Full download test bank at ebook textbookfull.

com

Industrial Process Identification

Perturbation Signal Design and

CLICK LINK TO DOWLOAD

https://textbookfull.com/product/industrial-
process-identification-perturbation-signal-
design-and-applications-ai-hui-tan/

textbookfull
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Digital Signal Processing Fundamentals and Applications

3rd Edition Lizhe Tan Ph.D. Electrical Engineering
University Of New Mexico

https://textbookfull.com/product/digital-signal-processing-
fundamentals-and-applications-3rd-edition-lizhe-tan-ph-d-
electrical-engineering-university-of-new-mexico/

Petroleum Refining Design and Applications Handbook

Rules of Thumb Process Planning Scheduling and
Flowsheet Design Process Piping Design Pumps
Compressors and Process Safety Incidents A Kayode Coker
https://textbookfull.com/product/petroleum-refining-design-and-
applications-handbook-rules-of-thumb-process-planning-scheduling-
and-flowsheet-design-process-piping-design-pumps-compressors-and-
process-safety-incidents-a-kayode-coker/

Industrial IoT Challenges Design Principles

Applications and Security Ismail Butun

https://textbookfull.com/product/industrial-iot-challenges-
design-principles-applications-and-security-ismail-butun/

Sustainable Design Through Process Integration:

Fundamentals and Applications to Industrial Pollution
Prevention, Resource Conservation, and Profitability
Enhancement 2nd Edition Mahmoud M. El-Halwagi
https://textbookfull.com/product/sustainable-design-through-
process-integration-fundamentals-and-applications-to-industrial-
pollution-prevention-resource-conservation-and-profitability-
Chemical Process Design and Simulation: Aspen Plus and
Aspen Hysys Applications Juma Haydary

https://textbookfull.com/product/chemical-process-design-and-
simulation-aspen-plus-and-aspen-hysys-applications-juma-haydary/

Carbon Nanotubes for Interconnects: Process, Design and

Applications 1st Edition Aida Todri-Sanial

https://textbookfull.com/product/carbon-nanotubes-for-
interconnects-process-design-and-applications-1st-edition-aida-
todri-sanial/

Alarm Management for Process Control a Best Practice

Guide for Design Implementation and Use of Industrial
Alarm Systems Rothenberg

https://textbookfull.com/product/alarm-management-for-process-
control-a-best-practice-guide-for-design-implementation-and-use-
of-industrial-alarm-systems-rothenberg/

Measurement while drilling (MWD): signal analysis,

optimization, and design Chin

https://textbookfull.com/product/measurement-while-drilling-mwd-
signal-analysis-optimization-and-design-chin/

Robot Systems for Rail Transit Applications 1st Edition

Hui Liu

https://textbookfull.com/product/robot-systems-for-rail-transit-
applications-1st-edition-hui-liu/
Advances in Industrial Control

Ai Hui Tan
Keith Richard Godfrey

Industrial Process
Identification
Perturbation Signal Design
and Applications
Advances in Industrial Control

Series editors
Michael J. Grimble
M. A. Johnson

Advisory Editor
Sebastian Engell, Technische Universität Dortmund, Dortmund, Germany

Editorial Board
Graham C. Goodwin, School of Electrical Engineering and Computing,
University of Newcastle, Callaghan, NSW, Australia
Thomas J. Harris, Department of Chemical Engineering, Queen’s University, Kingston, ON,
Canada
Tong Heng Lee, Department of Electrical and Computer Engineering, National University of
Singapore, Singapore
Om P. Malik, Schulich School of Engineering, University of Calgary, Calgary,
AB, Canada
Gustaf Olsson, Industrial Electrical Engineering and Automation, Lund Institute of
Technology, Lund, Sweden
Ikuo Yamamoto, Graduate School of Engineering, University of Nagasaki, Nagasaki, Japan

Editorial Advisors
Kim-Fung Man, City University Hong Kong, Kowloon, Hong Kong
Asok Ray, Pennsylvania State University, University Park, PA, USA
Advances in Industrial Control is a series of monographs and contributed titles
focusing on the applications of advanced and novel control methods within applied
settings. This series has worldwide distribution to engineers, researchers and
libraries.
The series promotes the exchange of information between academia and industry,
to which end the books all demonstrate some theoretical aspect of an advanced or
new control method and show how it can be applied either in a pilot plant or in
some real industrial situation. The books are distinguished by the combination
of the type of theory used and the type of application exemplified. Note that
“industrial” here has a very broad interpretation; it applies not merely to the
processes employed in industrial plants but to systems such as avionics and
automotive brakes and drivetrain. This series complements the theoretical and more
mathematical approach of Communications and Control Engineering.
Indexed by SCOPUS and Engineering Index.
Series Editors
Michael J. Grimble
M. A. Johnson
In-house Editor
Mr. Oliver Jackson Springer London, 4 Crinan Street, London, N1 9XW, United
Kingdom e-mail: [email protected]
Publishing Ethics
Researchers should conduct their research from research proposal to publication in
line with best practices and codes of conduct of relevant professional bodies and/or
national and international regulatory bodies. For more details on individual ethics
matters please see:
https://www.springer.com/gp/authors-editors/journal-author/journal-author-helpd
esk/publishing-ethics/14214

More information about this series at http://www.springer.com/series/1412

Ai Hui Tan Keith Richard Godfrey
•

Industrial Process
Identification
Perturbation Signal Design and Applications

123
Ai Hui Tan Keith Richard Godfrey
Faculty of Engineering School of Engineering
Multimedia University University of Warwick
Cyberjaya, Selangor, Malaysia Coventry, Warwickshire, UK

ISSN 1430-9491 ISSN 2193-1577 (electronic)

Advances in Industrial Control
ISBN 978-3-030-03660-7 ISBN 978-3-030-03661-4 (eBook)
https://doi.org/10.1007/978-3-030-03661-4

Library of Congress Control Number: 2018961195

© Springer Nature Switzerland AG 2019

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
To my parents
Ai Hui Tan
Series Editors’ Foreword

The series Advances in Industrial Control aims to report and encourage technology
transfer in control engineering. The rapid development of control technology has an
impact on all areas of the control discipline. New theory, new controllers, actuators,
sensors, new industrial processes, computer methods, new applications, new design
philosophies, new challenges, etc. Much of this development work resides in industrial
reports, feasibility study papers and the reports of advanced collaborative projects. The
series offers an opportunity for researchers to present an extended exposition of such
new work in all aspects of industrial control for wider and rapid dissemination.
Parameter estimation techniques for static and dynamical systems were probably
given a first thorough theoretical basis in the years around 1800. In about 1795,
Gauss claimed to have already developed and used the least squares estimation
method, a method that was given in definitive published form in 1805, but by
Legendre. Gauss’s own publication of the method he was using came a little later in
1809. All this initiated one of those scientific priority claim issues that is still
discussed today (See Gauss and the Invention of Least Squares by S. M. Stigler in
The Annals of Statistics, 1981, Vol. 9, No. 3, pp. 465–474).
Approximately one hundred and sixty years later, least squares came to be an
important tool in system identification for control systems design methods.
However, we like to think that before that, system identification for control design
began with the nonparametric methods of step responses and the ‘sustained oscil-
lation’ experimental methods of Ziegler and Nichols in the 1940s. In the early
1980s, these methods were given renewed relevance with the elegant ‘relay
experiment’ method of Åström and Hägglund. The Advances in Industrial Control
monograph series has many volumes on the related design of PID controllers and a
number of volumes dealing with the relay experiment paradigm, including:
• Autotuning of PID Controllers by Cheng-Ching Yu (ISBN 978-3-540-76250-8,
1999 [2nd edition, published as a stand-alone title: ISBN 978-1-84628-036-8,
2006]);
• Industrial Process Identification and Control Design by Tao Liu and Furong
Gao (ISBN 978-0-85729-976-5, 2012); and

vii
viii Series Editors’ Foreword

• Non-Parametric Tuning of PID Controllers by Igor Boiko (ISBN

978-1-4471-4464-9, 2012).
The advent of state-space system formalism for linear and nonlinear models
completely changed the landscape of control system theory. The work of Rudolf
Kalman and co-authors published in a key set of papers in the 1960s initiated a
different approach to systems theory and control systems design that was able to
subsume both time domain and frequency domain descriptions of dynamical pro-
cesses. The use of state-space system descriptions gave a significant impetus to
system identification methods for estimating the parameters of state-space and
transfer function models. The least squares algorithm adapted to both batch and
recursive computations became a fundamental tool in this new paradigm. The
sophisticated methods of system identification now include techniques such as
prediction error identification and subspace identification methods. This level
of theoretical maturity requires specialised authors to present the new ideas and
relevant Advances in Industrial Control monograph series and the Advanced
Textbooks in Control and Signal Processing series offerings include:
• Practical Grey-box Process Identification by Torsten Bohlin (ISBN
978-1-84628-402-1, 2006); and
• Identification of Continuous-time Models from Sampled Data edited by Hugues
Garnier and Liuping Wang (ISBN 978-1-84800-160-2, 2008); and
• System Identification by Karel J. Keesman (ISBN 978-0-85729-521-7, 2011).
Practical model identification for industrial systems can either use straightfor-
ward plant operational records or use an additional injected perturbation signal.
Operational data may not give very good estimation efficiency so the injected
identification signal is often the better strategy. But here operational constraints
may forbid any injected signal amplitude that causes observable end-product
quality degradation. Further, the identification signal should be designed to max-
imise estimation efficiency. This Advances in Industrial Control monograph enti-
tled Industrial Process Identification: Perturbation Signal Design and Applications
by authors: Ai Hui Tan and Keith Richard Godfrey reports their research on these
very issues. They design pseudorandom binary signals to avoid disturbing the
process performance significantly yet seek to optimise estimation accuracy. The
monograph has a good balance between theory and applications. Academic
examples illustrate the theory and a major attraction of the monograph is the
inclusion of two case studies. Chapter 6 reports the modelling of an electronic nose
system with direction-dependent properties. Chapter 7 presents an application to a
multivariable cooling system; from the process control domain. The monograph
closes with Chap. 8 that gives assessments and advice for using standard widely
available software in this field.

Glasgow, Scotland, UK Michael J. Grimble

M. A. Johnson
Industrial Control Centre
University of Strathclyde
Preface

Perturbation (input) signal design plays a very important role in system identifi-
cation for control applications. A good design can lead to maximally informative
experiments which reduce operational costs associated with identification tests.
Many different approaches to the design of perturbation signals have been described
over the years. However, the selection of the most suitable signal for a particular
industrial process application remains a challenging task. This monograph helps the
reader to understand the different designs that are particularly relevant in an
industrial context and guides the reader to make well-informed choices. The major
aspects of perturbation signal design such as the formulation of suitable specifi-
cations in the face of practical constraints, the classes of designs available, the
various objectives necessitating separate treatments when dealing with nonlinear
systems and extension to multi-input scenarios are considered. Two Case Studies
are included to provide a good balance between theory and practice. Several
software packages which are readily available on the Web are discussed, so that
readers can access designs for their particular application and do not have to go
through the time-consuming procedure of designing signals for themselves.
In the measurement of the dynamics of a system, a perturbation signal is applied
to the input of the system, and the response to it is measured at the output. These
input and output signals are then processed to give an estimate of the system
dynamics. This may be in the form of either a nonparametric estimate or a para-
metric estimate. Examples of nonparametric estimates include impulse responses or
step responses in the time domain or frequency responses in the frequency domain,
while examples of parametric estimates include continuous or discrete transfer
functions and state-space descriptions.
In practice, there are many important questions about the design of perturbation
signals that need to be considered. Among these are the following:
• If the system is noisy, then to obtain sufficiently accurate estimates of the system
dynamics, it is necessary either to average over a long period of time (which
may not be feasible in industrial systems) or to increase the perturbation signal
amplitude, which means that any assumption of linearity may no longer be

ix
x Preface

valid. Thus, are there signals that enable the effects of any nonlinearities to be
minimised, so obtaining a good linear model of the system?
• Is it possible to design perturbation signals that are particularly suitable for
characterising various aspects of the nonlinear behaviour itself?
• What are the possible signal designs for the general multi-input scenario as well
as for cases where the process is ill-conditioned?
This book is intended for industrial engineers especially control engineers, and
for researchers in process control and communication engineering; also for aca-
demics wishing to learn about the most recent results in perturbation signal design
and final year undergraduate and postgraduate students in electrical, mechanical and
communication engineering wishing to learn and apply perturbation signal design
in project work. It will be particularly helpful to those seeking guidance on
choosing identification software tools for use in practical experiments and Case
Studies.
The three most important aspects of this book are:
• Several software packages are discussed and explained so that the reader finds it
easy to get started on designing some common classes of signals.
• Experiments on real systems are discussed which provide readers with insights
on how to design suitable perturbation signals in practical scenarios.
• There is a lack of books on perturbation signal design and such a book has been
greatly overdue since many new designs have appeared in recent years.
There are eight chapters in this book. A summary of what is included in each
of these follows.
Chapter 1 sets the scene for the rest of this book. It starts with a short review of
early applications (in the 1960s) of system identification using periodic perturbation
signals to full-scale industrial processes and nuclear power plants; more recent
applications of identification techniques in industry feature in the remainder of this
book. The role of signal design in identification for control is then expounded. Time
domain and frequency domain specifications for the identification of linear systems
are explained. Performance measures for linear system identification are described.
The identification framework is then extended to nonlinear systems where the
concept of harmonic suppression is explained. The chapter concludes with a
comparison between periodic and nonperiodic signals.
Chapter 2 considers the design of pseudorandom signals for linear system
identification; these have fixed autocorrelation functions, and therefore fixed
spectra, dependent only on the class and period of the signal. The theory behind the
generation of several classes of pseudorandom signals is discussed and their
properties are explained. These include maximum length binary signals based on
shift register sequences, as well as several other classes of binary and near-binary
signals. The maximum length design extends to the multilevel case, where the
sequence-to-signal conversion determines not only the harmonic properties but also
the period of the resulting signal. An interesting class of truncated signals arises
Preface xi

from certain well-defined choices. The design of direct synthesis ternary signals and
the suboptimal direct synthesis ternary signals is discussed. The chapter concludes
with an application example for linear identification of a system in the presence of
nonlinear distortion.
Chapter 3 discusses the design of computer-optimised signals for linear system
identification. Unlike pseudorandom signals which have fixed spectra, the objective
now is to design a signal with a spectrum as close as possible to a specified
spectrum. The optimisation algorithms come in different forms, depending on the
class of signal, and particularly the number of signal levels. The classes of signals
considered include multisine sum of harmonics signals which can take any value
between their minimum and maximum, discrete interval signals which are either
binary or ternary and multilevel multiharmonic signals which have a small number
of signal levels, where this number is specified by the user. It is then shown that it is
also possible to combine the advantages of pseudorandom and computer-optimised
designs leading to a class of hybrid signals, which are generated as a combination of
pseudorandom signals and computer-optimised ones. The concept of optimal input
signals, where the power spectra are optimised based on initial models to satisfy an
application-related objective, is briefly described.
Chapter 4 describes perturbation signal designs for multi-input system identifi-
cation, giving the theory behind the design of sets of uncorrelated signals; these
allow the effects of the individual inputs to be easily decoupled at the system
output. The designs include Hadamard-modulated signals and signals with a zip-
pered spectrum. The latter may be designed using multisine signals or pseudo-
random signals. An application example on a simulated thermoelectric system is
presented. An alternative approach using a phase-shifting design is described; here
only one signal is generated and phase-shifted versions of this signal are applied to
the multiple inputs. The identification of ill-conditioned processes is then discussed.
The problem of ill-conditioning is caused by the singular values having widely
differing magnitudes. The method of virtual transfer function between inputs for the
identification of ill-conditioned systems is explained. Its effectiveness is illustrated
on a simulated multizone furnace.
In Chap. 5, the identification setting is turned to systems with nonlinearities and
time-varying properties. For systems with nonlinearities, the signal design is shown
to be highly dependent on the objectives of the identification test. The identification
of the best linear approximation is discussed next and is shown to be very useful
when a nonlinear process is to be linearised around its operating point. While the
best linear approximation depends on the perturbation signal applied, those with
Gaussian amplitude distribution are advantageous particularly in the identification
of block-oriented systems. For the measurement of Volterra kernels, it is shown that
multisine signals with specially designed harmonics enable the kernels to be
measured without interharmonic distortion. The chapter concludes with an expo-
sition of a method based on frequency domain analysis which allows the quan-
tification of the effects of nonlinearities, noise and time variation. The technique
xii Preface

requires only a single experiment and in the case of multi-input systems, makes use
of a set of uncorrelated perturbation signals. The effectiveness of the technique is
illustrated on a mist reactor system.
Chapter 6 is the first of two Chapters devoted to a Case Study. In this example,
the identification of an electronic nose system having direction-dependent proper-
ties, where the dynamics depend on whether the output is increasing or decreasing,
is described. It is shown that the use of maximum length binary signals allows the
detection of the nonlinearities through the input–output crosscorrelation function
due to the shift-and-multiply property. When inverse-repeat maximum length
binary signals are applied, the effects of even-order nonlinearities can be eliminated.
The detection of the even-order nonlinearities is also possible through analysis
of the output spectrum. The best linear approximation of the system is estimated
using various methods in the time and frequency domains. It is also shown that the
identification tests lead to the estimation of a Wiener model for the system and that
the use of an inverse-repeat perturbation signal results in higher accuracy in the
parameter estimates.
Chapter 7—in this second Case Study Chapter, the application of uncorrelated
multisine signals for the identification of a multivariable cooling system with
time-varying delay is illustrated. The system has two inputs, one associated with the
flow control system and the other with the Peltier system, and a temperature output.
The suppression of harmonic multiples of two and three in the signals confirms that
the system is largely linear. The application of several periods of the perturbation
signals leads to the identification of a time-invariant model for the channel from the
Peltier system to the temperature output and the detection of a time-varying com-
ponent for the channel from the flow control system to the temperature output.
A delay reconciliation technique is applied to remove the relative delays between
individual periods of the output so that the averaged period becomes more repre-
sentative of the individual periods. Continuous-time modelling of the time-invariant
part of the time-varying channel is further carried out as it facilitates online adaptive
identification of the variable delay using a gridding approach which accommodates
fractional delay values.
Chapter 8 discusses and explains several software packages for perturbation
signal design so that the reader finds it easy to get started on designing some
common classes of signals. These packages include prs (for maximum length
binary and other binary and near-binary pseudorandom signals as well as direct
synthesis ternary signals), GALOIS (for pseudorandom multilevel signals), func-
tions in the Frequency Domain System Identification Toolbox (for multisine sig-
nals, discrete interval binary and ternary signals, and optimal input signals),
multilev_new (for multilevel multiharmonic signals) and Input-Signal-Creator (for
pseudorandom multilevel signals and uncorrelated signal sets formed from these).
We would like to thank Prof. H. A. Barker, Chin Leei Cham and Dr. Timothy
Yap and for their excellent collaboration over many years; part of this collaborative
work appears in this book. We would also like to acknowledge our many other
Preface xiii

collaborators, colleagues and friends for their suggestions, support and encour-
agement. We would like to express our gratitude to the Series Editors Prof.
Michael J. Grimble and Prof. Michael A. Johnson, Springer Editor Oliver Jackson
as well as the staff in Springer for making this book a reality. Last but not least, we
would like to thank the IEEE, the IET and Elsevier for permission to reproduce
copyrighted materials in this book.

Cyberjaya, Malaysia Ai Hui Tan

Coventry, UK Keith Richard Godfrey
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 1
1.1 Historical Perspective of Industrial Applications
of System Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Role of Signal Design in Identification for Control . . . . . . . . . . . . 2
1.3 Specifications for the Identification of Linear Systems . . . . . . . . . . 4
1.3.1 Time Domain Specifications . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.2 Frequency Domain Specifications . . . . . . . . . . . . . . . . . . . . 6
1.4 Performance Measures for the Identification of Linear Systems . . . 8
1.4.1 Performance Index for Perturbation Signals . . . . . . . . . . . . 8
1.4.2 Frequency Domain Measure of Signal Quality . . . . . . . . . . 11
1.5 Harmonic Suppression in the Presence
of Nonlinear Distortions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 13
1.6 General Comparison Between Periodic
and Non-periodic Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.6.1 Impulse Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.6.2 Step Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.6.3 Random Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.6.4 Comparison Between Periodic
and Non-periodic Signals . . . . . . . . . . . . . . . . . . . . . . . ... 21
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 23
2 Design of Pseudorandom Signals for Linear System Identification . . . 25
2.1 Maximum Length Binary Signals . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Other Binary and Near-Binary Signals . . . . . . . . . . . . . . . . . . . . . . 30
2.2.1 Quadratic Residue Binary Signals . . . . . . . . . . . . . . . . . . . 30
2.2.2 Hall Binary Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2.3 Twin Prime Binary Signals . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2.4 Quadratic Residue Ternary Signals . . . . . . . . . . . . . . . . . . . 33
2.3 Multilevel Maximum Length Signals . . . . . . . . . . . . . . . . . . . . . . . 37
2.3.1 Pseudorandom Signals with Maximum Length . . . . . . . . . . 37
2.3.2 Truncated Pseudorandom Signals . . . . . . . . . . . . . . . . . . . . 38

xv
xvi Contents

2.4 Direct Synthesis Ternary Signals . . . . . . . . . . . . . . . . . . . . . . . . . . 47

2.5 Application Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3 Design of Computer-Optimised Signals for Linear
System Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.1 Multisine Sum of Harmonics Signals . . . . . . . . . . . . . . . . . . . . . . . 59
3.2 Discrete Interval Binary and Discrete Interval Ternary Signals . . . 66
3.3 Multilevel Multiharmonic Signals . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.4 Hybrid Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.5 Optimal Input Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4 Signal Design for Multi-input System Identification . . . . . . . . . . . . . . 95
4.1 Uncorrelated Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.1.1 Modulation with Rows of a Hadamard Matrix . . . . . . . . . . 96
4.1.2 Design of a Zippered Spectrum Using Multisine Signals . . . 101
4.1.3 Design of a Zippered Spectrum Using Pseudorandom
Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.1.4 Application Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.2 Phase-Shifting Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.3 Identification of Ill-Conditioned Processes . . . . . . . . . . . . . . . . . . . 111
4.3.1 Problem of Ill-Conditioning . . . . . . . . . . . . . . . . . . . . . . . . 111
4.3.2 Virtual Transfer Function Between Inputs . . . . . . . . . . . . . 114
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5 Signal Design for the Identification of Nonlinear
and Time-Varying Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.1 Objectives of Identification of Nonlinear Systems . . . . . . . . . . . . . 129
5.2 Identification of the Best Linear Approximation . . . . . . . . . . . . . . 129
5.3 Identification of Volterra Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.4 Quantification of Effects of Nonlinearities, Noise
and Time Variation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
5.4.1 Method Based on Frequency Domain Analysis . . . . . . . . . 141
5.4.2 Application Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6 Case Study on the Identification of a Direction-Dependent
Electronic Nose System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.1 Description of System and Experimental Setup . . . . . . . . . . . . . . . 153
6.2 Detection of Nonlinear Distortion . . . . . . . . . . . . . . . . . . . . . . . . . 154
6.2.1 Detection Through Step Tests . . . . . . . . . . . . . . . . . . . . . . 154
6.2.2 Detection Through Crosscorrelation Function . . . . . . . . . . . 155
6.2.3 Detection Through Output Spectrum . . . . . . . . . . . . . . . . . 161
Contents xvii

6.3 Identification of Linear Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . 162

6.4 Identification of Wiener Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
7 Case Study on the Identification of a Multivariable Cooling System
with Time-Varying Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
7.1 Description of System and Experimental Setup . . . . . . . . . . . . . . . 175
7.2 Identification of Linear Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
7.3 Delay Reconciliation Using Crosscorrelation . . . . . . . . . . . . . . . . . 179
7.4 Offline Identification of Invariant Dynamics . . . . . . . . . . . . . . . . . 185
7.5 Online Adaptive Identification of Variable Delay . . . . . . . . . . . . . . 187
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
8 Software for Signal Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8.1 prs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8.2 GALOIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
8.3 Frequency Domain System Identification Toolbox . . . . . . . . . . . . . 201
8.3.1 Generation of Multisine Signals . . . . . . . . . . . . . . . . . . . . . 203
8.3.2 Generation of Discrete Interval Signals . . . . . . . . . . . . . . . 207
8.3.3 Generation of Optimal Input Signals . . . . . . . . . . . . . . . . . 207
8.4 multilev_new . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
8.5 Input-Signal-Creator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Symbols, Operators and Abbreviations

Symbols

Only frequently occurring symbols and their main usage are listed.

Ap Amplitude of harmonic p
c Coefficients of feedback shift register
fresolution Frequency resolution
fs Sampling frequency
F Number of specified excited harmonics
g Primitive element of GF
G FRF of a system, or
Transfer function of a system
H Hadamard matrix, or
VTFBI
i Discrete-time index
pffiffiffiffiffiffiffi
j
1
k Harmonic number, or
General index
M Moment of a signal
M Number of signal levels of MLMH signals, or
Number of periods of a periodic signal, or
Number of inputs of multivariable system
n Discrete-time index, or
Degree of feedback shift register
N Period of a periodic signal
q GF
Ruu Autocorrelation function of the signal u
Ruy Crosscorrelation function between u and y
sq,n PRML sequence from GF(q) with primitive polynomial of degree n
t Continuous-time index

xix
xx Symbols, Operators and Abbreviations

ts Sampling interval
TN Measurement period = N  ts
u Input signal in the time domain
U(k), U(jx) Input signal in the frequency domain
U(s) Laplace transform of input signal
U(z−1) z-transform of input signal
x Angular frequency
y Output signal in the time domain
Y(k), Y(jx) Output signal in the frequency domain
Y(s) Laplace transform of output signal
Y(z−1) z-transform of output signal
Z Set containing integers
/p Phase of harmonic p
σ Singular value, or
Standard deviation

Operators

AT Matrix transpose
A−1 Matrix inverse
det Determinant
E Expectation
mod(v,w) Modulo operator giving the remainder after the division of v by w
sup Supremum
} Complex conjugate of U
U
var Variance
⊕q Modulo-q addition

Abbreviations

ARMAX Autoregressive moving average with exogenous input

ARX Autoregressive with exogenous input
CDMA Code division multiple access
DFT Discrete Fourier transform
DIB Discrete interval binary
DIT Discrete interval ternary
ELiS Estimator for Linear Systems
EMINE Effective minimum ratio between the actual amplitude and the specified
amplitude at any of the specified harmonics
FRF Frequency response function
Symbols, Operators and Abbreviations xxi

GF Galois field
GUI Graphical user interface
HAB Hall binary
MAE Mean absolute error
MLB Maximum length binary
MLMH Multilevel multiharmonic
MSE Mean square error
NID No interharmonic distortion
PAPR Peak-to-average power ratio
PIPS Performance index for perturbation signals
PIPSE PIPS (effective)
PRB Pseudorandom binary
PRML Pseudorandom multilevel
PWM Pulse width modulation
QRB Quadratic residue binary
QRT Quadratic residue ternary
RGA Relative gain array
RMS Root mean square
SNR Signal-to-noise ratio
TPB Twin prime binary
VTFBI Virtual transfer function between inputs
ZOH Zero-order hold
Chapter 1
Introduction

1.1 Historical Perspective of Industrial Applications

of System Identification

In the 1960s, there was considerable effort to improve the operation and control of
existing industrial processes, and the design of new plant. This led to increasing
interest in the measurement of system dynamics.
Time domain methods found increasing use, but step response and pulse response
measurements often suffered from the drawback that, in the presence of noise, either
the perturbation signal had to be so large that the normal operation of the process was
affected, or the experimentation time needed to be so long that there were difficulties
in holding the process steady over the necessary time.
This led to interest being focused on the use of pseudorandom signals applied at
the input, and input–output crosscorrelation methods being used to obtain estimates
of the impulse response of the system. A surprisingly large number of applications to
full-scale industrial plant (including nuclear reactors) were reported in the literature
during the 1960s. These were listed by Godfrey (1969a), with Table 3 in that paper
listing eight applications of pseudorandom signals in the process industries (chemi-
cal, electricity supply, oil, paper and steel) and Table 4 listing similar applications to
six different nuclear power systems. All applications except one used binary signals,
the exception being the work of Chang et al. (1968), who used a five-level pseudo-
random signal on a cold rolling mill at the Port Talbot Steelworks in Wales, with the
objective of reducing the effects of nonlinearities on the (linear) impulse response
estimates.
One of the examples of the application of the crosscorrelation technique using
pseudorandom binary (PRB) signals to a full-scale industrial plant was conducted at
the British Petroleum Oil Refinery in Belfast (Godfrey 1969b), as part of an agreement
between BP and the National Physical Laboratory, UK. The process chosen was the
30-plate distillate splitter column with its sidestream stripper, which was part of
the integrated crude oil distillation and distillate hydro-desulphurising unit at this

© Springer Nature Switzerland AG 2019 1

A. H. Tan and K. R. Godfrey, Industrial Process Identification, Advances in Industrial
Control, https://doi.org/10.1007/978-3-030-03661-4_1
2 1 Introduction

refinery. Three inputs (feed flow, reboiler temperature, and feed-heater temperature)
were initially chosen on the grounds that they were the most relevant to normal
operation of the column and they were perturbed separately using a 255-digit PRB
signal. Three column temperatures were chosen as the output variables. The column
and the rest of the refinery were operating normally under closed-loop control, and
it was found that in this case, perturbing the feed flow had little effect on the output
temperatures.
Measurements for the remaining six input–output pairs were analysed further.
In each case, pulse responses were estimated and compared with corresponding
pulse responses from simple step tests. The comparisons for two of the measured
temperatures were very satisfactory, but one of the temperature measurements was
very noisy and the crosscorrelation results proved much more satisfactory in this
case. Discrete (z-domain) pulse transfer functions of the form

b1 z −1 + b2 z −2 + · · · + bk z −k
z −r (1.1)
1 + a1 z −1 + a2 z −2 + · · · + ak z −k

were also estimated from the input–output data using a generalised least squares
procedure. It was found that, with the relatively large sampling interval, an initial
time delay of r  0 gave the best results in terms of error between model and process
output, while fourth-order models (k  4) gave negligible improvement over the
second-order models (k  2). The main practical problems encountered in the work
were that the perturbation signal amplitudes had to be kept very small so that the
normal operation of the refinery was not disturbed, and that it was not easy to keep
the unit completely stationary over long periods of time.
As mentioned earlier, more recent applications of identification techniques in
industry will feature in the remainder of this book, as the techniques are described.

1.2 Role of Signal Design in Identification for Control

The selection of perturbation signals is an important step in the design of an experi-

ment for system identification. A better design results in more accurate models; these
subsequently lead to better controller performance. If the uncertainty in the estimates
is small, the controller can be tuned more closely to its maximum performance since
a smaller safety margin can be accommodated. Advanced control techniques such as
model predictive control (Mayne 2014) which have gained increasing popularity in
the industry rely heavily on model accuracy in maximising their capability. In model
predictive control, the control signal is optimised at every time sample making use
of the process model in predicting the output. Further to this, a carefully designed
signal allows detailed analysis of linear contributions, nonlinear distortions and dis-
turbances. This provides insights leading to well-informed decisions on whether to
apply a linear controller to a dominantly linear process, a set of linear controllers to
1.2 Role of Signal Design in Identification for Control 3

a linear parameter-varying process or a nonlinear control strategy to a process with

significant nonlinear behaviour.
In light of time constraints in conducting identification tests in industry, the signal
should be designed to allow a maximum amount of information to be collected
within the available experimentation time. In a model predictive control project, the
plant test and subsequent model identification often consume more than 50% of
the total project time (Darby and Nikolaou 2012). In this aspect, broadband signals
which simultaneously excite several frequencies within the bandwidth in a single
experiment are advantageous compared to frequency-by-frequency testing. This is
because for every experiment conducted, time is required for transient effects to
decay and frequency-by-frequency testing requires a separate experiment for every
frequency at which the system response is to be measured.
The objective of the identification should be taken into account as different objec-
tives necessitate different signal designs. For instance, the identification of nonlin-
earities in a system requires a different signal than one used to identify the underlying
linear dynamics of the same system with the effects of nonlinearities minimised. A
trade-off is frequently needed in the light of conflicting requirements and practical
constraints. A common scenario is in the selection of signal amplitude. By increasing
the amplitude, the signal-to-noise ratio (SNR) at the output improves but doing so
may excite unwanted nonlinearities. Another example is the selection of the number
of signal levels. A larger number of levels lead to better frequency domain perfor-
mance which is closer to the desired specifications. However, this typically results
in lower power in the signal if the amplitude is kept unchanged. Thus, it is important
to have many different signal designs to cater for different situations.
The focus of this book is on general purpose periodic broadband signals, where
minimal knowledge of the system under test is assumed. These signals are widely
applicable in many applications and specifically in identification for control. For
pseudorandom signals, the classes considered belonging to PRB signals are maxi-
mum length binary (MLB), quadratic residue binary (QRB), Hall binary (HAB) and
twin prime binary (TPB) signals; those belonging to ternary signals are quadratic
residue ternary (QRT) signals, pseudorandom multilevel (PRML) signals having
three levels, and direct synthesis ternary signals; and those belonging to multilevel
signals are PRML signals generated from the appropriate Galois fields (GFs) which
may have three or more levels. For computer-optimised signals, multisine, discrete
interval binary (DIB), discrete interval ternary (DIT) and multilevel multiharmonic
(MLMH) signals are suitable for general purpose testing. However, some designs
which are optimised based on the availability of a larger amount of a priori informa-
tion are considered in Sects. 3.5 and 4.3; these are most easily implemented using
multisine signals. Optimal designs find applications in reidentification of existing
control loops as well as in the identification of ill-conditioned systems. The advan-
tages of using periodic signals will be explained in Sect. 1.6.
4 1 Introduction

1.3 Specifications for the Identification of Linear Systems

1.3.1 Time Domain Specifications

Several common time domain specifications for perturbation signals are as follows:
• amplitude,
• number of signal levels,
• amplitude distribution, and
• signal period.
The maximum and minimum values of a perturbation signal are determined by
the allowable range of the input transducer and that of the system output. The latter
is influenced by safety considerations (for example, the temperature of a reactor),
possible changes in the system at large amplitudes (for example, the output may enter
a nonlinear range such as a saturation range in amplifier circuits), and maximum
deviations from the nominal value (for example, in a production plant operating in
closed loop).
For the identification of linear systems, a binary signal offers the best possible
choice in terms of maximising signal power within amplitude constraints. However,
this choice also comes with the least flexibility in terms of frequency domain specifi-
cations, and a trade-off may be necessary between the two conflicting requirements.
Additionally, the number of signal levels may be limited by hardware constraints.
Actuator limitation is a common cause; Barker and Godfrey (1999) described an
application to a continuous hot-dip galvanising process for steel strip where a maxi-
mum of three levels could be applied. A similar point was made by Mohanty (2009),
who applied ternary signals to excite a flotation column. Easier processing is also
an important consideration. Pinter and Fernando (2010) utilised binary signals for
the estimation of code division multiple access (CDMA) networks. Ternary signals
were applied in optical CDMA networks allowing a simple encoder to be used in the
control base station (Yang 2008). They were also applied for channel identification
of fibre wireless uplinks in both single-user (Ng et al. 2011) and multi-user (Ng et al.
2016) scenarios. In Roinila et al. (2009), the use of a PRB signal allowed a simple cir-
cuitry to be implemented on a switched-mode converter. In Tan and Godfrey (2004),
physical construction of the electronic nose system limited the signals to having at
most four levels.
The amplitude distribution may be determined by the signal class in some cases,
but in other cases, may be optimised by the user. For linear system identification,
a binary or near-binary amplitude distribution will result in the highest SNR at the
system output. The number of signal levels and amplitude distribution for the classes
of signals considered in this book are summarised in Table 1.1.
The period of a perturbation signal depends on the sampling interval as well as
the measurement time per period. In many cases, some indication of the dynamics of
the system under study is available from preliminary step tests or historical records.
1.3 Specifications for the Identification of Linear Systems 5

Table 1.1 Number of signal levels and amplitude distribution for various classes of signals
Class Number of levels Amplitude distribution
MLB, QRB, HAB, TPB 2 Almost uniformly distributed
across the two levels
Inverse-repeat MLB, QRB, 2 Uniformly distributed across
HAB, TPB the two levels
QRT and inverse-repeat QRT 3 Near-binary, with very small
number of occurrences of the
level 0, and equal number of
occurrences of the other two
levels
PRML and truncated PRML Depends on GF, with some Depends on harmonic
user flexibility specification
Direct synthesis ternary 3 Uniformly distributed across
the three levels
Suboptimal direct synthesis 3 Almost uniformly distributed
ternary across the three levels
Multisine Practically infinite Gaussian for random-phase
design, increasing
resemblance to binary signals
with increasing crest factor
minimisation
DIB 2 Almost uniformly distributed
across the two levels
DIT 3 Depends on harmonic
specification
MLMH Small number of levels Depends on harmonic
specification
Gallev Small number of levels Depends on harmonic
specification

Several guidelines exist on the selection of the sampling interval and measurement
period. A simple choice is to select the sampling interval t s such that
minimum time constant
ts ≤ (1.2)
5
and the measurement period T N such that

TN ≥ 5 × maximum time constant. (1.3)

The signal period N is then computed from

N  TN /ts . (1.4)

The choice of sampling interval sets the sampling frequency f s since

6 1 Introduction

f s  1/ts (1.5)

whereas the choice of measurement period sets the frequency resolution f resolution
since

f resolution  1/TN . (1.6)

However, the effective frequency resolution may be larger if some harmonics are
suppressed.
For broadband signals with uniform discrete Fourier transform (DFT) magnitudes
such as PRB signals, it is recommended to set the sampling frequency equal to 2.5
times the maximum frequency of interest (Pintelon and Schoukens 2012). Note that
throughout this book, it is normally assumed that the sampling interval is the same
as the clock pulse interval. In any case where this is not so, it will be explicitly
mentioned.

1.3.2 Frequency Domain Specifications

Several common frequency domain specifications for perturbation signals are as

follows:
• bandwidth,
• harmonic spacing, such as linear or (quasi-)logarithmic,
• amplitude spectrum or DFT magnitude of the excited harmonics, and
• with or without zero-order hold (ZOH) pre-compensation.
Computer-optimised signals offer greater flexibility in terms of frequency domain
specifications compared with pseudorandom signals. Pseudorandom signals have
uniform DFT magnitude across the whole spectrum with linearly spaced excited
harmonics which may also incorporate harmonic suppression. (Note that it is pos-
sible, via an incorrect design, to obtain non-uniform spectrum, but this is not really
what these signals are intended for.) Computer-optimised signals such as multisine,
DIB, DIT and MLMH signals are able to accommodate low-pass and bandpass
designs as well as linearly spaced and (quasi-)logarithmically spaced excited har-
monics. Multisine signals can exactly meet any specification in terms of amplitude
spectrum. In particular, the amplitude spectrum can be shaped to match that of an
optimal design (see Sect. 3.5). DIB, DIT and MLMH signals are not able to exactly
satisfy specifications on amplitude spectrum due to their limited number of levels.
It is interesting to note that the ideal amplitude spectrum depends on whether a
parametric model or a nonparametric model is to be estimated. In the former case,
more power should be inserted at frequencies which contribute most to the system
parameters. This is typically where the system gain is large as the measurements at
these frequencies have smaller uncertainties and will result in more accurate para-
1.3 Specifications for the Identification of Linear Systems 7

signal
ZOH actuator plant
generator

anti- anti-
alias alias
filter filter

measured measured
input output

Fig. 1.1 Band-limited identification setting

metric estimates. In the latter case, power should be distributed so that a minimum
predefined accuracy can be obtained across the bandwidth of interest.
Computer-optimised signals such as multisine, DIB, DIT and MLMH signals
may be designed either with or without ZOH pre-compensation for the ampli-
tude spectrum. For continuous-time signals, the choice depends on the intersam-
pling behaviour. There are two popular assumptions (Pintelon and Schoukens 2012;
Schoukens et al. 2018). The first is band-limited assumption, where the band-limited
data have no power above the Nyquist frequency due to the presence of anti-alias
filters. The signal from the generator is sampled with a very high sampling frequency
so that the ZOH has little effect on the spectrum. As such, no ZOH pre-compensation
is required in the signal design. The identification setting is shown in Fig. 1.1 where
it can be seen that both the input and output signals are measured after the anti-alias
filters. Additionally, in a practical system, noise may enter at various locations in
Fig. 1.1.
The second assumption is ZOH assumption, where the signal is piecewise constant
between samples. This is the typical assumption when performing identification for
discrete controller design. The identification setting is shown in Fig. 1.2, where in
a practical system, noise may also enter at various locations. There will be high-
frequency components due to the ZOH assumption and no anti-alias filter is used.
Since the input signal at the generator is known, the identification is done for the whole
system which includes the ZOH, actuator and plant. For this setting, the input signal
may be specified to include ZOH pre-compensation, in order to compensate for the
(sin2 x)/x 2 power spectrum envelope of the ZOH. With the ZOH pre-compensation,
the required amplitude spectrum appears at the input to the actuator.
In the case of a discrete-time signal being required, intersampling behaviour is
not applicable and the signal should be designed without ZOH pre-compensation.
8 1 Introduction

signal
ZOH actuator plant
generator

known measured
input output

Fig. 1.2 ZOH identification setting

The focus of this book is on the design of continuous perturbation signals in the
ZOH identification setting using the assumption that the signal is piecewise constant
between samples, due to its relevance to process control. However, the design of
continuous perturbation signals with the band-limited assumption and the design of
discrete perturbation signals are also covered, to a lesser extent.

1.4 Performance Measures for the Identification of Linear

Systems

For the identification of the underlying linear dynamics of a system, the perturbation
signal should satisfy two main criteria. The first is that the signal amplitude should be
sufficiently small in order to minimise the effects of nonlinearities. The second is that
the power contained in the specified harmonics should be sufficiently large in order to
achieve an acceptable SNR. Since these are conflicting requirements, a prime purpose
in signal design is to maximise the power in the specified harmonics within the ampli-
tude constraints. For comparing different signals against this objective, it is essential
to have a suitable performance measure; this is discussed in Sect. 1.4.1. Note that
for the identification of nonlinear systems, modifications to the performance mea-
sure may be needed depending on the application. Additional requirements such as
persistency of excitation in both amplitude and frequency domains may be required.
A frequency domain measure is considered in Sect. 1.4.2.

1.4.1 Performance Index for Perturbation Signals

The dispersion of a stationary signal u (Godfrey et al. 1999) can be expressed by a

normalised measure λu defined as
1.4 Performance Measures for the Identification of Linear Systems 9

u max − u min u max − u min

λu    . (1.7)
2σu 2 u 2rms − u 2mean

The signal may be continuous, with u = u(t) or discrete, with u = u(i). In Eq. 1.7,
umax , umin , σ u , umean and urms denote the maximum, minimum, standard deviation,
mean and root mean square (RMS) of the signal u, respectively. In particular, umean
and urms are defined by

u mean  E[u(t)] or E[u(i)] (1.8)

and
 
u rms  E[u 2 (t)] or E[u 2 (i)]. (1.9)

The dispersion λu has a lower limit of unity but no upper limit. A signal with
the least possible dispersion is a binary signal with a uniform amplitude distribution
between the levels umax and umin . λu is independent of both the mean and the amplitude
scale of u, which makes it an advantage over other measures such as crest factor,
peak factor and form factor, defined by

crest factor  u peak /u rms , (1.10)

with u peak  max[|u|],

u max − u min
peak factor  √ , (1.11)
2 2u rms

and

form factor  u rms /E[|u|]. (1.12)

These measures were originally used to quantify sine wave distortions, and were
not designed with perturbation signals specifically in mind. They have the additional
drawback of being defined differently by different authors (Godfrey et al. 1999).
Crest factor, peak factor and λu also share the common disadvantage that they have
finite lower limits (for signals with the least possible dispersion), but no upper limits.
However, crest factor is a widely used measure.
It is preferable to use performance indices that range from 0% (worst possible
performance) to 100% (best possible performance, corresponding to least disper-
sion). A suitable measure, the performance index for perturbation signals (PIPS), is
inversely proportional to λu and is given by (Godfrey et al. 1999)

200 u 2rms − u 2mean
PIPS  100λ−1  % , u max > u min . (1.13)
u
u max − u min
10 1 Introduction

PIPS is independent of the signal mean and amplitude scale. It is 100% for a
signal with the best possible performance. A PIPS of 100% can be achieved by
binary signals for which umax and umin have equal duration or number of occurrences
(uniform amplitude distribution).
PIPS is applicable to all stationary signals, but the special case of periodic signals
is particularly important. If u is a periodic signal, with period N × t s when u is
continuous or period N when u is discrete, then the expectations in Eqs. 1.8 and 1.9
are obtained by averaging over a period, so that

N ts
1 
N
1
u mean  u(t)dt  u(i) (1.14)
N ts N i1
0

and
 
 
 N ts 1  N
 1
u rms  u (t)dt  
2 u 2 (i). (1.15)
N ts N i1
0

In this case, PIPS can also be expressed in terms of the frequency content of u. If
the DFT of the discrete signal u(i) is U(k), defined as


N 
j2π ki
U (k)  u(i) exp − (1.16)
i1
N

then using Parseval’s theorem

 N −1
1 
N
u 2 (i)  |U (k)|2 (1.17)
i1
N k0

leads to
N −1 N −1
1  2 1  1 
N
u 2rms  u (i)  2 |U (k)|2  2 |U (k)|2 + u 2mean . (1.18)
N i1 N k0 N k1

Substituting Eq. 1.18 into Eq. 1.13 results in

N −1
200 k1 |U (k)|2
PIPS  %. (1.19)
N (u max − u min )

The absence of the zero harmonic from Eq. 1.19 confirms that PIPS is independent
of the mean of u.
Another random document with
no related content on Scribd:
 Arnold Schoenberg, Musical Anarchist

Arnold Schoenberg, born in Vienna (1874), taught himself until he

was twenty. He then studied with Alexander von Zemlinsky, who
later became his brother-in-law. Zemlinsky once pointed him out
saying, “He is in his early twenties and I have taught him all I know;
he brought me an orchestral work recently for which he had to paste
two pieces of score paper together to write out his score, so large an
orchestra had he employed!” This was his tone poem, Pelleas and
Melisande, first performed in 1904. To this early period belong some
songs, a song cycle with orchestra on texts by Jens Peter Jacobsen,
Gurrelieder, and the sextet, Verklärte Nacht (Illumined Night).
Both Richard Strauss and Gustav Mahler were his friends, and
through Mahler’s efforts many of his compositions were performed.
His string quartet was played in America by the Flonzaley Quartet.
His Chamber Symphony, and his second string quartet, with solo
voice, performed (1924) at the Berkshire Chamber Music Festival,
belong to this same period.
So far, all that Schoenberg composed was based more or less on
models of the past, but being naturally an anarchist in music he tried
to escape from doing what others had done. Instead of writing works
that took fifty minutes to play like his string quartet (in one
movement), he wrote five orchestral pieces and piano pieces that
were mere suggestions of compositions, so short were they. He cut
out all development of themes, all old forms, all feeling for tonality,
writing in the twelve-tone scale which we explained as atonality; he
built his chords in intervals of fourths instead of thirds, and
purposely changed all the rules of harmony; he distorted all the
intervals, using a seventh or ninth instead of the octave, and making
every fourth and fifth a half step larger or smaller than was
customary. His melodies are marked by large skips and queer
intervals, but when one once knows his language, by its very
queerness, it is easily recognized as Schoenberg’s. Although he has
broken away from the slavery of old traditions, he may have “jumped
out of the frying-pan into the fire”!
 In Pierrot Lunaire, a cycle of twenty-one songs with chamber
music accompaniment, he uses a curious effect for the voice “which
must be neither sung nor spoken.” This same effect he uses in chorus
in his music drama, Die Glückliche Hand (The Lucky Hand) for
which he also wrote the libretto. Although this and another music
drama Erwartung (The Awaiting) were begun in 1909, they were
both performed for the first time in 1924 in Vienna. This long delay
was due to the prejudice against the work of this innovator, who on
the one hand has been laughed at, scorned, and reviled, and on the
other praised to the skies by a small group of disciples and imitators
whose works sound very much like their teacher’s.
Among these pupils are Egon Wellesz (1885) who more than the
other disciples has broken away from the master. He has gone his
way in writing music for the stage and combining the old ideas of
ballet and orchestral music with Greek drama in a modern dance
drama. He has also written interesting chamber and orchestral
music. Dr. Wellesz is also an authority on musical history; he has
written many books and articles on the subject, especially on early
opera, Byzantine and Oriental music. He has written a book on
Schoenberg (1921).
Alban Berg (1885), also a Schoenberg pupil, has written unusually
fine chamber music and a new opera, Wozzek, fragments of which
were played at a festival in Prague (Czecho-Slovakia) in May, 1925,
by the International Society of Contemporary Music, a movement
most valuable in encouraging and developing modern music. This
society holds yearly meetings in Europe, at which are heard the
works of all the young composers of the world, each country having a
branch, which sends its share of new works to make the festival’s
programs.
Others of the Schoenberg group are Anton von Webern (1883),
Paul Pisk, Karl Horovitz (1884–1925), Ernest Krenek, and Ernest
Toch (1887).
 Erich Korngold

Erich Wolfgang Korngold (1898) startled the musical world, just

before the War, by the astonishing compositions he wrote as a little
boy. Among these were orchestral works and a piano sonata of
extraordinary promise. He was born in Vienna and is the son of a
musician and musical critic. Young Korngold is known in America as
the composer of Die Todte Stadt (The Dead City) an opera in which
the soprano, Maria Jeritza, made her first appearance at the
Metropolitan Opera House. In many ways the opera goes back to the
old pre-Wagner form and is full of melody, unusual in a young 20th
century composer! He has written other operas bordering on the
lighter Viennese operetta and has kept away from the Schoenberg
influence.
 Modern German Music

Richard Strauss was the last of the great classic school of German
composers, which for two hundred years had led the world in music.
Curious as it seems, he has not influenced directly the younger
composers, who turned to Debussy, Busoni, Schoenberg and
Stravinsky. (Page 410.)
 Busoni the Great

Ferruccio Busoni (1866–1924) although an Italian, had a strong

influence in two fields of German music, that of piano playing and
composing. He lived in Berlin and was one of the brilliant thinkers
and musicians of the period. He left chamber music and orchestral
works, also several operas, one of which, The Harlequin, finished
just before his untimely death, combines traditional form with
radical ideas. His sonatinas for piano and a set of studies on
American Indian Themes are important. He made a deep study of all
methods, old and new, and gave his pupils the advantage of this wide
experience.
Although the young Germans are not copying the huge symphonic
form of Bruckner and Mahler, these two have gained greatly in
popularity and are serving as models. Hans Pfitzner (1869), opera
composer, is one of the most German of the living composers of the
pre-war period; Franz Schreker (1878), an Austrian, living in Berlin,
has taught many of the younger composers. He writes operas and
songs. Schoenberg, although in Vienna, is felt even in Berlin.
 Hindemith

Of the young Germans, Paul Hindemith is the most important. He

was born in 1895 and according to Riemann, “is the freshest and
most full-blooded talent among the younger German composers.” He
seems to satisfy the two factions, for he is not too radical for the Old
or too old-fashioned for the New, so as Lawrence Gilman says, “he
carries water on both shoulders.... He seems to be able to write
polytonally or atonally if he chooses, and also to write as the
Academics might observe, like a gentleman. Richard Strauss is
reported to have said to him: ‘Why do you write atonally when you
have talent?’”
Today he is viola player in the Amar Quartet, but he has played in
cafés, in the “movies,” dance halls, operetta theatres, and jazz bands!
Although only thirty, he has many chamber music pieces to his credit
and three dramatic works. His success has been tremendous.
A society to further an interest in the new music was founded by
Hermann Scherchen and Eduard Erdmann. Scherchen created a
sensation in Berlin just before the war by conducting Schoenberg’s
Pierrot Lunaire, and after having been a prisoner of war in Russia he
came back with renewed purpose of bringing the new music to the
public. He has published a few songs and a string quartet. His right-
hand man Erdmann, besides being a pianist, has written a
symphony, the first attempt of a youth without orchestral experience,
which astonished the audience as a combination of Richard Strauss,
Schoenberg, “to which is added a portion of genuine Erdmann
flavor,” says Hugo Leichtentritt.
Another young German is Heinz Tiessen (1887), who is writing
besides piano music in atonality, incidental music to a drama by
Hauptman, and songs.
Philipp Jarnach (1892), a pupil of Busoni, Spanish by birth,
educated in Paris, lives in Berlin and writes in the new style. Kurt
Weill (1900) is also a gifted Busoni-ite.
Ernst Toch (1887), Viennese by birth, who lives in Germany has
written string quartets, sonatas, concertos, and a symphony.
 Heinrich Kaminsky is accepted in Germany as the composer who
is trying to build a bridge from Bach to modern times. His Concerto
Grosso for double orchestra commands great respect.
 Hungary—Bartók and Kodály

Béla Bartók (1881) and his friend Soltan Kodály (1882) have done
much to bring Hungarian folk music into the modern world, for they
are steeped in folk tunes, which they use with skill and imagination.
Bartók has written a short opera, two ballets, orchestral works, string
quartets, violin sonata, and many piano compositions. His children’s
pieces are delightful, based as they are on Hungarian folk tunes.
We have spoken at length of the gypsy music of the Hungarians
brought to us by Brahms, Liszt and Sarasate (violinist and
composer). We also told you that the Hungarians were Magyars.
Adjoran Otvos, in the League of Composers Review says: “Bartók
and Kodály have accomplished a pioneer work of quite a different
nature, an exploration into the folk music of Hungary which has
yielded a collection of historic significance, the most important and
only authentic one made in that country.
“Bartók, poor and supported only by a scholarship, started in 1905,
an investigation of the music of his race. Spending a week with a
friend in the country, he heard a servant, while at work, singing a
tune quite different from the hybrid (mixed breed) gypsy airs which
pass for Magyar music, in Hungary and elsewhere. He contrived to
conceal himself and day after day, while the servant worked,
recorded a number of songs whose primitive character, he at once
recognized. With this impetus, he embarked on a tour which lasted
over two years, as long as his money held out. On his journeys among
the peasants he met Kodály, out on a similar mission of research.
Without previous inkling of each other’s aims, they proceeded
together, recording the ancient songs of the Magyars in the
compilation which is famous today.”
 Ernest von Dohnányi

Ernest von Dohnányi (1877) a noted pianist and composer of

Hungary has spent most of his life in Berlin and has toured Europe
and America in piano recitals. He has written many works for
orchestra, chamber music, piano and opera, all of which show more
influence of Brahms than of men of his own land. He has been
engaged as conductor of the State Symphony Orchestra of New York
for the season 1925–26.
A twenty-eight year old pupil of Béla Bartók, Georg Kosa, shows
decided gifts in his first orchestral work, Six Pieces for Orchestra.
 Czech School

The Czech school founded by Smetana and Dvorak and Zdenko

Fibich (1850–1900) and continued by Vitezslav Novak (1870), Josef
Suk (1874), and Vaclar Stepan (1889), has had a rebirth in the 20th
century. Leos Janacek, although over seventy, is the leading spirit;
Rudolf Karel (1881), a pupil of Dvorak, Bohuslav Martinu, a follower
of Stravinsky, and Ernest Krenek (1902), a pupil of Schreker, and
Alois Haba (1893), pupil of Novak and Schreker are the working
forces. (Janacek died in 1928.)
 The Quarter-Tone Man

Alois Haba first wrote chamber music, then he tried some

interesting experiments for which he is known as the “quarter-tone
man.” We have heard of quarter-tones among the Hindus and Arabs
(Chapter VI) and as the human ear has become more educated, the
possibility of dividing the scale into quarter-tones is much discussed,
and seems to be the next step in developing music along the line of
overtones (see above). Did you ever realize that as with eyes that are
far-sighted or near-sighted, ears may vary too, in the amount they
hear? Most people think that every one hears alike, but this is not so.
Stravinsky was one day sitting with a friend on the shore of a Swiss
Lake near which he lived. The friend said the water was calm and
still, but Stravinsky heard, a definite musical sound! Many of these
musical sounds unheard by our ears he has shown us in his music. In
the same way it is said that Haba has an extraordinarily keen ear and
in trying to express what he hears, he has written two string quartets
in the quarter-tone system. Stringed instruments are not in
tempered scales and lend themselves to any division of the interval,
into third-tones, as Busoni tried, and quarter-tones as Haba has
written. But he has gone further and has made a piano on which
quarter-tones may be played. This may prove to be the basis of music
of the future, or it may be merely one of the numerous experiments
without lasting value.
Arthur Honegger.
(Swiss-French)
Darius Milhaud.
(French)
Courtesy of “The
Musical
Quarterly.”

Béla Bartók.
(Hungarian)
 Photograph, Victor
Georg.

Louis Gruenberg.
(American)

Composers of Today.
Courtesy of “The
Musical Digest.”

G. Francesco Malipiero.
(Italian)
Courtesy of “The
Musical
Quarterly.”

Alfredo Casella.
(Italian)
Photograph,
Mendoza Galleries.

Arnold Bax.
(English)
 Photograph,
Bertam Pach.

Eugene Goossens.
(English)

Composers of Today.