10.1007@978 981 15 6403 1
10.1007@978 981 15 6403 1
10.1007@978 981 15 6403 1
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 **
Editors
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
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
General Chairs
Program Chairs
v
vi Organization
Steering Committee
Local Committee
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
xi
xii Contents
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.
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
e-mail: [email protected]
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.
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)
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:
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).
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.
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)
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.
d
V̄s = Rs ī s + ϕ̄s (10)
dt
d
0 = Rr īr + ϕ̄r (11)
dt
pM
Ce = Jm ī s · ϕ̄r∗ (14)
Lr
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)
We obtain:
1 L r d ϕ̄r
ī s = ϕ̄r + (18)
M Rr dt
p 2
ce = ρ α̇1 (19)
Rr
Therefore
ϕ̃r = ρe jα1 = B θ̈1 , θ̇1 , α1, α̇1 , Cr (21)
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)
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.
Sensorless Speed Control of Induction Motor … 9
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
d ω̂
dt (30)
dt
= JpM ρ I − Jf ω̂ − ĈJr + l2 ω̂ − ω
L r r e f sq
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
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.
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
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
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
−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
Sensorless Speed Control of Induction Motor … 13
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)
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)
Fig. 5 Schematic diagram the proposed flatness-based control scheme with the use of reduced
observer
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).
Sensorless Speed Control of Induction Motor … 15
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)
16 Y. Beddiaf et al.
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.
Sensorless Speed Control of Induction Motor … 17
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)
18 Y. Beddiaf et al.
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.
Sensorless Speed Control of Induction Motor … 19
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)
20 Y. Beddiaf et al.
Fig. 8 (continued)
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
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.
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:
• (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
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
Multi-UAVs Coverage Path Planning
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
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.
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.
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.
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.
Multi-UAVs Coverage Path Planning 31
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.
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)
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
Multi-UAVs Coverage Path Planning 33
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
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
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 .
2 Probabilistic Roadmap
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.
40 A. Belaid and B. Mendil
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
Reducing the Collision Checking Time in Cluttered … 41
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
42 A. Belaid and B. Mendil
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.
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
44 A. Belaid and B. Mendil
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
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
References
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
1 Introduction
M. Guesbaya · A. C. Megherbi
Laboratory of Identification, Command, Control and Communication, University of Mohamed
Kheider, Biskra, Algeria
e-mail: [email protected]
A. C. Megherbi
e-mail: [email protected]
H. Megherbi (B)
Department of Electrical Engineering, University of Mohamed Kheider, Biskra, Algeria
e-mail: [email protected]
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.
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
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:
50 M. Guesbaya et al.
(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)
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:
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:
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.
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:
vit+1 = vit + xit − x∗ f i (14)
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:
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:
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:
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.
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
54 M. Guesbaya et al.
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.
Random Scaling-Based Bat Algorithm for Greenhouse Thermal … 55
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
56 M. Guesbaya et al.
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
Random Scaling-Based Bat Algorithm for Greenhouse Thermal … 57
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
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
60 M. Guesbaya et al.
Fig. 9 Measured and predicted inside temperature of the 1st, 4th and 5th days
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
Random Scaling-Based Bat Algorithm for Greenhouse Thermal … 61
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
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
62 M. Guesbaya et al.
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.
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
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.
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)
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
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:
66 S. E. Farhi et al.
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 :
+
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
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.
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
where i sα = î sα − i sα ; i sβ = î sβ − i sβ
Defines the errors:
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):
70 S. E. Farhi et al.
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.
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
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. 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
74 S. E. Farhi et al.
small error in transient speeds only when stator resistance change, and good tracking
performance the motor torque with reference load torque.
6 Nomenclature
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 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
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
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.
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
[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
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.
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
80 A. Terki and H. Boubertakh
(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
(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)
(4) Power generating limit: The active powers that can be generated by the
generating units have minimum and maximum values.
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
Cuckoo Search Algorithm for Solving the Problem … 81
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)
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 .
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.
Cuckoo Search Algorithm for Solving the Problem … 83
(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
84 A. Terki and H. Boubertakh
N
et = u it pit − Dt (16)
i=1
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.
Cuckoo Search Algorithm for Solving the Problem … 85
5 Simulation Results
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 …
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 ($) = 5.5402e + 05
A. Terki and H. Boubertakh
Cuckoo Search Algorithm for Solving the Problem … 91
Table 3 Comparison
Method Total cost ($)
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
92 A. Terki and H. Boubertakh
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
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.
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
The hybrid scheme with wind turbine and photovoltaic system, doubly fed induction
generator models proposed is presented in this section.
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.
96 T. Boutabba et al.
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 )
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)
Isc,n + K I T
Io =
Voc,n +K V T
(8)
e aVt
−1
K V = voltage coefficients
K I = current coefficients.
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
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)
where,
D = duty cycle.
In this section the strategy of controls of the photovoltaic system and the DFIG are
discussed.
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
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).
100 T. Boutabba et al.
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.
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
A Comparative Study of MPPT Techniques for Standalone Hybrid … 101
VSS VS S M B VB VVB
VSS VS S M B VB VVB
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.
102 T. Boutabba et al.
VSS VS S M B VB VVB
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)
A Comparative Study of MPPT Techniques for Standalone Hybrid … 103
6 Simulation Results
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
104 T. Boutabba et al.
DC/DC Gat
DC/AC
AC/DC DC/DC
Superviseur S1
S2
Batterie
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)
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.
A Comparative Study of MPPT Techniques for Standalone Hybrid … 105
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)
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)
1000 12
800 10
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
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)
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.
A Comparative Study of MPPT Techniques for Standalone Hybrid … 107
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
108 T. Boutabba et al.
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
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.
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.
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.
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).
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. }
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.
Adapted Search Equations of Artificial Bee Colony … 113
- Pseudo code A
For c=1: parlimit
Generate j randomly between [1… NBFeatures];
Cycle =Max
Iter Best
More
Enhanced Employed Bee
feature
Pseudo code B
For c=1: parlimit
Generate j randomly between [1.. NBFeatures];
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() {
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
Generate j randomly between [1… NBFeatures];
(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:
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
Generate j randomly between [1… NBFeatures];
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
Adapted Search Equations of Artificial Bee Colony … 117
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 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.
118 H. Djellali et al.
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
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
Adapted Search Equations of Artificial Bee Colony … 119
Table 5 Performance of
FS methods Train SVM Test SVM # features
Proposed and Existing FS
accuracy (%) accuracy
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
FS methods Train SVM Test SVM # features
proposed methods on colon
accuracy (%) accuracy
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
120 H. Djellali et al.
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
Adapted Search Equations of Artificial Bee Colony … 121
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
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.
K. Menighed (B)
Department of Petrochemicals and Process Engineering, University of 20 août 1955, Skikda,
Algeria
e-mail: [email protected]; [email protected]
A. Chemori
LIRMM, University of Montpellier, CNRS, Montpellier, France
e-mail: [email protected]
B. Boussaid
MACS-LR16ER22, University of Gabes, Gabès, Tunisia
e-mail: [email protected]
J. J. Yamé
Université de Lorraine, CNRS, CRAN, F-54000, Nancy, France
e-mail: [email protected]
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].
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
126 K. Menighed et al.
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
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
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 .
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;
128 K. Menighed et al.
A. Interaction Prediction
Under Assumptions 1, at step k, the predictions of the interaction vectors are given
by
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)
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
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
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.
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
Networked Cooperation-Based Distributed Model … 131
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)
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
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 ).
132 K. Menighed et al.
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)
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
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 .
Networked Cooperation-Based Distributed Model … 133
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
−
i (l) x̂ i (k|k) −
i (l)u i (k − 1) − i (l) − v̂i (k + l|k − 1))) (21)
with
Upon obtaining the optimal parameter vector ηi , the control increment at time k
is as follows
134 K. Menighed et al.
⎡ ⎤
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
Under Assumptions 1, at step k, the predictions of the new interaction vectors are
given by
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).
Networked Cooperation-Based Distributed Model … 135
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,
136 K. Menighed et al.
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
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
138 K. Menighed et al.
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
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.
1 Introduction
L. Laggoun
Department of Industrial Engineering, University Abbas Laghrour, LSTEB Laboratory,
Khenchela, Batna, Algeria
e-mail: [email protected]
B. Kiyyour (B)
Department of Electrical Engineering, University of Biskra, Biskra, Algeria
e-mail: [email protected]
G. Boukhalfa · S. Belkacem · S. Benaggoune
Department of Electrical Engineering, University of Batna2, LEB Laboratory, Batna, Algeria
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.
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:
For studying the dynamic behavior, the following equation of motion was added:
dΩ
J = Te − Tr − f r Ω (4)
dt
142 L. Laggoun et al.
Vsc1
isc1
Vsc2
S c1 isc2
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.
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.
Direct Torque Control Using Fuzzy Second Order Sliding Mode … 143
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:
144 L. Laggoun et al.
Ψ s1 = Ψ ds1
2
+ Ψ qs1
2
Ψ s2 = Ψ ds2
2
+ Ψ qs2
2
(7)
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.
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)
Direct Torque Control Using Fuzzy Second Order Sliding Mode … 145
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:
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]
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
146 L. Laggoun et al.
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.
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:
Direct Torque Control Using Fuzzy Second Order Sliding Mode … 147
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.
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.
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
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.
Direct Torque Control Using Fuzzy Second Order Sliding Mode … 149
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
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
150 L. Laggoun et al.
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)
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
Direct Torque Control Using Fuzzy Second Order Sliding Mode … 151
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
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
152 L. Laggoun et al.
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
Direct Torque Control Using Fuzzy Second Order Sliding Mode … 153
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
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).
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.
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
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].
158 C. Boussalem et al.
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)
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].
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
160 C. Boussalem et al.
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
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)
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
Fractional Order Integral Controller Design Based … 161
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
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
k I = ωcγ (19)
Fractional Order Integral Controller Design Based … 163
H ref
0.14
H: γ = 1.1
H: γ = 1.33
0.12 H: γ = 1.5
H: γ = 1.66
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
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.
164 C. Boussalem et al.
• 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).
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
Fractional Order Integral Controller Design Based … 165
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
166 C. Boussalem et al.
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
10
0
0 50 100 150 200 250 300 350 400 450
Time(s)
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
Fractional Order Integral Controller Design Based … 167
15
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,
168 C. Boussalem et al.
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
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
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.
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
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.
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
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.
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.
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.
vcem = K t , And, T = K i
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
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)
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
Modeling of Torsional Vibrations Dynamic … 179
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.
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.
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.
Modeling of Torsional Vibrations Dynamic … 181
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
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).
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
Modeling of Torsional Vibrations Dynamic … 185
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
1 Introduction
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.
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:
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:
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,
m−1
r = g(xm ) − c j g x j = g(xm ) − K C,
j=0
⎡ ⎤
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 = K T −1 (10)
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)
where δi (t) is the i th generator rotor angle and δj (t) is the j th generator rotor angle.
192 Z. Jlassi et al.
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.
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
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.
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.
194 Z. Jlassi et al.
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
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
196 Z. Jlassi et al.
(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)
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
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.
Power System Generator Coherency Identification … 197
(a)
Mode of frequency 0.830 Hz Mode of frequency 0.774 Hz Mode of frequency 0.721 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 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 0.835 Hz Mode of frequency 0.782 Hz Mode of frequency 0.728 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 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(c)
Mode of frequency 0.829 Hz Mode of frequency 0.776 Hz Mode of frequency 0.724 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 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
198 Z. Jlassi et al.
is captured. Thus, for this test fault location, the power system is decomposed into
two coherent zones instead of three zones.
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)
Mode of frequency 0.721 Hz Mode of frequency 0.658 Hz
Mode of frequency 0.772 Hz
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
Amplitude Amplitude Amplitude
(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)
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
Power System Generator Coherency Identification … 199
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.
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.
200 Z. Jlassi et al.
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
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
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
202 Z. Jlassi et al.
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
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 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.
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
S2 UC S4
CONTROL
S2 2 UC S4
CONTROL
S2 UC S4
3
CONTROL
206 A. E. Toubal Maamar et al.
S1 4 S3
+ R L C
UE - i
S2 UC S4
CONTROL
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:
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]
208 A. E. Toubal Maamar et al.
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
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
Analysis and Experimental Validation … 209
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
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
3 Experimental Results
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%.
210 A. E. Toubal Maamar et al.
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.
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
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
60
RMS Voltage [v]
50
40
30
20
10
0
2 3 4 5 6 7 8
Frequency [KHz]
References
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.
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
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.
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
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.
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
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
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
T I T = In
−1 I
TI = TIT T . (7)
E A T = 0, E B = 0, E C T = 0. (8)
C. Contraction
(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
(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)
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 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)
(21)
(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 conditions of contraction are verified in [10, 11, 15], and its contracted
controller is found as:
p V p .
K = QpK (26)
N
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
226 M. Z. Doghmane and M. Kidouche
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
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
6 Simulation Results
• 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
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
Optimal Decentralized State Control of Multi-machine … 231
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.
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
dω
2H = (Pm − Pe )/ω (1)
dt
dω
θ= (2)
dt
Pm = Pr e f + D p (ωn − ω) (3)
1
Em = (Dq (Ur e f − Um ) + Q r e f − Q e ) (5)
Ks
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,
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].
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:
238 R. Aouini et al.
where |V0 | and |Icc | are the pre-fault bus voltage magnitude and fault current,
respectively.
– 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 ).
5 Simulation Results
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.
240 R. Aouini et al.
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
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%.
Large Synchronverter Integration in Power … 241
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
242 R. Aouini et al.
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
with a line length L = 300 km
Length 300 km
Cases CCT (s) I cc (p.u) E th (p.u) Z th (p.u) SCR
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
Table 3 CCTs, Current default Icc , Eth , Zth and the SCRs at POI for different ratio of SV integration
with a line length L = 200 km
Length 200 km
Cases CCT (s) I cc (p.u) E th (p.u) Z th (p.u) SCR
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
Large Synchronverter Integration in Power … 243
Table 4 CCTs, Current default Icc , Eth , Zth and the SCRs at POI for different ratio of SV integration
with a line length L = 50 km
Length 50 km
Cases CCT (s) I cc (p.u) E th (p.u) Z th (p.u) SCR
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
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
V L .I cc
E th = (17)
I cc − I L
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
Large Synchronverter Integration in Power … 245
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;
246 R. Aouini et al.
– 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
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.
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
e-mail: [email protected]
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
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 2 The desired trajectory x1d and its derivatives ẋ1d and ẍ1d are known,
continuous and bounded.
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
i ∂βi
θ̂˙i = −
xi
βi = i ϕi (χ )dχ + εi (6)
0
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.
1 T −1
n
W =
θ θi (9)
2 i=1 i i
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
∞
n
T 2
ϕi
θi dτ ≤ W (0) − W (∞) < ∞ (11)
0 i=1
Immersion and Invariance Based Adaptive Dynamic Surface Control … 251
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:
Step i(i = 2, · · · , n − 1): Define the ith surface error Si = xi − xid , and it’s time
derivative is defined 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:
Step n: Define the nth surface error Sn = xn − xnd , and its time derivative is
defined as:
252 Y. Soukkou et al.
1 T
u= −ϕn θ n + βn − ψn + ẋnd − K n Sn (21)
gn
where, K n > 0.
4 Stability Analysis
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:
For i = 2, · · · , n − 1:
For i = n:
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
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
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:
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
254 Y. Soukkou et al.
(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
1 2
y2 ẏ2 + y ≤ B2 (.)|y2 | (40)
τ2 2
1 2 y22 B 2 (.)
y2 ẏ2 ≤ − y2 + + 2 (41)
τ2 2 2
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
Immersion and Invariance Based Adaptive Dynamic Surface Control … 255
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
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,
the adaptive laws (4). Then, we choose the adaptive laws for θ i as:
∂βi
i
θ̂˙i = −
where, σθi are small design positive constants. The estimation dynamics are given
by:
256 Y. Soukkou et al.
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)
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
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.
5 Simulation Results
D q̈ + B q̇ + N sin(q) = I
M I˙ + H I + K m q̇ = V (57)
ẋ1 = x2
ẋ2 = f 2 (x̄2 ) + b2 x3
ẋ3 = f 3 (x) + b3 u (58)
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
u= x2 θ 3 + β3 + x3 + ẋ3d − K 3 S3 (61)
b3 M
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
Immersion and Invariance Based Adaptive Dynamic Surface Control … 259
2
x1d
1.5
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]
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]
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]
25
20 u
Input control voltage [V]
15
10
5
0
-5
-10
-15
-20
-25
0 5 10 15 20 25
Time [sec]
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.
Immersion and Invariance Based Adaptive Dynamic Surface Control … 261
References
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.
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
e-mail: [email protected]
K. Guesmi (B)
CReSTIC, IUT de Troyes, 09 rue de Quebec, 10026 Troyes, France
e-mail: [email protected]
M. Bougrine
LACoSERE Lab, Amar Telidji, University of Laghouat, Laghouat, Algeria
e-mail: [email protected]
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
||u(t)|| ≤ u m (5)
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:
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
l=1 i=1 μ Fil (x i )
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.
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)
1 2
V1 = e (15)
2 1
The derivative of V1 is
Then, the virtual control variable α1 and the derivative of e1 are given by
α1 = −k1 e1 + ẏd
ė1 = e2 − k1 e1 (17)
e2 = x2 − α1
e3 = x3 − α2 (18)
1 2
V2 = e1 + e22 (19)
2
The derivative of V2 is
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
ei = xi − αi−1
ei+1 = xi+1 − αi (22)
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
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
Design of an Adaptive Fuzzy Backstepping Synergetic Control … 269
1
u= −c(en − en−1 − kn−1 en−1 ) − f (x) − d̄ + α̇n−1
g(x)
1
−en−1 − σ (27)
τ
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)
θ̇ f = γ1 σ φ f (x) (32)
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
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
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)
g(x) − ĝ x, θg u = θ̃gT φg (x)u + ε2 (47)
d̄ − d̄ˆ θd̄ = θ̃d̄T φd̄ + ε3 (48)
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.
n−1
1 2
V̇ = − kl el2 + en−1 en − σ en−1 − σ + σε (50)
l=1
τ
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
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.
Design of an Adaptive Fuzzy Backstepping Synergetic Control … 273
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)
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)
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)
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
Design of an Adaptive Fuzzy Backstepping Synergetic Control … 277
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.
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
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
• 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 :
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
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
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:
Improvement of the Stability Performance of a Quad-Copter … 283
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ . ⎤ ⎡ . ⎤
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
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
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
Improvement of the Stability Performance of a Quad-Copter … 285
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:
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)
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.
Improvement of the Stability Performance of a Quad-Copter … 287
Ɵme(s)
Ɵme(s) Ɵme(s)
Ɵme(s) Ɵme(s)
Fig. 7 The influence of wind with a drag force, F_(trn) = 2 N (ANFIS command)
288 D. Baladji et al.
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
Improvement of the Stability Performance of a Quad-Copter … 289
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
αr e f t 5
α= 5
t5 t f − t
with: αr e f = 4
t5
ρ = hr 5
t5 t f − t
290 D. Baladji et al.
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.
Improvement of the Stability Performance of a Quad-Copter … 291
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
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.
1 Introduction
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
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
296 A. M. Taberkit et al.
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]
The Importance of Applying Artificial Intelligence … 297
discover the existence of many apparent clusters, then classify each data into the
corresponding one [38].
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
The Importance of Applying Artificial Intelligence … 299
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
300 A. M. Taberkit et al.
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].
The Importance of Applying Artificial Intelligence … 301
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
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
304 A. M. Taberkit et al.
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.
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].
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
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:
ż = w cos(θ ) − u sin(θ )
(3)
θ̇ = q
ze = zd − z (4)
308 M. Nasri et al.
• (x G , yG , z G ) = (x B , y B , z B ) = 0
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)
where
⎡ ⎤ ⎡ ⎤
0 u(t) 0 0
A(t) = ⎣ 0 0 1 ⎦, B(t) = ⎣ 0 ⎦
0 0 a(t) b(t)
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 à = 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.
Fuzzy H∞ Delay-Independent Stabilization of Depth Control … 309
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,
V1 −V1min
F11 = V1max −V1min
, F12 = 1 − F11
V2 −V2min (11)
F21 = V2max −V2min
, F22 = 1 − F21
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
Fuzzy H∞ Delay-Independent Stabilization of Depth Control … 311
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
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
r
U (t) = μi (ς (t))K i x(t) (16)
i=1
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
T T
ψ (τ )ψ(τ )dτ < γ
T 2
ϕ T (τ )ϕ(τ )dτ (18)
0 0
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
⎡ 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
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
In compact form
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
⎡ 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
r −1
Q̇ T ≥ φk (Q k − Q r ) (32)
k=1
Fuzzy H∞ Delay-Independent Stabilization of Depth Control … 315
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].
5 Simulation Results
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.
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.
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
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
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.
1 Introduction
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.
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.
⎧
⎨ 1, μ
ji ≥ dt h
t ij = Fi (xi )
(2)
⎩ 0, μ
ji < dt h
Fi (xi )
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
−
Petri Type 2 Fuzzy Neural Networks (PT2FNN) for Identification … 323
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)
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
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.
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
∂ 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:
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.
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
Petri Type 2 Fuzzy Neural Networks (PT2FNN) for Identification … 327
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:
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.
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
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.
1 Introduction
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
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.
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)
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.
Design Fractional Order PI Controller with Decoupler … 335
n t
α 1 d f (τ )
t0 Dt f (t) = dτ (8)
Γ (n − α) dt (t − τ )α−n+1
t0
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)
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
+∞
h(t) = e−ξ t μ(ξ )dξ (13)
0
+∞
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)
∂ 1
H oú H ( p) = a for 0 < a < 1 (16)
∂t p
sin(απ)
μ(ξ ) = ξ −α ξ >0
π
Design Fractional Order PI Controller with Decoupler … 337
+∞ v(ξ )
C(s) = ∫ dξ (18)
−∞ p+ξ
1 sin(απ) −α
v(ξ ) = K P δ(ξ ) + ξ (19)
Ti π
1
u(t) = K p e(t) + D −a e(t) (20)
Ti
where:
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)
1 KI
C(s)classical = K P 1 + = KP + = K P + K I s −1 (22)
TI s s
338 S. Laifa et al.
Fig. 1 Schematic diagram of the decentralized fractional P I a controller designed with decoupler
for a pilot plant batch distillation column
Design Fractional Order PI Controller with Decoupler … 339
+∞ v1 (ξ )
F O − P I1 = ∫ dξ (27)
−∞ p+ξ
+∞ v2 (ξ )
F O − P I2 = ∫ dξ (28)
−∞ p+ξ
∞
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
K ii
Q iiFOPDT (s) = e−L ii s (31)
Tii s + 1
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
1 Tii
K pii = (34)
K ii Tcii + L ii
Design Fractional Order PI Controller with Decoupler … 341
For more details of NDT and SIMC design methods (see in [10, 11]).
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
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
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
Design Fractional Order PI Controller with Decoupler … 343
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
-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.
6 Conclusion
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”
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,
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
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
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 .
(P1+P2)>PL
Mode 1 Mode 2
Gridon Gridoff
= =
(P1+P2)<PL
Mode 3
Unitentional
islanding
=
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.
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.
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)
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
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 the charging battery process “buck mode”: the control law corresponding this
mode is:
1
u bat1 = (1 − sign(s)) (13)
2
Direct Sliding Mode Control of Transient Power in Microgrid … 357
• 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].
4 Simulation Results
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.
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
Direct Sliding Mode Control of Transient Power in Microgrid … 359
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
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.
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
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
Vg
Vg PLL
reference
synthese
Voltage
(Eq 5)
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.
New Virtual Synchronous Generator Control Technique … 365
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
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:
⎧
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)
366 Y. Daili and A. Harrag
Ki
Vm = ( Q in − Q out ) (12)
s
From (8) to (10), the transfer functions of active and reactive powers have the
following expressions:
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.
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
New Virtual Synchronous Generator Control Technique … 367
Measured power
Reference power
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
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.
4 Simulation Results
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
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.
370 Y. Daili and A. Harrag
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
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
Time [Sec] Time [Sec]
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
Time [Sec] Time [Sec]
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.
New Virtual Synchronous Generator Control Technique … 371
Acknowledgment The Algerian Ministry of Higher Education and Scientific Research via the
DGRSDT supported this research (PRFU Project code: A01L07UN190120180005).
References
1 Introduction
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
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].
Guides Liner
Rs L1 I
W
Wire spool
Contact tube
ls
Voc CT
larc Varc
Workpiece
376 N. Hamouda et al.
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.
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
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 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
Design and Analysis of Robust Nonlinear Synergetic Controller … 377
k1 IW (s)
Vf (s)
Voc(s)
_ _
+ + Vp(s) 1 1
Ru(s) k0 Km larc(s)
+ + L1s+Rs + s
Rarc
Earc
them according to their physical relationship, one can obtain the configuration of the
control unit as shown in the block diagram of Fig. 2.
D1 D3 S1 S3
PMDC motor
Ra La
Vac Va A B
C
D2 D4 Ea
S2 S4
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].
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]
Design and Analysis of Robust Nonlinear Synergetic Controller … 379
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]
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
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)
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
380 N. Hamouda et al.
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)
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
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
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)
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
382 N. Hamouda et al.
5 Simulation Results
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.
300
250
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
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
Wire feed rate (IPM)
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
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. 6 Wire feed speed, armature current and angular displacement tracking responses of PMDC
motor in test 2
Design and Analysis of Robust Nonlinear Synergetic Controller … 385
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
Wire feed rate (IPM)
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
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. 7 Wire feed speed, armature current and angular displacement tracking responses of PMDC
motor in test 3
386 N. Hamouda et al.
350
300
50
0
0 0.05 0.1 0.15 0.2
Time (s)
demonstrates the high tracking performance and the strong robustness of the proposed
NSC control strategy against the model uncertainties.
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
Design and Analysis of Robust Nonlinear Synergetic Controller … 387
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
1 Introduction
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.
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
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.
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
A1
A2
q
B2
B1
392 E. Amirouche et al.
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.
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
L s0 = L sl + m (9)
Similarly, the mutual inductance between the phases can be expressed as:
Simulation Study of the Dual Star Permanent Magnet Synchronous … 393
∞
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.
dϕabc
[e] = − (15)
dt
Or
ei = vi − ri i (16)
394 E. Amirouche et al.
Pe
te = (17)
ω
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)
X dq = T1 X abc (20)
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)
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.
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).
396 E. Amirouche et al.
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
Starting from the instantaneous output power, the electromagnetic power as well as
torque can be derived.
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)
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].
Simulation Study of the Dual Star Permanent Magnet Synchronous … 397
⎡ ⎤
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
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].
398 E. Amirouche et al.
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
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)
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].
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
400 E. Amirouche et al.
⎡ ⎤
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)
7 Simulation Results
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,
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. 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.
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.
404 E. Amirouche et al.
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.
Simulation Study of the Dual Star Permanent Magnet Synchronous … 405
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
1 Introduction
Fig. 1 Synchronous
machine conventional
structure, with one damper
winding for the d- axis and
one damper winding for the
q-axis
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.
410 K.-E. Sari and B. Sari
Therefore:
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
Modeling and Experimental Identification of Salient-Pole … 411
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
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]:
412 K.-E. Sari and B. Sari
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
2
L f l = L rf l
3K 2f
2
L Dl = L rDl
3K D2
Modeling and Experimental Identification of Salient-Pole … 413
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).
Fig. 4 Open-circuit
characteristic
Modeling and Experimental Identification of Salient-Pole … 415
VL−L
Z d(unsat) = √ . (14)
3Isc
where:
• Rs : The stator resistance, and it is measured directly.
Fig. 5 Short-circuit
characteristic
416 K.-E. Sari and B. Sari
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
Finally, X q is obtained:
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]:
−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.
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.
Modeling and Experimental Identification of Salient-Pole … 419
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
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 .
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
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:
422 K.-E. Sari and B. Sari
⎡ ⎤
−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 .
6 Conclusion
References
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
1 Introduction
A. Ammar (B)
Signals and Systems Laboratory (LSS), Institute of Electrical and Electronic Engineering,
University of M’hamed BOUGARA of Boumerdes, Boumerdes, Algeria
e-mail: [email protected]
LGEB Laboratory, Electrical Engineering Department, Biskra University, Biskra, Algeria
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.
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
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).
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
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
Power Quality Improvement of PWM Rectifier-Inverter System … 431
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)
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 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.
To predict currents into the next step the Euler forward discretization equation is
used:
dx x(k + 1) − x(k)
≈ (11)
dt Tz
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.
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:
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.
Power Quality Improvement of PWM Rectifier-Inverter System … 435
Fig. 3 Diagram of predictive power/torque control for an induction motor drive using back to back
converter
5 Simulation Results
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
Q[VAR]
Q[VAR]
Q[VAR]
2000
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
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
100 100
100
90
90 90
Mag (% of Fundamental)
Mag (% of Fundamental)
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
Power Quality Improvement of PWM Rectifier-Inverter System … 437
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
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
Times [s] Times [s] Times [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
Motor-side-converter:
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
Power Quality Improvement of PWM Rectifier-Inverter System … 439
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
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.
M. Zareb (B)
Electronic Department, USTO-MB, University of Mascara, Mascara, Algeria
e-mail: [email protected]
B. Bakhti
Physics Department, University of Osnabrueck, University of Mascara, Osnabrueck, Germany
e-mail: [email protected]
Y. Bouzid
CSCS Laboratory, Ecole Militaire Polytechnique, Bordi El Bahri, Algeria
e-mail: [email protected]
H. Kadourbenkada
Electronic Department, University of Mascara, Mascara, Algeria
e-mail: [email protected]
K. Bouzgou · W. Nouibat
Electronic Department, LEPESA Laboratory, USTO-MB University, Oran, Algeria
e-mail: [email protected]
W. Nouibat
e-mail: [email protected]
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.
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
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
Novel Smart Air Quality Monitoring System … 445
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].
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
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
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]
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
U PID
+ e(t) Quadrotor
Trajectory PID Controller System
x, y, z, ϕ, θ and ψ
t
de(t)
UPID = K p .e(t) + K i e(τ ).dτ + K d . (1)
dt
0
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.
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:
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
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
Novel Smart Air Quality Monitoring System … 451
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.
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
60
50
40
30
0 500 1000 1500 2000 2500 3000
TIME [SEC]
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
Novel Smart Air Quality Monitoring System … 453
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.
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.
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
e-mail: [email protected]
N. Slimane (B)
Advanced Electronics Laboratory, Faculty of Technology, Electronics Department, Batna 2
University, Batna, Algeria
e-mail: [email protected]
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.
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).
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.
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).
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)
ṡ(t) = k p ė + ki e + kd ë (6)
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)
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.
GA Based Adaptive Fuzzy Sliding Mode for 3 DOF Spatial Arm … 459
Fig. 2 Diagram of adaptive fuzzy sliding mode (AFSMC) with surface PID
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).
460 R. Belloumi and N. Slimane
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)
4 Simulation Results
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 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
GA Based Adaptive Fuzzy Sliding Mode for 3 DOF Spatial Arm … 463
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
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.
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
e-mail: [email protected]
N. Zerari
Signal Processing Laboratory SP-Lab, Faculty of Sciences and Technology, Department of
Electronics, Constantine1 University, Constantine, Algeria
e-mail: [email protected]
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.
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:
⎡ ⎤ ⎡ ⎤
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.
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.
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
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)
Robust Adaptive Fuzzy Approach with Unknown Control … 469
g(x) − ĝ x, w∗g = w̃gT ξg (x) + Φg (14)
With
w̃ f = w f − w∗f (15)
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
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 +
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 +
ζ̇ = e T Pbu ns + α e T Pb2 (26)
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)
Robust Adaptive Fuzzy Approach with Unknown Control … 471
Φ̂˙ g = γg e T Pbu a − α N (ζ )e T Pb (32)
Where Φ f¯ = Φ f + d
⎡ ⎤
0 1 0 ··· 0
⎢ 0 0 1 ⎥
0 0
⎢ ⎥
⎢ .. .. .. ⎥
.. ..
Ak = ⎢ . . . ⎥ . .
⎢ ⎥
⎣ 0 0 ··· 0 1 ⎦
−kn −kn−1 · · · · · · −k1
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]
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
472 O. Bourebia and N. Zerari
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).
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],
..
.
Rule16 : if θ is PB and θ̇ is PB, Then y = w16
0.3
0.2
rad
0.1
-0.1
-0.2
-0.3
0 5 10 15 20 25 30 35
Temps (s)
0.1
0
(rad)
-0.1
-0.2
-0.3
-0.4
-0.5
0 5 10 15 20 25 30 35
temps (s)
La commande u
30
25
20
15
10
Force (N)
-5
-10
-15
-20
0 5 10 15 20 25 30
temps (s)
4 Conclusion
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)
References
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.
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
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
The dynamic of the process of CSTR considered in this paper can be written in the
following form [18, 19]:
480 Z. Nassira and C. Mohamed
⎧
⎪
⎪ 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
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.
where κ = 0.2785
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.
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 ].
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 .
482 Z. Nassira and C. Mohamed
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)
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:
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:
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.
Adaptive Neural-Network Control Design for Uncertain CSTR … 483
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:
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)
484 Z. Nassira and C. Mohamed
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
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).
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
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
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
W̃1 ≤ 2( V1 |t=0 + θ1 /β1 )/λmin Γ1−1
|τ̃1 | ≤ γ1 2( V1 |t=0 + θ1 /β1 ) (39)
488 Z. Nassira and C. Mohamed
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
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)
20
v(t) u(v(t))
10
-10
-20
-30
-40
0 5 10 15 20 25 30
Time (sec)
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
490 Z. Nassira and C. Mohamed
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)
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
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.
1 Introduction
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.
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
f k (x) = a k (2)
n k 1 mi
with μk (x) = μ
i=1 k, F
i ∈ Fi , . . . , Fi .
Fi
By introducing the concept of fuzzy basis functions [1], the output of fuzzy system
TS0 can be written as:
496 A. Bounemeur et al.
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)
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
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
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
498 A. Bounemeur et al.
⎧
⎪
⎨ 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) = fˆa 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) = fˆa x, θ ∗f a + ε f a (x) (16)
g(x) = ĝ x, θg∗ + εg (x) (17)
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.
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].
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 .
500 A. Bounemeur et al.
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) − fˆa (x, θ ) + y (n) + k T e + α ĝ(x, θ )N (τ )G (24)
ε0 + ĝ 2 (x, θ ) d
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)
·
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)
Adaptive Fuzzy Fault-Tolerant Control Using Nussbaum … 501
2
τ̇ = e T P Bu r b + αe T P B (36)
δ̇ = −σ0 δ (37)
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)
using (21)
becomes:
e(n) = yd(n) − f (x) − fˆ(x, θ )
− f a (x) − fˆa (x, θ ) − g(x) − ĝ(x, θ ) u c
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 ) − fˆa (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) − fˆa (x, θ ) − g(x)u r
ˆ (44)
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)
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 − 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)
⎡ ⎤
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
Adaptive Fuzzy Fault-Tolerant Control Using Nussbaum … 503
A T P + P A = −Q (49)
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
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
1
V̇ = − e T Qe + e T P B −w Tf (x) θ f − ε f (x) − w Tf a (x)
θ f a − ε f a (x) − wgT (x)
θg (u c − α N (τ )G)
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
V̇1 = 0 (57)
ηf
1 1 1
ε f a − ε f a ε̂˙ f a − ε g − ε g ε̂˙ g + δ̇δ
− (59)
ηf ηg σ0
+ ε g e T P B(u c − α N (τ )G) − δ 2
(60)
Adaptive Fuzzy Fault-Tolerant Control Using Nussbaum … 505
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
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
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
506 A. Bounemeur et al.
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,
4 Simulation Results
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)
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
Adaptive Fuzzy Fault-Tolerant Control Using Nussbaum … 507
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)
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)
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.
Adaptive Fuzzy Fault-Tolerant Control Using Nussbaum … 509
References
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
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.
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
516 S. Settoul et al.
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
Optimal Integration of Renewable Distributed Generation … 517
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]:
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:
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:
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)
518 S. Settoul et al.
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
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)
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)
520 S. Settoul et al.
−
→ −
→ −
→− →
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)
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
−
→ −
→ −
→− →
X (t + 1) = X rand − A . D (29)
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 *
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.
Optimal Integration of Renewable Distributed Generation … 523
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
524 S. Settoul et al.
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.
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
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.
528 S. Settoul et al.
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
Optimal Integration of Renewable Distributed Generation … 529
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.
530 S. Settoul et al.
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
532 S. Settoul et al.
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
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.
1 Introduction
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
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.
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.
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. 5 STATCOM
connected to a transmission
line [10]
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:
538 I. Boucetta et al.
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.
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
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).
540 I. Boucetta et al.
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.
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
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.
544 I. Boucetta et al.
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
• 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.
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
546 I. Boucetta et al.
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 Introduction
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].
• 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.
dV
T cosαT − D − wsinθ = m (2)
dt
Applying (1) in ⊥ direction to flight path
V2
F⊥ = ma⊥ = m
rc
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].
552 A. Lounis et al.
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
−1 ≤ Q(t) ≤ +1
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
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
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)
−1
p0k+1 = p0k − J p0k , h k .G p0k
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 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
q
∃ν1 , ν2 , . . . . . . . . . ..νq ∈ R/ p t f = νi ∇x Fi (x(t f )) + + p 0 ∇x φ t f , x f
i=1
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
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).
⎧
⎪
⎪
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
• 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).
Minimization of the Energy Consumption of an Aircraft 559
7 Conclusion
References
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.
1 Introduction
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. 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)”.
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].
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].
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),
Study to Improve the Technical Parameters … 569
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
Fig. 5 The 30 kV line from the photovoltaic power plant to El Aouinet substation
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
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
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.
574 B. Badache et al.
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:
Study to Improve the Technical Parameters … 575
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)
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
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)
578 B. Badache et al.
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
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.
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
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
Rt
λ= (4)
V
582 A. Abdellah et al.
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)
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:
Second-Order Super-Twisting Control of an Autonomous … 583
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]:
Fig. 2 Generator side control scheme of the direct drive wind turbine
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].
K (t) ≥ h m + η (22)
I f smc = F S MC (S (t),
S (t)) (24)
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)
Fig. 3 Membership
1 N Z P
functions of the inputs (S(t),
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
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
Second-Order Super-Twisting Control of an Autonomous … 587
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]:
Iqr ST = I1 + I2 (33)
588 A. Abdellah et al.
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.
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
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)
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)
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)
10
8
Mag
0
0 100 200 300 400 500
Frequency (Hz)
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
592 A. Abdellah et al.
10
8
Mag
0
0 100 200 300 400 500
Frequency (Hz)
10
8
Mag
0
0 100 200 300 400 500
Frequency (Hz)
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)
0
Iq AFSMC
Iq PI
Current Iq (A) -5 Iq STSMC
Iq*
-15
0 0.5 1 1.5 2
Time (s)
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
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
594 A. Abdellah et al.
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
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
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.
2 System Description
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].
Electrical
grid
IM
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
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
Torque equation
3 pM
Te = (φrd i sq − φrq i sd ) (9)
2L r
600 S. Khaled and G. Bachir
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
Open-Switch Faults Based on Six/Five-Leg Reconfigurable … 601
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.
602 S. Khaled and G. Bachir
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:
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.
Open-Switch Faults Based on Six/Five-Leg Reconfigurable … 603
Fig. 5 a Principle of PWM module for one converter; b for 5-leg converter when B3 legs are shared
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].
604 S. Khaled and G. Bachir
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
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).
6 Simulation Results
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.
606 S. Khaled and G. Bachir
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
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
608 S. Khaled and G. Bachir
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)
Open-Switch Faults Based on Six/Five-Leg Reconfigurable … 609
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)
610 S. Khaled and G. Bachir
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)
Open-Switch Faults Based on Six/Five-Leg Reconfigurable … 611
(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.
612 S. Khaled and G. Bachir
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
Open-Switch Faults Based on Six/Five-Leg Reconfigurable … 613
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
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.
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
e-mail: [email protected]
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
e-mail: [email protected]
M. Chadli
IBISC Lab.-UEVE, University of Paris-Saclay, 40 rue Pelvoux, Evry, France
e-mail: [email protected]
iL
L D
vin sw C
vC R vs
PMW d (t )
u b (t ) v (t )
d (t )
T on T off
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
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
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):
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:
where sat(u(k)) is the saturated control input u(k) of the boost converter, constrained
between 0 and 1 and is defined as:
A Robust Model Predictive Control for a DC-DC … 619
⎧
⎨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
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
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:
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]:
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)
φ T φ ≤ εδ T δ (19)
In which, Q(x) = Q(x)T , R(x) = R(x)T and S(x) are affine functions of x, and
is equivalent to
A Robust Model Predictive Control for a DC-DC … 621
or, equivalently
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
In this context, we consider that N = 2 and the parameter vector p∈[1/R] where:
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.
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
If the resulting closed-loop system for (24) is stable, x(∞|k) must be zero and
result in
A Robust Model Predictive Control for a DC-DC … 623
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)
[(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.
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 ) + (1 + α2−1 )φ̃(k + i|k )T R φ̃(k + i|k )
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
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
Y T G −1 − δmax I ≤ 0 (40)
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
A Robust Model Predictive Control for a DC-DC … 627
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.
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
A Robust Model Predictive Control for a DC-DC … 629
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
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.
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
Converter
DC-DC-
Energy Management System
Power
Desired
DC-DC-Bus
DC - AC
Inverter
Reference
Converter
DC-DC-
inputs
PV Source Vehicle
PMSM
SOC Battery
Energy Management Strategy for Hybrid Vehicle was explained. Before finishing,
the simulation results were showed. Finally, discussions and conclusions were given
(Fig. 1).
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].
Strategy for Optimization of Energy … 635
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
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].
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.
636 S. Belhadj et al.
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. 5 Membership
functions of fuzzy logic
controller
Power Demand
FC Reference 200
Current
Strategy for Optimization of Energy … 637
Reference Speed
125
124.95
Time [S]
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.
638 S. Belhadj et al.
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.
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
Strategy for Optimization of Energy … 639
Reference Speed
Vehicule Speed [Km/h]
50
0
0 200 400 600 800 1000 1200 1400
Time [S]
4
10
2
Required Power [W]
-2
-4
Time [S]
4
10
3
FC's Power [W]
0
0 200 400 600 800 1000 1200
Time [S]
0
0 200 400 600 800 1000 1200 1400
Time [S]
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,
642 S. Belhadj et al.
4
10
-2
-4
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
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.
1 Introduction
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
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
648 M. Zaouia et al.
−
→ 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)
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
Modeling and Analysis of an Electromagnetic Vibration … 649
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
¨
Mi j = υ ∇.α j dr dz
i ∇.α (9)
r
Ω
¨
Fi = (αi Js ) dr dz (10)
Ω
650 M. Zaouia et al.
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
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
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
652 M. Zaouia et al.
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
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
654 M. Zaouia et al.
cycle of vibration. Figure 9 depicts the magnetic force versus time corresponding to
the imposed cycle of vibration.
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. 12 Electromotive force versus time for different values of sprung mass
656 M. Zaouia et al.
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
Modeling and Analysis of an Electromagnetic Vibration … 657
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
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.
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].
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.
2 System 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,
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
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:
662 M. E.-A. Slimani et al.
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)
Experimental and Numerical Study of Hybrid PV/Thermal Solar … 663
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)
664 M. E.-A. Slimani et al.
4 Experimental Tests
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
666 M. E.-A. Slimani et al.
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)
600 12
G (W/m²)
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)
Experimental and Numerical Study of Hybrid PV/Thermal Solar … 667
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.
668 M. E.-A. Slimani et al.
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 (%)
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
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
670 M. E.-A. Slimani et al.
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.
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
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]:
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
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],
Maximum Power Point Tracking (MPPT) for a PV System … 675
1
vo = vi (7)
1− D
676 A. Djalab et al.
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.
Maximum Power Point Tracking (MPPT) for a PV System … 677
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
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
Return
10 0
10 -2
10 -4
0 20 40 60 80 100
119 Epochs
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).
150
120
110
50 100
90
80
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
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.
Maximum Power Point Tracking (MPPT) for a PV System … 681
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)
150
G = 1000 W/m2
100
80
G = 650 W/m2
60
50 G = 420 W/m2
40
20
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
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 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
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.
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
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.
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 )
Development of a Stand-Alone Connected PV System … 687
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
688 K. Rayane et al.
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
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:
5 Simulation Results
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
Fig. 8 PV voltage
Fig. 9 PV current
Development of a Stand-Alone Connected PV System … 693
Fig. 10 PV power
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
Development of a Stand-Alone Connected PV System … 695
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 multi- level 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
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.
1 Introduction
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.
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.
A. Experimental design
(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:
A Novel Hybrid Photovoltaic/Thermal Bi-Fluid … 701
• 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
702 M. E.-A. Slimani et al.
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
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)
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.
704 M. E.-A. Slimani et al.
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.
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
706 M. E.-A. Slimani et al.
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.
test day 80
60
40
20
0
06:00 08:00 10:00 12:00 14:00 16:00 18:00
Time (h)
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)
A Novel Hybrid Photovoltaic/Thermal Bi-Fluid … 707
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
708 M. E.-A. Slimani et al.
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
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.
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
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
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)
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
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
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].
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
and:
714 N. Hamouda and B. Babes
n
i=1 μ Fil (xi )
ξi (x) = L (9)
n
i=1 μ Fil (xi )
l=1
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
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
+
Vin _ D C R Vo
u
A DC/DC Buck Converter Voltage … 715
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
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)
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)
716 N. Hamouda and B. Babes
p ( p/q)−1 1
ë + α ė + β e ė = − σ (22)
q τ
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
where ξ f (x) and ξg (x) are the fuzzy basis functions vector, θ f and θg are the
parameters vector of fuzzy system.
A DC/DC Buck Converter Voltage … 717
∗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
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
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.
5 Simulation Results
40
Vref
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)
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)
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.
720 N. Hamouda and B. Babes
40
FTSC
AF-FTSC
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
A DC/DC Buck Converter Voltage … 721
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
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.
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
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.
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
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].
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 )
This formula is applicable for each node in both environments [7, 8].
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
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 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
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).
728 R. Abdelghani et al.
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.
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
Study of Electrical Field Distribution … 729
9
On the line passing through the walls
8 On the line passing through the center of the cavity
Case of a homogeneous insulation
7
Electric field E (KV/mm)
1
8 10 12 14 16 18 20
radial position ri (mm)
140
120
100
80
60
40
20
17.305 17.31 17.315 17.32 17.325 17.33 17.335 17.34
radial position ri (mm)
Fig. 5 Distribution of the electromechanical pressure under the influence of a cavity located
17.325 mm from the center of the conductive core
Study of Electrical Field Distribution … 731
700
On the line passing through the walls
On the line passing through the center of the cavity
600
Case of a homogeneous insulation
electrostatic pressure P (N/m 2)
500
400
300
200
100
0
8 10 12 14 16 18 20
radial position ri (mm)
450
On the line passing through the walls
400 On the line passing through the center of the cavity
Case of a homogeneous insulation
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
radial position ri (mm)
Fig. 8 Distribution of dielectric losses under the influence of a cavity located at 17.325 mm center
of the conductive core
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
Study of Electrical Field Distribution … 733
1600
On the line passing through the walls
1400 On the line passing through the center of the cavity
Case of homogeneous insulation
dielectric losses Pdiél (W/m 3)
1200
1000
800
600
400
200
0
8 10 12 14 16 18 20
radial position ri (mm)
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.
734 R. Abdelghani et al.
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
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.
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.
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.
The mechanical power extracted from the wind can be expressed as follows
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
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.
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
Self Tuning Fuzzy Maximum Power Tracking Control … 739
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
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.
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.
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.
740 C. Housseyn et al.
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.
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.
Self Tuning Fuzzy Maximum Power Tracking Control … 741
e(t) − e(t − 1)
e(t) = (18)
tsamp
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
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).
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:
Self Tuning Fuzzy Maximum Power Tracking Control … 743
e = G e × e. (21)
e = Ge × e (22)
u = (β ∗ G e ) × u (23)
where:
1
β = K1 + |e| (24)
m
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
744 C. Housseyn et al.
and frequency with the grid, but the main objective is to control the 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.
Self Tuning Fuzzy Maximum Power Tracking Control … 745
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.
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.
Self Tuning Fuzzy Maximum Power Tracking Control … 747
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.
748 C. Housseyn et al.
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
Self Tuning Fuzzy Maximum Power Tracking Control … 749
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.
A. Dali (B)
Centre de Développement des Energies Renouvelables, CDER, BP 62 Route de l’Observatoire,
16340 Bouzaréah, Algiers, Algeria
e-mail: [email protected]
S. Abdelmalek
Department of Electrical Engineering, Faculty of Technology, University of Medea, Medea,
Algeria
e-mail: [email protected]
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
e-mail: [email protected]
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
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
A = ⎣ 0 − RC
1
0 ⎦, χ = [x1 x2 x3 ]T
⎪
0 1 0 ⎩ = L2 V f c 0 0 T
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,
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
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:
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)
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]
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.
5 Simulation Results
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.
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
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
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
A Backstepping Controller for Interleaved … 761
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
1 Introduction
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.
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:
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:
Radiation reflected from the ground is assumed diffuse and obtained by [1].
Ig,T = Iρg (1 − cos β) 2 (3)
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)
ωs is the hourly angle of sunrise for a sloped surface and can be determined as
follows:
−1
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.
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.
The Impact of the Solar Radiation Profile on Sizing … 767
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
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.
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
The Impact of the Solar Radiation Profile on Sizing … 769
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 .
4 Sensitivity Analysis
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.
Indeed, some users could admit 100 h of shortage; they correspond to 16 min of daily
shortage.
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.
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:
24
Yr = IT /IT,ST C (KWh/KW) (15)
1
The Impact of the Solar Radiation Profile on Sizing … 773
Y f = E P Vu,d /P0 KWh/KWp,d (16)
L c = Yr − Ya KWh/KWp,d (17)
L S = Ya − Y f KWh/KWp,d (18)
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.
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.
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.
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.
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%.
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.
The Impact of the Solar Radiation Profile on Sizing … 777
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
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
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.
The performance indices like thermal gain and thermal yield are expressed below.
The useful thermal output of the system is given by following relationship:
782 O. Touaba et al.
While the overall thermal efficiency is calculated using the following relationship:
tf
dqth dt
ti
ηth = tf (3)
A ti Gi dt
Investigation and Prototyping Implementation … 783
3 Experimental Results
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.
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.
784 O. Touaba et al.
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
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
786 O. Touaba et al.
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
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
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
788 O. Touaba et al.
can be helpful. In the development of the field of solar heaters and conservation of
the environment successfully.
References
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
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.
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.
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
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.
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
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.
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
Piezoelectric Energy Harvesting … 795
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.
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
U Z e −Z em I
= (5)
F Z em −Z m v
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.
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.
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
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
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
e-mail: [email protected]
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.
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:
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)
xi → φ (xi ) (8)
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
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,
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)
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
where,
SPE(k) = x(k) − x̂(k)2 (25)
and,
SPE(k) = x(k) − x̂(k)2 (26)
SPE ≤ δ 2 (27)
n
l
(x)2 =
SPEφ = φ(x) − φ t2j − t2j (29)
j=1 j=1
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:
where γ = b/2a and h=2a 2/b, a and b are the estimated mean and variance of the
SPE [11, 12].
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
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
Fault Detection of Uncertain Systems Based on Interval Data Driven Approach 811
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
0.02
SPE
0.01
0
0 100 200 300 400 500 600 700 800 900 1000
0.04
SPE
0.02
0
0 100 200 300 400 500 600 700 800 900 1000
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.
Fault Detection of Uncertain Systems Based on Interval Data Driven Approach 813
5 Conclusion
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
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.
1 Introduction
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.
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
∞
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
∞
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
αr
δ= . (5)
2
θk = θr + (k − 1)αr . (6)
3 Machine 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
d[φr ]
[0] = [Rr ][ir ] + . (13)
dt
With
[L r s ] = [L sr ]T (15)
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]:
820 D. Kouchih et al.
⎡ ⎤
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)
μ0 rl
L kk = αr (1 − αr /2π ). (20)
g
μ0 rl
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.
Discrimination of Unbalanced Supply and Stator Interturn … 821
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
Common point
(o)
(b)
von
(c)
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.
4 Simulation Results
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.
U V
Discrimination of Unbalanced Supply and Stator Interturn … 823
2
10
fs
0
fsh1 fsh2
Current ias (A)
10
-2
10
-4
10
0 500 1000 1500 2000
Frequency (Hz)
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
100
80
Torque (N.m) 60
40
20
-20
-40
0 0.5 1 1.5
Time (s)
100
50
Current ias (A)
-50
-100
0 0.5 1 1.5
Time(s)
100 100
50 50
Current ibs (A)
0 0
-50 -50
-100 -100
0 0.5 1 1.5 0 0.5 1 1.5
4
x 10
1
-0.5
-1
0 0.5 1 1.5
Time (s)
30
20
10
-10
-20
-30
0 0.5 1 1.5
Time (s)
(a)
2
Common Point Voltage (V)
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
826 D. Kouchih et al.
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.
100
80
Torque (N.m)
60
40
20
-20
-40
0 0.5 1 1.5
Time (s)
100
Current ias (A)
50
-50
-100
0 0.5 1 1.5
Time (s)
100 100
0 0
-50 -50
-100 -100
0 0.5 1 1.5 0 0.5 1 1.5
Time (s) Time (s)
x 10
1
Rotor loop current (A)
0.5
-0.5
-1
0 0.5 1 1.5
Time (s)
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
828 D. Kouchih et al.
-50
0 0.5 1 1.5
Time (s)
(a)
2
Common Point Voltage (V)
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
Discrimination of Unbalanced Supply and Stator Interturn … 829
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
830 D. Kouchih et al.
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
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.
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.
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).
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
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
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
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.
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.
4 Simulation Results
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
teta (rad) teta (rad)
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
Impact of the Stator Winding Topology on the Fault Harmonic … 837
-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
teta (rad) teta (rad)
-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
teta (rad) teta (rad)
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
teta (rad) teta (rad)
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
teta (rad) teta (rad)
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
teta (rad) teta (rad)
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).
Impact of the Stator Winding Topology on the Fault Harmonic … 839
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
Fréquency (Hz) Fréquency (Hz)
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
Fréquency (Hz) Fréquency (Hz)
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
0
Amplitude (dB)
0
Amplitude (dB)
-50 -50
-100 -100
0 50 100 0 50 100
Fréquency (Hz) Fréquency (Hz)
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)
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
Fréquency (Hz) Fréquency (Hz)
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)
-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
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
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.
2 System Description
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:
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
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.
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).
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)
Sensor Fault Detection for Uncertain T-S DC Model … 847
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)
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.
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
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)
E = Ē + E L C̄ (18)
N = −Fi L (20)
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)
Sensor Fault Detection for Uncertain T-S DC Model … 849
⎡ ⎤
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)
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)
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
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
⎡ ⎤
−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
Sensor Fault Detection for Uncertain T-S DC Model … 855
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].
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
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)
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.
Detection of Stator Interturn Fault in Double Star … 861
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)
L ss Msr Rs 07∗3
L= R= (10)
MsrT L rr 03∗7 Rr
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
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.
Detection of Stator Interturn Fault in Double Star … 863
3 Simulation Results
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.
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.
5 Experimental Results
-50
-100
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz)
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)
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
Detection of Stator Interturn Fault in Double Star … 865
0
0
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
866 R. Hachelaf et al.
5
Stator Current (A)
-5
10
Stator Current (A)
-5
-10
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).
Detection of Stator Interturn Fault in Double Star … 867
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
868 R. Hachelaf et al.
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
870 R. Hachelaf et al.
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
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
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].
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
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]
874 A. Bouzekri and T. Allaoui
• 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
Rule with “AND” or “OR”: The example below shows how a few conditions are
related to AND.
Fuzzy-Expert System Fault Tolerant Control of IGBT … 875
⎧
⎪
⎪ IF condition. 1
⎪
⎪
⎪
⎪ AND IF condition. 2
⎪
⎨
.
(2)
⎪
⎪ .
⎪
⎪
⎪
⎪ AND IF condition. 2
⎪
⎩
THEN action
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
876 A. Bouzekri and T. Allaoui
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
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.
878 A. Bouzekri and T. Allaoui
Fig. 4 Membership functions of the input variables and their Output variable
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);
• Using the control orders of the faulty leg for the redundant one;
• Stopping the fault detection algorithm.
4 Simulation Results
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.
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.
882 A. Bouzekri and T. Allaoui
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
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.
1 Introduction
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
888 I. Attoui et al.
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 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:
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)
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.
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.
890 I. Attoui et al.
3- Reduction 3- Reduction
using LSDA using LSDA
Output
5- Recognize the the
fault type and results
its severity
Training
Diagnostic phase
phase
Multiclass Support Vector Machine … 891
(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.
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.
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.
892 I. Attoui et al.
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%.
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.
894 I. Attoui et al.
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
Multiclass Support Vector Machine … 895
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
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.
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
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
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
900 I. Laidani and N. Bourouba
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.
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
Processing Signal Parameters Based Fuzzy Inference System … 901
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
902 I. Laidani and N. Bourouba
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
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.
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
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.
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
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.
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
Processing Signal Parameters Based Fuzzy Inference System … 909
Fig. 7 a Membership functions of S0 for R1. b Output membership function for fault F52
Processing Signal Parameters Based Fuzzy Inference System … 911
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
912 I. Laidani and N. Bourouba
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.
Processing Signal Parameters Based Fuzzy Inference System … 913
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
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).
Nomenclature
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.
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].
918 Z. Gherabi et al.
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.
Line Current Spectrum Analysis as a Technique … 919
(a) (b)
(c) (d)
Fig. 3 Types of considered defects, a and b partial DMF, c and d complete DMF
Line Current Spectrum Analysis as a Technique … 921
Magnitude (db)
spectrum around the
fundamental frequency in the
presence of the DMF Lower Upper
sideband sideband
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.
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]:
922 Z. Gherabi et al.
|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.
6 Simulation Results
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
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.
924 Z. Gherabi et al.
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
Line Current Spectrum Analysis as a Technique … 925
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)
-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.
926 Z. Gherabi et al.
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)
-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.
Line Current Spectrum Analysis as a Technique … 927
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
928 Z. Gherabi et al.
Appendix B
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
Line Current Spectrum Analysis as a Technique … 929
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
930 Z. Gherabi et al.
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
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.
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
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.
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:
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
934 Y. Zennir et al.
a mg
Fx3 = Fx2 = · (5)
a+b 2
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:
4
mg
Fy = Fyi = u (sgn(ẏ1 ) + sgn(ẏ3 )) (8)
i=1
a+b
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
Fault-Tolerant Path-Tracking Control … 935
4 Simulink Model
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.
To calculate the rolling resistance torque, we the method in Stepwise Coast down
Methodology for Measuring Tire Rolling Resistance. The rolling resistance is a
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
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
Fault-Tolerant Path-Tracking Control … 939
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.
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
Fault-Tolerant Path-Tracking Control … 941
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
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.
Fault-Tolerant Path-Tracking Control … 945
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
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
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.
where:
Discrete Faults Diagnosis for the Multicellular Converter … 949
⎡ 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:
950 K. Haddadi et al.
⎡ 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.:
– There is one and only one discrete fault occurs at a given instant.
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.
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).
952 K. Haddadi et al.
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,
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
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
Discrete Faults Diagnosis for the Multicellular Converter … 953
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.˛
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.
N
y N(i) (0) i
y(t) ≈ y N (t) = t (8)
i=0
i!
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
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)
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
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.
Discrete Faults Diagnosis for the Multicellular Converter … 955
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).
956 K. Haddadi et al.
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.
Discrete Faults Diagnosis for the Multicellular Converter … 957
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)
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
Time (s) Time (s) Time (s)
Fig. 6 Blocking of u 1 = 1
958 K. Haddadi et al.
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
Time (s) Time (s) Time (s)
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
Time (s) Time (s) Time (s)
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
Discrete Faults Diagnosis for the Multicellular Converter … 959
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
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.
1 Introduction
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
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)
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:
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)
(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
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.
Fault Diagnosis of Uncertain Hybrid Actuators … 965
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)
where
T T
W = Wupper , Wlower , Z = Z upper
T
, Z lower
T
966 Z. Abdelmalek et al.
⎡ ⎤
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.
4 Simulation Results
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,
Fault Diagnosis of Uncertain Hybrid Actuators … 967
1.5
Reference
FDMPC Approach
0.5
-0.5
-1
0 1 2 3 4 5
Time [sec]
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.
968 Z. Abdelmalek et al.
0.4
0.2
Position Error [rad]
-0.2
-0.4
-0.6
-0.8
0 1 2 3 4 5
Time [sec]
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
Fault Diagnosis of Uncertain Hybrid Actuators … 971
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
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.
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
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:
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:
∞
J∞ (k) = X (k + j) + U (k + j) (4.b)
j=0
976 Z. Abdelmalek et al.
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:
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:
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)
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)
Fault Tolerant Predictive Control for Constrained … 977
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
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:
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
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)
(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
(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)
– Output Constraints:
f or : j ≥ 0, h = 1, 2, . . . , q
ymax = W
maxy(k)max ≥ maxCi A
j>0 j>0 max
>0 (19)
4 + F Fs A
A 5 T CiT P
End of proof.
980 Z. Abdelmalek et al.
4 Simulation Results
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)
Q0 = I, R0 = 1,
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.
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
984 Z. Abdelmalek et al.
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
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.
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
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
⎧
⎪
⎪ ẋ1 = x2
⎪
⎪
⎪
⎨ ẋ2 = x3
.. (1)
⎪ .
⎪
⎪
⎪
⎪ ẋ = f (x1 , x2 , . . . xn ) + g(x1 , x2 , . . . xn )u
⎩ n
y = x1
Or equivalently
y (n) = f (x) + g(x)u
(2)
y = x1
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
|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
With
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:
Active Adaptive Fuzzy Fault-Tolerant Control … 989
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)
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
990 A. Bounemeur and M. Chemachema
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:
with
where
μk (x)
w N (x) = N , k = 1, . . . , N (20)
j=1 μ j (x)
Active Adaptive Fuzzy Fault-Tolerant Control … 991
n
with .μk (x) = i=1 μ F̄ik ., F̄ik ∈ Fi1 , . . . , Fim i .
which represents the weight of rule Rk .
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
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)
|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)
u = û = w T θ + u r (24)
σ2
u r = sign e T P B ε̂u − (25)
eT P B
992 A. Bounemeur and M. Chemachema
θ̇ = γ e T P Bw(x) (26)
ε̂˙ u = n u e T P B (27)
σ̇ = −δ0 σ (28)
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)
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)
Active Adaptive Fuzzy Fault-Tolerant Control … 993
(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
With
⎡ ⎤ ⎡ ⎤
0 1 ··· 0 0
⎢ . .. ⎥ ⎢ .. ⎥
⎢ .. ⎥
A=⎢ . ··· 0 ⎥ B=⎢
⎢.⎥
⎥
⎢ ⎥ ⎣
⎣ 0 0 ··· 1 ⎦ 0⎦
−kn −kn−1 . . . −kn 1
A T P + P A = −Q (38)
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
994 A. Bounemeur and M. Chemachema
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
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
1 1
− e T Qe > − e T Qe ≤ 0 (48)
2g(x) 2g
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.
4 Simulation Results
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
Active Adaptive Fuzzy Fault-Tolerant Control … 995
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
μ Fi2 (xi ) = exp −
2 0.2
1 xi 2
μ Fi3 (xi ) = exp −
2 0.2
996 A. Bounemeur and M. Chemachema
-0.5
-1
-1.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)
1 xi − 0.4 2
μ Fi4 (xi ) = exp −
2 0.2
1 xi − 0.8 2
μ Fi5 (xi ) = exp −
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,
Active Adaptive Fuzzy Fault-Tolerant Control … 997
-0.5
-1
-1.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)
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).
998 A. Bounemeur and M. Chemachema
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
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
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
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.
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
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)
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
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:
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
1008 F. Demim et al.
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
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
Simultaneous Localization and Mapping Algorithm … 1009
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
5 5
4 4
20 20
10 15 10 15
10 10
0 5 0 5
0 0
-10 -10 -5 -10 -10 -5
⎛ ⎞ ⎛ ⎞
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
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
∇ 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
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)
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
Simultaneous Localization and Mapping Algorithm … 1013
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:
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)
1014 F. Demim et al.
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)
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.
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
Simultaneous Localization and Mapping Algorithm … 1015
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)
0.2
-0.2
0 50 100 150 200 250
Algorithm iterations
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
INS(XYZ)-with noise
0.2
0
Positions (m)
-0.2
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
Algorithm iterations Algorithm iterations
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
Algorithm iterations
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
Algorithm iterations Algorithm iterations
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
Algorithm iterations Algorithm iterations
Position Z
5.03
5.025
True path
5.02 Estimated EKF-SLAM
Z-uav(m)
5.01
5.005
5
4.995
0 10 20 30 40 50 60 70 80
Velocity U Velocity V
0.35 0.8
0.3 True path True path
V-uav(m/s)
Velocity W
0.04
W-uav(rad/s)
0.02
-0.06
0 10 20 30 40 50 60 70 80
Algorithm iterations
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
Algorithm iterations Algorithm iterations
Angle Psi
0.04
0.03
Psi-uav(rad)
0.02
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
Algorithm iterations Algorithm iterations Algorithm iterations
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
Algorithm iterations Algorithm iterations Algorithm iterations
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
Algorithm iterations Algorithm iterations Algorithm iterations
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
1020 F. Demim et al.
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
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.
1 Introduction
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.
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
1024 A. Zitouni et al.
+∞
Ca,b = x(t)ψa,b (t)dt, (1)
−∞
where
1 t −b
ψa,b (t) = √ ψ with a = 0 (2)
a a
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]:
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,
1026 A. Zitouni et al.
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)
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
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)
Comparison Between Gabor Filters and Wavelets Transform … 1029
Table 1 (continued)
(c3)
(d1)
(d2)
(d3)
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
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
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…
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.
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
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
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.
High Level Modeling and Hardware Implementation of Image … 1037
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).
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).
Fig. 5 Block diagram of the memory management strategy used for the implemented filters
High Level Modeling and Hardware Implementation of Image … 1041
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).
A. Simulation results
6 Conclusion
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
High Level Modeling and Hardware Implementation of Image … 1045
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
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.
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
e-mail: [email protected]
H. Zeraoula · K. Tourche
Département Radar, Ecole Supérieure -Ali Chabati, Reghaia Alger, Algérie
e-mail: [email protected]
K. Tourche
e-mail: [email protected]
M. Hamadouche
Faculté de Technologie, Univerisité de Boumerdes, Boumerdes, Algérie
e-mail: [email protected]
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)
According to the value of T, we deduce the concordance between F(x) and F*(x).
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:
1050 A. Rouabah et al.
– 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.
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
(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
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
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
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
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
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.
1054 A. Rouabah et al.
Fig. 6 Thresholds of CA-CFAR and KBS-Dynamic detectors in the absence of interfering targets
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.
Proposal for a Radar Detection Architecture … 1055
Fig. 7 CA-CFAR and KBS-Dynamic thresholds in case of interfering targets presence in cells with
different GIS
Figure 8 shows the presence of multiple targets in the reference window and with
the same level of 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.
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.
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.
Proposal for a Radar Detection Architecture … 1057
Fig. 9 KBS-ALC and KBS-Dynamic thresholds in presence of interfering targets in cells with
same GIS
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.
1058 A. Rouabah et al.
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.
8 Conclusion
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
1060 A. Rouabah et al.
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
1 Introduction
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”.
In order to accelerate spacecraft and artificial satellites in space, there exits two
solutions system design Chemical Propulsion (CP) and Electric Propulsion (EP).
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)
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
1
a= (ra + rb ) (6)
2
The radius to periapsis is given by:
1064 J. E. Benmansour et al.
h2 1
rp = (7)
μ 1+e
h2 1
rp =
μ 1 + rra −r
+r
p
a p
h ra + r p
2
= (8)
μ 2ra
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)
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
Comparative Analysis and Simulation of Propulsion and Energy … 1065
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.
1066 J. E. Benmansour et al.
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).
1
Z coordinate (ER)
0.5
-0.5
-1
-6 -4 -2 0 2 4 6
Y coordinate (ER)
4
Y coordinate (ER)
-2
-4
-6
-8 -6 -4 -2 0 2 4 6 8
X coordinate (ER)
35
30
Inclination (degrees)
25
20
15
10
0
0 10 20 30 40 50 60 70
Time (days)
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.
1068 J. E. Benmansour et al.
7.5
6.5
Velocity (km/s)
5.5
4.5
3.5
3
0 10 20 30 40 50 60 70
Time (days)
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)
4 Conclusion
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
1 Introduction
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
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
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.
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].
1076 R. Khelladi et al.
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.
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.
Narrowbad Sensor-Antenna Based on an Interdigital Capacitor 1077
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.
1078 R. Khelladi et al.
Fig. 9 Simulated/Measured
S11 parameter for the
unloaded case
Fig. 10 Simulated/Measured
S11 parameter for the Teflon
MUT
Fig. 11 Simulated/Measured
S11 parameter for the FR-4
MUT
Narrowbad Sensor-Antenna Based on an Interdigital Capacitor 1079
4 Conclusion
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
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.
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.
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
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
1084 O. Megherbi et al.
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.
Secure Color Image Transmission Based on the Impulsive … 1085
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
1086 O. Megherbi et al.
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
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
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
Secure Color Image Transmission Based on the Impulsive … 1087
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
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.
1088 O. Megherbi et al.
⎧
⎪ 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 .
4 Simulation Results
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
Secure Color Image Transmission Based on the Impulsive … 1089
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. 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.
Secure Color Image Transmission Based on the Impulsive … 1091
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
1092 O. Megherbi et al.
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
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
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.
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
Secure Color Image Transmission Based on the Impulsive … 1095
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
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.
1 Introduction
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).
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
(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
1100 A. Bounegab and M. Boulesbaa
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 sub- carriers 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.
Performance of FBMC for Backhauling 5G 1101
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
(a) (b)
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
1104 A. Bounegab and M. Boulesbaa
Fig. 9 BER based on the ITU model for SNR distribution RICE
Performance of FBMC for Backhauling 5G 1105
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.
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.
Performance of FBMC for Backhauling 5G 1107
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
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.
1108 A. Bounegab and M. Boulesbaa
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
1 Introduction
Input
Receiver
Output
MCIP Transmitter
(Configurable Microcontroller)
Backs. clock
Data Bus
Timer
PRNG
Program Data
Memory Memory Tag Memory
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
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
Fig. 3 Functions tree of the tag software with the program size of each function
Flexible FPGA-Based RFID Emulation Platform … 1115
results concern the implementation of both hardware and software parts. Note that
the tag program memory (software part) occupies about 7% of FPGA resources.
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
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.
Tag
RN16 (FM0)
Tag
Query
Reader
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].
1118 I. Mezzah et al.
Emulation controller
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.
Flexible FPGA-Based RFID Emulation Platform … 1119
6 Conclusion
References
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
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
e-mail: [email protected]
M. Belhocine
Centre de Développement des Technologies Avancées CDTA, Baba Hassen, Algiers, Algeria
e-mail: [email protected]
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
Dataset creation
Features extraction/preprocessing
Voting process No
Start interaction
Yes
Speech Recognition
Feeding Features
Action Recognition
Robot Reaction
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.
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
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
Similarly, let θ and be the rotation angles about the y-axis and z axis respectively:
1126 S. Kahlouche and M. Belhocine
⎡ ⎤
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
Time X Y Z θ φ ψ
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
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
biases of the network. ANN is applied on our dataset with ‘relu’ as activation
function and achieved 72.5% of accuracy (see Fig. 7).
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.
Figure 8 shows the confusion matrix of activity recognition using voting process,
which achieves accuracy of 97.1%.
B. Robot Reaction
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’.
(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
1132 S. Kahlouche and M. Belhocine
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
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
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.
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
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
1136 H. Belkhiri et al.
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
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
Security in the Internet of Things … 1139
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
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
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
Security in the Internet of Things … 1145
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 consumpti- on 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
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
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
A. Database description
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
1150 A. Tafsast et al.
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
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%
na
ni
Be
ig
al
M
Target Class
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%
n
ig
na
n
Be
ig
al
M
Target Class
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.
1154 A. Tafsast et al.
4 Conclusion
References
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
1 Introduction
R. Bendoumia (B)
Department of Electronics, Laboratory LATSI, University of Blida 1, Blida, Algeria
e-mail: [email protected]
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.
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:
The exact update relations of adaptive filters w12 (n) and w21 (n) by the FSAD
algorithm are given as follows [7]:
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
When the update formulas of the adaptive filter w21 (n) by the FSAD algorithm is
given as follow:
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
Advanced Feedforward-and-Feedback … 1161
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].
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]:
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
When the update formulas of the adaptive filter w21 (n) by the BSAD algorithm
is given as follow:
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
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
Advanced Feedforward-and-Feedback … 1163
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.
u2 (n)u1 (n)
w12 (n) = w12 (n − 1) + μ12 (n) (25)
p1 (n)p1 (n) +
T
ε
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:
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
v2 (n)v1 (n)
w12 (n) = w12 (n − 1) + μ12 (n) (33)
v1T (n)v1 (n) + ε
Advanced Feedforward-and-Feedback … 1165
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
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
6 Simulation Results
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
Advanced Feedforward-and-Feedback … 1167
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
Advanced Feedforward-and-Feedback … 1169
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
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.
1 Introduction
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.
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)
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
1174 N. Guidoum et al.
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
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 /
∞
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
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
1176 N. Guidoum et al.
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
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
New Log CFAR Radar for Compound Sea Clutter 1177
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.
1178 N. Guidoum et al.
Fig. 7 Variation of PFA versus the scale factor τ for the CIG distribution plus thermal noise and
two interfering targets
4 Conclusion
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
New Log CFAR Radar for Compound Sea Clutter 1179
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.
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.
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.
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
[φ11 (z − z 0 )] [φ12 (z − z 0 )]
[φ(z − z 0 )] = (4)
[φ21 (z − z 0 )] [φ22 (z − z 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]:
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
Stochastic Analysis of the Crosstalk Between Transmission Lines 1185
d
dz
[U (x, ξ )] = −[L(ξ )][I (z, ξ )]
(15)
dz (x, )]
d
[I ξ = −[C(ξ )][U (z, ξ )]
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
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}
∀y =∈ {0, . . . , p − 1}
# $
By solving the system (23), we find the coefficients α0 , . . . α P−1 , so can solve
Eq. (17).
Stochastic Analysis of the Crosstalk Between Transmission Lines 1187
4 Simulation
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
1188 H. Gheddar et al.
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
Stochastic Analysis of the Crosstalk Between Transmission Lines 1189
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]
0.2
MC PDF
F= 10 MHz
PC PDF
0.15
PDF
0.1
0.05
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
1 Introduction
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.
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.
1196 B. M. Akli et al.
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.
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).
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.
Moving Objects Detection and Tracking … 1197
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).
1198 B. M. Akli et al.
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
Moving Objects Detection and Tracking … 1199
level curve method. The first approach of active contours was improved by works
presented in [8, 9] based-on the geodesic Active contours.
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.
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
1200 B. M. Akli et al.
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 (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
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:
Moving Objects Detection and Tracking … 1201
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”.
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
1202 B. M. Akli et al.
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).
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.
1204 B. M. Akli et al.
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.
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.
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)
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
Moving Objects Detection and Tracking … 1205
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]:
−1
Kk+1 = Pk+1/k Pk/k HkT Hk Pk+1/k HkT + Rk (13)
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.
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
1206 B. M. Akli et al.
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
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)
Moving Objects Detection and Tracking … 1207
(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)
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.
1208 B. M. Akli et al.
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 ap- proaches 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 ori- ented 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
Moving Objects Detection and Tracking … 1209
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 stabi- lization. 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 parti- cle/kalman
filter for improved tracking of beam aspect targets,” IEEE Workshop on Statistical Signal
Processing, 2003, PP. 535–538
1210 B. M. Akli et al.
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.
1 Introduction
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
[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
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)
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.
1214 S. Tahraoui and M. Ouarzeddine
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.
Potential of Using Adaptive Complex Flower … 1215
• 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.
1216 S. Tahraoui and M. Ouarzeddine
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.
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.
Potential of Using Adaptive Complex Flower … 1217
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
γT v
150 30
gr ou n d t op ogap h y :φ 0
180 0
z
210 330
240 300
270
1218 S. Tahraoui and M. Ouarzeddine
(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
Potential of Using Adaptive Complex Flower … 1219
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
1220 S. Tahraoui and M. Ouarzeddine
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
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.
1 Introduction
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.
3 Proposed method
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
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)
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
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;
Intelligent Cooperative Method for Medical … 1227
– 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
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:
where:
1
Dissim (C, F) ≡ avgc [d E (c, F)] + avg f [d E ( f, C)] (6)
2
1228 K. Benhamza and H. Seridi
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
Image1 Image2
Image3 Image4
Image 5 Image6
Image 7
1230 K. Benhamza and H. Seridi
5 Conclusion
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)
Intelligent Cooperative Method for Medical … 1231
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
L. Messaouda (B)
Department of Science and Technology, University of Batna, Batna, Algeria
e-mail: [email protected]
Z. Messali
Department of Science and Technology, University of Bordj Bou Arreridj, Bordj Bou Arreridj,
Algeria
e-mail: [email protected]
R. Abdelghani
Electrical Engineering Department, Applied Automation and Industrial Diagnostic Laboratory,
Djelfa, Algeria
e-mail: [email protected]
S. LARBI
Department of Science and Technology, University of Laghouat, Laghouat, Algeria
e-mail: [email protected]
1 Introduction
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;
1236 L. Messaouda et al.
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)
where:
(1 − H (φ(x))).I (x)d x
u= (7)
1 − H (φ(x))d x
An Image Segmentation Model Using a Level … 1237
H (φ(x)).I (x)d x
v= (8)
H (φ(x))d x
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
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
1238 L. Messaouda et al.
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
An Image Segmentation Model Using a Level … 1239
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).
4 Experimental Results
40
60
80
100
120
20 40 60 80 100 120
Proposed method
(120 iterations)
20
40
60
80
100
120
20 40 60 80 100 120
Proposed method
(90 iterations)
40
60
80
100
120
140
160
180
200
Proposed method
(100 iterations)
20
40
60
80
100
120
140
160
180
200
Proposed method
(95 iterations)
Proposed method
(90 iterations)
Initial contour
Proposed method
(100 iterations)
Proposed method
(100 iterations)
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
Table 2 The coefficients: times, Dice, Hausdorff and MSSD of the three method s obtained for the
segmentation of Chest image
Methods Time (S) Dice coefficient Hausdorff coefficient MSSD coefficient
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
Table 3 The coefficients: times, Dice, Hausdorff and MSSD of the three method s obtained for the
segmentation of Arms_XRay image
Methods Time (S) Dice coefficient Hausdorff coefficient MSSD coefficient
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
Table 4 The coefficients: times, Dice, Hausdorff and MSSD of the three method s obtained for the
segmentation of Vessel_CTA1 image
Methods Time (S) Dice coefficient Hausdorff coefficient MSSD coefficient
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.
1246 L. Messaouda et al.
References
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.
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.
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].
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].
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
Fig. 1 Hardware
Convergence Architecture of
the 5G and Satellite
Networks
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
New Convergence Architecture Between 5G Mobile … 1253
Beginning
Else
Deactivate HARQ mechanism
End if
Sleep (03 s)
End while
End
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
1254 S. Sadouni et al.
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.
New Convergence Architecture Between 5G Mobile … 1255
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. 4 Comparison of the throughput received by the UE with the speeds 50 km/h and 100 km/h
6 Conclusion
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.
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
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.
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
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:
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:
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].
1262 S. Babahenini et al.
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
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
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
1264 S. Babahenini et al.
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
(b) PR Curve
(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
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.
1266 S. Babahenini et al.
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
(b) PR 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
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
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.
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
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.
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:
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:
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:
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:
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:
1274 I. Hassani et al.
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.
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:
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:
The Fast-NLMS Algorithm for Acoustic Echo … 1275
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:
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).
1276 I. Hassani et al.
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)
4 Simulation Results
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 Fast-NLMS Algorithm for Acoustic Echo … 1277
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
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
1278 I. Hassani et al.
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)
-2
0 1 2 3 4 5 6 7 8 9 10
4
Samples 10
-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
The Fast-NLMS Algorithm for Acoustic Echo … 1279
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
1280 I. Hassani et al.
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
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].
microstrip
line
2 Sensor Design
A. Theoretical aspect
CR LL
1284 S. Hocine et al.
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
Lp W
Wc
H Ht
Hc
We
g
Design of Sensor Based on CRLH … 1285
Port 1
Liquid Mixture
Port 2
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.
1286 S. Hocine et al.
-20
-30
-40
-50
-60
0.5 1 1.5 2 2.5 3 3.5
Frequency (Ghz)
-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
fo − f1
S= (5)
fo dεr
Design of Sensor Based on CRLH … 1287
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
1288 S. Hocine et al.
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