data or
application such as the presence of noise in the images, complex motions, non-rigid
and/or articulated objects, partial or total occultation, etc.
These problems, therefore, require a deformable model like geodesic active
contour. Geodesic active contours have been widely used as a solution for image
segmentation [1, 21]. It consists initially dividing the video into a succession of
images. Then, we segment the first image using geodesic active contour, as a result,
is a binary image that represents the detected object. The binary image will be used
as a mask for the next image until the end of the video sequences.
The principle of active contours is to evolve an initial curve towards the object of
interest. The evolution of this curve is done by the action of a force that is generally
derived from the minimization of a function. For more details refer to the work
proposed by [2]. Contour tracking methods are used if we want to extract the object
outline and if the latter deforms over time [22, 23]. These approaches are based on
a succession of segmentation [6]. They generally evolve the object contour at the
previous time towards his new position at the current time. This type of tracking
can be achieved with methods using state models or methods minimizing energy
functions on the contour whose term of attachment to data helps the positioning
of the estimated contour along with areas of the high gradient in the image [7,
24]. The geodesic method is an improvement of classical active contour methods.
This technique allows the simultaneous detection of external and internal borders
of several objects where the final contour does not depend on initialization. It is
independent on parameterization and the topology automatic change thanks to the
level curve method. The first approach of active contours was improved by works
presented in [8, 9] based-on the geodesic Active contours.
5 Moving Object Detection with Background

Compensation
Nowadays, there are many applications which work on the video research. Partic-
ularly, video from camera’s which are placed on the moving platform. But due to
uneven surface or camera vibration, the captured image tends to have undesired
jitters, shakes and blurs. This results in unlikable viewing experience and also affect
on image processing such as military application. Video stabilization is an essential
video improvement technology which is focusing on removing unnecessary camera
vibrations from image sequences [25, 26]. The proposed stabilization algorithm
consists of three steps as follows:
(1) Read Frames from a Movie File: Here, we read in the first two frames of a video
sequence.
(2) Collect Salient Points from Each Frame: Firstly, we collect points of interest
from both frames and then select likely correspondences between them. In this
step, we produce these candidate points for each frame. To have the best chance
that these points will have corresponding points in the other frame, we want
points around salient image features such as corners, For this, we use one of the
fastest corner detection algorithms.
(3) Select Correspondences Between Points: For each point, we extract a Fast Retina
Keypoint (FREAK) descriptor centered on it. The matching cost we use between
points is the Hamming distance since FREAK descriptors are binary.
5.1 Estimating Transform from Noisy Correspondences
Many of the point correspondences obtained in the previous step are incorrect. But
we can still derive a robust estimate of the geometric transform between the two
images using the M-estimator SAmple Consensus (MSAC) algorithm, which is a
variant of the RANSAC algorithm. When given a set of point correspondences, will
search for the valid Inlier correspondences. From these, it will then derive the affine
transform that makes the inliers from the first set of points match most closely with
the inliers from the second set. This affine transform will be a 3-by-3 matrix of the
form:
⎡ ⎤
a1 a3 0
⎣ a2 a4 0 ⎦ (1)
tx ty 1
Table 1 Properties of test

Test video Resolution Frame Rate (fps) Type
video
A 1280 × 720 30 .mp4
The parameters define scale, rotation, and sheering effects of the transform, while
the Parameters are translation parameters. This transform can be used to warp the
images Such that their corresponding features will be moved to the same image
location.
Given a set of video frames Ti , 0, 1, 2,…, etc., we can now use the above procedure
to estimate the distortion between all frames Ti and Ti + 1 as affine transforms Hi ,
Thus the cumulative distortion of a frame relative to the first frame will be the product
of all the preceding inter-frame transforms, or:
Hcumulative,i = Hi (2)
Hcumulative,i = Hi Pj=0 (3)
Hcumulative,i = Hi (4)
The all six parameters of the affine transform above are used, but, for numerical
simplicity and stability, we choose to refit the matrix as a simpler scale rotation-
translation transform. For this, only four free parameters compared to the full affine
transform’s six: one scale factor, one angle, and two translations. This new transform
matrix is presented as form:
⎡ ⎤
S cos(ang) −S sin(ang) 0
⎣ S sin(ang) S cos(ang) 0 ⎦ (5)
tx ty 1
We show this conversion procedure below by fitting the above-obtained trans-

form H with a scale-rotation-translation equivalent HSRT to show that the error of
converting the transform is minimal.
Run on the Full Video At each step we calculate the transform H between
the present frames. We fit this as an HSRT , and then, we combine this cumula-
tive transform,Hcumulative , which describes all camera motion since the first frame
(Table 1).
5.2 Experimental Results
The Peak Signal-to-Noise Ratio (PSNR) is used to evaluate the effectiveness of the
stabilization method. The PSNR between consecutive frames is defined as:
Fig. 3 Comparison between original and stabilized frames
2
Imax
PSNR(I1 , I0 ) = 10 log (6)
MSE(I1 , I0 )
where Imax is the maximum intensity value of a pixel and MSE is the mean square
error between consecutive frames, which is defined as:
1 N M
PSNR(I1 , I0 ) = (I1 (i, j) − I0 (i, j))2 (7)
N × M i=0 j=0
Nframe −1
1
ITF = PSNR(k) (8)
Nframe − 1
k−1
N and M represent the frame dimensions. PSNR measures the similarity between
frames. Because the stabilized sequence has more continuous frames than the original
jittered one, the PSNR of the stabilized sequence should be higher than that of the
original sequence (see Figs. 3 and 4). ITF (Inter-frame transformation fidelity) is
the average value of PSNR, which can be used to evaluate numerically how much a
sequence has been stabilized. It is defined as presented in “Fig. 3”.
6 Moving Object Tracking Methods
Object tracking is to track an object (or multiple objects) over a sequence of images. It
can be defined as a process of segmenting an object of interest from a video scene and
keeping track of its motion, occlusion, and orientation in order to extract the useful
Fig. 4 Frames of test video stabilization
information [27]. There are numerous approaches for object tracking (see “Fig. 5”).
These methods differ from each other based on the way they approach as: Which
object representation is suitable for tracking?, Which image features should be used?,
How should the motion, appearance, and shape of the object being modelled?. The
answers to these questions clearly, depend on the environment in which the tracking is
performed and the end use for which the tracking information is being sought. In this
paper moving object tracking is done by using a new method based-on the smooth
variable structure filter (SVSF) compared to Kalman filter is presented. Broadly
object tracking techniques can be classified as shown in “Fig. 7” (Fig. 6).
Fig. 5 Classification of object tracking techniques

Fig. 6 Target position and estimated errors based-on KF
Fig. 7 Accurate estimated positions based-on KF
6.1 Point Tracking
Objects detected in consecutive frames are represented by points, and the association
of the points is based on the previous object state which can include object position
and motion. This approach requires an external mechanism to detect the objects
in every frame. Point tracking is a complex problem particularly in the context of
occlusions, false detections of the object.
6.2 Kernel Based Tracking
Kernel tracking is usually performed by locating the moving object, which is repre-
sented by an embryonic object region, from one frame to the next. The object motion
is usually in the form of parametric motion such as translation, conformal, affine,
etc. These algorithms differ in terms of the present representation used, the number
of objects tracked, and the methods used for approximation the object motion. In
real-time, an illustration of the object using geometric shape is common. But one of
the restrictions is that some part of the objects may be left outside of the defined shape
while portions of the background may exist inside. This can be detected in rigid and
non-rigid objects. These are large tracking techniques based on the representation of
object, object features, appearance and shape of the object.
6.3 Silhouette Based Tracking
Some object will have a complex shape such as hand, fingers, shoulders that cannot be
well defined by simple geometrical shapes. Silhouette-based methods afford an accu-
rate shape description of the objects. The aim of a silhouette-based object tracking is
to find the object region in every frame by means of an object model generated by the
previous frames. It is capable of dealing with a variety of object shapes, Occlusion
and objects splitting and merging.
7 Moving Object Tracking Using Kalman Filter
The KF provides an elegant and statistically optimal solution for linear dynamic
systems in the presence of Gaussian white noise. It is a method that utilizes measure-
ments linearly related to the states, and error covariance matrices, to generate a gain
referred to as the Kalman gain [32, 34]. This gain is applied to the a priori state
estimate, thus creating an a posteriori estimate. The estimation process continues in
a predictor-corrector fashion while maintaining a statistically minimal state error
covariance matrix for linear systems. The following two equations describe the
system dynamic model and the measurement model used in general for (linear)
state estimation.
Xk+1 = Fk Xk + Wk (9)
Zk+1 = Hk+1 Xk+1 + Vk+1 (10)
The next five equations form the KF algorithm, and are used in an iterative fashion,
in conjunction with Equations “(1),” and “(2)”. Equation “(3),” extrapolates the a
priori state estimate and Equation “(5),” is the corresponding state error covariance.
The Kalman gain may be calculated by Equation “(6),” and is used to update the state
estimate and error covariance, described by Equations “(7),” and “(8),” respectively
[35]:
X̂k+1/K = Fk X̂k/k (11)
Pk+1/k = Fk Pk/k FkT + Qk (12)
−1
Kk+1 = Pk+1/k Pk/k HkT Hk Pk+1/k HkT + Rk (13)

X̂k+1/k+1 = X̂k+1/k + Kk+1 Zk+1 − Hk+1 X̂k+1/k (14)

Pk+1/k+1 = 1 − Kk+1 Hk+1 Pk+1/k (15)
where
• Q represents the process noise covariance matrix;
• R represents a covariance matrix;
• P defines the process covariance matrix;
• H represents a conversion matrix;
• I defines an identity matrix.
8 Moving Object Tracking Using SVSF Filter
An important element of the Kalman filter is the Kalman gain that minimizes the
errors in the estimated position [33]. It is the regulator between our estimate and the
measurement. An example of the target estimated position is shown in “Fig. 8”. The
prediction of Kalman filter converged towards the true position value faster than the
prediction of the position and the error was significantly lower. Nonlinearities and
noise in the system may cause the KF to become unstable, in order to overcome these
limits we use SVSF that deals with a model of the system either linear or nonlinear
Moving object tracking using Smooth Variable Structure Filters (SVSF).
The SVSF results in the state estimates converging to within a region of the state
trajectory, referred to as the existence subspace as can be seen in “Fig. 9”.
The SVSF method is model-based and applies to smooth nonlinear dynamic equa-
tions. The estimation process may be summarized by equations “(16),” to “(19),” and
is repeated iteratively [36]. An a priori measurement error estimate is calculated by
using an estimated model of the system. This a priori value is then used to calculate
an a priori estimate of the measurement, defined by equation “(16)”. A corrective
term, referred to as the SVSF gain, is calculated as a function of the error in the
predicted output, as well as, the gain matrix and the smoothing boundary layer width
Fig. 8 SVSF estimation concept
350 250
Epp
Eaa
300
200
250
150
200
150
100
100
50
50
Real traject
Predicted SVSF
Estimate SVSF
0 0
0 10 20 30 40 50 0 10 20 30 40 50
Fig. 9 Target position and estimated errors based-on SVSF
350 100
90
300
80
250 70
60
200
50
150
40
100 30
20
50
Real traject 10
KF KF
SVSF SVSF
0 0
0 10 20 30 40 50 0 10 20 30 40 50
Fig. 10 Target position and estimated errors based-on (KF and SVSF)
Fig. 11 Accurate estimated position using the SVSF
(ψ). The corrective term calculated in the Eq. “(17),” is then used in the Eq. “(18),”
to find the a posteriori measurement error estimate [37, 38].

X̂k+1/k = f X̂k/k , Uk (16)

Ẑk+1/k+1 = h X̂k+1/k (17)
o
Kk+1 = H + − ez,k+1/k + γ ez,k/k Sat ez,k+1/k , ψ (18)
X̂k+1/k+1 = X̂k+1/k + Kk+1 (19)
ez,k/k = Zk+1 − Ẑk+1/k+1 (20)
Two critical variables in this process are the a priori and a posteriori output error
estimates, defined by Eqs. “(15),” and “(18),” respectively. Note that Eq. “(15),” is
the output error estimate from the previous time step, and is used only in the gain
calculation [39, 40].
An example of the target position and estimated errors are shown in “Fig. 11”
and the errors of positions are shown in “Fig. 11”. Under normal condition when
the positions of KF are compared to the SVSF as can be seen in “Figs. 10” and
11, we found that the SVSF performed significantly better than the EKF in term
of position. In this context, the estimated positions are more accurately using the
SVSF and converged towards the true position value faster than KF. The RMSE was
significantly lower for the SVSF and it has the best result.
9 Conclusion
In this paper, we have studied and presented different methods of moving object
detection and tracking from video frames with camera motion compensation. Back-
ground subtraction is a widely used approach for detecting moving objects by a
static camera using simple algorithm, however, it is very sensitive to illumination
changes, non-rigid moving objects shadows and occlusion. In order to track a non-
rigid moving object and extract its contour information, we have proposed an effec-
tive method based-on geodesic active contour. An easy technique is presented for
object detection based on colour threshold to extract image information from a static
background. Moving object detection is relatively difficult for video captured by a
moving camera. To solve this problem, an algorithm for camera motion compensa-
tion is proposed using video stabilization technique. Thus, moving object tracking
with a moving camera is realized in three steps: firstly, the camera motion is compen-
sated. Secondly, a geodesic active contour is used to detect the moving target and
the last step is described by the robust smooth variable structure filter compared to
the Kalman filter. These filters are implemented to estimate and tracked the mobile
object position online. The proposed algorithm is validated using a different scenario
(static and dynamic background) and good results are obtained.
In future work, intelligent moving object detection and tracking algorithm will be
proposed based on learning invariant features transforms to solve the Moving object
detection.
References
1. L. Leal-Taixe, A. Milan, K. Schindler, D. Cremers, I. Reid, and S. Roth, “Tracking the trackers:
an analysis of the state of the art in multiple object tracking,” arXiv:1704.02781, 2017
2. B. Risse, M. Mangan, L. Del Pero, and B. Webb, “Visual Tracking of Small Animals in Cluttered
Natural Environments Using a Freely Moving Camera,” in Proceedings of the IEEE Conference
on Computer Vi- sion and Pattern Recognition, 2017, PP. 2840–2849
3. Bouwmans T (2014) Traditional and recent approaches in background modelling for
foreground detection: An overview. Computer Science Review 11:31–66
4. A. Yilmaz, O. Javed, and M. Shah, “Object tracking: A survey,” in ACM Computing Surveys
(CSUR), vol. 38(4), 2006
5. Asim R. Aldhaheri and Eran A. Ediris- inghe, “Detection and Classification of a Moving
Object in a Video Stream,” in Proceedings of the Intl. Conf. on Advances in Computing and
Information Technology-ACIT, 2014
6. M. Piccardi, “Background subtraction techniques: a review,” International Conference on
Systems, Man and Cybernetics, Oct. 2004, PP. 3099–3104
7. N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in Proceedings
of IEEE Computer Vision and Pattern Recognition, vol. 1, Jun 2005, PP. 886–893
8. S. Lefevre, E. Bouton, T. Brouard and N.Vin- cent, “A new way to use hidden Markov models
for object tracking in video sequences,” in Proceedings of IEEE International Conference of
Image Processing, ICIP’2003, Sept 2003, vol. 3, PP. 111–117
9. C.P. Papageorgiou, M. Oren and T. Poggio, “A general framework for object detection,” in
sixth IEEE international conference on Computer vision, Jan 1998, PP. 555–562
10. A. Yilmaz, O. Javed, and M. Shah, “Object tracking: A survey,” in ACM Computing Surveys
(CSUR), 2006, vol. 38(4)
11. P.D. Mahamuni, R.P Patil and H.S. Thakar, “Moving Object Detection Using Back- ground
Subtraction Algorithm using simulink,” in IJERT, vol. 3(6), June 2014
12. A. R. Aldhheri and E.A. Edirisinghe, “Detection and classisification of a moving object in
a video stream,” in Proc. of the Intl. Conf. on Advances in Com- puting and Information
Technology-ACIT, 2014
13. O. Khalifa, “Performance Comparison of Background Estimation Algorithms for Detecting
Moving Vehicle,” in the World Applied Sciences Journal of Mathematical Applications in
Engineering, PP.109– 114, 2013
14. Cheng H, Jiang X, Sun Y, Wang J (2001) Color image segmentation: Advances prospects. Pat.
Rec. 34(12):2259–2281
15. Wu J, Yan H, Chalmers A (1994) Color image segmentation using fuzzy clustering and
supervised learning. Journal of Elec. Imag. 3(4):397–403
16. P. Schmid, Segmentation of digitized dermatoscopic images by two-dimensional color
clustering, IEEE Trans. on Med. Image., Vol. 18, No.2, pp. 164âA˘ S¸ 171, Feb. 1999
17. Wilson, E., C. Lages, and R. Mah. “Improved research for overlapping segmentation based on
watershed algorithm,” 2002, Tulsa, OK, United States: Institute of Electrical and Electronics
Engineers Inc
18. Otsu N (1979) A threshold selection method from gray level histograms. IEEE Transactions
on Systems, Man and Cybernetics 9(1):62–66
19. Xu H, Liu T, Wang G (2014) Hybrid geodesic region-based active contours for image
segmentation. Computers Electrical Engineering 40(3):858–869
20. M. Kass, A. Witkin and D.S. Terzopoulus, “Active Contour Models,” in International. Journal
Of Computer Vision, vol. 1, PP.321–331, 1988
21. M.K. Pawar, S.S. Lokhande and N.N. Bapat, “Iris Segmentation using Geodesic Active
Contour for Improved Texture Extraction,” in Recognition.International Journal of Computer
Applications, vol. 47(16), PP. 0975–888, 2012
22. Waykole TS, Jain YK (2013) Detection and Tracking of Moving Objects from Video.
International Journal of Computer Applications 81(18):0975–8887
23. Casselles V, Kimmel R (1997) Geodesic Active Contours. Int J Comput Vision 22(1):61–79
24. C. Petitjean, “Contours actifs pour la segmentation d’indices visuels dans un visage,” 2000
25. Kim SK, Kang SJ, Wang TS, Ko SJ (2013) Feature point classification based global motion
estimation for video stabilization. IEEE Trans Consum Electron 59(1):267–272
26. Song C, Zhao H, Jing W, Zhu H (2010) Robust video stabilization based on particle filtering
with weighted feature points. IEEE Trans Consum Electron 58(2):570–577
27. Yang J, Schonfeld D, Mohamed M (2009) Robust video stabilization based on particle filter
tracking of projected camera motion. IEEE Trans Circuits Syst Video Technol 19(7):945–954
28. Bay H, Ess A, Tuytelaars T, van Gool L (2008) SURF: Speeded up robust features. Computer
Vis. Image Understanding 110(3):346–359
29. A. Alahi, R. Ortiz and P. Vandergheynst “FREAK: Fast Retina Key- point,” https://doi.org/10.
1109/cvpr.2012.6247715 in IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 2012
30. J. Li, T. Xu, and K. Zhang, “Real-Time Feature-Based Video Stabilization on FPGA,” https://
doi.org/10.1109/tcsvt.2016.2515238, IEEE Transactions on Circuits and Systems for Video
Technology, 2015
31. https://in.mathworks.com/discovery/affine- transformation.html
32. S. Haykin, “Kalman Filtering and Neural Networks,” John Wiley and Sons Inc., 2001
33. Y. Rui, and Y. Chen, “Better proposal distributions: object tracking using unscented particle
filter,” Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2001,
PP. 786–793
34. D.A. Zaugg, A.A. Samuel, D.E. Waagen and H.A. Schmitt, “A combined particle/kalman
filter for improved tracking of beam aspect targets,” IEEE Workshop on Statistical Signal
Processing, 2003, PP. 535–538
35. F. Demim, A. Nemra, K. Louadj, A. Boucheloukh, E. Kobzili, A. Allam, M. Hamerlain and

A. Bazoula, “Visual SVSF-SLAM Algorithm Based On Adaptive Boundary Layer Width,”
(Eds.): ICEECA’2017, Lecture Notes in Electrical Engineering book series (LNEE), vol. 522,
PP. 97–112, SPRINGER Nature Switzerland AG, September 2018
36. Demim F, Nemra A, Louadj K (2016) Robust SVSF-SLAM for Unmanned Vehicle in Unknown
Environment. Journal of ScienceDirect IFAC-PapersOnLine 49:386–394
37. Demim F, Nemra A, Louadj K, Hamerlain M, Bazoula A (2017) “Cooperative SLAM
for Multiple UGVs Using SVSF Filter”, Journal for Control, Measurement, Electronics.
Computing and Communications (Automatika) 58(1):119–129
38. Demim F, Nemra A, Louadj K, Hamerlain M, Bazoula A (2017) An Adaptive SVSF-SLAM
Algorithm to Improve the Success and Solving the UGVs Cooperation Problem. J Exp Theor
Artif Intell 1:1–26
39. Demim F, Nemra A, Louadj K, Hamerlain M (2018) Simultaneous Localization, Mapping,
and Path Planning for Unmanned Vehicle Using Optimal Control. Journal of Advances in
Mechanical Engineering 10(1):1–25
40. Demim F, Nemra A, Boucheloukh A, Hamerlain M, Bazoula A (2018) Cooperative Visual
SLAM Algorithm based on Adaptive Covariance Intersection. Journal of Advanced Engi-
neering and Computation 2(3):151–163
Potential of Using Adaptive Complex
Flower Pollination Algorithm
to Interferometric Coherence
Optimisation
Sofiane Tahraoui and Mounira Ouarzeddine
Abstract This paper investigates the potential of using Adaptive Flower Pollina-
tion Interferometric coherence optimization (AFPICO). The algorithm was proposed
previously and is based on complex Flower Pollination Algorithm (FPA). FPA is a
newly developed algorithm based on the pollination process of plants, that is used to
solve constrained and/or multi-objective optimization problems. In it, we highlight
the additional options when using AFPIC-Optimization instead of classical methods,
such as the optimization of specific scattering mechanisms, for example, surface,
dipole-like, or dihedral scatterings. Also, we show how to exploit the approach as a
for scattering-type classification.
Keywords Interferometric coherene · SAR · Optimization · Classification ·

Scattering mechanisms
1 Introduction
The electromagnetic scattering phenomenon of a medium can be completely

described using the coherence matrix [1, 2]. In general, the medium encountered
exhibits specific geometric properties at different polarization, which is detectable
through the associated covariance/coherence matrix form.
Being its strong dependence on to the polarization state and its highest sensitivity
to change of the electromagnetic scattering [3, 4], the interferometric coherence could
be used in different applications such as forests heights inversion, biomass estimation,
SAR tomography, and many other interferometric applications. Rather, it is used to
evaluates the feasibility of the technique itself by measuring the correlation between
a pair of the radar signal backscattered from the same target.
The optimization of the interferometric coherence is the key element for a lot
of these applications. Many algorithms have been proposed in this context [5–8].
S. Tahraoui (B) · M. Ouarzeddine

Faculty of Electronics and Computer Science, LTIR, USTHB, BP no. 32 El Alia, Bab Ezzouar,
Algiers, Algeria

https://doi.org/10.1007/978-981-15-6403-1_85
1212 S. Tahraoui and M. Ouarzeddine
Almost all of the methods are based on the definition of unknown scattering vectors,
aiming to combine the information coming from the polarimetric channels of the
two interferometric images. The physical interpretation of these scattering vectors is
a very important point, as though they represent the weighted proportions of three
orthogonal unitary pure scattering phenomenon within the resolution cell [9].
The methods mentioned above, use eigenvalue/eigenvector-based optimization
(Eigen-based) which give abstracts scatterings vectors. Recently in [4], a new coher-
ence optimization method has been proposed, which is based on a complex-valued
Flower Pollination Algorithm [10] called” AFPICO” Adaptive Flower Pollination
for Interferometric Coherence Optimization. This algorithm has the ability to find the
optimized scattering vectors that represent the dominant mechanism phenomenon.
And also, able to optimize a specific scattering mechanism.
In this paper, we try to highlight the advantages of using AFPICO instead of
Eigen-based optimization methods, and how this can help open up a new perspective
in polarimetric SAR interferometry, such as the ability to optimize specific scattering
mechanisms (i.e. surface scattering, dihedral scattering, and volume scattering). And
how to exploit this new optimization approach as a for land classification.
The remaining of this paper is structured as follows. Section two presents the
optimization method AFPICO, and the associated mathematical development. In the
third section, we show how to interpret the scattered vector in terms of scattering
phenomenon. Then we show how to exploit this ability as a new classification method.
Finally, some conclusions are drawn.
2 Basic Notions
2.1 The Interferomatric Coherence
Polarimetric SAR Interferometry (PolInSAR) is based on the coherent combina-

tion of SAR acquisitions at different polarizations [11]. The complete information
measured by the PolInSAR system can be represented in form of three 3 × 3 complex
matrices [T11 ], [12 ], and [T22 ] formed using the outer products of scattering vector
from each antenna K p1 , and K p2 represented in Pauli basis as:
[T11 ] = K p1 .k∗p1
[T22 ] = K p2 .k∗p2
[12 ] = K p1 .k∗p2 (1)
[T11 ], and [T22 ] are Hermitian coherency matrices that describe the polarimetric
properties for each acquisition separately, however, and [12 ] is a non-Hermitian
complex matrix which contains the interferometric phase information.
Potential of Using Adaptive Complex Flower … 1213
By Introducing two unity complex vectors ω 1 and ω

2 which may be interpreted
as generalized scattering mechanisms, we are able to generate two complex scalar
images Im1 and Im2 by projecting the scattering vectors K p1 and K p2 into this vectors,
as:
1∗T .k∗p1
Im1 = ω
2∗T .k∗p2
Im2 = ω (2)
1 and ω
The interferogram related to the scattering mechanisms ω 2 is then given
by
∗T
Im1 Im2 = ω 1∗T .k∗p1 ω 2∗T .k∗p2 1∗T [12 ]ω
=ω 2 (3)
And the corresponding interferometric phase follow as:

∗
∗T
φ = arg Im1 Im2 = arg ω
1 [12 ]ω
2 (4)
The normalized complex correlation between Im1 and Im2 estimated at different
1 and ω
polarizations for an arbitrary choice of scattering mechanisms ω 2 may be
derived as:
1∗T [12 ]ω
ω 2
γ̃ = = γ e jφ (5)
1∗T [T11 ]ω
ω 2∗T [T22 ]ω
1ω 2
where γ is the module of the complex coherence γ̃ , and ϕ is the interferogram phase.
The angular brackets () are standing for spatial averaging over several samples, which
aim to reduce the speckle noise, hence providing a more reliable estimate of the degree
of correlation [12].
The coherence optimization methods aim to find the maximum of coherence
magnitude with respect to ω 1 and ω 2 . Analytically, these two vectors could be
interpreted as a particular dominant scattering mechanism. Hence, their geometric
parametrization can be written as [9]:
⎛ ⎞
cos αi
i = ⎝ sin αi cos βi e jδi ⎠
ω (6)
sin αi sin βi e jχi
where i ∈ {1, 2}, α corresponds to the scattering types (i.e. from surface scattering
α = 0o , passing by dipole-like mechanism when α = 45o , to double bounce scat-
tering when α = 90o ). The angle β characterizes the polarization orientation, δ angle
indicates the phase difference between the (S H H + SV V ) and (S H H − SV V ) terms.
And the angle δ is the phase difference between the (S H H + SV V ) and (S H V ) terms.
2.2 Coherence Optimization Using AFPICO
AFPICO uses a Flower Pollination algorithm FPA [10], to find the optimized vectors
that represent the dominant scattering mechanism, and maximize the module of the
coherence.
The Flower Pollination Algorithm (FPA) is a nature-inspired optimization
algorithm that imitates the pollination behavior of flowering plants.
This process can occur at two different search scales, namely local scale, when the
pollinator seeks for pollen on the neighboring. And at global scale when Pollinators
jump or fly distant steps. The switching between global and local scale occurs in a
probabilistic way. FPA has been inspired by this process and implemented using the
following equations:

xit+1 = xit + L xit − g∗ (7)
This equation simulates global step, where xit is the ith pollen, in other words is
the suggested solution vector at the ith iteration, g∗ is the optimal solution at this
iteration, and the parameter L imitates the global step size. This parameter gives the
ability to move over a long various distance’s steps. These distances are imitated
based on Lévy-flight1 .

xit+1 = xit + ε x tj − xkt ε ∈ [0, 1] (8)
This equation simulates both local pollination and flower constancy in a limited
neighborhood when x tj and xkt represent pollens from different flowers of the same
plant species.
Mathematically, x tj and xkt are selected from the same population, this equivalently
becomes a local random walk if the parameter ε is drawn from a uniform distribution
between 0 to 1 [10].
The use of FPA for Interferometric coherence optimization, allows the optimiza-
tion of a specific scattering, such as surface scattering, double bounce scattering, and
volume scattering [13]. So, the goal is to determine the two weight vectors ω 1 and
ω
2 that represent a real scattering mechanism and enhance the coherence module
in Eq. (5). To this end, some FPA parameters should be defined in this context as
follow:
• Population Size (SzP): Since the solution to optimization problem in FPA is based
on population scope (SzP). This parameter simulates the number of pollinators
(for example insects) which contributes in the pollination process, where each
pollinator brings a random pollen (i.e. random couple of vectors).
1 The interested reader is referred to Chap. 11 of the book entitled” Nature-Inspired Optimization
Algorithms” by Xin-She Yang [10] for the methodological framework and the implementation of
the method.
• Switching Probability (SwP): Although FPA process can occur at both global
and local scales, as in nature, its frequency of occurrence at local scale, is higher
than at global scale. This, is because the insect tends to find pollen in its adjacent
vicinity. This feature is controlled by switching probability (SwP).
In summary, the optimization process using AFPICO is restated by the following
Algorithm 1.
After the initialization of the algorithm parameters, we define a pixel scattering-
type by defining the angle α using H/A/α decomposition.
Then the optimization turns around the remaining angle of the scattering vector
(6), namely β,δ, and χ . Thus, the algorithm generates a defined number (i.e.”Sz P”)
of these angles, then after finding the best solution in the initial population that
maximizes the coherence, this result is improved by using local or/and global polli-
nation during a number of iterations. The Switching Probability ‘’SwP” controls the
probability of switching between global and local scales search.
In order to speed up the optimization process, we have proposed a substantial

modification in the original algorithm [4], where the switching between global and
local search is controlled every loop by updating the switching probability factor
under the criterion of improvement or worsening of the solution. That is, as long
as the solution still getting improved we increase the probability of seeking in the
neighboring (i.e. local pollination), and vice versa. After testing the whole generated
vector, the best solution is saved for the pixel under process. This operation is repeated
a number of iterations in order to refine the solution.
3 The Added Values
In order to highlight the added value of the proposed optimization method (AFPICO),
we have used the E-SAR repeat pass PolInSAR data acquired at L band over the
Oberpfaffenhofen area in Germany (Fig. 2a) by the German Aerospace Centre using
E-SAR radar. The temporal baseline is within one day. The site contains various
properties, namely residential areas, forested area, bare soil areas, and an aircraft
runway.
3.1 Interferometric Coherence Improvement
Figure 2d, shows the histograms of coherence associated with AFPICO, and clas-
sical coherence namely HH, HV, and VV, of the small area (red dashed rectangle
in Fig. 2). The general assessment reveals the improvement of the optimized coher-
ence compared to lexicographic ones. Regardless of the improvement of coherence
product, and phase stability improvement [4].
The added-value of AFPICO compared to the classical methods, is that the opti-
mized vectors ω 1 and ω
2 are linked physically by imposing the α parameter means
between the two acquisition. That mean that we force the optimization of the same
scattering in the resolution cell to be optimized. Note that the parameter α (see Fig. 2c)
is related to the scattering type, namely surface, dihedral, or volume scattering.
Also, in contrast to previous methods, the optimization is performed in the two
images by taking into account the geometric deformation induced by high spatial
baseline length, which may bias the scattering center. Furthermore, to the ability to
perform an ordinary optimization, AFPICO gives additional possibilities in contrast
to the classical optimization methods such as:
• Optimize specific scattering: The ability to optimize a particular scattering mech-
anism amongst others, such as surface scattering, dihedral scattering, dipole-like
scattering…etc.
• Optimize the dominant scattering: Performs predefinition of dominant scat-

tering mechanism using H/A/α decomposition. After defining parameter α, the
optimization turns around the remaining angle of the scattering vector (6).
• Know the remaining angles: Possibility to know the detailed information about
the optimized vector (6), such as the orientation angle β, and the phases difference
δ and χ.
• A simple implementation of assumptions: In case of the assumptions such as;
one scattering mechanism in the two images ω 1 = ω 2 , which is the case for
example in RVOG model, the optimization is easy to implement and gives good
outcomes in terms of the product of coherence and phase quality.
3.2 RGB Coherence for Classification
At this stage, one may ask: What is the uniqueness and importance of AFPICO
optimization method compared to the previous ones?
The response to this question is the RGB coherence of Fig. 2b which strongly
highlights the benefit and the uniqueness of the method.
As we aforementioned, AFPICO, gives the possibility to control the optimized
vectors ω 1 and ω
2 . Before we presenting this new concept, it is important to point out
that the behavior of the coherence changes from one scatterer to another, and from
polarization to another [14], That could be observed in the pattern of the distribution
of its loci in the complex plane (see Fig. 1). The angular distribution of the coherence
corresponds to the height dispersion of the phase centers. These latter are substantially
governed in one side by the density and the width of the target, and in the other side by
the radar frequency and polarization [15]. Whereas the radial dispersion corresponds
Fig. 1 Coherence loci in the z 90

complex plane
120 60
γT v
150 30
gr ou n d t op ogap h y :φ 0
180 0
z
210 330
240 300
270
to the coherence magnitude. A value of coherence close to 1 indicates that there is

almost no penetration.
This fact has incited us to exploit this property for land classification since we
can optimize specific scattering using AFPICO. Figures 2b, displays the RGB color
composite image of a combination of three coherence from different scattering mech-
anisms, namely dihedral/helix (red color), dipole like (green), and surface (blue
color). This is realized by setting α term for each mechanism (i.e. from surface scat-
tering α = 0o , passing by dipole- like mechanism when α = 45o , to dihedral scattering
when α = 90o ). So, the optimization can be used as a decomposition process.
Note that the RGB coherence gives a high level of discrimination between the
scattering class. For example, the aircraft runway appears with darker blue compared
to other surfaces. The urban areas appear in white color, whereas the forested regions
in black. The areas with the purple color, which is a mixture of green that represents
the vegetation and blue that represents the bare surfaces, indicating that these regions
contain equivalent proportions of surface and vegetation, that is, grassy surface.
(a) (b)
(c) (d)
Fig. 2 a Optical image of test data in Oberpfaffenhofen. (source: Google Earth). b RGB coherence
of Oberpfaffenhofen; red: α = 90o , green: α α = 45o , blue:α α = 0o . c The angle α retrieved from
the optimized vector this parameter represents scattering mechanism type. d Histograms associated
with AFPICO (Blue color), lexicographic coherence
Indeed, the final aim is to use this coherence decomposition in height retrieval,
which may help to increase its reliability.
4 Conclusion
In this paper, we tried to highlight the potential of using a newly proposed inter-
ferometric coherence method named AFPICO, which is based on complex Flower
Pollination Algorithm FPA. AFPICO has been recently evaluated and its perfor-
mance was compared to classical optimization through several tests including their
statistics, their interferograms, their phase profiles as well as their number of residues,
showing its good outcomes [4]. Complementary notes have been summarized in this
paper, showing the benefits that could be drawn when using this method.
Regardless That AFPICO is the one that gives the best interferometric phase
product and less rate of residues compared to the classical method. AFPICO gives
extra possibilities such as the ability to optimize a one scattering mechanism amongst
others, the ability to optimize only the dominant scattering mechanism by defining
parameter α beforehand, the Knowledge of the detailed information about the opti-
mized scattering, such as the orientation angle β, and the phases difference δ and
χ, and exclusivity of use it for land classification. All this plus its simplicity of
implementation.
This research axis is indeed very recent and requires more attention, and with the
evolution of parallel programming and a quantum computer, the obstacle of execution
time in iterative algorithm will no longer arise, especially in the case of a pixel by
pixel processing. In the field of interferometric coherence optimization, AFPICO’s
performance demonstrates its suitability for the PolInSAR applications and emerges
as a new competitive approach to the classical method.
Acknowledgments The authors would like to thank the European Space Agency (ESA) for the
data provision.
References
1. Freeman A, Durden S (1998) A three-component scattering model for polarimetric SAR data.
IEEE Trans Geosci Remote Sens 36(3):963–973
2. Nghiem SV, Yueh SH, Kwok R, Li FK (1992) Symmetry properties in polarimetric remote
sensing. Radio Sci 27(5):693–711
3. Tahraoui S, Ouarzeddine M, Souissi B (2013) Interferometric Coherence Optimization: A
Comparative Study. In Eighth International Conference on Broadband and Wireless Computing,
Communication and Applications (BWCCA), Oct. 2013, pp 427–431
4. S. Tahraoui and M. Ouarzeddine, “Adaptive Complex Flower Pollination Algorithm for Interfer-
ometric Coherence optimisation.” in 2018 International Conference on Applied Smart Systems
(ICASS), Nov. 2018, pp. 1–6
5. Cloude SR, Papathanassiou KP (1998) Polarimetric SAR interferometry. IEEE Trans Geosci
Remote Sens 36(5):1551–1565
6. E. Colin, C. Titin-Schnaider, and W. Tabbara, “Investigation on different interferometric
coherence optimization methods,” in Applications of SAR Polarimetry and Polarimetric
Interferometry, vol. 529, Jan. 2003, pp. 1–6
7. S. Guillaso, A. Reigber, L. Ferro-Famil, “Evaluation of the ESPRIT approach in polarimetric
interferometric SAR”, in Proceedings (2005) IEEE International Geoscience and Remote
Sensing Symposium, 2005. IGARSS’05., vol. 1. Seoul, Korea: IEEE, Jul. 2005:32–35
8. Colin E, Titin-Schnaider C, Tabbara W (2006) An interferometric coherence optimization
method in radar polarimetry for high-resolution imagery. Geoscience and Remote Sensing,
IEEE Transactions on 44(1):167–175
9. Cloude SR, Pottier E (1997) An entropy based classification scheme for land applications of
polarimetric SAR. IEEE Trans Geosci Remote Sens 35(1):68–78
10. X.-S. Yang, Nature-Inspired Optimization Algorithms, paperback reprint of hardcover 1st ed.,
2014 edition ed. Amsterdam Boston Heidelberg London New York Oxford Paris San Diego
San Francisco Singapore Sydney Tokyo: Elsevier, Sep. 2016
11. S. R. Cloude and K. P. Papathanssiou, “Polarimetric optimisation in radar interferometry,”
Electronics Letters, June 1997, Vol. 33, No. 13, pp. 1176–1178, Jun. 1997
12. Lee J-S, Cloude SR, Papathanassiou KP, Grunes MR, Woodhouse IH (2003) Speckle filtering
and coherence estimation of polarimetric SAR interferometry data for forest applications. IEEE
Trans Geosci Remote Sens 41(10):2254–2263
13. Cloude S, Pottier E (1996) A review of target decomposition theorems in radar polarimetry.
IEEE Trans Geosci Remote Sens 34(2):498–518
14. M. Tabb and R. Carande, “Coherence region shape extraction for vegetation parameter esti-
mation in polarimetric SAR interferometry,” in Geoscience and Remote Sensing Symposium,
2002. IGARSS’02. 2002 IEEE International, vol. 5, 2002, pp. 2596–2598 vol.5
15. Treuhaft RN, Siqueira PR (2000) Vertical structure of vegetated land surfaces from interfero-
metric and polarimetric radar. Radio Science 35(1):141–177
Intelligent Cooperative Method
for Medical Image Segmentation
Karima Benhamza and Hamid Seridi
Abstract This paper presents a robust segmentation scheme for noisy medical
images exploiting hybrid cooperative segmentation. Two competing process use
duality between edge detection and region extraction methods to improve their mutual
results. Parallel execution and information exchange are combined to reduce the error
rates generated by each of the methods executed independently. Proposed adaptive
scheme is evaluated using Mean Squared Error (MSE) and Shannon entropy methods.
The obtained results show the effectiveness of the method applied on real clinical
data.
Keywords Medical image · Edge detection · Region segmentation · Cooperative

segmentation
1 Introduction
Many imaging modalities such Magnetic resonance imaging (MRI), computed

tomography (CT), digital mammography provide valuable details about the anatomy
of a subject and are essential technological means for diagnosis and planning
treatment.
Medical images are essential parts of diagnosis. Indeed, Object shape and appear-
ance are two important sources of information that when integrated, provide more
reliable image interpretation. Image segmentation plays a decisive role in many
medical applications by automating the extraction of anatomical regions of interest
such localization of pathology and study anatomical structures.
However, medical images are limited by low contrast, high noise that can alter the
intensity of a pixel such that its classification becomes uncertain. In addition, images
K. Benhamza (B) · H. Seridi

LabSTIC, 8 Mai 1945 Guelma University, BP 401, Guelma, Algeria
H. Seridi

https://doi.org/10.1007/978-981-15-6403-1_86
1222 K. Benhamza and H. Seridi
exhibit intensity inhomogeneity and the level intensity of a single tissue class varies
gradually over the extent of the image intensity.
Thus, medical images tend to pose greater difficulty in shape recovery compared
with other types of images and greatly complicate the task of segmentation. To
address these problems, various extensions of existing segmentation methods have
been proposed to have more efficient and accurate results [1].
Although many algorithms have been proposed in this field, medical image
segmentation continues to be a complex and challenging problem.
This paper proposes a hybrid segmentation technique integrating in the same
scheme, a mutual cooperation segmentation and a results cooperation segmentation.
The aim is to reduce the error rate generated by each of the methods used inde-
pendently and to perform a reliable segmentation in noisy medical images. These
images are usually subject to no-uniform and fadeout illumination and loss of focus.
This paper is organized as follows: In Sect. 2, related medical image segmentation
works are presented. Then proposed method is detailed in Sects. 3 and 4 comparative
results are exposed. Conclusion is given in Sect. 5.
2 Medical image segmentation Review
Medical images segmentation can relate to different tasks such as delineating

the contours of an organ or extracting its internal structures that provide useful
information for many clinical decisions.
Numerous researches has been devoted to medical image segmentation in the
literature [2–4]. Thresholding offers an effective tool for obtaining a segmentation
in images where different structures have contrasting intensities. Several algorithms
[5] have been developed. However, spatial characteristics of an image are not taken
into account in these algorithms.
In the other hand, region growing is a technique for extracting image region based
on some predefined criteria. This method can also be sensitive to noise, causing
separate regions or many holes in extracted regions [6, 7].
Edge detection methods are the most common methods in medical image segmen-
tation [8–10]. The limits of edge based method in medical field are diverse. Indeed,
edge performance is affected by the presence of noise causing the presence fake and
weak edges in the detected edge image.
Markov Random Fields (MRF) are widely used to model segmentation classes,
texture properties and intensity inhomogeneity that occur in medical images [11]. A
difficulty associated with MRF models is proper selection of the parameters control-
ling the strength of spatial interactions. Too high a setting can result in an exces-
sively smooth segmentation and a loss of important structural details. In addition,
they usually require computationally intensive algorithms.
Nevertheless, these methods require the setting parameters and thresholds whose
optimal values are difficult to choose in X-ray or ultrasound images. MRF may fail
Intelligent Cooperative Method for Medical … 1223
to separate foreground and background in the presence of intensity inhomogeneity

and it often generates small holes inside the segmented object.
Recently, there has been a new tendency toward the cooperation of edge and region
segmentation in order to produce better results Segmentation based on region-edge
cooperation permits one to exploit the two pieces of complementary information in
order to integrate them into the same segmentation scheme and, thus, contribute to a
better segmentation [12].
Many cooperative approaches have been developed, but three categories can be
identified: sequential cooperation, mutual cooperation and results cooperation [13].
These approaches differ by how the regions and edges are integrated. Sequential
cooperative methods produce good results when used methods in the processing chain
produce a satisfactory results. However, there is no generic method that effectively
treat all types of images. In mutual cooperation, the two segmentation techniques
exchange information in parallel. In the other hand, the cooperative approach of
results is delicate because it offers a redundancy of information. This would lead to
a poor quality image segmentation. Few works used the cooperative approaches in
medical image segmentation.
In work [14], authors propose a hybrid segmentation method based on level set
integrating boundary and region information of images. The main contributions of
this work are to modify the velocity function for the boundary-based level set method,
and to design a novel energy function as a stopping criterion.
This velocity function is modified according to the statistical characteristics of
the segmented regions during the evolution. Thus, the stopping criterion depends on
not only the boundary information of the image but also the statistical characteristics
of the segmented regions, which can overcome the over-segmentation effectively.
In the paper [15], authors intend to combine watershed algorithm with Geometric
Vector Flow (GVF) snake model to reduce the computational complexity, to improve
the insensitiveness to noise, and capture range. The image is segmented firstly through
watershed algorithm and then the edges produced will be the initial contour of GVF
model. This allowed to improve boundaries by coupling the smoothness of the edge
map obtained due to watershed algorithm. The proposed method was compared with
hybrid segmentation algorithm based on watershed and balloon snake.
In [16], a hybrid region-based active contour model is presented to segment
medical images. An energy functional model has been proposed, consisting of three
weighted terms: global, local and regularization terms. This total energy is incorpo-
rated into a level set formulation with a level set regularization term, from which a
curve evolution equation is derived for energy minimization.
Luo et al. proposed a segmentation scheme to combine both region- and edge-
based information into the graph-based (RGB) segmentation method [17]. The only
interaction required is to select two diagonal points to determine a region of interest
(ROI) on the original image.
The ROI image is smoothed by a bilateral filter and contrast is improved by
histogram equalization. Next, the enhanced image is filtered by pyramid mean shift
to improve homogeneity.
With the optimization of particle swarm optimization (PSO) algorithm, the

RGB segmentation method is performed to segment the filtered image. Finally, the
segmented image is processed by morphological opening and closing to refine the
tumor contour.
The aim of the work [18] is to design a system applied to magnetic resonance
brain images (MRI) that took advantage of two complementary approaches: edge
segmentation and region segmentation. The most important part of the system is
the use of duality between edge detection and region extraction. Proposed model is
based on this work.
3 Proposed method
This article proposes a hybrid segmentation technique, integrating in the same

scheme, a mutual cooperation segmentation and a results cooperation segmentation.
The aim is to reduce the error rate, generated by each of the methods used inde-
pendently, and to perform a reliable segmentation of noisy medical images. General
scheme of proposed method is presented in Fig. 1. The main steps are described
below.
Developed system is composed of two modules: One devoted to the extraction of
uniform regions and the second to contour extraction.
The model is based on iterative cooperation between the two methods: edge detec-
tion and region growing. Edge detection allows the most evident frontiers of homo-
geneous regions to be found. The region-growing procedure serves to close contours
and to obtain more precise segmentation.
Fig. 1 Proposed cooperative segmentation scheme

Two types of cooperation are used: mutual and results cooperation. Both methods
are iterated until they converge towards coherent and stable results.
The compatibility of results is achieved by minimizing a dissimilarity measure
between the contours and the boundaries of regions which provide an optimal solution
in the sense of the compatibility between the results of the two segmentations.
1. Edge detection
We used Canny operator [19] as first step to obtain preliminary edge map. Canny
detector provides a simple and smart technique for locating the objects and extracting
their shape information.
Canny edge detection algorithm includes following steps: Smoothing image in
order to remove the noise. Then, calculation of the amplitude and direction of the
gradient. The third step is inhibition of non-maxima for the gradient amplitude to get
rid of spurious response to edge detection. Then, the detection with dual threshold
method to determine potential edges.
In this step, Otsu’s method [20] can be used to generate the high threshold. The
low threshold is typically set to half the high threshold. Finally last edge connection
step tracks edge by hysteresis (Fig. 2).
2. Region growing
Region growing is the popular algorithm for region based medical image segmen-
tation. This method presents many advantages: It can correctly separate the regions
that have the same properties or clear edges. It has a simple principle and noise sensi-
tivity is less than thresholding. Region growing has been widely used to extract the
potential lesions from background. However, the result of region growing depends
on the seed pixels selection and can produce holes in the extracted shape [21].
Implementation starts by selecting starting points called seed pixels. Edge infor-
mation is used to choose correct seeds emplacement. The growth of a region has to
start from inside it. Seed pixels are automatically chosen as local gray value extrema
of the image and where the gradient remains small. Then, the region grows by adding
similar neighboring pixels according to a homogeneity criterion (contrast between
selected pixel and region), increasing step by step the size of the region. The vector
of characteristics of the region constitutes of its mean intensity, its standard deviation
and the number of its pixels.
In this work, homogeneity criterion consists of the weighted sum of the contrast
between the region and the pixel, and the value of the modulus of the gradient of the
Original Gaussian Gradient amplitude & Non-maxima

image smoothing direction computation suppression
Output edge detection Edge connection Dual thresholding
Fig. 2 Flow chart of Canny edge detector [19]

pixel [22]. Combining the region growing mechanism with the gradient information
can be expressed by the following formulas:
N +1 N 1
X av = XN + f (i, j) (1)
N + 1 av N + 1
N
x(i, j) = X av − f (i, j) (2)
z(i, j) = (1 − ϕ)x(i, j) + ϕG(i, j) (3)
N
where X av is the average gray value of the region which is updated pixel by pixel.
The contrast of the current pixel with respect to the region is denoted by x(i, j).
Parameter ϕ controls the weight of gradient, G(i, j). Finally, the sum of the local and
the global contrast is the final homogeneity measure, z(i, j).
A low value of z(i,j) denotes the convenience of aggregating the pixel to the region.
Otherwise, local maximum of the gradients on neighborhood of the current pixel is
examined. During the course of the region expansion, when a pixel is added to a
region, its model is then updated. The procedure stops at the pixel with the local
gradient maximum.
3. Edges correction using the extracted regions
a. False edges reduction
To eliminate the false edges linked to the presence of noise and to confirm those that
are significant, the edge detection process exploits the information from the regions:
– An edge point is validated if its gradient belongs to the class of strong contrast
gradients.
– It is rejected if the regions that border it have the same labels and if its gradient
belongs to the class of the weak contrast gradients.
– If the edge point belongs to the class of medium contrast gradients, it is validated
when the regions situated on either side of the edge possess different labels and at
least one of its neighboring points has a gradient belonging to the strong contrast
class.
Short edges (one to three pixels) are eliminated within homogeneous resulted
regions. The small generated regions are also merged with parent region having
closest grey level.
b. Edge linking process.
This step consists in filling the discontinuities between points that represent the
same contour. The first stage consists of examining the neighboring chains of edge
points by applying criteria given through:
– Proximity: two chains cannot be merged if their distance is superior to a threshold;
– Continuity: the contour often being rectilinear, two chains have a strong
probability of merging if they are in the prolongation each from the other.
4. Regions’ boundaries correction using the detected edges
In addition to an imprecise location of the borders, the procedure of formation of

the regions can generate sub-segmentations in spite of the use of weak thresholds of
aggregation.
These errors are mainly due to noises that are superimposed on borders and have
gray levels similar to the region being formed.
To correct this defect, a second criterion involving the gradient of the examined
point was introduced in the aggregation procedure. So, in addition to the similarity
criterion used, the pixel must have a low gradient to be aggregated to the region.
5. Compatibility of results methods
The aim is to provide an adaptive model with compatible results of both segmen-
tations (edge and region). It is carried out by minimizing a dissimilarity measure
between detected edges and the boundaries of the regions [23]. The Hausdorff
distance is used to measure the similarity (error measuring) between edge and
boundary maps [24]. A short overview of its definition is presented bellow.
Initially conceived as a measurement between two sets of points A and B in
a metric space E, it is considered as a dissimilarity measure between two binary
images A and B, the Hausdorff distance can be defined as follows [24]:
Given two finite sets of points A=(a1 ,a2 …,an ) and B=(b1 ,…, bm ), and an
underlying distance d, the Hausdorff distance is given by:
DH (A, B) = max[h(A, B), h(B, A)] (4)
where:
h(A, B) = maxa∈A {minb∈B [d(a, b)]} (5)
Thus, the Hausdorff distance H(A,B) is the maximum of h(A,B) an d h(B,A). It

measures the degree of mismatch between two sets by measuring the distance of the
point of A that is farthest from any point of B and vice versa.
Let C be the map of edges and let F be that of the boundaries of the regions. The
adopted approach consists of calculating, for each edge (resp. boundary) point in
the image C (resp. F), the minimal deviation, which separates it from a boundary,
dE (c,F); (resp. from an edge, dE (f,C) point in the map F (resp. C).
The averages of the minimal deviations are then calculated for each map. The
final dissimilarity measure corresponds to the overall average of two averages.
It is given by:
1
Dissim (C, F) ≡ avgc [d E (c, F)] + avg f [d E ( f, C)] (6)
2
where avgc [dE (c,F)] (resp. avgf [dE (f,C)]) is the average of the minimal deviations
dE (c,F) (resp. dE (f,C).
4 Results Discussion
We have presented a method that integrates region and edge information to improve
both region segmentation and edge detection. The proposed method has been evalu-
ated on several medical images selected from MedPix-Medical image database [25].
MedPix is a free open-access online database of medical images, integrating nearly
59,000 images and textual metadata. The content is organized by disease location,
pathology category and by image classification.
Mean Squared Error (MSE) is used to compare results of proposed method with
results of region growing method. MSE denotes a signal fidelity measure. It provides
a quantitative score that describes the degree of similarity/fidelity or, conversely, the
level of error/distortion between two signals.
Suppose that A = { Aij | i = 1, 2, .., N and j = 1, 2,.., M } and B = { Bij | i =
1, 2,.., N and j = 1, 2, .., M } are two finite-length, discrete signals (visual images),
where N is the number of signal samples (pixels) and Aij and Bij are the values of
the th
ij samples in A and B, respectively.
MSE formula is expressed as follows:
1 M N 2
MSE = Ai j − Bi j (7)
MN x=1 j=1
Small MSE values mean lesser error.

Furthermore, performance of the proposed method is evaluated based on the accu-
racy in edge detection. Shannon [26] entropy is used to evaluate resulted edge detec-
tion map compared to Canny operator. This measure was introduced by Shannon as
a basic concept in information theory, measuring the average missing information
on a random source.
The Shannon’s entropy function is given by the following equation:
n
H(I) = − pi log pi (8)
i=0
Where, I represent the image whose entropy is to be measured. pi is the frequency

of pixels with intensity i. A higher entropy value denotes a better performance of
method.
Table 1 shows MSE values for selected medical images (Fig. 3). Table 2 shows
the entropy values for resulted edge detectors. It is observed that the proposed method
yields higher entropy as compared to Conventional Canny edge detector. On the other
hand, the small MSE values of the proposed method indicate a good extraction of
the region.
Table 1 MSE values

Images\Methods Region growing Proposed method
Image1 0.1125 0.0594
Image2 0.1246 0.0548
Image3 0.1061 0.0342
Image4 0.1021 0.0315
Image5 0.1535 0.0510
Image6 0.1102 0.0858
Image7 0.1137 0.0599
Fig. 3 Test Images
Image1 Image2
Image3 Image4
Image 5 Image6
Image 7
Table 2 Entropy values

Images\Methods Canny Proposed method
Image1 0.3221 0.9967
Image2 0.3484 0.9753
Image3 0.2084 0.6372
Image4 0.2126 0.6495
Image5 0.6357 0.9436
Image6 0.4245 0.8585
Image7 0.3168 0.9631
5 Conclusion
In this paper, hybrid cooperative segmentation system applied to medical images is

proposed. The system combines three concepts:(i) The integration of the informa-
tion resulting from two complementary segmentation methods: edge detection and
region extraction. This allows us to exploit the advantages of each. (ii) The active
information exchanges between these methods. This permits the adjustment of the
control parameters of the used methods. (iii) Introduction of a mechanism, which
checks the compatibility of results through a comparison of the two segmentation
maps.
The proposed method has been evaluated on several medical images selected from
MedPix-Medical image database. Mean Squared Error (MSE) and Shannon entropy
are used to compare results of proposed method with region growing and Canny
operator. Obtained results are encouraging and denote a large reduction of broken
edges even in noisy images. Parallel execution and information exchange between
the two methods have allowed a good edge and region detection in most clinical
images.
References
1. Norouzi A, Rahim MSM, Altameem A, Saba T, Rad AE, Rehman A, Uddin M (2014) Medical
image segmentation methods, algorithms, and applications. IETE Technical Review 31(3):199–
213
2. Withey DJ, Koles ZJ (2008) A review of medical image segmentation: methods and available
software. International Journal of Bioelectromagnetism 10(3):125–148
3. Masood, S., Sharif, M., Masood, A., Yasmin, M., & Raza, M. (2015). A survey on medical
image segmentation. Current Medical Imaging Reviews, 11(1), 3-14.]
4. Litjens G, Kooi, T., Bejnordi, B. E., Setio, A. A. A., Ciompi, F., Ghafoorian, M., & Sánchez,
C. I. (2017). A survey on deep learning in medical image analysis. arXiv preprint arXiv:1702.
05747
5. Senthilkumaran, N., & Vaithegi, S. (2016). Image segmentation by using thresholding tech-
niques for medical images. Computer Science & Engineering: An International Journal,
6(1)
6. Oghli MG, Fallahi A, Pooyan M, Automatic region growing method using GSmap and
spatial information on ultrasound images, Electrical Engineering (ICEE), 2010 18th Iranian
Conference on; 2010: IEEE
7. Szénási S (2014) Distributed region growing algorithm for medical image segmentation.
International Journal of Circuits, Systems and Signal Processing 8(1):173–181
8. Liu X, Duan Z, Wang X, Xu W (2016) An image edge detection algorithm based on improved
wavelet transform. International Journal of Signal Processing, Image Processing and Pattern
Recognition 9(4):435–442
9. Benhamza, K. and Seridi, H. Canny edge detector improvement using an intelligent ants
routing. Evolving Systems, pp.1-10, 2019. https://doi.org/10.1007/s12530-019-09299-0
10. Lin WC, Wang JW (2018) Edge detection in medical images with quasi high-pass filter based
on local statistics. Biomedical Signal Processing and Control 39:294–302
11. Li SZ, Singh S. (2009). Markov random field modeling in image analysis: Springer Science &
Business Media
12. Cufi X, Munoz X, Freixenet J, Marti J (2003) A review of image segmentation techniques
integrating region and boundary information. Advances in imaging and electron physics 120:1–
39
13. Sebari I, He DC (2007) Les approches de segmentation d’image par coopération régions-
contours. Revue télédétection 7:499–506
14. Wang, G.; Liu, H.; Zhang, S.; Liang, J. A, (2009) Novel Hybrid Segmentation Method for
Medical Images Based on Level Set. In Proceedings of the 2nd International Congress on
Image and Signal Processing, Tianjin, China, 17–19 October 2009; pp. 1–5
15. Jayadevappa, D.; Kumar, S.; Murthy, D. A Hybrid segmentation model based on watershed
and gradient vector flow for the detection of Brain tumor. Int. J. Signal Process. Image Process.
Pattern Recognition 2009, 2, 29–42
16. Liu, T., Xu, H., Jin, W., Liu, Z., Zhao, Y., & Tian, W. (2014). Medical image segmentation based
on a hybrid region-based active contour model. Computational and mathematical methods in
medicine, 2014
17. Luo, Y., Liu, L., Huang, Q., & Li, X. (2017). A novel segmentation approach combining
region-and edge-based information for ultrasound images. BioMed research international, 2017
18. Cherfa Y, Cherfa A, Kabir Y, Kassous S, Jaillard A, Dojat M, Garbay C (2007) Segmentation
of Magnetic Resonance Brain Images Using Edge and Region Cooperation Characterization
of Stroke Lesions. Int. Arab J. Inf. Technol. 4(3):281–288
19. Canny J., A Computational Approach to Edge Detection, IEEE Transactions on Pattern Analysis
and Machine Intelligence, PAMI-8(6), pp. 679-698, 1986
20. Otsu N (1979) A threshold selection method from gray-level histograms. IEEE Trans Systems,
Man and Cybernetics 9(1):62–66
21. Rabottino G, Mencattini A, Salmeri M, Caselli F, Lojacono R (2011) Performance evalua-
tion of a region growing procedure for mammographic breast lesion identification. Computer
Standards & Interfaces 33(2):128–135
22. Munoz X, Freixenet J, Cufı X, Martı J (2003) Strategies for image segmentation combining
region and boundary information. Pattern recognition letters 24(1–3):375–392
23. Benhamza K, Seridi H (2018) Hybrid Cooperation Scheme for Medical Image Segmentation,
1st International Conference on Artificial Intelligence and its Applications (AIAP’18), pp.170–
173, December 04-05, 2018, El Oued, Algérie
24. Dubuisson, M. P., & Jain, A. K. (1994). A modified Hausdorff distance for object matching.
In Proceedings of 12th international conference on pattern recognition . IEEE .Vol. 1,pp. 566-
568
25. MedPix, Medical image database, https://medpix.nlm.nih.gov/home
26. Shannon CE (1948) A mathematical theory of communication. Bell System Technical Journal.
27:379–423
An Image Segmentation Model Using
a Level Set Method Based on Improved
Signed Pressure Force Function SPF
Larbi Messaouda, Zoubeida Messali, Rouini Abdelghani,

and Samira LARBI
Abstract Segmentation is a vital aspect of medical imaging. It aids in the visual-

ization of medical data and diagnostics of various dieses. In this paper, we present
a novel Level Set Method driven by New Signed Pressure Force function (SPF)
for image segmentation. The advantages of this method are Firstly, the Function of
Signed Pressure can efficiently stop the contours at weak or blurred edges, Second,
exterior and interior boundaries can be detected no matter where the initial contour
starts. We compare this method with Chan Vese (C-V) model and Geodesic Active
Contours (GAC). The performance of each method can be evaluated either visually,
or from similarity measurements between a reference and the results of the segmenta-
tion. We have applied each method for different medical images. Through simulated
results, we have demonstrated that the best results are achieved by the our method.
Keywords Image segmentation · Geodesic active contours (GAC) · Chan vese

(C-V) · Medical images · New function of signed pressure (SPFn)
L. Messaouda (B)
Department of Science and Technology, University of Batna, Batna, Algeria
Z. Messali
Department of Science and Technology, University of Bordj Bou Arreridj, Bordj Bou Arreridj,
Algeria
R. Abdelghani
Electrical Engineering Department, Applied Automation and Industrial Diagnostic Laboratory,
Djelfa, Algeria
S. LARBI
Department of Science and Technology, University of Laghouat, Laghouat, Algeria

https://doi.org/10.1007/978-981-15-6403-1_87
1234 L. Messaouda et al.
1 Introduction
Accurate image segmentation is a testing research point of expanding interest, partic-

ularly in the most recent decade. It has been, and still, a vital errand in the fields
including PC vision, restorative images, and image investigation. The segmentation
approaches vary starting with one methodology of medical images [1–3], then onto
the next. In other words, the procedure for segmentation ought to be done, as indi-
cated by a methodology of obtaining (scanners, radiography, Magnetic Resonance
Images).Different methods have been proposed for image segmentation including
the Active Contour Models (ACM) [4, 5]. The fundamental idea of ACM is to
evolve a curve under some constraints from a given image to reach the boundaries
of the desired objects, by minimizing energy functional. According to the energy,
ACM are classified into two main categories: edge-based models [6–8] and region-
based models. Edge-based models rely on the image gradient information to stop the
evolving curve on the desired object boundaries [9–14]. Geodesic Active Contour
(GAC) is one of the well-known models in this category. Although GAC has been
successfully applied for images with high variation in gradient at the object’s bound-
aries, it meets difficulty when dealing with the objects having blurred or discrete
boundaries and it hardly detects the object corrupted by noise. Unlike to edge-based
models, the region-based models do not utilize the image gradient; they utilize image
statistics inside and outside the curve to control the evolution with better performance
including weak edges and noise. One of the famous region-based models is the Chan
and Vese (C-V) model which is a simplified Mumford-Shah model. The C-V model
utilizes the global information of homogeneous regions. Thus, it has good segmen-
tation result for the objects with weak or discrete boundaries but often has erroneous
segmentation for images with intensity inhomogeneity.
To ameliorate the robustness to contour initialization, we proposed a hybrid model
combining the local and global intensity information, which utilizes a region-based
Signed Pressure Force (SPF) function to update curve evolution to regularize the
Level set function. This technique named Selective Binary and Gaussian Filtering
Regularized Level Set (SBGFRLS) method [17], is based on Chan Vese (C-V) model
[8] and Geodesic Active Contours (GAC) [7]. Both these models are based on active
contour model. The SBGFRLS model is able to overcome the problems of the GAC
and the C-V model but still it is not good for medical images. It is not able to detect all
boundaries. So here in the proposed model a new SPF function (SPFn) is developed
to overcome this problem. Proposed model also takes less time to converge.
The paper is structured as follows: In Sect. 2, we describe two distinguished classes
of segmentation approaches the classic GAC and C–V models. Section 3 describes the
formulation of our proposed methods. In particular we show how Gaussian smoothing
kernel is used to regularize a level set function. Result discuss of implementation
issues of the considered methods is presented in Sect. 4. We evaluate the behavior
of the methods using biomedical images. The main conclusion of this work is given
in Sect. 5.
An Image Segmentation Model Using a Level … 1235
2 Methods of Segmentation
In conventional level set formulations, the level set function usually builds up irreg-
ularities through its progress, which could cause numerical errors and may destroy
the stability of the improvement. So, a numerical solution, named re-initialization, is
usually practical to commonly substitute the degraded level set function by a signed
distance function. The main contribution of Chamming Li et al. [15] is the elimination
of the re-initialization procedure thanks to their proposed variational formulation for
geometric active contours. In this variational formulation, the regularity of the level
set function is basically kept through the level set progress. The level set progress is
resulting as the gradient flow that reduces energy functional with a distance regular-
ization expression and an external energy that makes the movement of the zero level
set to preferred position. The distance regularization expression is identified by a
potential function such that the resultant level set progress has a single forward-and-
backward (FAB) distribution effect, which is capable to preserve a preferred form
of the level set function, mainly a signed distance profile close to the zero level set.
This produces a new kind of level set progress named distance regularized level set
evolution (DRLSE) [15]. This technique can be simply realized through simple finite
difference scheme and is computationally more competent than the usual level set
methods. The energy functional is:

1
E g (φ) = μ (|∇φ| − 1)2 d xd y + λ g δ(φ)|∇φ|d xd y + ν g H (−φ)d xd y
2

(1)
where denotes the image domain and φ is the level set function which is defined
as follows:
⎧
⎨ −ρ x ∈ 0 − ∂ 0
φ (x, t = 0) = 0 x ∈ ∂ 0 (2)
⎩
ρ x ∈ − 0
ρ = constant > 0
0 ∈ : The subset in the image domain.
∂ 0 : The boundary of 0 .
2 (|∇φ|
1
− 1)2 d xd y: As a metric to characterize how close a function φ is to a

signed distance function in ⊂ 2 . This metric will play a key role in our variational
level set formulation.
μ: is a parameter controlling the effect of penalizing the deviation of φ from a
signed
distance function,
λ g δ(φ)|∇φ|d xd y + ν g H (−φ)d xd y: is a certain energy that would drive

the motion of the zero level curve of φ.
λ, ν Are positives constants;
H (.) : Is a Heaviside function and its derivative the Dirac function δ

g: Is the edge indicator function defined by :
1
g= (3)
1 + |∇G σ ∗ I |2
where G σ is the Gaussian kernel with standard deviation σ. The convolution in (3)
is applied to smooth the image to decrease the noise. I is the image to be segmented.
This technique has two weaknesses: (1) It is responsive to the first location of the
curve; (2) This method is not robust against noise since the border indicator function g
is based on the image gradient. So, a step of smoothing process should be introduced
before segmentation on the image to be segmented. However, this step deteriorates
the weak boundaries.
A. Chan and Vese Method
Chan and Vese [10] proposed a widely used region-based model, namely Chan-Vese
model (CV model). The CV model is based on the Mumford–Shah (MS) segmenta-
tion technique [9]. It has been successfully implemented in binary phase segmenta-
tion. The CV model employs the statistical information within and outer the contour
with the purpose of controlling the progression. The CV model is formulated by:
The evolution equation is as follows [10]:
∂φ
(x) = δ(φ(x))((I (x) − v)2 − (I (x) − u))2 + λ δ(φ(x))k (4)
dt
The Energy criterion:

E(φ) = F(I (x), φ(x))d x + λ δ (φ(x))∇φ(x)d x (5)

F(I (x), φ(x)) = H (φ(x))(I (x) − v)2

+ (1 − H (φ(x)))(I (x) − u)2 (6)
where:
δ the Dirac function.

H the Heaviside function

(1 − H (φ(x))).I (x)d x

u= (7)
1 − H (φ(x))d x


H (φ(x)).I (x)d x

v= (8)
H (φ(x))d x

u and v are two parameters updated at each iteration.

The C-V model has good performance when segmenting images with weak edges
and noise. Thus, the use of global information of the image leads to inaccurate
segmentation results for images with intensity Inhomogeneity.
B. Geodesic Active Contours(GAC)
Geodesic Active Contours (GAC here) is a model that theoretically combines explicit
Active Contour Models (Snakes) with the Implicit Active Contour Models (Level
Sets). It was simultaneously introduced by Kichenassamy et al. [16] and Caselles
and Kimmel [6] . The model starts from the Snakes formulation and shows that it is
equivalent to finding a geodesic curve in a Riemannian space with a metric based on
the image content. Level set formulation for GAC model is given by:

∂φ ∇φ
= g(|∇φ|) div + α + ∇g.∇φ (9)
∂t |∇φ|
where φ is the level set function, ∇ is the gradient operator, α is real constant called
the balloon force which controls the curve evolution and g is the edge based function
and defined as in Eq. (4)
I
g(|∇ I |) = (10)
I + |∇G σ ∗ I |2
where Gσ is a Gaussian Kernel with standard deviation σ.

Because GAC model is based on gradient information so it is not suitable for
images with weak edges. In some cases balloon force, which is very difficult to
design has been used. The other weak balloon force doesn’t allow the contour to
pass through narrow part of the object and in case of the large balloon force, contour
will pass through the weak edges. When contour is far from the object boundary then
also it is difficult to find interior or exterior boundaries of the object.
3 The Proposed Method
This method utilizes the advantages of both GAC and C-V model. In the substitution
of Edge Stooping Function (ESF), a region based Singed Pressure Force function
has been (SPFn) developed. This SPFn function controls the direction of evolution.
Opposite signs (range of SPF function is [−1 1]) around the boundaries of the object
in this function make the contour to able to expand when it is inside the boundary
and to shrinks when it is outside the boundary.
SPFn function proposed in this model is as in Eq. (11):

(u + v) ∗ I (x, y) − u×v
sp f n (I (x, y)) = 2
, x, y ∈ (11)
max (u + v) ∗ I (x, y) − u×v
2

where:
u and v are two parameters updated at each iteration, in Eqs. (7) and (8)
respectively.
On the substitution of SPF function in Eq. (9), the level set formulation takes the
form as in Eq. (12)

∂φ ∇φ
= sp f (I (x, y)). div + α |∇φ| + ∇sp f (I (x, y)).∇φ, x, y ∈
∂t |∇φ|
(12)

∇φ
The regular terms div |∇φ| |φ| and ∇sp f.∇φ can be removed because the method
utilizes the statistical in formation of the regions. Thus the final level set model is
given as in Eq. (12)
∂φ
= sp f (I (x, y)).α|∇φ|, x, y ∈ (13)
∂t
The main steps of our Proposed Method are as well shown
Input: Image of size n l × n c .
Image I : →
Where:
∈ R n is the domain of the image.
Step 1 Initialize the level set function φ as:
⎧
⎨ −ρ x ∈ 0 − ∂ 0
φ(x, t = 0) = 0 x ∈ ∂ 0 (13)
⎩
ρ x ∈ − 0
ρ = constant > 0
0 ∈ : subset in the image domain .
∂ 0 : the boundary of 0 .
Step 2 Compute the coefficients u and v using eq. (7) and
(8) respectively.
Step 3 Evolve the level set function according to Eq. (13)
Step 4 Let φ = 1 if φ > 1, otherwise φ = −1
Step 5 Regularize the level set function with a Gaussian
filter, i.e. φ = φ ∗ G σ
Step 6 Check whether the evolution of the level set function
has converged. If not, return to step 2.
(end of procedure).
In this part, we discuss the performances of our proposed method. A quantitative

comparative study is also established to highlight the superiority of our method.
These methods have been applied to medical images with different modalities. For
the sake of comparison, the initial contour used is the same for all the methods.
The comparison is carried out in term of several criteria, namely: the Dice criterion,
MSSD, Hausdorff Distance (HD), and the Time. The considered criteria Dice coeffi-
cient, MSSD and HD measure the closeness between the reference segmentation and
the segmentation provided by each selected algorithm. The smallest HD,MSSD and
time, the more efficient our LSM method. Dice coefficient should be between 0 and
1. When Dice coefficient reaches 1, the segmentation is perfect.All the numerical
experiments were run on a PC with 3.40 GHz Core i5 CPU, 8 GB RAM and Matlab
(R2016b). For each measure and each image, all methods were conducted two times
by utilizing these two different initial positions of the initial curve: position1 (P1)
and position 2 (P2) are computed and appeared in Tables 1, 2, 3 and 4. We use four
examples of medical images, namely the Brain image, the chest image, Arms_XRay
and Vessel_CTA1 respectively.
Example 1 (Fig. 1).
Example 2 (Fig. 2).
Example 3 (Fig. 3).
Example 4 (Fig. 4).
Four experiments are conducted to validate the adequacy of the Proposed method
. Four representatives are taken as baselines on medical images. Chan and Vese, the
GAC model and Proposed method (This method utilizes the advantages of both GAC
and C-V model). For the sake of comparison, the initial contour used is the same for
all the methods.
Experiment1 applied the three methods to medical image with Brain image, as
shown in Fig. 1. The left column in the first row of Fig. 1 shows the initial contour,
which is inside the Brain, the others column of the first row and second row show
the corresponding segmentation results of Proposed method, GAC method and Chan
and Vese algorithm, respectively. We have found that, the evolution of the Proposed
method and GAC model converge at 120 iterations, while for the Chan & Vese
algorithm, the evolution converge at 132 iterations. Proposed method and Chan &
Vese algorithm detect the exterior whereas the snakes model fails to detect the brain.
The left column in the third row shows the initial contour, which is around the
brain image, the others columns of third and fourth row show the corresponding
Fig. 1 Segmentation of Initial contour

Brain image
20
40
60
80
100
120
20 40 60 80 100 120
Proposed method
(120 iterations)
GAC model Chan&Vese

(120 iterations) (132 iterations)
Initial contour 200iterations
20
40
60
80
100
120
20 40 60 80 100 120
Proposed method
(90 iterations)
GAC model Chan & Vese


chest image 20
40
60
80
100
120
140
160
180
200
20 40 60 80 100 120 140 160 180 200 220
Proposed method
(100 iterations)
GAC model Chan&Vese

Initial contour
200iterations
20
40
60
80
100
120
140
160
180
200
20 40 60 80 100 120 140 160 180 200 220
Proposed method
(95 iterations)
GAC model Chan&Vese


Arms_XRay image
Proposed method
(90 iterations)
GAC model Chan&Vese

Initial contour
Proposed method
(100 iterations)


Vessel_CTA1 image
Proposed method
(100 iterations)
GAC model Chan&Vese

Initial contour
Proposed method
(105 iterations)

Table 1 The coefficients: times, Dice, Hausdorff and MSSD of the three method s obtained for the
segmentation of Brain image
Methods Time (S) Dice coefficient Hausdorff coefficient MSSD coefficient
GAC 11.57 0.32 51.44 436.17
Chan and Vese 3.18 0.78 25.62 42.68
Proposed method 1.10 0.87 12.78 24.17
segmentation results. We see that the contour of the brain is accurately detected by
Proposed method and Chan and Vese algorithm but the GAC method fails to detect
the contour.
In Experiment 2, compares the Proposed method, the GAC method and Chan and
Vese algorithm by applying them to a chest image. The left column in the first row
shows the initial contour which is inside the chest, the initial contour of third row is
around the chest, the others columns of first and second row show the corresponding
segmentation results of the three models .The segmentation by Proposed method
and Chan &Vese algorithm are obviously more accurate than the corresponding
segmentation results of the GAC model.
Experiment 3 applied the three methods on a medical image containing
Arms_XRay image as shown in Fig. 3. The left column in the first row shows
the initial contour which is inside the Arms_XRay, the other columns show the
corresponding segmentation results of the three models. The results show that GAC
method missed the boundaries because its image gradient-based stopping function
cannot well deal with the weak boundaries; Chan and Vese also failed to detect the
boundaries, Proposed method almost correctly outlined the object with a smooth
contour .
In Experiment 4, these three methods were applied to an Vessel_CTA1 image,
as shown in Fig. 4. The left column in the first row shows the initial contour which
is inside the Vessel_CTA1 (of the left Ort). The initial contour of the third row is a
middle of the Vessel_CTA1, the other columns show the corresponding segmentation
results of the three models. The obtained results illustrate that Proposed method (for
Initial contour 1) detected the part of boundaries around the mass, but the result is not
complete; GAC method failed to detect the complete region; Chan and Vese method
obtained the same curve outlining the mass to be captured but the Proposed method
gives the best results.
From Tables 1, 2, 3 and 4, it can be see that Proposed method succeed for
segmentation the image:
5 Conclusion
In this work, we have established a complete quantitative comparison of segmenta-

tion algorithms based on level set method for medical images. We have used four
segmentation of Chest image
GAC 13.88 0.32 55.10 142.11
Chan&Vese 5.15 0.80 8.00 22.30
Proposed method 1.20 0.95 11.24 13.23
segmentation of Arms_XRay image
GAC 12.78 0.51 56.10 152.00
Chan&Vese 5.16 0.60 26.00 12.30
Proposed method 2.71 0.92 24.10 11.22
segmentation of Vessel_CTA1 image
GAC 14.86 0.62 86.10 135.00
Chan and Vese 3.16 0.75 36.00 34.30
Proposed method 2.00 0.94 15.18 13.54
test images with two different modalities. The considered level set approaches are
compared in the sense of four metrics. From the experimental results, we easily
conclude that Proposed method give the best results than Chan and Vese model and
GAC method.; The Proposed method reduces the expensive re-initialization of the
traditional level set method to make it more efficient. This method combines the
merits of the traditional GAC and C–V models, which possesses the property of
local or global segmentation.
References
1. Qiucheng S, Yueqian H, Qingchang T, Chunjing L, Ming L (2014) A robust edge detection

method with sub-pixel accuracy. Optik 125:3449
2. Havaei M, Davy A, Warde-Farley D, Biard A, Courville A, Bengio Y, Pal C, Jodoin PM,
Larochelle H (2017) Brain tumor segmentation with deep neural networks. Med Image Anal
35:18–31
3. Wang Xueqin, Li Shurong, Li Jiaojiao, Wang Jiayan, and Zhang Minyi. A novel segmentation
model with dual level set function based on chan-vese and local binary fitting models. 2016
3rd International Conference on Systems and Informatics (ICSAI), pages 837–842, 2016
4. Wang Huafeng, Li Lihong C, Wei Xinzhou, Liu Wanquan, Wang Yuehai, Liang Zhengrong
(2017) A new framework for detection of initial flat polyp candidates based on a dual level set
competition model. ComputerAided Diagnosis, In Medical Imaging
5. Caselles V, Kimmel R, Sapiro G (1997) Geodesic active contours. Int. J. Comput. Vis. 22(1):61–
79
6. Kass M, Witkin A, Terzopoulos D (1988) Snakes: Active Contour Models. Int J Comput Vision
1(4):321–331
7. R.Buxton.: Introduction of Functional Magnetic Resonance Imaging, Principles and tech-
niques. Cambridge, UK: Cambridge University Press.2003
8. Xu C, Prince J (1998) Snakes, shapes, and gradient vector flow. IEEE TransI mage Process
7(3):359–369
9. Chan TF, Vese LA (2001) “Active contours without edges”,IEEE Trans. Image Process.
10(2):266–277
10. X. Chang, et al., “A Study of Image Segmentation Based on Level Set Method”, in International
Conference on Advanced Computer Science and Electronics Information, pp. 360–363, 2013
11. B. Wang., X. Gao, D. Tao., X. Li.: A unified tensor level set for image segmentation. IEEE
Trans. Syst., Man, Cybern. B, Cybern., vol. 40, no. 1, pp. 857–867, Feb. 2010
12. Vese LA, Chan TF (2002) A multiphase level set framework for image segmentation using the
Mumford and Shah model. Int. J. Comput. Vis. 53(3):271–293
13. M. Rousson., T. Brox., R. Deriche.: Active unsupervised texture segmentation on a diffusion
based feature space. In Proc. IEEE Conf.Comput. Vis. Pattern Recognit., 2003, pp. 699–704
14. Paragios N, Deriche R (2002) Geodesic active regions and level set methods for supervised
texture segmentation. Int J Comput Vis 46(3):223–247
15. C. Li, C. Xu, C. Gui, and M. D. Fox, “Level Set Evolution without Re-Initialization: a New
Variational Formulation,” Proc. IEEE Conf. Computer Vision and Pattern Recognition, vol. 1,
pp. 430–436, 2005
16. Y. Chen, D. Tseng.: Medical image segmentation based on Bayesian level set method. in Proc.
Med. Imag. Info., 2007, pp. 25–34
17. Zhang K, Zhang L, Song H, Zhou W (2010) Active contours with selective local or global
segmentation: A new formulation and level set method. Image Vis Comput 28:668–676
New Convergence Architecture Between
5G Mobile Telecommunication Networks
and Satellite Networks to Enhance Their
Capacities and Improve Their
Performance
Salheddine Sadouni, Malek Benslama, André-luc Beylot,

Mohammed Nassim Lacheheub, and Ouissal Sadouni
Abstract The subscribers’ growing demand in terms of mobile traffic requires the
development of new architectures for the purpose of reinforcing the capacities, the
performance and the geographic expansion of telecommunication mobile networks.
As a matter of fact, the so-called fifth generation networks (5G) seem to be the
best solution in order to respond to the increase of the traffic volume when satellite
telecommunication networks bring large geographic coverage areas. With this aim
in mind, in the current paper, we will focus on the establishment of new convergence
architecture between both telecommunication networks: 5G and Satellite; in order
to boost their capacity, their performance, as well as expanding their coverage areas.
The intersection of the latter gives birth to a specific area called Handover Zone,
where mobility management becomes totally different to ensure transparent current
communication. This new architecture is essentially comprising two parts: hardware
and software. Hardware sets up physical interconnection between both 5G mobile
and satellite networks; while software integrates a specific algorithm that takes into
charge the activation oftwo distinct mechanisms in propitious time to guarantee and
optimal quality of service (QoS) when passing from one network to the other.
Keywords Convergence architecture · 5G mobile network · Satellite

telecommunication network · Handover · Quality of service
S. Sadouni (B) · M. Benslama

Electronics Department, University of Constantine 1, Constantine, Algeria
A. Beylot
Institut de Recherche en Informatique de Toulouse, University of Toulouse, Toulouse, France
M. N. Lacheheub · O. Sadouni
Software Technologies and Information Systems Department, LIRE Laboratory, University of
Constantine 2, Constantine, Algeria

https://doi.org/10.1007/978-981-15-6403-1_88
1248 S. Sadouni et al.
1 Introduction
Mobile communication has been crucial to the development of our modern societies
and their current lifestyles. Since the emergence of these mobile communication
systems, the way of accessing and exchanging information has drastically changed.
However, mobile communication networks ought to constantly evolve in order to
handle the mounting requirements of their users. The latter are accentuated by the
proliferation of smart devices as well as flow-intensive applications. Therefore, new
generations of more efficient and effective networks must be developed. In this regard,
fifth generation mobile communication systems come into sight, with a view to
handling a wide range of current and future challenges [1].
A 5G network should be able to stand all of the components that seek to connect
with one another in order to access different types of services; whether they are mobile
devices subscribers, IoT devices (Internet of Things), or even smart gadgets within
a D2D (Device-to-Device) communication context [2]. The fifth generation network
must have massive connectivity that allows it to simultaneously take into charge all of
the former components. This same network must be able to provide great reliability
and higher availability by suggesting wider coverage areas. For this purpose, one
of the solutions consists in reusing existing telecommunication technologies which
cover wide areas of service such as satellite telecommunication.
In fact, satellite telecommunication is the only system which is capable of
providing instant coverage of important Earth services without having to look at
their geography by using large bandwidths. On the other hand, these satellite systems
possess an architecture which is very different from that of the fifth generation
mobile networks. The former facts lead us in this paper to suggest a new architecture
of convergence between fifth mobile networks and satellite telecommunication in
order to pool the resources of both networks and to increase their capacity and their
easement area.
To this end, this paper will be divided into several sections. In the first section,
we will focus on the 5G telecommunication networks and their main characteristics.
The second section will be devoted to the presentation of satellite telecommunication
networks and their attributes in a one-to-one relationship with our research task. The
third section of this paper will shed the light on the elaboration of a new architecture
of convergence of both 5G and satellite networks. Afterwards, we will incorporate the
simulation parameters and the obtained results in Sect. 4. Finally, we will terminate
our research task in the sixth section with a conclusion and suggestions for future
research activities.
2 5G Fifth Generation Telecommunication Networks
In recent years, 5G networks have been reported in several scientific investigations

which aimed at developing new applications based on these networks. The infatuation
New Convergence Architecture Between 5G Mobile … 1249
over 5G networks could be related to technical performances required and defined by

the European project METIS (Mobile and Wireless Communications Enablers for the
2020 Information Society) [2]. These requirements have to do with the optimization
of the following elements:
• The volume of mobile data should be a thousand-time higher per coverage zone.
• Increasing user data rates from 10 to 100 times.
• Increasing the number of simultaneously connected devices from 10 to 100 times.
As far as we are concerned, the most important requirement in this paper has
to do with increasing the capacity of the 5G network and its coverage area. This
requirement could be achieved by the contribution of satellite networks that have
been already deployed through a new architecture of convergence [3].
A. 5G Network Architecture
The 5G network comprises a radio access (NG-RAN: Next Generation RAN) and
a core network (5G Core).
(1) 5G Radio Access: 5G radio access consists of new generation base stations
which make up the mobile connection node together with the 5G core network
(5GC). UE mobiles (user equipment) communicate with the base stations via
a 5G radio link through base stations of the type gNB (next Generation Node
Base Station).
(2) 5G Core Network: The 5G core mainly relies on distributing tasks between
the user plan and the control plan. As far as we are concerned, we will focus
particularly in this article on the user plan. The latter will allow us to establish
our architecture of convergence. The user plan is made up of two entities:
• AMF entity (Access and Mobility Management Function). It deals primarily with
the localization of UE as well as the establishment of connection.
• SMF entity (Session Management Function). It is responsible for the management
of PDN sessions based on control plan gateways SGW-C and PGW-C.
Regarding the control plan, it consists of several gateways to manage the connec-
tions with the data networks (PDN). The latter are controlled by SGW-U and
PGW-U gateways of the user plan. There are evidently multiple entities which split
into the 5GC control plan such as:UPF (User Plan Function), UDM (Unified Data
Management), and NSSF (Network Slice Selection Function). These entities realize
diverse functions, but they do not constitute a stakeholder in the development of our
convergence architecture [3].
3 Satellite Telecommunication Networks
Satellite telecommunication networks were originally designed to play the role of a

long-distance communication relay between an emitter and a receiver on Earth. In
fact, a geostationary satellite has a large coverage area which could attain a third
of the Earth’s surface. From a technical and an economic point of view, satellite
development has over time contributed to the emergence of new varied satellite
applications and to increasing the number of their users. These applications have
also improved the implementation of several satellite broadcast standards DVB-S
(Digital Video Broadcast-Satellite), and its second generation DVB-S 2 for television
services [4].
(1) DVB-RCS Norm: As far as the interactive satellite services are concerned, the
DVB-RCS norm (Digital Video Broadcast Return Channel via Satellite) estab-
lishes a return channel from the users to the satellite. Hence, The DVB-RCS
norm is considered as a DVB-S complement. In fact, the DVB-S streams possess
different information to manage the global system in terms of access, signaling
and terminals synchronization. Regarding the DVB-RCS, it defines à system
of satellite two-way large band for the VSAT terminals (Very Small Aperture
Terminals). In such a system, the up-link uses the DVB-S (or DVB-S2) and the
down-link uses the DVB-RCS norm. The latter reduces the cost of the connec-
tion suggesting a satellite return link to the users without monopolizing the
resources in vain.
(2) The Architecture of Satellite Telecommunication Systemm Via the DVB-RCS
Norm: Every subscriber’s communication transits from the users’ terminals
(RCST: Return Channel Satellite Terminal) in the direction of the sattellite
via satellite gateways and the NCC (Network Control Centre). Therefore, the
topology of the used satellite network is considered to be a star network. As
a matter of fact, in such a configuration, an RCST can send and receive data
while the Earth’s network which is managed by satellite gateways is playing the
role of an interface between the subscriber, the satellite, and the other network
like internet. Thus, the satellite center of control and management sets a policy
of access to the satellite network resources,and takes into charge the signalling
and different other services [5].
4 A New Architecture of Convergence for Both 5G

and Satellite Networks
Nowadays, the integration of a 5G mobile network into a satellite communication

system appears to be an essential solution to solve various connectivity problems.
Among these problems is the expansion of the 5G coverage area, using the satellite’s
one. To serve subscribers in remote areas such as desert regions or in ships in the
middle of the oceans. In addition, we find in the literature various studies that focus on
the issue of integration of 5G mobile networks on satellite communication systems.
There are also new solutions in this field that can meet the requirements of 5G; mega
satellite constellations and satellite high-throughput (HTS) offer significant data rates
compared to conventional satellites.
Among the most promising initiatives for the integration of a 5G mobile network
into a satellite telecommunications system are the following three:
• the 5GPPP project and its SAT5G program, which emerged from a partnership
between public and private providers, is funded by the European Union EU [6];
• the initiative of the European Space Agency (ESA), which has launched the
ARTES programme for the implementation of a series of tests, as part of the
process of developing a 5G satellite interconnection [6];
• the specifications of the Third Generation Partnership Project (3GPP), which sets
out the minimum requirements for the provision of 5G network services in satellite
environments [7].
A. The Architecture of Convergence Advantages
A network architecture of convergence primarily contributes to the commercial and

the technical efficiency of the 5G and satellite networks. Since it sets up a pre-
integration policy of the components of both networks in order to pool their resources.
This same architecture must take in charge an element which is very important in
automating the process of switching from a process to another without data loss or
connection failures [8]. This could be achievable through the success of the mech-
anism of 5G-satellite inter vertical handover. Such an architecture of convergence
could have two main advantages:
• increasing the capacity of both 5G and satellite networks via the policy of pre-
integration of the components;
• expanding the coverage areas of 5G networks by using the coverage areas of
satellite networks;
• responding to emergencies (natural disasters, military missions, etc.) [9].
B. Establishing the 5G Satellite Architecture of Convergence
The implementation of our architecture of convergence in networks and terminals

could be based upon two parts: hardware and software [10].
As far as hardware is concerned, it is essentially made up of a gateway called
S5GW (Satellite 5 generation Gateway) which is responsible for the interconnection
of both 5G and satellite networks, as well as the mobility of the UE from one network
to another. S5GW has a direct link with the 5G mobile network throughout its core
5GC and across the PDN which is controlled by the SMF. The latter grants to the
users a set of sessions that have the AMF’s approval to access the 5G network
RAN. As a result, the S5GW routes the users’ data streams that are coming from
the satellite network to a UPF (User Plan Function) of the 5GC which sends them
to the appropriate ngNB. On the other hand, the S5GW gateway is connected to the
satellite network via the NCC which guarantees the resources management and the
access of the users coming from the 5G network (Check the Fig. 1).
As far as software is concerned, it guarantees a transparent passage without
connection failure for the user who moves from one network to another. As a matter
Fig. 1 Hardware
Convergence Architecture of
the 5G and Satellite
Networks
of fact, the significant architectural differences existing between the 5G mobile

networks and satellite networks involve the determination of a transition policy. The
latter uses two mechanisms: The Inter-network Vertical Handover Mechanism: This
concept comes into effect within an intersection area between both 5G and satellite
networks (zone HO) which consists of implementing two UE radio links with both
networks at the same time [11]. The UE resumes its normal state (it communicates
with only one network through only one radio link) on the edge of the HO area or
when leaving it. The second mechanism called HARQ (Hybrid Automatic Repeat
reQuest), comes also into effect within the HO area but only for the satellite network
HARQ combines the data retransmission with the error-correcting codes with the aim
of ensuring mobile transmission reliability through satellites since satellite telecom-
munication suffers from a propagation delay and very high error rates [12]. We
combine hardware functioning in the following algorithm.
Algorithm: HO-Area
Input: terminal settings
Output: terminal settings-HO
Link-list: list of UE links
Rx-min5G, Rx-minSat, Rx-Sat, Rx5G: signal power
Beginning
while (Rx5G <= Rx-min5G) do

Link-list = Sat;
Link-list = 5G;
If (Rx5G < Rx-min) and (Rx-Sat > Rx-minSat) then

Activate HARQ mechanism
Else
Deactivate HARQ mechanism
End if
Sleep (03 s)
End while
End
In this paper, we have focused on the development of a convergence architec-

ture rather than an integration architecture. There are several research studies in
the literature that focus on the embedding of a 5G mobile network into a satellite
communications network, however, all these studies are still at the experimental stage.
Because, the realization of a complete integration between the two above-mentioned
networks, which have significant architectural distinctions, requires several elements
to be taken into account at the same time. In order to guarantee proper behavior when
a subscriber moves from one network to another without disconnection or loss of
data. The first element to be taken into account is the management of the hori-
zontal handover; this task is not as easy as we think because a 5G mobile network
implies a very short delay whereas the one provided by a satellite network is very
high. The second element concerns the management of subscriber mobility, which is
becoming more complicated in handover areas, to ensure the continuity of commu-
nications, data transfer and resource allocation, etc. Another very important aspect is
the security management and authentication of subscribers as well as the allocation
of services according to subscribers’ profiles. All these and many other elements
must be considered in order to ensure the full integration of a 5G mobile network
into a satellite communications network.
5 Simulation Parameters and Results
A. Simulation
According to our study, we recommend to bring to light the advantages of this new
convergence architecture in terms of 5G mobile and satellite networks capacity. With
this aim in mind, we will do all in our power to determine the effects of subscribers’
mobility in a coverage area from one network to another and vice versa on the QoS
[10]. In effect, when a UE finds itself in an HO area, our new convergence architecture
must take the relays to ensure a transparent passage i.e. without connection failure
or data loss.
This leads us to dissect the functioning of this new convergence architecture
which is chiefly based upon two mechanisms: Handover and HARQ. In addition,
our convergence architecture must consider the subscribers’ speed of displacement,
since the latter heavily affects the process of mobility management and the quality
of service QoS.
To that end, we simulate a UE which moves between two coverage areas of the 5G
and the satellite networks. Once arrived at the HO area, it is the S5GW gateway that
takes over activating the HO-Area algorithm. Therefore, we set up two cases: The
UE moves with a speed of 50 km/h, or it moves with a speed of 100 km/h. The UE
moves from the 5G network to the HO area in the direction of the satellite coverage
area. Afterwards, starting from the latter, the UE goes back to the 5G easement area
passing by the HO area. We have put for this simulation three different scenarios.
• The First Scenario: in the HO area, S5GW only activates the HO mechanism.
• The Second Scenario: in the HO area, S5GW only activates the HARQ
mechanism.
• The Third Scenario: in the HO area, S5GW activates both mechanisms meanwhile.
In order to determine the performances brought by our new convergence archi-
tecture, we use the average throughput that is received by the UE as performance
criteria [13]. Check the Simulation parameters in the Table 1.
B. Results
C. Results Discussion
In the Fig. 2, we note that the throughput received by the UE is more important
when it is in the 5G mobile network coverage area. It is also relatively weak in the
satellite coverage area. Passing by the handover area, we observe a freefall of the
throughput which attains a zero value in the second scenario, and an average value
in the first scenario. While in the third scenario, it keeps the throughput steady and
close from the one received by the UE in the satellite coverage area.
In the Fig. 3, we observe some developments that are almost identical for the three
curves recorded in the 2nd figure. Except for the throughput which is relatively low
in the case when the speed is of 100 km/h compared when it is of 50 km/h.
Figure 4 perfectly illustrates the effects of the displacement speed of the UE on the
QoS, because with a speed of 50 km/h, we note an important throughput especially
in the HO area and in the area covered by the satellite.
Table 1 Simulation
Parameters of the satellite networks
parameters
Orbit Low-earth
altitude (km) 1500
Cycle time 60
Jammer band base Frequency 30
Jammer bandwidth 10
Number of Frequencies 65
Parameters of the UE
Wireless lan MPE2-TS data rate (Mbps) 100
Transmit Power (w) 0.005
Buffer size (bits) 256000
Long retry limit 7
Short retry limit 4
Parameters of the 5G mobile network
Bandwidth (MHz) 10
Central frequency (MHz) 2110
Maximum latency allowed (s) 3
K(rx, 2) W/Mbps 2.86 × 10−3
Overhead 15%
Fig. 2 The throughput received by the UE with a speed of 50 km/h

Fig. 3 The throughput received by the UE with a speed of 100 km/h
Fig. 4 Comparison of the throughput received by the UE with the speeds 50 km/h and 100 km/h
6 Conclusion
Because of the architectural and functional differences between both networks: 5G

mobile and satellite, the establishment of a new convergence architecture turns out
to be very helpful and complexe at the same time. En effect, this convergence archi-
tecture expands the coverage areas of the 5 g and satellite networks by improving
the passage of the subscribers from one network to another. However, it must take in
charge the management of the mobility in the handover area to ensure a transparent
passage e.i. without connection failure or data loss.
Managing the mobility within the handover area well implies the activation of two
mechanisms in propitious time. The first is the handover mechanism which avoids
current connection clean cut. The second mechanism is the Hybrid Automatic Repeat
reQuest which greatly increases the QoS of the satellite mobile communications.
Finally, based on this study, we recommend certain suggestions for future research
tasks:To focus on the concept of IoT (Internet of Things) and the concept of Big Data
in a 5G mobile heterogeneous environment. To take into consideration the aspect of
security of communications within this same type of environment to the advantage
of the subscribers.
Acknowledgments Special acknowledgement to SADOUNI Nesrine for her great contribution to

the writing of this scientific paper.
References
1. Berger, and L. Dittmann, “Cloud RAN for Mobile Networks – ATechnology Overview,” IEEE
Communications Surveys Tutorials, vol. 17,no. 1, pp. 405–426, Firstquarter 2015
2. Oliva DL, Perez XC, Azcorra A, Giglio AD, Cavaliere F, Tiegelbekkers D, Lessmann J,
Haustein T, Mourad A, Iovanna P (2015) Xhaul: Toward an Integrated Fronthaul/Backhaul
Architecture in 5G Networks. IEEE Wirel Commun 22(5):32–40
3. Oliva A, Hernandez JA, Larrabeiti D, Azcorra A (2016) AnOverview of the CPRI Specification
and its Application to C-RAN-basedLTE Scenarios. IEEE Commun Mag 54(2):152–159
4. Aris S, Sadouni S, Benslama M (2018) A new concept of Satellite Telecommunication Mobility
scenarios with the context of intra-handover control management. International journal of
Communications 3:P6–P10
5. R.Ravishankar et al., “High Data Rate and Bandwidth EfficientDesigns for Satellite Commu-
nication Systems,” Proc.AIAA Int. Commun. Satell. Systems Conf., 2017, Oct. 2017,Trieste,
Italy
6. G-PPP Project; http://www.5G-PPP.eu
7. GPP, “Technical Specification Group Services and System Aspects; Service Requirements for
the 5G System; Stage 1,” Release 15, TS 22.261 V15.0.0, Mar. 2017
8. S. Sadouni, M. Benslama, A. L. Beylot, “New SCTP protocol extension for more robust Ad
Hoc mobile networks”. 4th International Conference on Control Engineering & Information
Technology (CEIT). pp, 1–6, 2016
9. Vasavada Y et al (2016) Architectures for Next Generation HighThroughput Satellite Systems.
Int. J. Satell. Commun. Networking 34:523–546
10. R. Gopal, “Convergent Architectures for Multi-Orbit SatelliteCommunications,” Proc. AIAA
Int. Commun. Satell. SystemsConf. 2017, Oct. 2017, Trieste, Italy
11. S. Sadouni, M. Benslama, A. L. Beylot, “ Increasing the performances of UMTS and
WIMAX heterogeneous mobile networks by the integration of Multihoming”. 4th International
Conference on Control Engineering & Information Technology (CEIT), p. 1 – 6, 2016
12. Szczecinski L, Khosravirad SR, Duhamel P, Rahman M (2013) Rate allocation and adaptation
for incremental redundancy truncated HARQ. IEEE Trans Commun 61(6):2580–2590
13. S. Sadouni, M. Benslama, A. Mebarki, A. Beylot, “ SCTP-WSN new extension for more
reliable sparse wireless sensor networks”. 13th International Wireless Communications and
Mobile Computing Conference (IWCMC). pp. 1–6, 2109-2114, 2017
Comparative Study of Noise Robustness
in Visual Attention Models
Sarra Babahenini, Foudil Cherif, and Fella Charif
Abstract Humans are experts at quickly and accurately Identifying the most visually
noticeable foreground object in the scene,. Several models of saliency detection are
now available that have shown successful applications in various fields. Using visual
attention techniques, we examine whether the effects of noise have an influence on
their recognition rate. To answer this question, we select three recent computational
techniques of visual saliency: Saliency Detection Using human saliency map tech-
nique, Saliency Detection Using Contourlet Transform technique (CT) and Saliency
Detection Using low-rank and Structured sparse Matrix Decomposition (LSMD)
technique, and we have evaluated the performance techniques based on the MAE,
WF, AUC, PR curves and F-measure c. While the experimental results on the testing
set show that LSMD technique has a better performance compared to the other
techniques.
Keywords Visual · Saliency · Feature maps · Saliency map · Contourlet

transform · LSMD · Noise robustness
1 Introduction
The quantity and variety of visual information and the volume of data that reaches our
eyes at every moment are gigantic. To cope with this huge influx of incoming data,
the human visual system (SVH) has the ability to extract the relevant information that
is in the visual field, among all the other data existing in the scene. This is referred
S. Babahenini (B) · F. Cherif

Computer Science Department, LESIA Laboratory, Biskra University, Biskra, Algeria
F. Cherif
F. Charif
Department. of Electricity & Telecommunications, Ouargla University, Ouargla, Algeria

https://doi.org/10.1007/978-981-15-6403-1_89
1260 S. Babahenini et al.
to as visual attention; visual saliency is a process that identifies important places in

a visual field.
Determining objects of interest in unfamiliar environments is an important task for
humans and machines. Visual saliency methods are computer processes that guide
an agent’s attention to places that may be relevant in an image or scene. We distin-
guish between two concepts of visual saliency: the bottom-up mechanism guides the
viewer’s attention to stimulus based salient regions while the top-down mechanism
involves semantic context for the scene, without any task-based knowledge for the
scene.
Salient region detection is one of a pre-processing method, Its goal is to detect
salient regions in an image in terms of a saliency map, Many previous studies have
shown that salient region detection is useful, and it has been applied to many appli-
cations including segmentation, object recognition, image retargeting, photo rear-
rangement, image quality assessment, image thumbnailing, and video compression
[10].
In this paper, we have evaluated the effects of noise and their influence in recog-
nition rate by selecting three techniques of visual attention. Our main contribution
is to corrupt an image by noise with different variance and study its effects with five
metrics.
The rest of the paper is organized as follows. Section 2 presents the three tech-
niques of detection of saliency with which we made the comparative study. In Sect. 3,
we conduct the experiment study and present the results. Finally Sect. 4 discusses
and concludes this paper.
2 Saliency Detection Techniques
A. Saliency Detection Using Human Saliency Map
Most models of saliency are biologically inspired and based on a bottom-up computa-
tional model. However, this saliency detection approach has been shown to correlate
with visual attention. This technique uses the local energy of the steerable pyramid
filters [1], include features in a simple saliency model described by Torralba [2] and
Rosenholtz [3] based on subband pyramids also takes intensity, orientation and color
contrast as important features for bottom-up saliency and include the three chan-
nels corresponding to these image features as calculated by Itti and Koch’s saliency
method [4]. Finally include the values of the red, green and blue channels, as well
as the probabilities of each of these channels as features and the probability of each
color as computed from 3D color histograms of the image filtered with a median
filter at six different scales [5]
B. Saliency Detection Using Contourlet Transform (CT)
Comparative Study of Noise Robustness in Visual Attention Models 1261
Ilyass et al. in [6] try to obtain the saliency map by using the Contourlet Transform.
They use the color image as input and generate two saliency map (local and global)
in order to combine them into a final saliency map, so the feature maps is generated
by the Contourlet Transform. Local saliency map is the result of the combination of
the feature maps on each level. Feature maps fcs (x, y) are used to calculate the local
saliency map:

SL (x, y) = arg max( f SL (x, y), f Sa (x, y), f Sb (x, y) ∗ I K x K (1)
Global saliency map calculate the global distribution of local features to obtain
the global saliency map. From fcs (x, y), a location (x, y) can be represented as a
feature vector with a size of (3 × N) from all feature maps (24 features for each
location). The likelihood of the features at a given location can be defined by the
probability density function (PDF) with a normal distribution. Using the PDF, the
global saliency map can be computed as:
SG (x, y) = (log( p( f (x, y))−1 )) 2 ∗ I K x K

1
(2)
The final saliency map is the result of the combination of the two maps SG (x,y)
and SL (x, y). The combination is defined as:
S(x, y) = M(SL (x, y) × e(SG (x,y)) ) ∗ I K x K (3)
C. Saliency Detection Using low-rank and Structured sparse Matrix Decomposition

(LSMD) model
Another saliency detection approach is introduced by Houwen [7], they proposed

Low-rank and Structured sparse Matrix Decomposition (LSMD) model for salient
object detection. In this technique, a tree-structured sparsity-inducing norm regu-
larization is firstly introduced to provide a hierarchical description of the image
structure to ensure the completeness of the extracted salient object. The similarity of
saliency values within the salient object is then guaranteed by the ‘∞-norm. Finally,
high-level priors are integrated to guide the matrix decomposition and enhance the
saliency detection.
A meaningful index tree that store underlying structure information of a natural
image, and impose it on S as a structured constraint. Thus, a general tree structured
sparsity regularization can be written as:
minL∗ + λS1 s.t.F = L + S (4)

L ,S
Where F is the feature matrix of an input image, that can be decomposed as a low-
rank matrix L plus a sparse matrix S corresponding to the non-salient background
and the salient object, respectively [7].
3 Comparative and Experimental Results
This section is dedicated to a comparative study of the three models that we

have detailed previously, Saliency Detection Using human saliency map tech-
nique, Saliency Detection Using Contourlet Transform technique (CT) and Saliency
Detection Using low-rank and Structured sparse Matrix Decomposition (LSMD)
technique.
A. Data set
In this comparative study, we have used the DUT-OMRON dataset [9]. This dataset
is introduced to evaluate salient object detection algorithms on images with more
than a single salient object and relatively complex background. It contains 5,168
high quality natural images, where each image is resized to have a maximum side
length of 400 pixels. Annotations are available in forms of both bounding boxes
and pixel-wise binary object masks. Furthermore, eye fixation annotations are also
provided making this dataset suitable for simultaneously evaluating salient object
localization and detection models as well as fixation prediction models.
B. Evaluation Metrics
We evaluate the performance using the five measures used in [8] based on the over-
lapping area between ground truth annotation and saliency prediction, including the
PR (precision-recall) curve, the ROC (receiver operating characteristic) curve and
the AUC (Area Under ROC Curve) score. They are defined as follows:
FP
FPR = (5)
FP + T N
TP
FPR = (6)
T P + FN
1 (ζ 2 + 1) × pr ecision × r ecall
N
Fζ = (7)
N i=1 ζ 2 × pr ecision + r ecall

H
W
|S(i, j) − G(i, j)|
i=1 j=1
M AE = (8)
H ×W
Precision corresponds to the percentage of salient pixels correctly assigned, and

recall is the fraction of detected salient pixels belonging to the salient object in the
ground truth. For a grayscale saliency map, whose pixel values are in the range
[0, 255], we vary the threshold from 0 to 255 to obtain a series of salient object
segmentations. The PR curve is created by computing the precision and recall value
sat each threshold. The ROC curve can also be generated based on true positive rates
and false positive rates obtained during the calculation of the PR curve.
Fig. 1 Qualitative Comparison: Column 1: RGB images, column 2: ground truth, column 3:
contourlet map, column 4: Human saliency map, column 5: LSMD map
C. Comparison with ground truth

(1) Qualitative Comparison: The qualitative comparison results by the ground
truth and the three methods are shown in Fig. 1.
It’s shown that most saliency detection methods can handle well simple images
with relatively homogenous background.
They can generate high-quality saliency map but the best one as we see in Fig. 1
and almost similar to ground truth is the method of “LSMD”.
(2) Quantitative Comparison: Fig. 2 shows the quantitative results of the three
methods. It is known that the method “LSMD” it perform competitively and
it is better than the other methods in terms of the P-R curve, ROC curve and
F-Measure curve. Especially, the precision can remain above 80% within a large
threshold range.
D. Analysis of saliency map to Noise
The images under study were corrupted by Gaussian noise with different variances,
including 35/255, 25/255, 15/255 and 5/255. According to Table 1, Figs. 3,4 and 5
Fig. 2 Quantitative
comparisons of saliency
maps produced by different
approaches with data set
“DUT-OMRON” in terms of
ROC, PR curves and
F-measure c
(a) ROC Curve
(b) PR Curve
(c) F-mesure Curve

Table 1 Experimental results with different noise

Variance INDEX Contourlet Human saliency map LSMD
NO noise MAE↓ 0,26712 0,36353 0,16115
AUC↑ 0,71403 0,81245 0,80639
WF↑ 0,17529 0,2091 0,4403
OR↑ 0,17742 0,15898 0,47398
5/255 MAE↓ 0,28596 0,36948 0,18796
AUC↑ 0,68093 0,80893 0,72159
WF↑ 0,15485 0,20696 0,33167
OR↑ 0,12371 0,1402 0,33594
15/255 MAE↓ 0,30148 0,37608 0,20632
AUC↑ 0,66339 0,80692 0,69476
WF↑ 0,15217 0,20511 0,28371
OR↑ 0,099769 0,12515 0,26781
25/255 MAE↓ 0,30559 0,38041 0,20632
AUC↑ 0,64823 0,80447 0,69476
WF↑ 0,15078 0,20403 0,28371
OR↑ 0,090345 0,11631 0,26781
35/255 MAE↓ 0,3104 0,38605 0,2103
AUC↑ 0,6478 0,80412 0,64204
WF↑ 0,15017 0,203 0,23827
OR↑ 0,079005 0,10195 0,21051
(the five quantitative evaluation metrics), we can see that when variance increases
gradually, the model of human saliency map can resist to noise however LSMD
model is affected by the noise, It’s shown that Human saliency map model is robust
to noise
4 Conclusion and Future Work
In order to improve the results obtained from several saliency detection models, we
have implemented two computational models of visual saliency. The first is saliency
detection using Contourlet transform, this model combined two saliency maps: local
and global, these are combined to produce a single saliency map. The second that
we called Low-rank and Structured sparse Matrix Decomposition (LSMD) model
which has an excellent effectiveness and achieves the superior performance on the
public benchmark data set and finally Classification model of visual saliency « human
saliency map » based on bottom-up computation and top-down image semantics to
match actual eye movements.
Fig. 3 Quantitative
comparisons of saliency
maps with noise produced by
different approaches with
data set “DUT-OMRON” in
terms of ROC, PR curves
and F-measure c
(a) ROC Curve
(b) PR Curve
(c) F-mesure c Curve

Fig. 4 Column 1: RGB + noise images, column 2: ground truth, column 3: contourlet map, column
4: Human saliency map, column 5: LSMD map
We evaluated the performance models based on the ROC, PR curves and F-

measure c, the overall performance of the model LSMD is good and gives better
results compared to other models, however when the image is polluted by noise we
see that the model of LSMD is very sensitive compared with the rest of models. Based
on the preceding discussion and comparative study of the techniques presented, the
work in this paper needs more in-depth study in the future, and can be extended in
the following directions:
• make investigation on new techniques based on machine learning algorithms,
especially on CNN architectures.
• add more datasets in testing.
• propose an approach combining two techniques LSMD and CT then try to improve
the performance, such as taking advantage of the merits.
Fig. 5 Feature importance

across three different models
References
1. T. Judd, K. Ehinger, and A. Torralba, “Learning to Predict Where Humans Look Fr, 2010
2. A.OlivaandA.Torralba. Modeling the shape of the scene: A holistic representation of the spatial
envelope. International Journal of Computer Vision, 42:145–175, 2001
3. Rosenholtz R (1999) A simple saliency model predicts a number of motion pop out phenomena.
VisionResearch 39(19):3157–3163
4. Itti L, Koch C (2000) A saliency-based search mechanism for overt and covert shifts of visual
attention. Vision Res 40:1489–1506
5. T. Avraham and M. Lindenbaum. Esaliency: Meaningful attention using stochastic image
modeling. IEEE Transactions on Pattern Analysis and Machine Intelligence, 99(1), 2009
6. Abouelaziz I, El Hassouni M (2015) “New models of visual saliency: Contourlet transform
based model and hybrid model”, 2015 Intell. Syst. Comput. Vision, ISCV
7. H. Peng, B. Li, R. Ji, W. Hu, W. Xiong, and C. Lang, “Salient object detection via Low-rank and
Structured sparse Matrix Decomposition,” Proc. 27th AAAI Conf. Artif. Intell. AAAI 2013,
no. 1, pp. 796–802, 2013
8. Borji, A., Cheng, M. M., Hou, Q., Jiang, H., & Li, J. Salient object detection: A survey. Compu-
tational Visual Media, 5(2), 117–150. (2019).https://doi.org/10.1007/s41095-019-0149-9
9. Xiang Kui Ruan and Na Tong and Huchuan Lu. How far we away from a perfect visual saliency
detection- DUT-OMRON: a new benchmark dataset. Inproceedings Ruan 2014 HowFW. http://
saliencydetection.net/dut-omron/#org96c3bab
10. Kim J, Han D, Tai YW, Kim J (2016) Salient region detection via high-dimensional color
transform and local spatial support. IEEE Trans Image Process 25(1):9–23. https://doi.org/10.
1109/TIP.2015.2495122
The Fast-NLMS Algorithm for Acoustic
Echo Cancellation with Double-Talk
Detection
Islam Hassani, Madjid Arezki, and Ahmed Benallal
Abstract In this paper, we introduce a Fast Normalized Least Mean square algo-
rithm (FNLMS) to eliminate the acoustic echo in the presence of double talk signal.
When the near-end speech is present the performance of the adaptive filtering is
degraded, thus, a double-talk detector (DTD) must be used for controlling the
Acoustic Echo Cancellation (AEC). A Normalized cross correlation (NCC) method
applied to FNLMS algorithm is proposed in this paper. Experimental results using
different input signals and real measured impulse response (audio-conference room),
show a fast convergence and detection of double-talk of the proposed algorithm
compared with NLMS algorithm.
Keywords Acoustic echo cancellation · Double talk detector · Fast-NLMS ·

Adaptive filters
1 Introduction
Acoustic echo cancellers are used to suppress acoustic echoes [1] in modern commu-
nication systems. An acoustic echo canceller (AEC) is generally implemented by an
adaptive finite impulse response (FIR) filter. The presence of the near-end speech
makes the adaptation of the echo cancellation filter problematic. A strong near-end
speech acts as a large disturbance to the adaptive filtering algorithm and may cause
the echo cancellation filter to diverge [2]. It is for this reason that the update of the
echo cancellation filter should be turned off when near-end speech is detected. An
I. Hassani (B) · M. Arezki · A. Benallal

Signal Processing and Image Laboratory (LATSI), Department of Electronics, University of
Blida1, Blida, Algeria
M. Arezki
A. Benallal

https://doi.org/10.1007/978-981-15-6403-1_90
1272 I. Hassani et al.
algorithm that detects the presence of near-end speech is called a double talk detector
(DTD).Several methods of double-talk detection have been proposed in the literature.
Such as, a level energy based DTD: the well-known Geigel algorithm, presented in
[3], uses an amplitude comparison between near-end and far-end speech signals, and
a method based on the signal envelope is presented in [4], Holder inequality based
method is presented in [5]. DTD using coherence is developed in [6], also, methods
based on cross-correlation (CC) and normalized cross correlation (NCC) algorithms
are proposed in [7] and [8]. In this paper, we propose to use Fast NLMS algorithm
with Normalized Cross-Correlation method for acoustic echo cancellation applica-
tion. We focused our interest on the evaluation of the mean-square-error MSE. The
rest of this paper is organized as follows: in Sect. 2, we describe the AEC system
and the adaptive filtering, and then Double-Talk detection methods. The proposed
algorithm is given in Sect. 3. The simulation results are presented in Sect. 4 and
finally a conclusion is given in Sect. 5.
2 Acoustic Echo Cancellation System
The principle of an echo cancellation is shown in Fig. 1. the echo signal received on
the microphone is obtained by the convolution between the room impulse response
h and the input signal x(n), then the echo signal is:
y(n) = hT x(n) (1)

T
where h = h 0 h 1 , . . . , h L−1 , L is the length of the echo path, the superscript ()T
denotes transpose of a vector and x(n) = [x(n)x(n − 1), . . . . . . , x(n − L + 1)]T is
the vector contains the last L samples of the far-end speech signal x(n). The desired
signal is:
d(n) = y(n) + s(n) + v(n) (2)
Fig. 1 Principle of echo

cancellation
The Fast-NLMS Algorithm for Acoustic Echo … 1273
where s(n) is the near-end speech signal and v(n) is the background noise.
w(n) is a finite impulse response (FIR) adaptive filter.
The estimated echo ŷ(n) is created by the convolution of the coefficient vector
T
of adaptive filter w(n) = w0 (n)w1 (n), . . . . . . , w L−1 (n) , with the received input
signal x(n). The estimated echo is:
ŷ(n) = w T (n − 1)x(n) (3)
The adaptive algorithms estimate h by w(n) using a priori estimation error, is

written, for each n:
e(n) = d(n) − ŷ(n) (4)
The filter is updated at each instant by feedback of the estimation error proportional
to the adaptation gain, denoted as g(n), and according to:
w(n) = w(n − 1) + g(n)e(n) (5)
A. Adaptive filtering
The different adaptive algorithms are differentiated by the adaptation gain calcula-
tion.
The Least Mean Square (LMS) algorithms derived from the gradient [9], based
on the minimization of the mean-square error. For the Normalized LMS (NLMS)
algorithm, the adaptation gain is given by:
μ
g(n) = x(n) (6)
Lπx (n) + c0
where μ is the adaptation step and c0 is a small positive constant used to avoid
division by zero in absence of the input signal and πx (n) is the estimated power of
input signal.
The Recursive Least Square (RLS) algorithm, for which minimizes a deterministic
sum of squared errors [9]. Fast versions of these algorithms are derived from the RLS
by using forward and backward linear prediction analysis over the signal x(n), the
adaptation gain is given by:
g(n) = γ (n)c̃(n) (7)
where the variables γ (n) and c̃(n) indicate respectively the likelihood variable and
normalized Kalman vector.
Recently, a new adaptive algorithm with fast convergence and low complexity
is proposed [10]. This algorithm FNLMS derived from the FRLS algorithm where
the adaptation gain is obtained by discarding completely the forward and backward
predictors. Thus, in this algorithm, a simplified adaptation gain is proposed:
ε(n)

c̃(n) − λα(n−1)+c
= 0 (8)
c(n) c̃(n − 1)
where ε(n) is the first order prediction error, α(n) is the forward prediction error
variance and λ is the exponential forgetting factor 0 < λ ≤ 1.
ε(n) = x(n) − a(n)x(n − 1) (9)
where a(n) is the prediction parameter and it is estimated by the following equation:
r1x (n)
a(n) = (10)
r0x (n) + ca
where r1x (n) is an estimate of the first lag correlation function of x(n), r0x (n) an
estimate of the input signal power and ca is a small positive constant used to avoid
division by zero. The forward prediction error variance is now evaluated by:
α(n) = λα(n − 1) + ε2 (n) (11)
B. Double talk-detection
In Acoustic Echo Cancellation, the most difficult problem is to handle with the
situation of Double-talk presence. Double-talk occurs when far-end and near-end
talk at the same time, as a result, the far-end speech signal is corrupted by near-end
signal.
To solve this problem, one introduces the Double-talk Detector. The task of DTD
is to freeze the adaptation step during filtering in case of near-end speech present to
avoid the divergence of adaptive algorithm. Without DTD, when the near-end talking
would make the system estimation process fail and produce extremely erroneous
results. Typically, the DTD calculates a variable decision ξ (n), and double-talk is
declared if is lower than a given threshold value T [11]. The optimum variable
decision ξ (n) for double-talk detection will behave as follows:
• If s(n) = 0 (double talk is not present),ξ (n) ≥ T .
• If s(n) = 0 (double talk is present), ξ (n) < T .
The control of the adaptive filter by DTD is defined as:

ξ (n) ≥ T, DT D = 0, update f ilter coe f f icients
decision =
ξ (n) < T, DT D = 1, f r eeze adaptation
Geigel is one of the most popular algorithms of DTD, which compares the magni-
tude of the microphone signal with the recent history of the far-end signal, where the
variable decision of this latter is defined as:
max{|x(n)|, . . . , |x(n − L + 1)|}

ξG (n) = (12)
|d(n)|
A method of DTD based on a Cross-Correlation (CC) between the far-end

and error signals is proposed in [12]. Moreover, approximate versions, such as a
Normalized Cross-Correlation method (NCC) are developed in [8, 13]. This method
calculates a variable decision ξ N CC (n) which is defined by:
r̂ed
ξ N CC (n) = 1 − (13)
σ̂d2
where r̂ed is the estimate of red , red = E{e(n)d(n)} is the cross-correlation between
e(n) and d(n).E{.} denotes the mathematical expectation and σd2 is the variance of
d(n).
This variable decision is based on the estimates r̂ed and σ̂d2 where are found using
the recursive form:
r̂ed (n) = λ DT D r̂ed (n − 1) + (1 − λ DT D )e(n)d(n) (14)
σ̂d2 (n) = λ DT D σ̂d2 (n − 1) + (1 − λ DT D )d(n)d(n) (15)
where λ DT D is the exponential weighting factor (λ DT D < 1).
3 The Studied DTD for AEC
In this section, we propose to use the Fast-NLMS type adaptive algorithm with DTD,
the method used in this article is known as Normalized Cross-correlation technique. It
is based on the calculation of the estimates using the exponential recursive weighting
algorithm to obtain the values of the cross-correlation (rem ) between error signal and
microphone signal and the variance (σm2 ) of the microphone signal. And achieve the
decision statistic (ξ DT D ) from these values. Finally, we compare this value to the
threshold (T ) to make the decision of Double-talk Detector.
Mean Square Error (MSE):
The purpose of the adaptive filter is minimizing the Mean Square Error:

M S E(d B) = 10log10 E |e(n)|2 (16)
Therefore, the values and graph of this quantity will be essential to evaluate
the performance of the adaptive filter. If the adaptive algorithm works well, after
convergence time, the value of MSE should be reduced gradually to zero (for the
case of no near-end signal).
Table 1 Fast-NLMS with

Initialization:
DTD based on NCC
L = length(h);
w(0) = c̃(0) = 0, γ (0) = 1, r1 (0) = 0,
λ = 0.99, λa = 0.9985,
α(0) = r0 (0) = E 0 ,where E 0 is an initialization constant.
T = 0.92; Set the threshold
λ DT D = 0.95; to calculate decision statistic of DTD
d(n) = y(n) + s(n) + v(n);
For each instant of time n = 1, 2…
Prediction error:
r1 (n) = λa r1 (n − 1) + x(n)x(n − 1)
r0 (n) = λa r0 (n − 1) + x 2 (n)
r1 (n)
a(n) = r0 (n)+ca
ε(n) = x(n) − a(n)x(n − 1)
α(n) = λα(n − 1) + ε2 (n)
Adaptation gain:

ε(n)
c̃(n) − λα(n−1)+c
= 0
c(n) c̃(n − 1)
x(n)ε(n)
δ(n) = c(n)x(n − L) + λα(n−1)+c0
γ (n) = 1+γγ(n−1)δ(n)
(n−1)
Filtering Part:
e(n) = d(n) − w T (n − 1)x(n)
—————- NCC for DTD————————
threshold(n) = T;
compute
r̂ed (n) = λ DT D r̂ed (n − 1) + (1 − λ DT D )e(n)d(n)
σ̂d2 (n) = λ DT D σ̂d2 (n − 1) + (1 − λ DT D )d(n)d(n)
r̂ed
ξ N CC = 1 − σ̂d2
if ξ N CC > threshold
μ = 1, update filter coefficients.
end
——————————————————
w(n) = w(n − 1) − μe(n)γ (n)c̃(n)
The proposed algorithm is demonstrated in the Table 1.
We used in our simulation two different types of input signals. The first one is a
stationary correlated noise, with a spectrum equivalent to the average spectrum of
speech. It is usually called USASI (USA Standards Institute, now ANSI) noise in
the field of acoustic echo cancellation see Fig. 3 (upper). Since in real situation the
input signal is nonstationary, we also ran the algorithm on a second signal, which
is a typical speech signal see Fig. 5 (upper). These two signals, sampled at 16 kHz,
are filtered by a real acoustic impulse response measured in a real audio-conference
(denoted ACN) room to obtain the desired signals see Fig. 2.
Figures 3 and 5 (Lower), show the Near-end speech signal used in our simulations
and Double-Talk periods. As we can see, the periods (from 0 to 30000) and (from
90000 to the last sample) represent the Double-Talk periods. Figure 4 shows Mean
1.5
1
Normalized magnitude
0.5
-0.5
-1
-1.5
0 100 200 300 400 500 600 700 800 900 1000 1100
Samples
Fig. 2 ACN impulse response with length 1024
Far-end signal x(n)

2
1
magnitude
Normalized
-1
-2
0 2 4 6 8 10 12 4
10
Samples
Near-end signal s(n)
4
magnitude
Normalized
-2
-4
0 2 4 6 8 10 12
4
Samples 10
Fig. 3 The used signals x(n):USASI noise, s(n): speech signal used for Double-Talk
10
NLMS with DTD
0 FNLMS with DTD
-10
Mean Square Error (MSE)
-20
-30
-40
-50
-60
-70
-80
-90
-100
0 2 4 6 8 10 12
4
Samples 10
Fig. 4 Comparative performance of NLMS with DTD and FNLMS with DTD using the data of
Fig. 3 NLMS (μ = 1) FNLMS (λ = 0.99,λa = 0.9985)
Far-end signal x(n)

2
Normalized
magnitude
-2
0 1 2 3 4 5 6 7 8 9 10
4
Samples 10
Near-end signal s(n)

2
magnitude
Normalized
-2
0 1 2 3 4 5 6 7 8 9 10
4
Samples 10
Fig. 5 The used signals x(n):speech input signal, s(n): speech signal used for Double-Talk
square error performance of the NLMS and the FNLMS algorithms when Normalized
Cross Correlation detector is used as DTD and using USASI noise input signal. We
have noticed that FNLMS with NCC detector converges faster than NLMS with
DTD when there is no Double-Talk, and the adaptation is slowed down during the
Double-Talk periods.
The same results of Fig. 4 are obtained in Fig. 6 and using a non-stationary signal
(speech) as input, both algorithms are freezed or slowed down during Double-Talk
20
NLMS with DTD
10 FNLMS with DTD
-10
Mean Square Error (MSE)
-20
-30
-40
-50
-60
-70
-80
0 1 2 3 4 5 6 7 8 9 10
4
Samples 10
Fig. 6 Comparative performance of NLMS with DTD and FNLMS with DTD using the data of
Fig. 5, NLMS(μ = 1) FNLMS (λ = 0.99, λa = 0.9985)
periods, and the FNLMS algorithm with DTD has a high convergence speed in the
absence of Near-end signal.
5 Conclusion
In this paper, we have presented the FNLMS algorithm with DTD based on NCC,
where the main purpose of this DTD is to halt the update of filter coefficients during
the double-talk periods. DT periods are determined from decision statistic depending
on the estimated variance of near-end signal and cross-correlation between error
signal and microphone signal. The simulation results demonstrate a good perfor-
mance with FNLMS-type algorithm compared to NLMS algorithm with different
input signals and a real acoustic impulse response, in terms of convergence speed
and detection of near-end signal.
References
1. Christoph, Markus. “Acoustic echo cancellation.” U.S. Patent No. 8,712,068. 29 Apr. 2014
2. Benesty, Jacob, et al. “Double-Talk Detection.” A Perspective on Stereophonic Acoustic Echo
Cancellation. Springer, Berlin, Heidelberg, 2011. 71–79
3. Duttweiler D (1978) A twelve-channel digital echo canceler. IEEE Trans Commun 26(5):647–
653
4. Szwoch Grzegorz, Czyżewski Andrzej, Kulesza Maciej (2008) A low complexity double-talk
detector based on the signal envelope. Sig Process 88(11):2856–2862
5. Paleologu, Constantin, et al. “Class of double-talk detectors based on the holder inequality.”
Acoustics, Speech and Signal Processing (ICASSP), 2011 IEEE International Conference on.
IEEE, 2011
6. Gansler, Tomas, et al. “A double-talk detector based on coherence.” IEEE Transactions on
Communications 44.11 (1996): 1421–1427
7. Benesty Jacob, Morgan Dennis R, Cho Jun H (2000) A new class of doubletalk detectors based
on cross-correlation. IEEE Trans Speech Audio Process 8(2):168–172
8. Gänsler Tomas, Benesty Jacob (2006) The fast normalized cross-correlation double-talk
detector. Sig Process 86(6):1124–1139
9. Haykin, Simon O. Adaptive filter theory. Pearson Higher Ed, 2013
10. A. Benallal, M. Arezki, “A fast convergence normalized least-mean-square type Algorithm for
Adaptive Filtering”, International Journal Adaptive Control and Signal Processing, 2013 John
Wiley & Sons, https://doi.org/10.1002/acs.2423
11. M. Hamidia, A. Amrouche, “Double-talk detector based on speech feature extraction for
acoustic echo cancellation,” In proceeding of IEEE, the 22nd International Conference on
Software, Telecommunications and Computer Networks, SoftCOM, pp. 393–397, 2014
12. Benesty J, Morgan DR, Cho JH (2000) A new class of doubletalk detectors based on cross-
correlation. IEEE Trans on speech and audio processing 8(2):168–172
13. Benesty J, Gänsler T (2002) A multichannel acoustic echo canceler double-talk detector based
on a normalized cross-correlation matrix. European Tran on telecommunications 13(2):95–101
Design of Sensor Based on CRLH
for Liquid Mixture Application
Sekkache Hocine, Mohamed Lashab, and Salim Ouchtati
Abstract A high sensitivity sensor loaded with composite right/left-handed (CRLH)

resonator is investigated in this paper. The proposed sensor offers many advantages,
as high sensitivity, small sample volume, short assay time, low-cost, and lab-on-
a-chip compatibility. To achieve higher sensitivity, the sensor must have a sharp
resonance in its frequency response and a high concentration of electric field to
allow the detection of small changes in dielectric environment. The proposed sensor
can detect the liquid identification from the resonance frequency shift caused by the
liquid sample under test. The operating frequency is from 2 to 3.5 GHz. The sensor
is designed and simulated using HFSS (High-Frequency Structure Simulator), three
types of liquid are tested; methanol, ethanol and distilled water. The obtained results
are very promising compared to the literature.
Keywords Microwave sensor · Metamaterial · CRLH · Liquid mixture ·

Sensitivity
1 Introduction
Metamaterials are effective media that can provide an engineered response to elec-
tromagnetic radiation that is not available from the range of naturally occurring mate-
rials. They consist of arrays of structures in which both the individual elements and
the unit cell (e.g., SRR, CSSR, CRLH…) are small compared with the wavelength
of operation [1].
S. Hocine · S. Ouchtati (B)

University of 20 Aout 1955,- Skikda, Skikda, Algeria
S. Hocine
M. Lashab
University of Larbi Ben M’hidi, Route de Constantine, Oum El Bouaghi, Algeria

https://doi.org/10.1007/978-981-15-6403-1_91
1282 S. Hocine et al.
The CRLH-TL (composite right/left-handed transmission line) technology is

based on a combination of right hand (RH) and left hand (LH) electric equivalent
circuit, leading to a resonant frequency of the antenna or the sensor independently of
its dimension [2]. Left hand metamaterials (LHMs) based on TL approach general
presents wider band and lower losses [3]. Design of CRLH-TL is based on the inser-
tion of inter-digital capacitors and shorted stub inductors into the designed sensor, by
adjusting the dimension of these components one can have desired resonant frequency
[4].
It has been shown that by increasing the shunt inductance, the zero-order resonant
frequency can be decreased, so that the sensor size can be made much smaller [5].
Zero order resonators (ZOR) are based on composite right left handed (CRLH)
transmission line (TL) features considered as metamaterial behavior, first introduced
by Caloz and itoh in [6, 7]. The concept of composite right/left-handed (CRLH)
MTM, introduced by Caloz et al. [8].
In the literature many research works have been carried out on liquid mixture
classification, a dielectric characterization based on measurement probably the
first one presented by Bao [9]. Dielectric identification of liquid mixture such as
Methanol–ethanol, water–methanol, and water–ethanol have been measured with
the open-ended coaxial-probe method, and analyzed with Debye function over a
wide frequency range. Similar work on complex dielectric was given by Meissner
and Wentz [10], in this work the temperature behavior was added. Dielectric material
characterization based on experimental measurement carried out by Saeed [11], the
sensor is based on rectangular cavity. A two ports sensor loaded with SRR meta-
materials unit cells is presented by Withayachumnankul [12], the sensor is used to
identify dielectric of liquid mixture such as ethanol and methanol. A sensor loaded
with metamaterial CSSR unit cell is proposed by Ebrahimi [13], the sensor has
two ports and can be used to microfluid identification with high sensitivity. A two-
dimensional displacement and alignment sensor is proposed by Horestani [14], the
sensor is based on two open-ended transmission lines, and each loaded with a split
ring resonator (SRR). The sensor proposed by Benkhaoua [15] is a DD-SRR, double
sided SSR, with only one port, it used also for liquid mixture. The sensor proposed
by Albishi [16], is two ports type, based on CSSR, it used for crack detection in
metallic materials.
This work was preceded by another work in [17], in which we used the SRR
unit cell metamaterial. The sensor consists of a microstrip line loaded on both sides
with two SRR unit cells, where the addition of the SRR unit cell of the sensor has a
positive role in raising its sensitivity (Fig. 1).
In this paper a sensor with two ports has been designed, the sensor is loaded with
CRLH-TL resonator, first to increase its sensitivity, second to enable miniaturization
effect. The paper is based only on simulation by HFSS, the obtained results show
high sensitivity of the sensor for liquid mixture by dielectric identification, three
types of liquid are tested; methanol, ethanol and distilled water.
Design of Sensor Based on CRLH … 1283
Fig. 1 Sensor with two SRR

unit cells
SRR unit cell
microstrip
line
2 Sensor Design
A. Theoretical aspect
Microwave structure based on transmission line (TL) technology, allow fabricating

a planar antenna with a resonant frequency that has no dependence on the physical
dimension of the structure. Sensors or antenna based on TL technology contains
an infinite wavelength at its fundamental mode which is required to realize a ZOR
antenna. A practical realization of a LH-TL, which includes unavoidable RH effects,
known as a CRLH-TL is able to deliver an infinite wavelength. CRLH-TLs are
obtained by cascading a sub-wavelength unit cell constituted of a series capacitance
and a shunt inductance as the LH MTM structure [18].
The properties of the CRLH metamaterial can be analyzed using the transmission
line theory [19]. The equivalent circuit model of the CRLH TL unit-cell is shown
in Fig. 2. The resonance condition of such open-ended resonator is given by the
dispersion relation [6], for n equal to zero the CRLH-TL resonator is obtained.
The homogeneous model of a CRLH lossless transmission line shown in Fig. 2
consists of an inductance LR in series with a capacitance CL and a shunt capacitance
CR in parallel with an inductance LL . The propagation constant of TL is given by:
√
γ = α + jβ = ZY (1)
Fig. 2 CRLH-TL resonator, LR CL

electric circuit
CR LL
where Z and Y are the per-unit length impedance and per-unit length admittance,
respectively. Z and Y of CRLH TL are defined as:

1
Z (ω) = j ωL R − (2)
ωC L

1
Y (ω) = j ωC R − (3)
ωL L
B. Proposed Sensor
The proposed sensor is presented in Fig. 3, the substrate is an FR4 with relative
permittivity εr = 4.4 and dielectric loss tangent δ = 0.02, the thickness of the
substrate is 1.6 mm. The sensor is composed of strip metallic line with two ports, the
sensor presents a coplanar structure, ground at the top side. The proximity coupling
is used as the feed network to achieve a good impedance matching to 50 . The
sensor is loaded with CRLH-TL resonator as shown in Fig. 4, the zigzag strip is
Fig. 3 Proposed sensor

E Wt He
Lp W
Wc
H Ht
Fig. 4 CRLH resonator
Hc
We
g
Table 1 Parameters of the proposed sensor in (mm)

L W H E Lp He
16 10 6 2.8 2 2.15
Ht Wc Wt We Hc g
3.6 4 0.4 2.35 0.25 2
Port 1
Liquid Mixture
Port 2
Fig. 5 Sensor loaded with liquid mixture
considered as the admittance Y, and the inter-digital capacitor considered as the Z

impedance. The geometrical dimension is given by Table 1 (Fig. 5).
3 Results Analysis
The transmission coefficient of the sensor without liquid mixture is given in Fig. 6,
the figure show two resonant frequencies in the range from 0.5 GHz to 3.5 GHz; the
main one is at 3.05 GHz. The sensor loaded with liquid mixture is presented in Fig. 7,
we notice that the transmission coefficient of the water is shifted down in frequency
to 1.15 GHz, the resonant frequency of the methanol is at 2.28 GHz, the resonant
frequency of the ethanol is at 2.66 GHz.
The complex permittivity of the tested liquid mixture at 25 °C in the range of
1.4 GHz to 1.9 GHz are given as follow; methanol (ε’ = 34, ε” = 2) ethanol (ε’ =
26, ε” = 3) and distilled water (ε’ = 79, ε” = 0.1), where ε’ is the real part of the
permittivity and ε” is the complex part of the permittivity, this is given by [9].
To determine the characteristics and the accuracy of the proposed sensor, addi-
tional simulation were conducted using distilled water/ethanol mixture in a range
of 0–100% ethanol concentration with the step size of 10% as liquid samples.
Fig. 6 Transmission S21, Sensor without Liquid Mixture

coefficient without liquid 0
air
mixture
Transmission coefficient S21 (dB)

-10
-20
-30
-40
-50
-60
0.5 1 1.5 2 2.5 3 3.5
Frequency (Ghz)
Fig. 7 Transmission S21, Sensor with Liquid Mixture

0
coefficient S21 of the sensor
with Liquid mixture -5
Transmission coefficient S21 (dB)
-10
-15
-20
-25
-30
Methanol
-35 Ethanol
Water-distilled
-40
-45
0.5 1 1.5 2 2.5 3
Frequency (Ghz)
According to the Debye relaxation model, the real part of the mixture permittivity
can be expressed as follows [9]:
εs − ε∞
ε = ε∞ + (4)
1 + ω2 τ2
Where εs is the low-frequency permittivity, ε∞ is the high-frequency permittivity,

τ is the relaxation time, ω is the radiated frequency [9].
Due to the values stability of the permittivity around 1.5 GHz, the dispersion is
not taken into account. We can express the sensor sensitivity by:
fo − f1
S= (5)
fo dεr
Where fo is the frequency of test and f1 is a near frequency to fo , and d εr is the

variation in the relative permittivity.
The simulated sensitivity has a maximum value of 0.57%, this depends on the
type of liquid used. Compared to the literature this is a very high sensitivity.
4 Conclusion
In this paper, a new type of sensor based on CRLH is designed for liquid mixture
identification. The use of this sensor is of great interest for many fields of application,
such as chemistry, microbiology, biomedical, medicine, and personal safety.
The sensor is simulated on HFSS, three types of liquid are test; distilled water,
methanol and ethanol. The obtained results have shown that the sensitivity of the
sensor is very high compared to the literature and the dimension of the sensor is very
compact as well the testing surface.
Future work will focus primarily on improving sensor sensitivity in order to
distinguish between different liquid mixtures more accurately.
References
1. Capolino F. Theory and phenomena of metamaterials. CRC press, 2009 by Taylor and Francis
Group, LLC. pp 14–1
2. Li L, Jia Z, Huo F, Han W (2013) A novel compact multiband antenna employing dual-
band CRLH-TL for smart mobile phone application. IEEE Antennas and Wireless Propagation
Letters, 12
3. Jordi Bonache, Marta Gil, Ignacio Gil, Joan García-Garcíaand Ferran Martín, “On the Electrical
Characteristics of Complementary Metamaterial Resonators,” IEEE Microwave And Wireless
Components Letters, Vol. 16, No.10, Oct 2006
4. Naeem Ahmad Jan, Mohamed Lashab, Chemseddine Zebiri, Djouablia Linda, Benabdelaziz
Fatiha, R.A.Abd-Alhameed, “Compact CPW Antenna Loaded With CRLH-TL & EBG for
Multi-band and Gain Enhancement,” LAPC’16 Loughborough, 2016
5. Mohamed Lashab, Naeem Ahmad Jan, Chemseddine Zebiri, “The I shape Antenna Loaded
With ZOR For WLAN and WiMax Application,” LAPC’15 Loughborough, 2015
6. Sanada A, Caloz C, Itoh T (2003) “Zeroth Order in Composite right/left-handed transmission
line resonators”, in Proc. Asia-pasific Microwave Conf, Seoul Kouria Nov
7. C. Caloz and T. Itoh, “Transmission line Approach of left-handed (LH) Materials and Microstrip
Implementation of an Artificial LH Transmission Line,” IEEE Transaction on Antennas And
Propagartion, Vol. 52, No. 5, May 2004
8. C. Caloz and T. Itoh, “Electromagnetic Metamaterials: Transmission Line Theory and
Microwave Applications,” 2006 by John Wiley & Sons, Inc. pp. 16
9. Bao Jian-Zhong, Swicord Mays L, Davis Christopher C (1996) Microwave Dielectric Charac-
terization of Binary Mixtures of Water, methanol, and ethanol. J. Chem. Phys. 104(12):22
10. Thomas Meissner and Frank J. Wentz “The Complex Dielectric Constant of Pure and Sea
Water From Microwave Satellite Observations,” IEEE Transaction on Geoscience And Remote
Sensing, Vol. 42, No. 9, Sep. 2004
11. Kashif Saeed, Roger D. Pollard and Ian C. Hunter, “Substrate IntegratedWaveguide Cavity
Resonators for Complex Permittivity Characterization of Materials,” IEEE Transaction on
Microwave Theory and Techniques, Vol. 56, No. 10, Oct 2008
12. Withawat Withayachumnankul, Kata Jaruwongrungsee, Christophe Fumeaux and Derek
Abbott, “Metamaterial-Inspired Multichannel Thin-Film Sensor,” IEEE Sensors Journal, Vol.
12, No. 5, May 2012
13. Amir Ebrahimi, Withawat Withayachumnankul, Said Al-Sarawi, Derek Abbott, “High-
Sensitivity Metamaterial-Inspired Sensor for Microfluidic Dielectric Characterization,” IEEE
Sensors Journal, Vol. 14, No. 5, 2014
14. A.K. Horestani, J. Naqui, D. Abbott, C. Fumeaux and F. Martín, “Two-dimensional displace-
ment and Alignment sensor based on reflection Coefficients of open microstrip lines loaded
With split ring resonators,” Electronics Letters 10th April 2014 Vol. 50 No. 8 pp. 620–622
15. Larbi Benkhaoua, Mohamed Taoufik Benhabiles, Mohamed Lahdi Riabi & smail Mouissat,
“Miniaturized Quasi-Lumped Resonator for Dielectric Characterization of Liquid Mixtures,”
IEEE Journal & Magazine, Vol. 16, Iss. 06, 2016
16. Ali M. Albishi, Omar M. Ramahi, “ Microwaves-Based High Sensitivity Sensors for Crack
Detection in Metallic Materials,” IEEE Tansaction on Microwave Theory and Techniques, Vol.
65, Iss. 5, 2017
17. Hocine Sekkache, Mohamed Lashab, “Study and design of a sensor containing Meta-materials,
Application into biomedical,” IEEE Xplore Digital Library, 2018
18. K. Fujimoto and H. Morishita, “Modern Small Antennas,” Cambridge University Press, 2013.
pp. 217
19. C. Caloz and T. Itoh, Electromagnatic Metamaterials (Transmission Line theory and microwave
application), Wiley IEEE Press 1st edition, 2005

10.1007@978 981 15 6403 1

Uploaded by

Copyright:

Available Formats

10.1007@978 981 15 6403 1

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

10.1007@978 981 15 6403 1

Uploaded by

Copyright:

Available Formats

Lecture Notes in Electrical Engineering 682

More information about this series at http://www.springer.com/series/7818

Salim Ziani Ivan Zelinka

Proceedings of the 4th

Salim Ziani Ivan Zelinka

ISSN 1876-1100 ISSN 1876-1119 (electronic)

The International Conference on Electrical Engineering and Control Applications

Salim Ziani, University of Frères Mentouri Constantine 1, Algeria

Soﬁane Bououden, University of Khenchela, Algeria

Soﬁane Bououden, University of Khenchela, Algeria

ICEECA established a local committee from the department of electronics as an

International Advisory Committee

Abedelfetah Charef, University of Frères Mentouri Constantine 1, Algeria

Scientiﬁc Program Committee

ICEECA established an international committee of selected well-known experts in

Abdelaziz Hamzaoui, University of Reims, France

Mohamed Boumehraz, University of Biskra, Algeria

This event is supported by the University of frères Mentouri Constantine 1,

Khenchela, Algeria Sofiane Bououden

Control and Systems Engineering (CSE)

Networked Cooperation-Based Distributed Model Predictive Control

The Importance of Applying Artiﬁcial Intelligence on Unmanned

GA Based Adaptive Fuzzy Sliding Mode for 3 DOF Spatial Arm

Renewable Energy (RE)

Strategy for Optimization of Energy Management Based on Fuzzy

Investigation and Prototyping Implementation of a Novel Solar

Faults Diagnosis-Faults Tolerant Control (FTC)

Fault-Tolerant Path-Tracking Control with PID Controller

Signal and Communications (SC)

Narrowband Sensor-Antenna Based on an Interdigital Capacitor . . . . 1071

An Image Segmentation Model Using a Level Set Method Based

Yassine Beddiaf, Fatiha Zidani, and Larbi. Chrifi-Alaoui

Keywords Induction motor · Differential flatness · Sensorless vector control

© Springer Nature Singapore Pte Ltd. 2021 3

2 Differential Flatness Theory

Consider a dynamic system defined by the following state equation:

we define the following variables

y(t) = X 2 (t) and U (t) = ẏ(t)

X 2 (t) = ẏ(t) and U (t) = ÿ(t)

2.1 Control Law Synthesis Methodology for Flat Systems

2.2 Trajectory Planning

In the hypothesis of a perfect model, we will then have, for t from t0 to t f ,

2.3 Stabilization Around Reference Trajectories

Fig. 1 Bloc diagram of

3.1 Analysis of the Flatness of IM Model

The IM modeling stationary reference frame is given by the following equations:

ϕ̄s = L s ī s + M īr e j pθ1 (12)

ϕ̄r = L r īr + M ī s e j pθ1 (13)

where * represents conjugated complex variables.

From Fig. 2 we can write

ϕr = ϕ̄r e j pθ1 = ρe jα1 (17)

Fig. 2 Angle spotting

From Eq. (15) we define the following equation:

If we choose y = (α1 , θ1 ) as a flat output, we obtain:

ϕ̃r = B(y, ẏ, ÿ) (22)

ϕ̄s = L s ī s + Me j pθ1 īr

3.2 Design of Flatness Control Based of IM

L r J θ̈1−r e f + f θ̇1−r e f + Ĉr

where Ĉr represents the estimated charge torque.