Computer-aided Nonlinear Control System Design

Amir Nassirharand

Computer-aided Nonlinear
Control System Design
Using Describing Function Models

123
Amir Nassirharand
Faculty of Engineering
The University of Nottingham
Malaysia Campus
Jalan Broga
43500 Semenyih
Malaysia

ISBN 978-1-4471-2148-0 e-ISBN 978-1-4471-2149-7

DOI 10.1007/978-1-4471-2149-7
Springer London Dordrecht Heidelberg New York

British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library

Library of Congress Control Number: 2011942925

© Springer-Verlag London Limited 2012

Apart from any fair dealing for the purposes of research or private study, or criticism or review, as
permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced,
stored or transmitted, in any form or by any means, with the prior permission in writing of the publishers,
or in the case of reprographic reproduction in accordance with the terms of licenses issued by the
Copyright Licensing Agency. Enquiries concerning reproduction outside those terms should be sent to
the publishers.
The use of registered names, trademarks, etc., in this publication does not imply, even in the absence of a
specific statement, that such names are exempt from the relevant laws and regulations and therefore free
for general use.
The publisher makes no representation, express or implied, with regard to the accuracy of the information
contained in this book and cannot accept any legal responsibility or liability for any errors or omissions
that may be made.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

To Shabnam, Sam, and Daniel who suffer as
I do my research
Of course to Nicole, Jessie, Kaydence, and
Kiera
To my mother, Shokat Vahdat, and my late
father, Hassan, who are proud of this work
Preface

The aim of this book is to give an in-depth treatment of computer-aided design

of nonlinear control systems using describing function models. The approach
is systematic, and it is based on a describing function platform. A large class
of nonlinear systems is covered with no restrictions on the number, type, and
arrangement of nonlinear terms, system order, and the number of inputs and/or
the outputs. Unlike other books in the area of nonlinear control that are highly
mathematically oriented, this book is software oriented.
It is believed that today’s tight control system performance requirements do not
allow a linear frame of mind for the designer. The designer’s first choice to linearize
a nonlinear model places so many unrealistic restrictions on the real model that
drives up the costs of final product.
A full nonlinear frame of mind for design of nonlinear feedback control systems
requires a computer-aided design approach. Many experts argue that such an
approach would not guarantee stability. Here, without loss of generality, stability is
demonstrated using the fact that if the sinusoidal input describing function models
of a nonlinear system is obtainable, then that nonlinear system is stable around the
operating regimes of interest.
I like to emphasize that having a full nonlinear frame of mind does not mean that
designers should throw away other available linear and more restricted nonlinear
control literature. On the contrary, one should capitalize on others’ intellectual
developments to bring good low-cost products for mankind. This book is a vehicle to
account for nonlinear effects, and it simultaneously enables the designer to interface
with other control system design works.
A person with a thorough knowledge of calculus and an elementary knowledge
of numerical analysis and software engineering would most benefit from this book.
This book should be of interest to aerospace, chemical, electrical, mechanical, and
mechatronics students and practicing control engineers as well as other practical
engineers and managers who are simply interested in design of low-cost real-life
nonlinear feedback systems.

vii
viii Preface

®
Many key MATLAB functions are developed and presented to allow easy design
®
of nonlinear feedback systems. The algorithms and the corresponding MATLAB
functions are demonstrated by solving example problems of the sort encountered in
mechatronics and aerospace. A versatile controller design software based on the
®
presented describing function design platform (using MATLAB Graphics User
Interface (GUI)) is under development which will be commercially available by
early 2013.

Semenyih, Malaysia Amir Nassirharand

Acknowledgments

First, I like to sincerely thank my mentor, teacher, friend, counselor, adviser,

and academic father, Professor James H. Taylor, for his continued support and
encouragement throughout my academic and professional life. He has always been
there to answer my questions.
I also like to thank my students. Special thanks to Professor Abhijit Patwardhan
whose contributions on search of the parameter in simultaneous stabilization theory
shed light on this important problem while he was pursuing his Master’s degree at
the University of Kentucky; my apologies if I was too hard on you! Mr. Li See Chew,
my Ph.D. student, who is working on extension of some of the works presented here
to the unstable case in a direct way and who is a pioneer in experimental work among
my group members, deserves special thanks for putting up with my supervision.
I sincerely thank Mr. Sze Hong Teh (University of Nottingham – Malaysia Campus,
(UNMC)) for extending my Ph.D. work on dual-range linear controller design from
the single-variable nonlinear case to the multi-variable nonlinear case. His other
works on idle speed control of automobile engines, bank-angle control of aerospace
vehicles with hard nonlinearities, and initiation of projects in chemical engineering
and continuing to play with the spacecraft attitude control problem have contributed
to preparation of this monograph. His collaboration in developing a versatile
software package to support this monograph is acknowledged. I certainly have been
lucky to have the opportunity to work with you. Support of Brandon (my Master’s
degree student at UNMC) is appreciated; I appreciate his aggressiveness to do his
thesis on application of describing function techniques to nonlinear multivariable
problem of satellite attitude control. Finally many thanks to all those undergraduate
Mechanical Engineering and Mechatronics students at UNMC (Ahmed, Ajitpal,
Aimen, Arun, Athira, Awaii, Bartholomew, Bushra, Chan, Chin, Gift, Hammad,
Khawja, Lee, Lim, Mohd, Muhammad, Prem, Saeid, Sameer, Sanz, Sean, Shola,
Tan, Tee, Yuvanraj) who selected my projects in the dynamic systems and control
field as their final year theses, final year design projects, and final year group
development projects.
Many thanks to Dr. Karimi and his research group (Norouz, Seyed Mehdi,
Seyed Reza, Ali, Davood, : : : ) at K.N.T. University of Technology who taught me

ix
x Acknowledgments

the practical aspects of modeling, simulation, and control of engineering systems;

they substantially contributed to my professional growth that is reflected in this
monograph. Contributions of Seyed Reza in solving the linear and nonlinear H1
control problems and nonlinear lead-lag design problem given in this monograph
are greatly appreciated. Contributions of Norouz in application of algebraic linear
multivariable controller design approach of C. T. Chen to describing function-based
single range controllers are also acknowledged.
I like to express my sincere appreciation to Mr. Hadjirasouliha (former President
of the Farab Company) and other personnel (Ardavan, Ata, Farshid, Farzan,
Massoud, Mehdi, Mohamad Reza, Pouria, Samira, : : : ) at Farab for supporting me
and allowing me to do my research while I was doing my engineering tasks.
Special thanks to Professor Seamus Garvey and Professor Michael Wood of
University of Nottingham (UK Campus) for their invaluable support of my teaching
in terms of providing their lecture notes and powerpoint slides as well as answering
my detailed questions on Systems Engineering, Design of Experiments, Dynamic
Systems, Reliability, and Robustness; this helped me allocate the extra time required
to do my research and complete this monograph.
I appreciate the help, guidance, and supervision of Dr. Stewart McWilliam and
Andrew Spowage while at UNMC. Special thanks to Professor Mike Cloke (former
Dean) and Professor Ian Pashby (CEO) for giving me the opportunity to continue
my academic career at UNMC.
The support, guidance, direction, and critique of Professor James Rowland
(University of Kansas) of my initial proposal for my Ph.D. studies, which led
to my pursuit of controller design for nonlinear systems, are acknowledged. His
undergraduate text on control systems inspired me to pursue and develop closed-
form solutions for design of linear lead-lag, linear PID, and other classical linear
controllers (that are given in the Appendices) in an algorithmic and systematic way.
I also wish to express my gratitude to Mr. Oliver Jackson (Springer-Verlag
London) for keeping an open mind in making the decision to go ahead with
publication of this monograph. His understanding and support for the need for
a software-oriented approach to design of nonlinear feedback control systems is
noteworthy. It is noted that without the help and support of Ms. Charlotte Cross
(Springer-Verlag London), this manuscript would not have been completed in time;
special thanks to Ms. Cross for her help in obtaining the required permissions for
use of copy righted material.
Last but not least, supports of my sisters Zohreh and Gita are appreciated. I also
thank Mr. Hassan Bojnordiazad and Nasrin Rayat as they have touched my life in a
positive way. Their encouragement keeps me going. I like to thank John and Patsy
Hiser for contributing to my spiritual growth.
Thank you all!
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.1 What’s New? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.2 Design Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.3 Objectives .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3
1.4 Software.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4
1.5 Organization of the Monograph . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4
2 Frequency Domain Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5
2.1 Early Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5
2.1.1 Example: Limiter . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6
2.2 Modern Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9
2.2.1 Single-Input Single-Output Case . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9
2.2.2 Multi-input Multi-output Case . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 13
3 Single-Range Controller Design . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 23
3.1 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 23
3.2 Example – Single-Variable Case: A Decoupled Liquid
Propellant Engine.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 25
3.3 Example – Multivariable Case: Idle Speed Control Problem . . . . . . . . 34
4 Dual-Range Controller Design . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 43
4.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 43
4.2 Controller Synthesis .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 44
4.3 Software.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 46
4.4 Example 1: A Single-Variable Servo System . . . . .. . . . . . . . . . . . . . . . . . . . 46
4.5 Example 2: Multivariable Bank Angle Control Problem .. . . . . . . . . . . . 48
5 Multirange Nonlinear Controller Design . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 57
5.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 57
5.2 Controller Synthesis .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 57
5.2.1 Describing Function Inversion (Nassirharand 2009a)* .. . . . . 60
5.3 Software.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 66
5.3.1 Software for Step 5 . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 66

xi
xii Contents

5.3.2 Software for Describing Function Inversion

(Nassirharand 2009a) .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 69
5.4 Example: Single-Variable Case: Servo System . . .. . . . . . . . . . . . . . . . . . . . 72
5.4.1 Nonlinear PID Controller Design . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 72
5.4.2 Nonlinear Lead-Lag Controller Design
(Nassirharand and Mousavi Firdeh 2008) .. . . . . . . . . . . . . . . . . . . 75
5.4.3 Nonlinear H1 Controller Design. . . . . . . . .. . . . . . . . . . . . . . . . . . . . 78
5.5 Example: Multivariable Case: Bank Angle Control of a UAV . . . . . . . 82
6 Experimental Study: Unstable SISO Systems . . . . . . . .. . . . . . . . . . . . . . . . . . . . 87
6.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 87
6.2 Experimental Rig .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 87
6.3 Nonlinear PID Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 88
6.4 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 95
Appendix A Closed-Form Solution to Linear Classical Controllers.. . . . . 97
A.1 Linear Lead-Lag Controller Design . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 97
A.1.1 Software.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 99
A.1.2 Demonstration Example Problem.. . . . . . . .. . . . . . . . . . . . . . . . . . . . 100
A.2 Linear PID Controller Design.. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 103
A.2.1 Software.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 104
A.2.2 Demonstration Example Problem.. . . . . . . .. . . . . . . . . . . . . . . . . . . . 104
A.3 Proportional Plus Rate Feedback Controller Design.. . . . . . . . . . . . . . . . . 105
A.3.1 Software.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 109
A.3.2 Demonstration Example Problem.. . . . . . . .. . . . . . . . . . . . . . . . . . . . 111
A.4 Proportional-Integral Plus Rate Feedback Controller Design . . . . . . . . 112
A.4.1 Software.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 116
A.4.2 Demonstration Example Problem.. . . . . . . .. . . . . . . . . . . . . . . . . . . . 116
Appendix B Algebraic Linear Multivariable Controller Design . . . . . . . . . . 121
B.1 Theory.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 121
B.1.1 Discussion (Nassirharand and Karimi 2004c)** . . . . . . . . . . . . . . 125
B.1.2 Systematizing the Design Approach .. . . . .. . . . . . . . . . . . . . . . . . . . 127
B.2 Software.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 127
Appendix C Additional Example Problems for Multivariable
Single-Range Controller Design . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 129
C.1 Example Problem: One Combustion Chamber Liquid
Propellant Engine.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 129
C.2 Example Problem: Five Combustion Chamber Liquid
Propellant Engine.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 137
C.3 Example Problem: An Unstable Multivariable Robot Arm . . . . . . . . . . 146
Appendix D Dual-Range Controllers and Simultaneous
Stabilization Theory. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 153
D.1 Bézout’s Identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 153
D.1.1 Multivariable Case. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 154
Contents xiii

D.1.2 Single-Variable Case . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 155

D.1.3 Software.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 159
D.2 DRLCD Using Simultaneous Stabilization Theory .. . . . . . . . . . . . . . . . . . 159
D.2.1 Theory and Algorithm .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 159
D.2.2 Software.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 165
D.2.3 Example Problems.. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 165
D.2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 171
References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 175

Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 179
Chapter 1
Introduction

1.1 What’s New?

This is the first monograph on Computer-Aided Control System Design (CACSD)

of nonlinear systems based on the describing function design platform that was
originally proposed by Taylor in 1983. Since then, this platform has been developing
as reported in the open literature. Other key characteristics that make this a new
monograph follow:
• The considered class of nonlinear systems is of the following form:

x.t/
P D f .x; u; t / (1.1)

y.t/ D g .x; u; t / (1.2)

where x is the state vector of dimension n, u is the input vector of dimension m,

y is the system output vector of dimension l, and t is the time variable. There are
no restrictions on the number, arrangement (e.g., could be in forward or feedback
paths), and type of nonlinear terms (e.g., could be continuous or discontinuous),
system order, and the number of inputs and/or the outputs (e.g., could be square
or rectangular).
• Controller design for the considered class of nonlinear systems is describing-
function-based.
• All design methods are presented in an easy to follow algorithmic format.

1.2 Design Platform

The design platform is sinusoidal-input describing-function-based, and the structure

of the final closed-loop system is depicted in Fig. 1.1.

A. Nassirharand, Computer-aided Nonlinear Control System Design: Using Describing 1

Function Models, DOI 10.1007/978-1-4471-2149-7 1,
© Springer-Verlag London Limited 2012
2 1 Introduction

Fig. 1.1 The structure of the Command

considered nonlinear
feedback system

Error Input Output

Controller f p (.)

Linear or Nonlinear Plant

Nonlinear

Describing functions are used for frequency domain modeling. Then, depending
on the number of frequency domain models to be considered, a linear controller or
a nonlinear controller design approach is outlined. In summary, the design platform
considers three different cases: (1) single-range linear controller design (SRLCD),
(2) dual-range linear controller design (DRLCD), and (3) multi-range nonlinear
controller design (MRNCD).
When a designer starts designing a controller for a nonlinear system, he or she
is not sure if a linear controller would suffice or a nonlinear controller should be
designed. For this reason, it is recommended that designer should consider SRLCD,
DRLCD, and MRNCD in that order. In other words, at first SRLCD is tried; if
satisfactory results are not obtained, DRLCD should be executed. If DRLCD fails,
MRNCD is tried.
With reference to items noted below (Taylor 1983; Taylor and Strobel 1985;
Taylor and O’Donnell 1990; Nassirharand and Karimi 2006; Teh and Nassirharand
2010a), the following fact may be concluded. Sinusoidal-input describing function
(SIDF) models are the best of the kind of models that characterize the input-output
behavior of a nonlinear plant.
• Small-signal models, in contrast to the nonlinear plant, do not depend on
amplitude level of the excitation signal.
• Random input or other types of describing function models are not able to
characterize the dependency of the nonlinear plant on various frequencies of
interest.
• Design based on one or more SIDF models, which covers the range of excitation
amplitude levels, provides a solid basis for a robust design. This is because the
sensitivity of the plant behavior to the amplitude level of the excitation signal is
an important consideration in design of robust nonlinear feedback systems. The
input-output SIDF models, without sacrificing the performance of the feedback
system, may be utilized to assure robust system stability.
1.3 Objectives 3

• One parameter is used in characterizing the input-output SIDF models; this

parameter is the amplitude level of the excitation signal. This method of model
parameterization is in contrast to that of linearizing methods. In linearizing
methods, for each nonlinear term, one parameter, whose value is determined
by taking the derivative of the nonlinear term, is used; however, amplitude
of excitation signal characterizes each SIDF model. It should be noted that
robust design methods for nonlinear systems based on small-signal linear models
become very difficult to apply. In this case, the end result of the design is a
feedback system that reduces system performance in order to achieve absolute
stability.
• The controller synthesis method is applicable to nonlinear systems whose input-
output models in the frequency domain are sensitive to the amplitude level of the
excitation signal.
The primary goal of the controller synthesis is to create a closed-loop system
whose behavior is fairly insensitive to the amplitude level of the excitation signal.
In other words, the closed-loop system behavior would be insensitive to large
perturbations of the plant including uncertainties and those caused by disturbances.

1.3 Objectives

Limited efforts in the past have been undertaken to provide a versatile CACSD
setting for nonlinear control system design. This is primarily due to the fact that
a large body of nonlinear control literature considers special classes of nonlinear
systems. For example, geometric approaches assume Jacobian information is
available (Hunt et al. 1987). A good part of sliding mode control literature considers
nonlinear feedback systems that must be in companion form (Utkin et al. 1999).
Quantitative feedback theory (QFT) considers a similar class of nonlinear systems
to that considered in this monograph; however, its original formulation results in
high-order linear controllers (Horowitz 1976). Nonlinear controller design using
QFT approach is still developing.
The main objective of this monograph is to enable control engineering students
and/or practicing engineers to design robust nonlinear feedback systems using
a newly developed CACSD platform. Considering the fact that commercially
® ®
available control engineering software utilities such as MATLAB and MATRIXx
have advanced control engineering education and practice, it is safe to assume
control engineers feel comfortable solving problems using a personal computer.
Another goal of the monograph is to generate interest in advancing the presented
concepts to enable design of controllers for a larger class of systems. For example,
consideration of higher harmonic effects is a good future work.
4 1 Introduction

1.4 Software

The required software for design of robust nonlinear feedback systems (in form of a
MATLAB® Graphics User Interface) is under development. However, the available
software that is used to solve the example problems is outlined throughout the
monograph. The listings for some of the developed functions are also given.

1.5 Organization of the Monograph

In Chap. 2, describing function generation from state-variable models is dis-

cussed. Example problems for single-variable and multivariable cases are solved. In
Chap. 3, an algorithm is given for design of single-range controllers using describing
function models followed by algorithm description and numerical examples for both
single-variable and multivariable cases. In Chap. 4, an algorithm is given for design
of dual-range controllers using two describing function models followed by example
problems for both single-variable and multivariable cases. In Chap. 5, multi-
range nonlinear controllers are designed for both single-variable and multivariable
nonlinear systems; demonstration example problems of the sort encountered in
robotics and aerospace are given to see the typical results that may be achieved.
Finally, in Chap. 6, an experimental study of the nonlinear controller synthesis
procedure outlined in Chap. 5 is presented with application to an unstable single-
variable inverted pendulum problem.
Appendix A gives the closed-form solutions to the design of linear classical
controllers that are used to arrive at the nonlinear controllers described in Chap. 5.
Appendix B outlines an algebraic procedure for design of single-variable and
multivariable linear plants. Appendix C includes the solution to other example
problems encountered in single-range controller design. Appendix D includes
application of the dual-range linear controller design problem to linear systems with
integrity and nonlinear systems; the problems associated with the application of
simultaneous stabilization theory to nonlinear systems are discussed.
Chapter 2
Frequency Domain Modeling

2.1 Early Formulation

Here, a frequency domain model of a nonlinear system is defined to be the sinusoidal

input describing function (SIDF) model of that nonlinear system. For linear systems,
frequency domain models are only a function of excitation frequency. Generally
speaking, a frequency domain model of a nonlinear system is a function of both
amplitude and frequency of excitation (Gelb and Vander Velde 1968; Atherton
1975). This dependency on the amplitude of excitation is the most important
characteristic of a nonlinear system that must be taken into account when designing
robust closed-loop feedback systems.
The basic version of the SIDF modeling approach is based on the following
assumptions:
1. One single separable nonlinear term in the process.
2. The nonlinear term is time invariant. In other words, autonomous nonlinear
processes are considered.
3. The input to the nonlinearity term may be characterized by a sinusoid of the
following form:

u.t/ D A sin.!t /; (2.1)

where u is the input, A is the amplitude of excitation, ! is the frequency of

excitation, and t is the time variable.
4. Only the fundamental component of the output signal may be considered. In
other words, it is assumed that higher harmonic effects are negligible.
5. The nonlinearity is odd. This means that nonlinearity is symmetric about the
origin; e.g., f .u/ D f .u/.
With the above assumptions, the following approach may be used to obtain the
frequency domain model of a nonlinear process (Slotine and Li 1991). From the

A. Nassirharand, Computer-aided Nonlinear Control System Design: Using Describing 5

Function Models, DOI 10.1007/978-1-4471-2149-7 2,
© Springer-Verlag London Limited 2012
6 2 Frequency Domain Modeling

third assumption, assume a sinusoidal input. Then, approximate the output signal,
y.t/, by Fourier series.
1
a0 X
y.t/ D C Œan cos .n!t/ C bn sin .n!t/; (2.2)
2 nD1

where
Z
1
an D y.t/ cos .n!t / d .!t/; (2.3)



Z
1
bn D y.t/ sin .n!t / d .!t/: (2.4)



With reference to the fourth and fifth Assumptions, Eq. 2.2 becomes

y.t/  b1 sin .!t / : (2.5)

By definition, the SIDF model of a nonlinear process is the ratio of the funda-
mental component of the nonlinear process output by that of the input sinusoid.

b1 sin .!t/ b1
N.A/ D D : (2.6)
A sin .!t / A

Therefore, with the basic SIDF assumptions, the SIDF model for static operations
is not a function of frequency.

2.1.1 Example: Limiter

Consider the limiter nonlinearity shown in Fig. 2.1. The nonlinear element has two
regions. In order to find the describing function model of this nonlinearity, first
assume a sinusoidal input of the form u.t/ D A sin .!t/. Then, evaluate Fourier
integrals. Since the nonlinearity is odd, it is concluded that a0 D 0; a1 D 0.

y.t/ D b1 sin.!t/; (2.7)

where
Z
1
b1 D y.t/ sin .!t / d .!t /: (2.8)


2.1 Early Formulation 7

Fig. 2.1 The limiter

nonlinearity y
md

Asin w t u y(t)
d

For A  ı;

y.t/ D mu.t/ D mA sin .!t/ ; (2.9)

From the expression for b1

Z
mA
b1 D mA sin .!t/ sin .!t/ d .!t/ D mA: (2.10)



From Eq. 2.6,

b1 mA
N.A/ D D D m D N: (2.11)
A A

Therefore, for u.t/  ı, the gain of a saturation is simply the constant m.

For A > ı;
8
< mA sin .!t / 0  !t  !t1
y.t/ D  ; (2.12)
: mı !t1  !t 
2

where !t1 D sin1 .ı=A/. Hence,

Z Z2
1 1
b1 D y.t/ sin .!t/ d .!t/ D y.t/ sin .!t/ d .!t /
 
 0

Z!t1 Z=2
4 4
D mAsin .!t / d .!t / C
2
mı sin .!t / d .!t/
 
0 !t1
8 2 Frequency Domain Modeling

Fig. 2.2 The describing

function model of the limiter

Z!t1
4mA 1 4mı =2
D Œ1  cos .2!t/ d .!t/  cos .!t/ j!t1
 2 
0

2mA 2mA 4ma

D .!t/ j!t
0 
1
sin .2!t/ j!t
0 
1
Œ0  cos .!t1 /
  
2mA mA 4mı
D !t1  2 sin .!t1 / cos .!t1 / C cos .!t1 /
  
2mA 2mA ı 4mı
D !t1  cos .!t1 / C cos .!t1 /
  A 
" r #
2mA ı ı2
D !t1 C 1 2 : (2.13)
 A A

Again, using Eq. 2.6 with !t1 D sin1 .ı=A/,

"   r #
b1 2mA 1 ı ı ı2
N.A/ D D sin C 1 2 : (2.14)
A  A A A

The corresponding describing function plot is shown in Fig. 2.2. In this figure,
the axis labeled by “normalized amplitude” corresponds to A=ı, and the axis labeled
by “normalized gain” corresponds to N.A/=m (Gelb and Vander Velde 1968).
Now, let us interpret the results. The saturation element has a constant gain over
the linear region. This is shown by a horizontal line segment on the describing
function graph (Fig. 2.2). However, inside the saturation nonlinearity, gain decreases
as the amplitude of the excitation increases. Note that since gain multiplied by
input must equal a constant, gain must decrease as the amplitude of the excitation
increases. This behavior is shown in Fig. 2.2 by the curved region.
2.2 Modern Formulation 9

2.2 Modern Formulation

With the modern formulation (Taylor 1983; Taylor and Strobel 1985; Nassirharand
1987), the restrictions of the earlier formulation would be lifted. Also, the usual
SIDF modeling limitations (such as the requirement that the input to each nonlin-
earity must be sinusoidal) do not exist with the modern formulation. The approach
is based on simulation and evaluation of Fourier integrals. In majority of cases, the
mathematical model of a system would be required before the task of controller
design could begin. Even in those cases that a mathematical model is not required
for controller design purposes, a mathematical model would be required to verify
or validate design. However, with the presented design platform, if a mathematical
model is not available, then experimental results may be used (Li and Nassirharand
2010).
In reality, the implemented mathematical models of real-life systems are hardly
represented in an academic form. For example, the model would have if : : : then
: : : type statements that do not usually fall into an academic class. In other words,
most practical models are in terms of a computer code or a simulation model. In
general, most of the models may be represented in the general framework given by
Eqs. 1.1 and 1.2.
The computer models comprised of 5,000 lines of code are not unusual (Karimi
et al. 2003). Once the computer model is developed, one faces a question. This
question is how would one put this model in a form that available elegant controller
design techniques would apply? The first easy way out is to obtain a small-signal
model for the developed nonlinear model. In many cases, a linear model of a real-
life system does not exist (e.g., see the robotic system studied in Sect. 2.2.1.2
(Nassirharand 1987)). This is primarily because real systems include multivalued
and/or discontinuous nonlinearities whose derivatives do not exist; hence, a linear
model (by replacing each nonlinear term with a term whose gain is the slope of the
nonlinearity at the operating point) would not be obtainable.
One successful approach to put the complicated mathematical models (or
nonlinear models whose small-signal models do not exist) in a form suitable for
controller design purposes has been the describing function approach. The problem
statement follows. Given the state variable model of a nonlinear system (see Eqs.
1.1 and 1.2), how would one characterize the input/output behavior of that system
in the frequency domain?

2.2.1 Single-Input Single-Output Case

The stable plant is excited by a sinusoid of the following form:

u.t/ D u0 C a cos.!t/; (2.15)

10 2 Frequency Domain Modeling

where u is the input to the plant, u0 is the DC value of the input, a is the amplitude
of the excitation, and ! is the frequency of excitation. The equations of motion are
numerically integrated to obtain the output, y.t/. Once the output reaches steady
state, the Fourier integrals for period k are evaluated; these integrals are of the
following form:

ZkT
Im;k D y.t/e j m!t dt; (2.16)
.k1/T

where k D 1; 2; 3; : : : is the period index, m D 0; 1; 2; : : : is the harmonic index,

and T D 2=! is the period. I0;k is the constant or DC component of the response,
and the harmonic dependent transfer functions, Gm;k , are given by the following
relation:

Gm;k .j m!I u0 ; a/ D !Im;k =a: (2.17)

The Fourier integrals are assumed to have converged if the error in magnitude
and phase of the computed pseudo frequency response in two consecutive cycles
are small. The first harmonic SIDF models are given by G1;k . The flow chart for this
procedure is depicted in Fig. 2.3 (Nassirharand 1987).

2.2.1.1 Software
® ®
A MATLAB function is available whose primary inputs are (1) the SIMULINK
file name corresponding to the model of the nonlinear plant, (2) an array of
excitation amplitudes, and (3) an array of excitation frequencies. The primary
outputs are the real and imaginary parts of the describing function models as a
function of excitation amplitudes and frequencies. This software is based on an
earlier work of Nassirharand in 1987.

2.2.1.2 Example: A Servo System

Consider the servo system shown in Fig. 2.4, which was originally considered by
Taylor and Strobel 1985 and later by Nassirharand et al. 1988. The mathematical
model of the nonlinear plant is given by Eqs. 2.18–2.21.

xP 1 D x2 (2.18)

xP 2 D Tm =J (2.19)
2.2 Modern Formulation 11

Fig. 2.3 Flow chart for

describing function generator
START

FOR ALL DESIRED EXCITATION D

AMPLITUDES

FOR ALL DESIRED EXCITATION C

FREQUENCIES

K=K+1
(K = Period Index)

• EXCITE THE PLANT WITH

a cos (wt)
A
• INTEGRATE EQUATIONS OF
MOTION
• OBTAIN SYSTEM OUTPUT

8
< Te  fv xP 1  fc Sign .xP 1 / if jTe j > fc
Tm D Te  fv xP 1  fc Sign .xP 1 / if xP 1 ¤ 0 (2.20)
:
0:0 if jTe j < fc and xP 1 D 0

m1 Vin if jVin j  ı
Te D (2.21)
Sign.Vin /: .m1 ı C m2 .jVin j  ı// if jVin j > ı;

wherefc Coulomb friction, 0:1Nm; fv coefficient of viscous friction, 0:1Nms=rad;

J moment of inertia, 0:01Kg  m2 ; m1 a slope, 5Nm =V ; m2 a slope, 1Nm =V ; ı
saturation break point, 0:5V.
12 2 Frequency Domain Modeling

Fig. 2.3 (continued)

A

• INTEGRATE THE FOURIER

INTEGRALS

FOURIER N
INTEGRALS B
CONVERGED?

CALCULATE AND STORE

GAIN AND PHASE

D C

END

The computer model (Taylor and Strobel 1985) of this process in terms of a
®
FORTRAN subroutine is given by Nassirharand 1987. By executing the MATLAB
function described in the previous section, the generated describing function models
are obtained, and they are depicted in Figs. 2.5 and 2.6. The corresponding three-
dimensional plots are given by Figs. 2.7 and 2.8.
From examination of Figs. 2.7 and 2.8, it is apparent that the gain and phase
surfaces as a function of amplitude are not flat; hence, the system is not linear.
2.2 Modern Formulation 13

Vin Te
Saturation Stiction

Tm
x1 (t)
Nonlinear Plant Load 1/s

Fig. 2.4 The schematic model of the servomechanism example

Fig. 2.5 The pseudo Bode gain plot (a is the amplitude of excitation)

Fig. 2.6 The pseudo Bode phase plot (a is the amplitude of excitation)

2.2.2 Multi-input Multi-output Case

There are two important key points that must be kept in mind when characterizing
the input/output behavior of a multivariable nonlinear process. The first one is that
principle of superposition does not hold here; therefore, one is not able to determine
the effects of each input on the corresponding outputs and then sum the effects.
14 2 Frequency Domain Modeling

Fig. 2.7 The gain surface of the pseudo frequency response

Fig. 2.8 The phase surface of the pseudo frequency response

Hence, all input channels must all be simultaneously activated. The second key point
is that excitation frequencies of inputs could not be the same because the separate
effects of each input on each output would be indeterminate.
The characterization approach proceeds in the following fashion. The nonlinear
multivariable system given by Eqs. 1.1 and 1.2 is excited by input signals of the
following general form:
2.2 Modern Formulation 15

 
up .t/ D u0;p C ap cos !p t C p ; p D 1; 2; :::m; (2.22)

where u is the input, u0 is the DC component of the input, a is the amplitude of

excitation, ! is the frequency of excitation,  is the phase shift, and p is the input
channel index. Then, the dynamic equations of motion are numerically integrated
to obtain the outputs as a function of time, yq .t/; q D 1; 2; :::; l. Then, the matrix
h;k
Fourier integrals, denoted Iq;p , are evaluated.

ZkT
  
h;k
Iq;p D yq .t/ exp j h !p t C p dt; (2.23)
.k1/T

where h denotes the harmonic index, k denotes the period index of the Fourier
integral, q is the output channel index, and T is the overall period of the input
1;k
vector. Finally, the pseudo frequency response estimate, denoted by matrix Gq;p is
calculated as follows:
  2 1;k
1;k
Gq;p j !q I u0 ; a;  D I : (2.24)
ap T q;p

In order to analyze the higher harmonic effects, one sets h D 2; 3; :::.

The Fourier integrals are said to have converged if the following conditions are
satisfied:
ˇ ˇ
ˇ M k  M k1 ˇ
ˇ q;p q;p ˇ
ˇ ˇ < "M ; (2.25)
ˇ k
Mq;p ˇ
ˇ ˇ
ˇ k k1 ˇ
ˇ'q;p  'q;p ˇ < "' ; (2.26)

k k k
where Mq;p and q;p represent the log magnitude and phase of Gq;p .j!I u0 ; a;  /.

2.2.2.1 Software
®
A MATLAB function similar to the single-variable case is available whose primary
®
inputs are (1) the SIMULINK file name corresponding to the model of the
multivariable nonlinear plant, (2) a matrix of excitation amplitudes, and (3) a matrix
of excitation frequencies. The primary outputs are the matrices of real and imaginary
parts of the describing function models as a function of excitation amplitudes and
®
frequencies. At the nucleus of this software is a MATLAB function that generates
the matrix pseudo frequency response data for one amplitude and one frequency.
This function is given in Fig. 2.9 (Nassirharand and Karimi 2002).
16 2 Frequency Domain Modeling

Fig. 2.9 Listing for the multivariable sinusoidal input describing function generator. Source:
(Nassirharand and Karimi 2002) © Elsevier
2.2 Modern Formulation 17

Fig. 2.9 (continued)

18 2 Frequency Domain Modeling

Fig. 2.9 (continued)

2.2 Modern Formulation 19

Fig. 2.9 (continued)

2.2.2.2 Example: An Automobile Engine

An idle speed control model used by Bengea et al. 2004 is utilized to demonstrate the
idle speed controller design procedure and the typical results that may be achieved.
This model is given below:

dPm .t/ v Vd
C N.t/Pm .t/ D K10 ˛.t/; (2.27)
dt 4Vm
20 2 Frequency Domain Modeling

Fig. 2.10 Describing function model for the automobile engine – matrix gain plot. Source:
(Teh and Nassirharand 2010a) © Sage

dN.t/
J C N.t/ D k Pm .t   / C kı ı.t/ C f .t/; (2.28)
dt

whereJ engine inertia (0.0843 N  ms2 =rad); kı a coefficient relating engine

torque to the spark advance (2 ˙ 0.05 Nm =deg ); k a coefficient relating engine
torque to the intake manifold pressure (0.57143 ˙ 0.015 Nm =kPa ); K10 a constant
relating the idle throttle to the air mass flow rate past the throttle plate (110 ˙ 2.5
.kPa =s / =deg ); N.t/ engine speed, dependence on t may be omitted for simplicity
(rad =s ); Pm .t/ intake manifold pressure, dependence on t may be omitted for
simplicity (Pa); t time variable (s); Vd engine displacement (0.0046 m3 ); Vm
manifold volume (0.0029 m3 ); ˛.t/ idle bypass valve opening (deg); ı.t/ spark
timing advance (deg); v volumetric efficiency of the pumping process (0.55 ˙
0.025 dimensionless);  3.14 : : : .;  induction to torque delay at nominal idle speed;
N0 , (0.1 s); f disturbance torque .N  m/;  lumped damping coefficient (0.592 ˙
0.0125 Nm  s).
The parameter values and the associated uncertainties given in the table above
correspond to those of the Ford V8 4.6 Liter engine. This model and the corre-
sponding parameter values have been used quite frequently by many researchers
(e.g., Bengea et al. 2004) as a benchmark problem to test their controller design
procedures. Here, the same model parameters are used for illustration purposes.
®
A SIMULINK model of Eqs. 2.27 and 2.28 is developed. Then, the output time
histories for various sinusoidal inputs with u0;1 D 17:04deg, u0;2 D 0:2676 deg,
a1 D 0:01u01 rad =s , a2 D 0:01u02kPa, !1 D Œ0:1; 0:2; 0:3; 0:5; 0:8; 1; 2; 4; 7; 11;
2.2 Modern Formulation 21

Fig. 2.11 Describing function model for the automobile engine – matrix phase plot. Source:
(Teh and Nassirharand 2010a) © Sage

19; 32; 55; 94; 150rad =s , !2 D Œ0:2; 0:3; 0:4; 0:6; 0:9; 2; 3; 5; 8; 12; 20; 33; 56; 95;
151rad =s , and 1 D 2 D 0 deg are generated. Generally speaking, the user
should avoid frequencies that are simple multiples of one another (e.g., one is
twice as much as the other); however, our experience with various problems in
aerospace, Mechatronics, and automotive has revealed that in those specific cases
no major flaws in the Fourier analysis were present when isolated simple multiples
of frequencies were used. In continuation, the Fourier integrals are evaluated when
the outputs reach steady state; then, the SIDF model of the I.C. engine is generated.
This quasi-linear model along with system identification results is depicted in
Figs. 2.10 and 2.11 (Teh and Nassirharand 2010a). In these figures, the small circles
correspond to SIDF models, and the solid lines correspond to the linear fits obtained
®
using invfreqs function of MATLAB .
Chapter 3
Single-Range Controller Design

3.1 Procedure

In this part, a systematic procedure for design of single-range controllers is

presented; this algorithm is a modified version of its original version outlined by
Taylor (1983). The nonlinear and dynamic mathematical model of the system,
which may include both soft (continuous) and hard (discontinuous) nonlinearities,
is assumed to be available. The systematic controller design procedure is based on
describing function models of the plant coupled with any linear controller design
technique. It should be emphasized that, in general, small-signal models for this
class of nonlinear systems do not exist.
The control system design procedure is systematic, and it involves six steps as
given by the following algorithm.
Step 1 relates to user-defined performance measures, and the user must identify
the model of the process that he wants to mimic. In most cases, the desired
linear model would be a second-order transfer function that possesses the desired
dominant poles, and satisfy the desired steady state error conditions. The user may
translate the time and/or frequency performance measures to the desired natural
frequency and the desired damping ratio by considering the performance specifica-
tion equations noted in Rowland (1986); then the steady state error specifications
determine if a zero is also required. The zero may be determined from relations for
definition of a steady state error.
Step 2 requires the knowledge of the operating regimes of interest. Note
that unlike operating points, operating regimes are characterized by the range of
expected amplitudes and frequencies of the excitation signal. For example, for the
case of regulating the pressure of the combustion chamber of a liquid propellant
engine, one may use the steady-state values of the positions of the regulator
valve piston such that those values result in the desired values of the combustion
chamber pressures (Nassirharand and Karimi 2004a). Note that the mentioned
steady-state values of the position of the regulator valve piston are known by
design; alternatively, the control system designer may determine the desired steady

A. Nassirharand, Computer-aided Nonlinear Control System Design: Using Describing 23

Function Models, DOI 10.1007/978-1-4471-2149-7 3,
© Springer-Verlag London Limited 2012
24 3 Single-Range Controller Design

state values of the position of the regulator valve control piston by simulation and
by noting those values of the position of the regulator valve control piston that
result in the desired values of the combustion chamber pressure. The expected
frequencies of excitation are usually determined from the knowledge of the natural
frequency of the system. As a rule of thumb, the lower range would be about
two decades below the natural frequency of the system, and the upper frequency
range would be about one decade above the natural frequency of the system.
Alternatively, the user may initially set the lower and upper frequency ranges to
0.1 rad=s and 50 rad=s, respectively. Then, the pseudofrequency response plots are
generated. If the low-frequency response portion is not characterized, then the lower
frequency range may be lowered by one decade and so on until the low-frequency
response portion is characterized. The upper frequency range is also increased if
the high-frequency gain of the pseudofrequency response is substantial. Then the
Fourier-based approach described in the previous chapter would be used to obtain
the describing function models.

Algorithm 3.1
1. Identify the desired reference linear model.
2. Obtain the describing function models of the plant, and select a nominal
model.
3. Identify a linear model of the nominal frequency domain model of the
previous step (optional).
4. Design a controller based on any available linear control theory.
5. Verify design, and use optimization to tune the controller gains as neces-
sary.
6. If the results are satisfactory, then stop; otherwise consider designing a
dual-range linear controller.

In step 3, once the describing function models are obtained, one of these models
is selected as the nominal model. Normally, the nominal model would be the case in
which the excitation signal would correspond to the normal operating conditions of
the plant in the absence of disturbances, in the absence of loss of sensory devices,
or in the absence of any perturbations of the plant. Note that assumption is that one
describing function model of the system adequately describes the dynamic behavior
of the system; otherwise, a dual-range or a multi-range controller design approach
may have to be employed (see Chaps. 4 and 5). In any case, it is recommended
to study the effectiveness of a single-range controller before a more complicated
controller design procedure such as a dual-range or a multi-range nonlinear control
approach is adopted. In applications that nonlinearity effects are not dominant, the
spread in describing function models would not be significant, and therefore, a
single-range linear controller would suffice.
Step 4 requires identification of a linear model whose frequency response data
matches that of the nominal describing function model of the previous step. Since
3.2 Example – Single-Variable Case: A Decoupled Liquid Propellant Engine 25

the describing function models are representation of nonlinear systems, the standard
relation between the two components of the frequency response data that exists for
linear systems does not hold for describing function models. Therefore, care must
be taken when fitting the pseudofrequency response data. The outcome of this step
is a linear model described in terms of a transfer function.
Step 5 is to use the linear model of the previous step, and to interface with any
one of the available linear controller design techniques to design a linear controller.
For example, one may use factorization theory coupled with analytical optimization,
and obtain a controller design equation of the following form (see Appendix B for
details):

A0 X D B (3.1)

where A0 and B are known, and they are in terms of coefficients of coprime factors
of the linear model of the previous step, the coefficients of coprime factors that are
solution to the Bezout identity, the coefficients of the desired linear model of step 1,
and frequency range of interest; vector X contains the coefficients of the function
parameter, r.s/, that must be substituted in the celebrated Youla parameterization
equation (see Eq. B.5) to obtain the desired controller (Youla et al. 1976).
Finally, in step 5, the design is verified; in some cases, the controller parameters
may have to be tuned. In such cases, one may use optimization for this purpose.

3.2 Example – Single-Variable Case: A Decoupled Liquid

Propellant Engine

A case study of the sort encountered in combustion chamber pressure control of a

liquid propellant engine (LPE) is considered. A simplified schematic diagram of the
LPE is shown in Fig. 3.1. The operation of the engine starts by an electric command
that starts combustion in the starter. This combustion generates pressurized gas that
opens the diaphragm valves (not shown) below the fuel and oxidizer tanks, and
it also turns the turbine. Consequently, the shaft of the turbopump starts rotating,
and the fuel and oxidizer pumps start pumping fuel and oxidizer into the system.
Now, consider the two fuel and oxidizer paths. In the fuel path, fuel enters the
stabilizer control valve, and the output of this valve is a controlled amount of fuel
toward the combustion chamber and the gas generator. At the same time, in the
oxidizer path, oxidizer passes through the shut-down valve (not shown), and it enters
the combustion chamber. This fuel and oxidizer that have entered the combustion
chamber produce the required thrust via combustion. This thrust must be controlled.
The amount of oxidizer that enters the gas generator indirectly controls the thrust,
and the desired amount of oxidizer that must flow through the gas generator is
utilized via the regulator control valve. This oxidizer along with the fuel that have
entered the gas generator from the fuel path produce a combustion; this combustion
results in a pressurized gas that keeps the turbine running when the starter leaves
26 3 Single-Range Controller Design

Fig. 3.1 A simplified

schematic diagram of the
liquid engine (Key is given
in Table 3.1)
2
11 3 13
12
10 9

4
5 6

8 7

Table 3.1 Key for Fig. 3.1 No. Description

1 Combustion camber
2 Gas generator
3 Starter
4 Turbine
5 Fuel pump
6 Oxidizer pump
7 Regulator valve
8 Stabilizer valve
9 Starter combustion products
10 Gas generator combustion products
11 Electric command
12 Oxidizer tank
13 Fuel tank

the circuit. Depending on the value of the angular speed of the turbine, controlled
amounts of fuel and oxidizer flow toward the combustion chamber.
The combustion chamber pressure control system is one of the most important
parts of a liquid propellant engine. A class of regulators, which is used to control
the combustion chamber pressure, is composed of complicated hydraulic control
3.2 Example – Single-Variable Case: A Decoupled Liquid Propellant Engine 27

valves. These control valves have many disadvantages (Nassirharand and Karimi
2004a). For example, development of nonlinear and dynamic mathematical model
of such control valves is very time consuming and difficult. When such models are
developed and incorporated with the rest of the engine model, the execution time
of the simulation code is considerably increased. Another disadvantage is that the
control valve may accept only one set point (the set point is the desired pressure
value of the combustion chamber that the control system must maintain).
Therefore, it is desirable to design an alternative pressure control system which
would involve microprocessor-based regulators. Such pressure regulators could
accept various set points by software. The advantages of regulators that could
accept various set points are (1) the final desired pressure value could be reached at
specified stages in order to reduce the effects of water-hammer as well as excessive
overshoot of the pressure inside the combustion chamber and (2) to reduce the
generated impulse after the engine shut down command is issued; again, in this case,
the combustion chamber pressure is reduced from its steady-state value to zero at
specified stages.
The design of the regulator control loop is fully based on dynamic and nonlinear
mathematical model of the engine (Karimi et al. 2003). In order to develop the
mathematical model of the engine, four groups of elements are identified. Those
element groups are (1) valve, (2) turbopump, (3) combustion chamber and gas
generator, and (4) connections (e.g., pipes). In this treatment, the mathematical
model of the engine is put into perspective by selecting one element from each
group, and the mathematical model of that element is developed.
Without loss of generality, an orifice model could be used for valves and
connection pipes. The fuel pump model follows, and the oxidizer pump equations
are similar:

P 5 D fu Q5
m (3.2)

5 H5 Q5
T Q5 D (3.3)
!5 5
 
H5 D a5 C b5 Q5  c5 Q52 =g (3.4)

a5 D A5 !52 (3.5)
!
  H5;p
A5 D 0:97 C 0:8q5;p 2
(3.6)
!5;p
1
q5;p D (3.7)
2H5;p b5;2 tan.ˇ5;2 /
1C Q5;p !5;p K5;z 5;zp

b5 D B5 !5 (3.8)
28 3 Single-Range Controller Design

! 1
  H5;p Q5;p
B5 D 0:325  0:8q5p 2
(3.9)
!5;p !5;p

c5 D C5 (3.10)
 
2
H5;p =!5;p
C5 D 0:296  2 ; (3.11)
Q5;p =!5;p

where
A5 D an intermediate variable for fuel pump head equation
a5 D coefficient of the fuel pump head equation
B5 D an intermediate variable for fuel pump head equation
b5 D a coefficient of the fuel pump head equation
b5;2 D the exit width of fuel pump impeller
C5 D an intermediate variable for fuel pump head equation
c5 D a coefficient of the fuel pump head equation
H5 D the fuel pump head
H5;p D the nominal fuel pump head
K5;z D the effect coefficient of fuel pump blade
mP 5 D the mass flow rate of fuel pump
Q5 D the volumetric flow rate of fuel pump
Q5;p D the nominal volumetric flow rate of fuel pump
q5;p D an intermediate parameter in terms of nominal values of fuel pump
T Q5 D the moment of fuel pump
ˇ5;2 D exit angle of fuel pump blade
5 D f u :g
5 D the efficiency of fuel pump
5;p D the nominal efficiency of fuel pump
5;zp D the hydraulic efficiency of fuel pump
!5 D the turbopump angular speed
!5;p D the nominal angular speed of turbopump
f u D the fuel density.
A simple combustion model is used as below:

dP1
˛ C ˇP1 D m
P 1;f u .t  1 / C m
P 1;ox .t  1 / ; (3.12)
dt
where
P 1;ox D the oxidizer mass flow rate
m
P1 D the combustion chamber pressure
˛; ˇ D functions of pressure and mixture ratio; for a given P1 and mixture ratio,
the values of ˛ and ˇ are computed off-line from energy equations, and
3.2 Example – Single-Variable Case: A Decoupled Liquid Propellant Engine 29

they are incorporated as tables of values into the computer model. Then,
as simulation time marches ahead, the needed mentioned values for a given
P1 and mixture ratio are obtained by linear interpolation.
1 D the combustion delay time.
In this part, a describing function approach coupled with factorization theory is
used for design of the combustion chamber pressure control system. The primary
reason for this selection is that engine model is of the form given by state variable
equations of the form given by Eqs. 1.1 and 1.2, and describing function approach
is inherently capable of handling such nonlinear models.
The problem statement for the pressure controller of the combustion chamber
can be stated as follows – given the computer model of a liquid propellant engine,
how does one design a controller that would control the pressure of the combustion
chamber?
Algorithm 3.1 is applied to this problem. The nonlinear and dynamic computer
model of the liquid propellant engine is utilized to design a controller for the
combustion chamber. The controller design procedure steps are executed as follows.
In step 1, the desired transfer function is identified in terms of a linear second-order
transfer function that exhibits the desired behavior. The identified transfer function
is of the following form:

100
y;u D
hD : (3.13)
s 2 C 36s C 100

In step 2, the describing function models are obtained as outlined in Chap. 2, and
these models are depicted in Fig. 3.2 (Nassirharand and Karimi 2004a). Notice that
spread in these models is not significant, and therefore, it may be concluded that
a single-range linear controller might suffice. The describing function model that
is close to the nominal operating conditions was selected for system identification
®
purposes of the next step. In step 3, the MATLAB function invfreqs is used to
identify a linear model. The identified linear model is of the following form:

138:52
G.s/ D : (3.14)
s 2 C 52:81s C 253:47

The quality of this fit is shown in Fig. 3.3 (Nassirharand and Karimi 2004a).
In step 4, Eq. B.26 is solved with

m D n D 3; !1 D 0:01; !2 D 5:

The following controller is designed:

0:7219s 2 C 38:12s C 183:00

C.s/ D : (3.15)
s 2 C 36s
30 3 Single-Range Controller Design

Fig. 3.2 SIDF models of the liquid propellant engine. Source: (Nassirharand and Karimi 2004a)
© Sharif University of Technology

Fitting of a Linear Model

0

−10
Gain, db

−20

−30
10−1 100 101 102
Frequency, Rad./Sec.
0
Phase, Degrees

−50

−100

−150
10−1 100 101 102
Frequency, Rad./Sec.

Fig. 3.3 Linear system identification. Source: (Nassirharand and Karimi 2004a) © Sharif Univer-
sity of Technology
3.2 Example – Single-Variable Case: A Decoupled Liquid Propellant Engine 31

Normalized C.C. Pressure vs. Normalized Time

1.4
Simulation (Orifice Regulator) Set Point=0.935
1.2
1.177
Set Point=1.177
1
0.935
Normalized Pressure

Simulation (New Regulator)

0.8
Experiment (Present Regulator)
0.6

0.4

0.2

-0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized Time

Fig. 3.4 Design verification. Source: (Nassirharand and Karimi 2004a) © Sharif University of
Technology

The verification results are shown in Fig. 3.4 (Nassirharand and Karimi 2004a).
In the beginning, the pressure set point is set to one unit, and later this set point is
arbitrarily changed to 1.177 and 0.935 units in order to verify that regulator could
accept various set points. In Fig. 3.4, there are three curves: (1) the experimental
results of the engine with the present hydraulic regulator, (2) the simulation results
of the engine with the present regulator modeled as an orifice, and (3) the simulation
results of the engine with the new regulator. By comparison of these three curves,
the following important point may be noted. One of the problems associated with
the present pressure control system is that response overshoots. The curve corre-
sponding to the new regulator design with second set point demonstrates that initial
overshoot is not due to the regulator. In fact, the initial overshoot is due to the design
of the starter of the engine. With the new design, one may redesign the starter to
bring the combustion chamber pressure to a lower value, and then the new regulator
would be able to bring the pressure value to its desired value without any significant
amount of overshoot as is demonstrated in Fig. 3.4 (see the behavior corresponding
to the second set point command of the curve entitled “Simulation C New Regula-
tor”). The robustness test is performed by causing a 20% reduction in pumps’ heads;
the results are shown in Fig. 3.5 (Nassirharand and Karimi 2004a). From this figure
it is concluded that design is also robust because describing function models are one
of the most effective models to achieve a robust design.
For comparison purposes, in step 4 a H1 controller is designed (Nassirharand
and Mousavi Firdeh 2009a). The structure of the H1 control feedback system is
32 3 Single-Range Controller Design

Robustness Test
1.4
Simulation (New Regulator)
1.2 Simulation (Orifice Regulator)

1
Normalized Pressure

0.8 Pumps' Heads Were Reduced by 20%

Experiment(Present Regulator)
0.6

0.4

0.2

−0.2
0 0.2 0.4 0.6 0.8 1
Normalized Time

Fig. 3.5 Response with pumps’ failures. Source: (Nassirharand and Karimi 2004a) © Sharif
University of Technology

Fig. 3.6 Structure of the H1 feedback system. Source: (Nassirharand and Mousavi Firdeh 2009a)
© Elsevier

shown in Fig. 3.6. In this figure, W s are the weights, Z1 is the weighted disturbance
output signal, Z2 is the weighted control signal, Z3 is the weighted error signal, U
is the control signal, and Y is the error signal. The following remarks are in order:
1. In order to achieve a bandwidth of 4 rad/s, a prefilter, Wprfltr D 4.sC250/
250.sC4/
,
is specified; the zero of the prefilter is specified to satisfy the required H1
constraints (Doyle et al. 1989).
3.2 Example – Single-Variable Case: A Decoupled Liquid Propellant Engine 33

Fig. 3.7 Controller frequency response comparison. Source: (Nassirharand and Mousavi Firdeh
2009a) © Elsevier

2. It is desired to have a steady-state error of less than 1%; therefore, a performance

weighting function, Wperfrmnc D 0:1.sC4/
.sC0:01/
, acting on the error signal is selected.
3. Considering a multiplicative uncertainty, a disturbance weighting function,
Wdist D 2.sC9/
.sC90/
, for robust stability is specified. This reflects a 20% uncertainty for
frequencies under 9 rad/s and a 200% uncertainty for frequencies above 90 rad/s.
4. In order to compensate for erosion and high levels of friction in mechanical
parts, a control signal weighting function, Wcntrl D 3:5.sC12/
.sC95/ , is considered. This
weighting function reflects the fact that is desired to weaken the control signals
above approximately 15 Hz by a factor of 30%; it also reflects that it is desired
to consider a weak penalty for low-frequency control signals.
A reduced order model of the H1 controller that is designed using the hinfsys
®
function of MATLAB software follows:
182:3s C 1521
C.s/ D : (3.16)
s 2 C 99:4s

The performance of the H1 controller and that of the controller obtained using a
factorization approach with the nonlinear plant is compared in Figs. 3.7, 3.8, 3.9, and
3.10 (Nassirharand and Mousavi Firdeh 2009a). From these figures, the following
conclusions may be drawn:
1. The frequency response of the H1 controller (see Fig. 3.7) is more desirable
because it attenuates the high-frequency signals; therefore, it is capable of being
fairly insensitive to high-frequency noise in the system as demonstrated by
Fig. 3.8. It is apparent that control signal of the H1 controller is less noisy than
that of the factorization controller.
2. With reference to Fig. 3.9, the step response of the H1 controller is more
desirable because it does not overshoot. Pressure overshoot could be harmful
for the operation of the combustion chamber. Of course, the overshoot in the
34 3 Single-Range Controller Design

Fig. 3.8 Control signal comparison at the presence of measurement noise. Source: (Nassirharand
and Mousavi Firdeh 2009a) © Elsevier

Fig. 3.9 Normalized step response comparison. Source: (Nassirharand and Mousavi Firdeh
2009a) © Elsevier

factorization design could be removed at the expense of an increase in the

corresponding settling time.
3. With reference to Fig. 3.10, the control signal of the H1 controller is more
desirable because it does not have the initial abrupt decrease in value.

3.3 Example – Multivariable Case: Idle Speed

Control Problem

Algorithm 3.1 is applied to the idle speed control of uncertain automobile engines
with time delay (Teh and Nassirharand 2010a). It is shown that results compete
with the observer-based design of Bengea et al. (2004). Stability is demonstrated by
successful generation of describing function models of the final nonlinear closed-
loop system.
3.3 Example – Multivariable Case: Idle Speed Control Problem 35

Fig. 3.10 Control signal comparison without measurement noise. Source: (Nassirharand and
Mousavi Firdeh 2009a) © Elsevier

This problem may be classified as nonlinear, multivariable, uncertain, and

distributed at the presence of load torque disturbances. Many researchers have
developed various solution methods (see references in Teh and Nassirharand 2010a).
These methods consider the following approaches:
(a) Linearizing the model and applying linear control theory
(b) Ignoring the time delay
(c) Approximating the time delay
(d) Utilizing an estimator or an observer
(e) Applying robust control techniques
(f) Applying neural networks
(g) Applying fuzzy control method
(h) Applying sliding mode or variable structure methods
(i) Applying optimal control technique
(j) Applying adaptive control theory
Except the standard methods mentioned in (a)–(c) above, all mentioned methods
are relatively complicated. Those methods also require a high level of expertise in
the utilized control method. These design methods usually involve invoking heuris-
tics, and they require subjective judgment of the designer. The problems associated
with the standard technique of linearizing the plant are well known. For example,
the final control system would be highly sensitive to the operating point of interest
(i.e., for standard plant perturbations, the closed-loop system may become unstable).
The use of mentioned complicated controller design methods is usually justified
if the classical control system design methods do not produce adequate level of
robustness, stability, and/or performance (Nassirharand et al. 1988). In continuation,
a comprehensive explanation about the time delay of the IC engine during the
design procedure is given. The presence of the time delay turns the plant into
a distributed system. Such a system description would fit the class of nonlinear
36 3 Single-Range Controller Design

systems considered herein as demonstrated by Eqs. 1.1 and 1.2. The effects of time
delay and distributive nature of the plant are captured in the output time domain
histories mentioned in step 2. Then these output time domain histories are trans-
formed to their frequency domain counterparts by evaluation of the Fourier integrals
as described earlier. Hence, the generated frequency domain model includes the
effects of time delay. This frequency domain model is the basis of controller design
in step 4. Finally, approximations made in obtaining the frequency domain model
(including those related to the time delay approximations) are compensated for by
tuning the gains of the elements of the matrix controller as described in step 5.
Finally, the performance of the final synthesized controller with the actual nonlinear
plant with the time delay is verified.
Algorithm 3.1 is executed as follows:
Step 1 – With reference to the work by Bengea et al. (2004), and the multivariable
design approach outlined in Appendix B, the desired transfer function is selected for
both main control loops:

225
hd .s/ D : (3.17*)
s 2 C 27s C 225

Step 2 – The SIDF models are generated as outlined in Chap. 2, and these results
along with linear system identification task of the following steps are shown in
Figs. 2.10 and 2.11.
®
Step 3 – MATLAB is used, and the following matrix transfer function is identified:

g11 .s/ g12 .s/
G.s/ D ; (3.18*)
g21 .s/ g22 .s/

where
23:6799s C 59:1867
g11 .s/ D (3.19*)
s 2 C 9:4742s C 19:3981

137:0035
g12 .s/ D (3.20*)
s 2 C 1:3259s C 9:0378

10:8771
g21 .s/ D (3.21*)
s 2 C 11:2327s C 34:1192

111:4211
g22 .s/ D : (3.22*)
s C 4:8593

*Source: (Teh and Nassirharand 2010a) © Sage

3.3 Example – Multivariable Case: Idle Speed Control Problem 37

Step 4 – From the application of linear lead-lag controller design approach outlined
in Appendix A (with !1 D 0:01 and !2 D 10), the following compensator is
designed:

2:654s 2 C 1744s C 11930

c11 .s/ D : (3.23*)
190:4s 2 C 4828s

Similarly, with hd .s/ as before, another compensator denoted c22 .s/ is designed:

13:64s C 59:886
c22 .s/ D : (3.24*)
6:52s 2 C 176:1s

Now, use Eq. B.2 and obtain the off diagonal terms of the 2 by 2 matrix
compensator transfer function, C.s/, as follows:

18:69s 3  259:1s 2  1140s  1592

c12 .s/ D (3.25*)
1:544s 5 C 47:6s 4 C 178:6s 3 C 549:9s 2 C 941:9s

2:789s 3 C 1911s 2 C 22200s C 63070

c21 .s/ D : (3.26*)
2121s 4 C 77620s 3 C 676600s 2 C 1835000s

Step 5 – Design verification reveals that controller gains must be tuned. To tune the
gains, the following objective function is used:

E D ŒN1 .t/  Nd .t/2 C ŒN2 .t/2

C ŒPm1 .t/  Pmd .t/2 C ŒPm2 .t/2 : (3.27*)

The idea behind the selection and applicability of the above optimization criterion
is explained next. The first and the third terms of the objective function assure
tracking of the command signals of the two channels (speed and manifold pressure);
the second and fourth terms of the objective function assure decoupling of the speed
and manifold pressure axis. This objective function is also used in engine propulsion
control as demonstrated in Nassirharand and Karimi (2004b).
For this example, C.s/ is of the following form:

K1 c11 .s/ K2 c12 .s/
C.s/ D : (3.28*)
K3 c21 .s/ K4 c22 .s/

In this research, the simplex optimization method implemented by the fminsearch

®
function of MATLAB is utilized.

*Source: (Teh and Nassirharand 2010a) © Sage

38 3 Single-Range Controller Design

Fig. 3.11 Design comparison – speed axis. Source: (Teh and Nassirharand 2010a) © Sage

Fig. 3.12 Design comparison – pressure axis. Source: (Teh and Nassirharand 2010a) © Sage

The gains of the controller are tuned by minimizing objective function (3.27)
with N2 .t/ D Pm2 .t/ D 0. The optimum values of these gains follow:

K1 D 0:6933; K2 D 0:1915; K3 D 0:0015; K4 D 0:5172:

Finally, the design is verified by a digital simulation of the linear controller and
the nonlinear plant; the results are compared with those obtained by the method of
Bengea et al. (2004), as shown in Figs. 3.11, 3.12, 3.13, and 3.14. It is apparent that
the classical design presented herein competes with the more complicated observer-
based approach. Furthermore, the design is also verified to be robust with respect
to engine parameter uncertainties and torque load disturbances. This is done by
simulating the linear controller and the nonlinear plant with a disturbance load
torque of 10 N  m as well as nominal, upper limits, and lower limits of the engine
parameters. These results are shown in Figs. 3.15 and 3.16.
3.3 Example – Multivariable Case: Idle Speed Control Problem 39

Fig. 3.13 Design comparison of control signals – spark timing advance. Source: (Teh and
Nassirharand 2010a) © Sage

Fig. 3.14 Design comparison of control signals – idle bypass valve opening. Source: (Teh and
Nassirharand 2010a) © Sage

Fig. 3.15 Normalized speed as a function of time with disturbance in the presence of parametric
uncertainties. Source: (Teh and Nassirharand 2010a) © Sage
40 3 Single-Range Controller Design

Fig. 3.16 Normalized manifold pressure as a function of time with disturbance in the presence of
parametric uncertainties. Source: (Teh and Nassirharand 2010a) © Sage

Fig. 3.17 Stability test – gain matrix of the closed-loop system (A1 is speed and A2 is manifold
pressure). Source: (Teh and Nassirharand 2010a) © Sage
3.3 Example – Multivariable Case: Idle Speed Control Problem 41

Fig. 3.18 Stability test – phase matrix of the closed-loop system (A1 is speed and A2 is manifold
pressure). Source: (Teh and Nassirharand 2010a) © Sage

Stability of the final nonlinear closed-loop system is verified by successful

generation of the corresponding SIDF models at different operating regimes.
Stability and insensitivity to amplitude of excitation of the closed-loop system is
examined by generating the closed-loop SIDF models at various command speeds
and manifold pressures. These results are shown in Figs. 3.17 and 3.18 (Teh and
Nassirharand 2010a). It is apparent that closed-loop system is stable and robust with
respect to magnitude of the command signals.
Chapter 4
Dual-Range Controller Design

4.1 Background

In the early 1980s, the following question was considered. When is it really required
to design a nonlinear controller for a highly nonlinear system? The need for a
reply to this question received some attention (Taylor 1983). With the advancement
of numerous robust control techniques for linear and nonlinear systems, the
need to have a systematic procedure to answer the above question became even
more evident. Taylor (1983) introduced the idea for design of dual-range linear
controllers. In that work, he hypothesized that if one uses two quasi-linear models
for controller design purposes, then controllers with a higher performance than those
of their single-range counterparts may be designed. Consequently, it was suggested
that if a dual-range linear controller did not provide adequate level of robustness,
then a nonlinear controller design procedure should be considered. This hypothesis
was verified for highly nonlinear single-variable systems (Nassirharand et al. 1988;
Nassirharand 1991).
The original proposed approach for design of dual-range linear controllers for
nonlinear systems called for application of the simultaneous stabilization theory of
Vidyasagar and Viswanadham (1982). In that approach, one would parameterize the
class of all stabilizing controllers in terms of a (matrix) transfer function parameter
R.s/ followed by a search for the optimum controller. One iterative procedure for
single-variable systems is proposed in Garvin and Mathew (1996). It has been
discovered that searching for R.s/ results in high-order controllers and a direct
search for the controller results in appreciably lower-order controllers (Nassirharand
et al. 1988; Nassirharand 1991; Teh 2010). For example, when searching for R.s/
for a nonlinear plant with two integrators, a fourth-order controller is designed;
a direct search results in a first-order simultaneously stabilizing controller (see
Sect. 4.4). Furthermore, with this approach there is no guarantee that the controller
that stabilizes two linear approximations of the nonlinear plant would also stabilize
the actual nonlinear system. For this reason, in the new algorithm for design of
dual-range controllers, there is a step that calls for tuning of the simultaneously

A. Nassirharand, Computer-aided Nonlinear Control System Design: Using Describing 43

Function Models, DOI 10.1007/978-1-4471-2149-7 4,
© Springer-Verlag London Limited 2012
44 4 Dual-Range Controller Design

stabilizing controller. Examine Appendix D for a demonstration example problem

that highlights the controller high-order problem as well as the stability problem
mentioned above. It could be argued that appropriate coprime factorization could
also result in competitive results; however, that would require further research which
may not be justified under the conditions that a direct search for the simultaneously
stabilizing controller is shown to be more straightforward.
It is also discovered that, when using the factorization theory, a high-order
controller (i.e., of the order of 50) is required to stabilize only one of the plants given
in the example problem of Sect. 4.5 (Teh 2010); furthermore, for more complex
problems (such as quad-rotor helicopter problem (Al-Kindi and Nassirharand
2010)), MATLAB runs to numerical overflow problems. For this reason, it is
recommended not to use the simultaneous stabilization theory of Vidyasagar and
Viswanadham (1982) for applications to nonlinear systems. The preferred approach
is to select a controller of a relatively low order (e.g., 0–4), with arbitrary coefficients
that at least stabilizes one of the linear approximations, as a starting solution; then,
optimization would be used to determine the coefficients of the controller. See
Appendix D for justifications.

4.2 Controller Synthesis

The following algorithm is developed to obtain a desired dual-range linear controller

(Teh and Nassirharand 2010b).
Steps 1–3 are similar to the single-range controller design discussed in the
previous chapter. Step 4 is to select any controller that results in a stable closed-
loop system for any one of the linear plants of step 3. For the purposes of this
task, any linear design approach would suffice. In step 5, the fminsearch function
of MATLAB® (which uses a Simplex (Reklaitis et al. 1983; Nelder and Mead
1965) search approach) may be utilized, and (for the cases that decoupling is not
required) the following objective function may be minimized (Nassirharand et al.
1988; Patwardhan 1988; Nassirharand 1991; Teh 2010):

F D ˛F0 C .1  ˛/ F1 ; (4.1)

where
   X i 2
Fi D ˇ i
ymax d;i
 ymax i
C < ymin >2 C yj  yjd;i ; (4.2)
j
( i
2 i
i
ymin if ymin <0
< ymin >2 D ; (4.3)
i
0 if ymin 0
4.2 Controller Synthesis 45

Algorithm 4.1

1. Derive the mathematical model of a reference linear model that the final
closed-loop system is to mimic its dynamic and static behavior.
2. Obtain the describing function models of the nonlinear plant at various
operating regimes (see Chap. 2).
3. Select two of the models of the previous step, and obtain linear fits.
4. Arbitrarily select a controller (of a low order, e.g., 0–4) with arbitrary
coefficients that preferably stabilize at least one of the linear fits of the
previous step.
5. Apply optimization, and optimize the coefficients of the dual-range linear
controller for the two linear plants of step 3.
6. Apply optimization and tune the gains of the linear controller when placed
to operate with the actual nonlinear plant.
7. Verify design of the final nonlinear closed-loop system. If satisfactory
results are not obtained, proceed to design a nonlinear controller as
outlined in Chap. 5; otherwise, stop.

d D a superscript indicating the desired value(s),

F D the objective function,
i D a superscript index (0 for nominal condition and 1 for the other),
j D a subscript index for unit step-response array,
max D a subscript indicating the maximum value,
min D a subscript indicating the minimum value,
y D the discretized closed-loop unit step response,
˛ D a weighting factor, 0  ˛  1,
ˇ D a weighting factor, ˇ  0.
In Eq. 4.2, the terms in brackets avoid designs with excessive amounts of
overshoot and undershoot.
In cases that decoupling is required, objective function (4.1) with the following
alternative definitions is minimized (Nassirharand and Taylor 1990; Mousavi-Firdeh
et al. 2006; Nassirharand and Karimi 2004b; Teh and Nassirharand 2010c; Teh
2010):
8 9
<X  2 X   =
2
Fi D yji 0  yjd;i0 C yki ; (4.4)
: 0 ;
j k

where j 0 D 1; 2; :::; l, k D 1; 2; :::; l; k ¤ j 0 , l D the number of outputs.

The rest of the parameters and variables are defined as before. It should be
noted that one alternatively could modify Eq. 4.4 to penalize excessive amounts
of overshoot and undershoot similar to Eq. 4.2.
46 4 Dual-Range Controller Design

In step 6, the controller gains are tuned by the same optimization routine as that
used in the previous step. The experience of the authors in designing controllers
for nonlinear plants based on a fitted transfer function (to the corresponding SIDF
data) reveals that further tuning of the controller gains may be required before the
controller could perform well when connected to the actual nonlinear plant. For this
reason, the controller of the previous step is used as a starting solution, and a similar
objective function to that presented by Eq. 4.1 is minimized in order to tune the
controller parameters. In this case, y would be the normalized discretized closed-
loop step response of the actual nonlinear plant when connected to the controller;
also, index i would refer to the particular operating regime of interest.
Step 7 is to verify design. For stability, the following argument is used. If the
describing function models of the final closed-loop system at various operating
regimes are obtainable, then it may be concluded that Fourier integrals (see Eqs.
2.16 and 2.23) have converged; consequently, this suggests that the outputs must
have been bounded. Hence BIBO (bounded-input bounded-output) stability may be
demonstrated in this fashion. For verification of the performance of the closed-loop
system, normalized step-response plots by exciting the plant with step inputs of
various magnitudes may be obtained.

4.3 Software

A MATLAB® function is available that automates the design in step 5; the primary
inputs to this function are the two linear plants identified in step 3 of Algorithm
4.1 as well as the reference linear model identified in step 1; the primary output
®
is the dual-range controller. Another MATLAB function is available that uses the
Simulink model of the nonlinear plant and tunes the controller parameters obtained
in step 5 (see Teh 2010 for details).

4.4 Example 1: A Single-Variable Servo System

Consider the SISO servo system that was also considered in Sect. 2.2.1.2. It is
desired for the closed-loop system to mimic a second-order system with a damping
ratio of 0.307 and a natural frequency of 43.22; similar performance measures
were also achieved by Taylor and Strobel (1985) when designing a nonlinear PID
controller. Algorithm 4.1 is executed as follows:
1. The reference linear model follows:

!n2
Gr .s/ D ; (4.5)
s 2 C 2&!n s C !n2

where !n D 43:22 rad=s and & D 0:307.

4.4 Example 1: A Single-Variable Servo System 47

2. The describing function models are generated as outlined in Sect. 2.2.1. See Figs.
2.5, 2.6, 2.7 and 2.8 for pseudofrequency response plots.
3. The two linear models corresponding to small and large excitation amplitudes are
selected. It is recommended to use the two linear models whose gain characteris-
tics bound those of all the others in class; this corresponds to describing function
models with excitation amplitudes of 0.25 V and 0.8 V. A linear fit is obtained,
and the two linear models follow (Nassirharand et al. 1988; Nassirharand 1988):

0:003989s C 1:231
G0 .s/ D ; (4.6*)
0:003761s 2 C 0:05475s
0:040314s C 1
G1 .s/ D : (4.7*)
0:23641s 2 C 0:22521s

4. The following simultaneously stabilizing (SS) controller is selected as the

starting solution for the optimization:

1 .s C 2/
Cssstart .s/ D : (4.8)
sC4
®
5. The MATLAB function DRLCD, which is outlined in Sect. 4.3, is utilized,
and objective function (4.1) with Eq. 4.2 and setting ˛ D 0:1 and ˇ D 0 is
minimized; the following controller is obtained:

3:1671s C 1:1469
Cssuntuned .s/ D : (4.9)
0:0229s C 0:1089

6. The controller parameters of the previous step are tuned by optimization. The
final synthesized controller is of the following form:

0:0301 .1:1814s C 51:660/

Csstuned .s/ D : (4.10)
0:0017s C 0:1996

7. Design validation consists of generating the normalized closed-loop responses

for different step magnitudes as well as generating the describing function
models of the final nonlinear closed-loop system at various operating regimes.
The latter is executed to demonstrate stability. The results are depicted in
Figs. 4.1 and 4.2. The normalized step-response plots with a nonlinear PID
controller are depicted in Fig. 5.6 (see Chap. 5 and Taylor and Strobel 1985).
It is evident that the performance of the dual-range linear controller competes
with a nonlinear PID controller.

*Source: (Nassirharand et al. 1988) © ASME

48 4 Dual-Range Controller Design

Fig. 4.1 Normalized step-response plots of the final closed-loop system with the dual-range linear
controller

Fig. 4.2 Stability test – the generated describing function models of the final closed-loop system
with the dual-range linear controller

4.5 Example 2: Multivariable Bank Angle Control Problem

Consider the unmanned aerial vehicle bank angle control problem that is studied in
Colgren 2004; the corresponding schematic block diagram is depicted in Fig. 4.3.
The problem statement is to design a dual-range linear controller for this nonlinear
system that would force the bank angle response of the corresponding nonlinear
feedback system to mimic a second-order system with a damping ratio of 0.9 and
a natural frequency of 1.11 rad=s; these are approximately the same performance
4.5 Example 2: Multivariable Bank Angle Control Problem 49

Fig. 4.3 The schematic diagram of the UAV and the structure of the closed-loop system. Source:
(Teh 2010) © S.H. Teh

measures that are achieved when designing a multivariable nonlinear controller as

outlined in Chap. 5 (Nassirharand 2010a). The details of the application of the
design procedure given in the previous section follow (Teh 2010):
1. With reference to Fig. 4.3 and the problem statement, the following reference
linear model is considered:
2 3
!n2 s
6 s 2 C 2&!n s C ! 2 0 7
Gr .s/ D 6 n 7; (4.11*)
4 2
!n 5
0
s 2 C 2&!n s C !n2

where !n D 1:11rad=s and & D 0:9.

2. In this step, the input/output behavior of the nonlinear plant at various operating
regimes is to be characterized. The sinusoidal-input describing function models
of the nonlinear plant are obtained for eight input amplitude values in the
frequency range of 0.1–1,000 rad=s. The input amplitude values considered were
0.75ı , 1ı , 2ı , 4ı , 8ı , 16ı , 32ı , 64ı , and 128ı. The resulting frequency response
plots are depicted in Fig. 4.4.
3. The two quasi-linear models corresponding to input amplitude levels of 1ı and
8ı are arbitrarily selected for system identification purpose. The MATLAB
®

function invfreqs is utilized for this purpose. The identified models are given
by the following equations; the quality of fit may be examined by considering
Figs. 4.5 and 4.6.
2 3
70200
6 s 2 C 23:08s C 96:81 7
G0 .s/ D 6
4
7;
5 (4.12*)
70200
2
s .s C 23:08s C 96:81/

*Source: (Teh 2010) © S.H. Teh

50 4 Dual-Range Controller Design

Fig. 4.4 UAV open-loop pseudofrequency response plots. Source: (Teh 2010) © S.H. Teh

Fig. 4.5 Linear system identification at a low excitation amplitude. Source: (Teh 2010) ©
S.H. Teh
4.5 Example 2: Multivariable Bank Angle Control Problem 51

Fig. 4.6 Linear system identification at a high excitation amplitude. Source: (Teh 2010) ©
S.H. Teh

Fig. 4.7 Stability test. Source: (Teh 2010) © S.H. Teh

2 3
9606
6 s 2 C 22:15s C 92:54 7
G1 .s/ D 6
4
7:
5 (4.13*)
9606
s .s C 22:15s C 92:54/
2

*Source: (Teh 2010) © S.H. Teh

52 4 Dual-Range Controller Design

Fig. 4.8 Normalized step-response plots of the synthesized dual-range linear controller and the
nonlinear UAV. Source: (Teh 2010) © S.H. Teh

Fig. 4.9 Normalized roll-rate plots of the synthesized dual-range linear controller and the
nonlinear UAV. Source: (Teh 2010) © S.H. Teh

4. Arbitrarily, select the parameters of the starting solution of the controller. In

order to appreciate the developments in this chapter, a starting solution using
the application of simultaneous stabilization theory is obtained as detailed in Teh
2010. It is not necessarily recommended to use this approach to obtain a starting
solution, and it is given here for illustration purposes only. The following starting
solution for the controller is selected:
4.5 Example 2: Multivariable Bank Angle Control Problem 53

Fig. 4.10 The normalized step-response plots of the closed-loop system with a single-range linear
controller and the nonlinear UAV. Source: (Teh 2010) © S.H. Teh

Fig. 4.11 The normalized roll-rate plots of the closed-loop system with a single-range linear
controller and the nonlinear UAV. Source: (Teh 2010) © S.H. Teh

    
995:4 s 2 C 5:8s C 147:3 4:6 s 2 C 16270:9s  474819
C start
.s/ D :
s 2 C 43751:9s C 8:8E6 s 2 C 43751:9s C 8:8E6
(4.14*)
®
5. The MATLAB function described in Section 4.3 is used, and objective function
(4.1) with Eq. 4.2 is minimized to obtain the following simultaneously stabilizing
controller:

*Source: (Teh 2010) © S.H. Teh

54 4 Dual-Range Controller Design

Fig. 4.12 The step-response plot of a nonlinear H1 controller and the nonlinear UAV. Source:
(Colgren 2004) © R. Colgren

Fig. 4.13 The roll-rate plot of a nonlinear H1 controller and the nonlinear UAV. Source: (Teh
2010) © S.H. Teh

2 3T
0:48s 2  1:94s C 2:45E  4
6 0:67s 2 C 7:20E  4s 7
6 7
Cssuntuned .s/ D 6 7 : (4.15*)
4 2:86E  4s 2  0:40s C 4:36E  5 5
0:11s 2 C 1:24E  4s

*Source: (Teh 2010) © S.H. Teh

4.5 Example 2: Multivariable Bank Angle Control Problem 55

6. The parameters of the controller in Eq. 4.15 are tuned as outlined in Sect. 4.2.
2   3T
1:50 0:72s 2 C 3:0s C 6:61E  4
6 0:26s 2 C 5:5E  4s 7
Csstuned .s/ D 6 7
4 15:53 2:62E  4s 2 C 0:30s C 5:1E  5 5 : (4.16*)

0:21s 2 C 1:01E  4s

7. The describing function models of the closed-loop system at various operating

regimes are depicted in Fig. 4.7. It is apparent that the system is stable. For
verification of performance measures, the normalized step-response plots of
the bank angle and roll-rate are given in Figs. 4.8 and 4.9, respectively. The
normalized step-response plots of the bank angle and roll-rate of a single-range
linear controller (Nassirharand and Mousavi-Firdeh 2009b), the normalized
step-response plots of the bank angle and roll-rate of a multi-range nonlinear con-
troller (Nassirharand 2010a), and one step-response plot of the bank angle and
roll-rate of a nonlinear H1 controller (Colgren 2004) are shown in Figs. 4.10,
4.11, 4.12, 4.13, 5.20 and 5.21, respectively. It is apparent that the performance
of the dual-range linear controller is superior to that of the single-range linear
controller, and it competes with the two other nonlinear controllers.

*Source: (Teh 2010) © S.H. Teh

Chapter 5
Multirange Nonlinear Controller Design

5.1 Background

The first attempt at design of nonlinear PID controllers based on sinusoidal-input

describing function models and nonlinearity inversion using inverse describing
function technique is reported in Taylor and Strobel (1985) and later in Nassirharand
and Karimi (2006). Competitive work (using random-input describing functions)
is reported in Suzuki and Hedrick (1985) as well as Nanka-Bruce and Atherton
®
(1990). A series of MATLAB functions is also developed in Nassirharand and
Karimi (2006) that automates the design procedure of Taylor and Strobel (1985).
Following this, design of nonlinear lead-lag and nonlinear H1 controllers is
developed in Nassirharand and Mousavi Firdeh (2008) and in Nassirharand and
Mousavi Firdeh (2009c), respectively. See Nanka-Bruce and Atherton (1990) for
competitive nonlinear lead-lag controller design, and see Colgren (2004) for an
alternate nonlinear H1 controller design.

5.2 Controller Synthesis

The following algorithm may be utilized to obtain a desired nonlinear controller

for a highly nonlinear plant characterized by Eqs. 1.1 and 1.2 (Taylor 1983;
Nassirharand and Karimi 2006; Nassirharand and Mousavi Firdeh 2008, 2009c).
Steps 1 and 2 are similar to the single-range controller design discussed in Chap.
1. In step 3, any linear controller design technique may be used to design a linear
controller. For example, one could fit a linear model to the nominal pseudofrequency
response data followed by application of the cad controller software (Nassirha-
rand 2009b) to design a PID (Nassirharand et al. 2003.), lead-lag (Nassirharand
and Karimi 2004c), or a nth-order transfer function (Nassirharand 2008a). Also,
®
MATLAB may be used and a linear H1 controller may be designed. Furthermore,
it is recommended to check the step response of the designed linear controller

A. Nassirharand, Computer-aided Nonlinear Control System Design: Using Describing 57

Function Models, DOI 10.1007/978-1-4471-2149-7 5,
© Springer-Verlag London Limited 2012
58 5 Multirange Nonlinear Controller Design

and the actual nonlinear plant at nominal conditions before proceeding to the next
step. The parameters of the linear controller may have to be tuned similar to that
described in step 6 of Algorithm 4.1. Step 4 is self-explanatory. In step 5, in order to
compute Eq. 5.1, one must place the describing function generator software inside
the optimization loop. This will cause the execution time of the simulation code to
unreasonably increase, and design for industrial systems with the usual accessible
computers becomes almost impossible. In order to lift this restriction, to compute
C Gi , one could compute Ci .j!/ :G .j!; ei jCi .j!/j/, where
Ci D the controller at operating regime i
ei D amplitude of the sinusoid at the input to the controller
G D the pseudofrequency domain model of nonlinear plant at a specific operating
regime
p
j D 1
! D frequency, rad=s
To accomplish this, the G .j!; ai / data obtained in step 2 coupled with interpola-
tion is utilized. This computation method considerably increases the execution speed
®
of step 5, and minimizing E .j!/ via the fminsearch function of MATLAB for
complicated industrial systems via standard computers may easily be accomplished.
Step 6 is described in the next subsection, and in step 7, normalized step-response
plots with various step magnitudes of interest are obtained. For demonstration of
stability, the describing function models of the final closed-loop system at various
operating regimes may be generated.

Algorithm 5.1

1. Derive the mathematical model of a reference linear model that the final
closed-loop system is to mimic its dynamic and static behavior.
2. Obtain the describing function models of the nonlinear plant at various
operating regimes (see Chap. 2).
3. Select one of the input-output models of the previous step as the nominal
model, and design a linear controller for this nominal model. The structure
of the linear controller could be a PID, a lead-lag, a H1 controller, or
simply a nth-order transfer function.
4. Place the model of the designed controller of the previous step in series
with the nonlinear plant. Using the technique outlined in Chap. 2, generate
the corresponding open-loop describing function model. This generated
open-loop model is the desired model that the open-loop system is to
mimic at all operating regimes. It should be noted that in this step, the
excitation signal is obtained by dividing the amplitude which corresponds
to the selected nominal model used in step 3 by the gain of the linear
5.2 Controller Synthesis 59

controller at a specific frequency. For example, the controller gain at

the phase crossover frequency may be used (Taylor and Strobel 1985);
for a type zero controller like a lead-lag controller or a controller that
does not have a crossover frequency, the low-frequency gain may be used
(Nassirharand and Mousavi Firdeh 2008, 2009c).
5. In this step, a set of linear controllers are designed such that the following
objective function over the frequency set of interest is a minimum.

E .j!/ D jC  G   C Gi j (5.1)

where
C Gi D the pseudofrequency response model of the controller and the
nonlinear plant at operating regime i
C  G  D the desired pseudofrequency response model obtained in
step 4
E D a measure of the error between the desired open-loop pseud-
ofrequency response model obtained in step 4 and the open-loop
pseudofrequency response model comprised of the controller and the
nonlinear plant at operating condition i
It should be noted that E .j!/ measures the sensitivity (or robust-
ness) of the open-loop system, and minimizing E .j!/ at various op-
erating regimes of interest results in the required amplitude-dependent
behavior of a minimum sensitivity controller. The output of this step is
a set of controller parameters for each ei .
6. Interpret the table of values for each controller parameter as a function of
ei (which was obtained in the previous step) as the describing function
model for that controller parameter; by using the SIDF model inversion
technique described below (Nassirharand 2009a), invert the describing
function model for each controller parameter to identify the desired
nonlinear function for that controller parameter.
7. In this step, a model of the nonlinear controller which incorporates the
synthesized nonlinear functions of the previous step is constructed, and
the design is verified. For verification, the describing function mod-
els of the synthesized nonlinear closed-loop system may be generated
to demonstrate stability; in other words, if the Fourier integrals (see
Eqs. 2.16 and 2.23) converge at all operating regimes, then stability
may be concluded. Furthermore, the normalized step-response plots
with various magnitudes of command signal may also be generated
to verify time-domain behavior. Additionally, in order to quantify the
amount of achieved reduction in sensitivity, the followings are evaluated
numerically:
60 5 Multirange Nonlinear Controller Design

2! 3
X Z ˇˇ ˇ
2
ˇ
0 D 4 ˇ1  Gi .j!; ai / ˇ d! 5 (5.2)
ˇ G  .j!; a / ˇ
i !1
2! 3
2ˇ ˇ
X Z ˇ  ˇ
1 D 4 ˇ1  i ˇ d! 5 (5.3)
ˇ  ˇ
i !1

where
ai D amplitude of the sinusoid at the input to the nonlinear plant at
operating regime i
a D amplitude of the sinusoid at the input to the nonlinear plant at
nominal conditions
i D the pseudofrequency domain model of the open-loop system
comprised of the designed nonlinear controller and the actual nonlinear
plant at operating regime i
  D the pseudofrequency domain model of the open-loop system
comprised of the designed linear controller and the actual nonlinear
plant at nominal conditions (see step 3)
0 D measure of sensitivity of the nonlinear plant
1 D measure of sensitivity of the designed open-loop system
The closer k (k D 0; 1) is to zero, the smaller is the level of sensitivity.
The percent amount of reduction in sensitivity would be given by the
followings relation:
0  1
RD  100 (5.4)
0

5.2.1 Describing Function Inversion (Nassirharand 2009a)*

The inversion procedure starts by guessing the initial values for the independent
variables of an assumed nonlinear function. Then, the describing function model
of the nonlinearity is numerically generated. This is followed by minimizing an
objective function that yields the desired parameters of the nonlinear function.
Identification of linear systems from frequency response data is not a new
subject, and there is considerable literature that covers this subject (e.g., see
Nassirharand 1988 and references therein). Unlike nonlinear systems, this problem

*Source: (Nassirharand 2009a) © Elsevier

5.2 Controller Synthesis 61

for linear systems is independent of amplitude. One application of identification of

systems from frequency response data is in the area of controller design.
The focus of this section is on identification of nonlinear models or nonlinear
functions from their corresponding describing function models. What is common in
most identification works is that a model for the function/system has to be assumed,
and then the parameters of that model are identified. This development is of no
difference, and a model for the function has to be assumed whose parameters may be
identified using a newly developed algorithm. For identification of nonlinear system
structure, see Petrovic et al. (2000), Haber and Keviczky (1999), and Bongard and
Lipson (2005).
Lyashevskiy and Chen (1997) and Besancon-Voda and Blaha (2002) used
describing functions to identify the parameters of nonlinear functions of a nonlinear
system. Those works are based on imposing harmonic balance conditions, which
lead to a set of nonlinear algebraic equations that must be solved numerically.
One problem associated with that approach is that input to each nonlinear term is
assumed to be sinusoidal; such assumptions limit the application of the technique. In
this work, simulation and evaluation of Fourier integrals are utilized, and there are
no assumptions that input to each nonlinearity is sinusoidal. A method is also devel-
oped that uses describing function approach for expressing the nonlinearity in the
structure (Ozer et al. 2009); then, using measured responses, one can calculate the
describing function model and perform identification. Unlike this work, this requires
one to isolate the nonlinear term within the system. A method for identification
of friction using describing functions has also been developed by Kim and Chung
(2006); unlike this work, the method of Kim and Chung (2006) is limited to iden-
tification of friction elements within servo systems. Prazenica et al. used Volterra
theory to identify nonlinear system dynamics from measured input-output data. This
approach is useful if the system includes polynomial-type nonlinear terms, and (un-
like this work) it may not be applied to systems with discontinuous nonlinear terms.
An algorithm for identification of a system divided into cascaded blocks of dynamic
linear, static nonlinear, and dynamic linear subsystems has also been developed
(Shi and Sun 1990). A Kernel-based version of recursive least-squares algorithm
for nonlinear system identification purposes was developed by Vaerenbergh et al.
(2007). A nonlinear receptance coupling approach (NRCA) was used to identify a
nonlinear force function of a device for evaluation of machine dynamic characteris-
tics (Dhupia et al. 2008). Multiple integration method was used to identify nonlinear
single-variable systems from measured data (John 2001); note that in principle the
developed algorithm is applicable to identification of multivariable systems as well.
However, its application is left for future work. There are other nonlinear system
identification techniques based on neural networks (Ge et al. 2008) and fuzzy theory
(Li et al. 2010). See Kerschen et al. (2006) for a literature survey on the subject. The
primary differences between this work and other noted literature are: (1) the proce-
dure is computer-aided, (2) the procedure is not limited to inversion of memoryless
®
nonlinear terms, and (3) the corresponding MATLAB software is presented.
For static nonlinear terms the describing function model is simply the ratio of
the fundamental output (obtainable by a Fourier analysis described in Chap. 2) to
62 5 Multirange Nonlinear Controller Design

the amplitude of the applied sinusoidal input. The inverse problem considered here
is simply one of being given the describing function and finding the corresponding
nonlinearity. The analytical solution to this problem is given by Gelb and Vander
Velde (1968) and Colgren (2004), and the interested reader is referred to that
literature; the analytical solution will not be repeated here. The analytical solutions
are good only for the class of considered nonlinear functions, and for a complete an-
alytical solution, one should theoretically develop a catalogue of infinite analytical
solutions. As an alternative approach, a numerical technique may be considered.
In this approach, the user assumes the structure of the nonlinear function, and
the model parameters are identified. Therefore, the form of the assumed nonlinear
function may also have memory, and the fact that nonlinear function may have
memory does not limit the application of the procedure.
The inversion process in step 6 of Algorithm 5.1, has received limited attention
(Atherton 1970). At the first sight, it might appear that if a static output versus
input graph consisting of piecewise linear segments is to be inverted, it can be
accomplished readily by reordering the axis without the algorithm developed in this
section. To clarify this matter, assume that in step 6 of Algorithm 5.1, the nonlinear
function describing one of the controller gains as a function of the excitation
amplitude is determined to be given by Fig. 2.2. As was mentioned, the nonlinear
function in Fig. 2.2 is interpreted as the describing function model of an as yet
unknown nonlinear function; this unknown function describes the controller gain
as a function of the error or the excitation signal. This nonlinear function is given
by Fig. 2.1. In other words, Fig. 2.1 is the inversion of describing function model
given by Fig. 2.2. In step 6 of Algorithm 5.1, Fig. 2.1 must be obtained by inversion
of the describing function model given in Fig. 2.2. It is clear that reordering the
axes, without the algorithms developed in this section, would not result in the
desired nonlinear function. Furthermore, unlike inversion of Fig. 2.2, it is generally
not easy to obtain the inverted format as noted in Atherton (1970) and research
in Nassirharand and Mousavi Firdeh (2008, 2009c), Taylor and Strobel (1985),
and Nassirharand and Karimi (2006). In Gelb and Vander Velde (1968), early
need for inversion of describing function models was noted. Then Atherton (1970)
and Colgren (2004) gave directions and procedures for inversion of memoryless
describing function models.
The procedure for inversion of describing function models is computer-aided.
Similar to other available inversion methods, the user must select a model for
the nonlinear function. Previous works demonstrate that a general piecewise linear
function appears to be a good starting point (Taylor 1983; Taylor and Strobel 1985;
Nassirharand and Karimi 2006; Nassirharand and Mousavi Firdeh 2008, 2009c).
This choice has the following features:
1. Allows linear gain at small amplitudes where the plants usually behaves as a
linear system
2. Allows sudden jumps or reduction in controller gain at intermediate operating
regimes
3. Allows gradual increase or reduction of gains at large amplitudes
5.2 Controller Synthesis 63

Fig. 5.1 A simple candidate

nonlinear function for
inversion (Taylor 1983) f1(e) m1

m2

d
Error, e

The designer may also choose to examine the catalogue of nonlinear functions
given in the “Appendix” of Gelb and Vander Velde (1968), and select the model that
seems to suite his problem. The approach is independent of the selected model.
Once a model is selected, it must be described in terms of its parameters as a
®
MATLAB function. Then, a set of initial values for parameters of this model are
selected, and the describing function model of the assumed nonlinear function is
digitally obtained (Nassirharand and Karimi 2002). Finally, an objective function
measuring the error between the desired describing function model and that of
the actual nonlinear function is minimized. This minimization is done by using
®
fminsearch function of the MATLAB software. The developed algorithm for
describing function inversion is given below.
In step 1, the user must define a model for the nonlinearity. To accomplish
this task, the user could plot the given describing function model. Then, he could
compare this plot with the catalogue of describing function models given in
“Appendix” of Gelb and Vander Velde (1968). Finally, he would select the model
that may exhibit the desired gain/phase-amplitude behavior. The state of art for
controller synthesis applications suggests a nonlinear function similar to those in
Figs. 5.1, 5.2, and 5.3.
Maybe in future, nonlinear gain functions with memory will be needed. In step
2, only a guess is required. Again for controller synthesis applications, by looking at
the desired gain-amplitude plot, one could easily determine an initial guess for the
parameters. In step 3, the MATLAB function that is given in Chap. 2 is modified to
fit SISO models, and this function is used to obtain the describing function model
of the assumed nonlinear function. In step 4, the objective function is evaluated, and
64 5 Multirange Nonlinear Controller Design

Fig. 5.2 A complicated

nonlinear function with an
abrupt change at origin for f2 (e) m3
inversion (Nassirharand and
Mousavi Firdeh 2008)
d3

m2

d2
m1
d1 d2
d1

Error, e

then minimized using an optimization routine. Specifically, the fminsearch function

®
of the MATLAB that uses a Simplex (Reklaitis et al. 1983; Nelder and Mead 1965)
search technique is used here. In step 5, the inversion results are verified.

Algorithm 5.2
1. Assume a model for the desired actual nonlinear function.
2. Determine a set of initial values for the parameters of the model in step 1.
3. Use the getsidfinv2 function developed here, and obtain the describing
function (DF) model of the assumed nonlinear function.
4. Minimize the following objective function:
X
ED jGi;D .j!/  Gi;A .j!/j2 (5.5)
i
5.2 Controller Synthesis 65

m4
f3 (e)

d2

m3

d1

m2

m1 d2 d3

d1
Error, e

Fig. 5.3 A complicated nonlinear function with linear gain at origin for inversion (Nassirharand
and Mousavi Firdeh 2009c)

where
Gi;D .j!/ is the given describing function model at operating regime i ,
Gi;A .j!/
p is the actual describing function model, !is the frequency, and
j D 1.
If the assumed nonlinear function is memoryless, then Ewould be
independent of frequency.
5. Verify the inversion results. If satisfactory results are obtained, then stop;
otherwise, change the initial guess and/or the assumed model of step 1 and
go to step 3.
66 5 Multirange Nonlinear Controller Design

5.3 Software

5.3.1 Software for Step 5

An example software for optimization of the gains of the controller (i.e., a PID
controller) is presented in Fig. 5.4. The main program reads in the values of the
pseudofrequency responses of the plant and the desired open-loop model; initial
guesses for the controller gains are set, and finally the gains are optimized. Refer to
the comments inside the required functions for details.

Fig. 5.4 Listing of the software for optimization of the gains

5.3 Software 67

Fig. 5.4 (continued)

68 5 Multirange Nonlinear Controller Design

Fig. 5.4 (continued)

5.3 Software 69

5.3.2 Software for Describing Function Inversion

(Nassirharand 2009a)

An example software to achieve the inversion task is presented in Fig. 5.5. In this
figure, a main program, the function that calculates the Objective Function (5.5)
(i.e., errorinv), the function that generates the describing function model of the

Fig. 5.5 Listing of the software for describing function inversion. Source: (Nassirharand 2009a)
© Elsevier
70 5 Multirange Nonlinear Controller Design

Fig. 5.5 (continued)

nonlinear function (i.e., getsidfinv2), and the function that calculates the output of
the assumed nonlinear function (i.e., getyinv2) are given.
In the main program, the gain vs. amplitude data is supplied, and then the
®
fminsearch function of the MATLAB is invoked to minimize Objective Function
(5.5). This followed by verification of the results, and the main program ends.
The function errorinv first imposes any penalty functions that may be associated
with the undesirable values for the parameters of the assumed nonlinear function.
Then, the describing function model of the assumed nonlinear function is generated
followed by evaluation of Objective Function (5.5). Function getsidfinv2 calculates
the describing function model of the nonlinear function; the user must supply the
function getyinv2 that calculates the outputof the nonlinear function that is required
5.3 Software 71

Fig. 5.5 (continued)

for evaluation of the Fourier integrals as discussed in Chap. 2. Function getyinv2

is user-supplied, and it calculates nonlinearity function for the assumed sinusoidal
input.
72 5 Multirange Nonlinear Controller Design

Fig. 5.5 (continued)

5.4 Example: Single-Variable Case: Servo System

5.4.1 Nonlinear PID Controller Design

Consider the SISO servo system that was also considered in Sect. 2.2.1.2. It is
desired for the closed-loop system to mimic a second-order system with a damping
ratio of 0.307 and a natural frequency of 43.22; this approximately corresponds to a
settling time of 0.3 s and an overshoot of 37%. These performance measures were
originally achieved by Taylor and Strobel (1985) when designing a nonlinear PID
controller. For comparison purposes, the same performance measures are selected.
Algorithm 5.1 is executed as follows. Steps 1 and 2 are similar to that presented in
Sect. 4.4. After execution of steps 3 and 4, the results of step 5 are summarized in
Table 5.1 (Nassirharand and Karimi 2006).
5.4 Example: Single-Variable Case: Servo System 73

Fig. 5.5 (continued)

Table 5.1 The discrete No. ei KP KI KD

controller gains at various
operating regimes. Source: 1 0.0628 5:97 37.4 0.043
(Nassirharand and Karimi 2 0.0804 5:13 36.0 0.045
2006) © Elsevier 3 0.1005 4:71 36.0 0.050
4 0.1608 4:38 39.5 0.069
5 0.3215 6:81 55.9 0.129
6 0.6430 8:89 64.0 0.167
7 1.2811 9:84 67.7 0.193
8 2.5621 10:5 68.5 0.206

Table 5.2 The parameters of S1 S2 D ı

the synthesized nonlinear
function gains. Source: KP 5:10 11:16 0.0 0.189
(Nassirharand and Karimi KI 36:77 72:81 0.0 0.126
2006) © Elsevier KD 0:047 0:215 0.0 0.132

With reference to Fig. 5.1, the results of describing function inversion (step 6)
are summarized in Table 5.2 (Nassirharand and Karimi 2006).
74 5 Multirange Nonlinear Controller Design

Fig. 5.6 The normalized step responses with various amplitudes of excitation signal

Fig. 5.7 Stability test for the nonlinear PID design

Finally, the design is verified, and the results of the last step are shown in Figs. 5.6
and 5.7.
As was mentioned before, Eqs. 5.2 and 5.3 may be evaluated for frequencies
between 0 and 150 rad=s to measure the percent reduction in sensitivity; in this
case, 0 D 1924, 1 D 221, and R D 88:5. In other words, 88.5% reduction in
sensitivity is achieved.
5.4 Example: Single-Variable Case: Servo System 75

Fig. 5.8 The SIDF model of the open-loop system comprised of the nominal compensator, C  ,
and the nonlinear plant

5.4.2 Nonlinear Lead-Lag Controller Design (Nassirharand

and Mousavi Firdeh 2008)

The problem statement is the same as that given in Sect. 5.4.1 with the difference
that a nonlinear lead-lag controller is required to be designed. Steps 1 and 2 are
identical to those described in Sect. 4.4. In step 3, the middle model (at excitation of
0.325) is selected as the nominal model; a linear model is identified using invfreqs
® ®
function of MATLAB , and the MATLAB function described in Sect. A.1.1 is
used to design a lead compensator. The compensator gain is adjusted to obtain a
satisfactory step response when the resulting nonlinear system is simulated. The
designed lead compensator is of the following form:

C  D 1:5 .0:3514s C 3:36/ = .0:03303s C 1/ (5.6*)

In step 4, the desired open-loop behavior is determined by obtaining the

pseudofrequency response of the open-loop system comprised of the nominal
compensator, C  , and the nonlinear plant. The desired open-loop system behavior
is shown in Fig. 5.8.
This quasi-linear model serves as the target open-loop pseudofrequency response
model at all operating regimes; this target model is used in optimization in order to
obtain the parameters of the lead compensator at various operating regimes that
cause insensitivity of the nonlinear open-loop system to various amplitude levels of

*Source: (Nassirharand and Mousav Firdeh 2008) © Springer

76 5 Multirange Nonlinear Controller Design

Table 5.3 Parameters of the ı1 ı2 d1 d2

synthesized nonlinear
functions. Source: A0 0.1486 0.2811 0.1822 0.1452
(Nassirharand and Mousav A1 0.1762 0.5761 0.0079 0.1111
Firdeh 2008) © Springer B1 0.0583 0.2373 0.0006 0.0001
d3 m1 m2 m3
A0 0.3085 1.4951 1.1223 6.9811
A1 0.3338 0.4475 0.4105 1.2591
B1 0.0005 0.0282 0.0168 0.0286

excitation. The target model has approximately a 30-degree phase margin and an
approximate gain margin of 3.
In step 5, the set of linear lead compensators at various excitation amplitudes
is obtained by minimizing Objective Function (5.1). The controller parameters are:
A0 D f6.1141, 5.0390, 4.3959, 3.3293, 3.3834, 5.0765, 5.9311, 6.3986, 6.7062,
6.8369g, A1 Df0.6507, 0.6001, 0.5737, 0.5447, 0.7374, 1.0177, 1.1620, 1.2460,
1.3322, 1.1587g, and B1 Df0.0423, 0.0380, 0.0344, 0.0241, 0.0235, 0.0259, 0.0271,
0.0272, 0.0264, 0.0302g. Where, the parameters of the lead compensator are defined
as below:
C D .A1 s C A0 / = .B1 s C 1/ (5.7*)

The constants are nonlinear function of the error signal: (a) for a given value of
the error signal, the nonlinearities (obtained in step 5) are evaluated and this gives
®
a set of constants; (b) in SIMULINK implementation, the values obtained in (a)
are divided by error value; and (c) the values obtained in (b) sit in place of the lead
and/or lag coefficients and the resulting controller is excited by the error signal.
In step 6, the parameters of the synthesized nonlinear functions (see Fig. 5.2) for
compensator parameters are defined in Table 5.3.
The quality of fit is shown in Figs. 5.9, 5.10, and 5.11. These figures show
the results of describing function (DF) inversion process for A0 , A1 , and B1 .
The nonlinearity parameters are optimized in a fashion that DF model of that
nonlinearity mimics the desired amplitude-dependent gains of the compensator
parameters.
Finally, in step 7 the design is verified using numerical simulation. The normal-
ized step-response plots of the closed-loop system at various operating regimes
of interest are depicted in Fig. 5.12. This figure demonstrates the results of
design verification step. The normalized step responses at various amplitudes of
reference signals indicate that closed-loop system behavior is fairly insensitive to
the amplitude level of the command signal.
For stability, the describing function models of the synthesized nonlinear closed-
loop system at various operating regimes are generated. The stability test is shown

*Source: (Nassirharand and Mousav Firdeh 2008) © Springer

5.4 Example: Single-Variable Case: Servo System 77

Fig. 5.9 The results of describing function inversion process for A0 . Source: (Nassirharand and
Mousav Firdeh 2008) © Springer

Fig. 5.10 The results of describing function inversion process for A1 . Source: (Nassirharand and
Mousav Firdeh 2008) © Springer

in Fig. 5.13. Also, it is apparent that closed-loop system is fairly insensitive to the
amplitudes of excitations. As was mentioned in step 7 of Algorithm 5.1, 0 and
1 may be evaluated to measure the percent reduction in sensitivity; in this case,
0 D 2308, 1 D 322, and R D 86. In other words, 86% reduction in sensitivity is
achieved.
78 5 Multirange Nonlinear Controller Design

Fig. 5.11 The results of describing function inversion process for B1 . Source: (Nassirharand and
Mousav Firdeh 2008) © Springer

Fig. 5.12 Normalized step-response plots of the synthesized nonlinear control system. Source:
(Nassirharand and Mousav Firdeh 2008) © Springer

5.4.3 Nonlinear H1 Controller Design

The problem statement is the same as that given in Sect. 5.4.1 with the difference
that a nonlinear H1 controller is required to be designed. The structure of the
nonlinear H1 controller is shown in Fig. 5.14 (Nassirharand and Mousavi Firdeh
2009c). Steps 1 and 2 are identical to those described in Sect. 5.3.1. In step
3, the model with the highest gain amplification effect (at excitation of 0.8) is
5.4 Example: Single-Variable Case: Servo System 79

Fig. 5.13 Stability test for the nonlinear lead-lag controller. Source: (Nassirharand and Mousav
Firdeh 2008) © Springer

Fig. 5.14 The considered

nonlinear fractional
transformation. Source:
(Nassirharand and Mousavi
Firdeh 2009c) © Wiley

selected as the nominal model; a linear model is identified using invfreqs function of
® ®
MATLAB , and the MATLAB function hinfsyn is used to design a H1 controller.
The controller gain is adjusted to obtain a satisfactory step response when the
resulting nonlinear system is simulated. The designed H1 controller is of the
following form:

1447s C 2:154e4
C  D 5:0 (5.8)
270:3s C 1:019e4

In step 4, the desired open-loop behavior is determined by obtaining the open-

loop sinusoidal-input describing function model comprised of C  and the nonlinear
plant. In step 5, the set of linear H1 controllers at various excitation amplitudes are
obtained as outlined previously; those are given below:
80 5 Multirange Nonlinear Controller Design

Table 5.4 Parameters of the ı1 ı2 ı3 d1 d2

identified nonlinear functions
A0 0.0102 0.0229 2.1452 0.0029 0.0685
A1 0.0011 0.0477 0.0599 0.2709 0.3943
B0 0.0120 0.0679 2.1806 0.0047 0.0001
B1 0.0327 0.0328 0.9938 0.0097 0.0233
m1 m2 m3 m4
A0 4.3368 2.6090 1.8661 5.0720
A1 0.0025 7.4417 0.2044 9.0010
B0 0.3462 1.1701 0.7885 1.8399
B1 3.8129 14.9105 1.1075 0.6282

 
1:0537; 0:8967; 1:0580; 1:1145; 1:4041; 1:6809; 1:9423; 1:8791;
A0 D 1e5
2:2523; 3:5026
 
7:3540; 6:7826; 6:8800; 7:0189; 7:9919; 8:7454; 9:1093; 8:7004;
A1 D 1e3
8:6104; 8:9454
 
0:6893; 0:6927; 0:8584; 1:0524; 0:8955; 0:8370; 0:8690; 0:8058;
B0 D 1e4
0:9209; 1:3263
 
383:8594; 367:8280; 385:5476; 266:7148; 172:5880; 155:4505;
B1 D
138:2746; 121:1289; 80:5201; 80:5201

Where the parameters of the H1 controller are defined as below:

A1 s C A0
C.s/ D (5.9)
B1 s C B0

In step 6, controller parameters are inverted. With reference to Fig. 5.3, the
parameters of the synthesized nonlinear functions for controller parameters are
defined in Table 5.4. The quality of fit is shown in Figs. 5.15, 5.16, 5.17, and 5.18.
Finally, in step 7, the design is verified using numerical simulation. For stability
test, see the sinusoidal-input describing function models of the final synthesized
closed-loop system as depicted in Fig. 5.19. Also, the normalized step-response
plots of the closed-loop system at various operating regimes of interest are depicted
in Fig. 5.20. From examination of this figure, it may be concluded that amplitude
insensitivity objective is satisfied. As was mentioned before, Eqs. 5.2–5.4 may be
evaluated to measure the percent reduction in sensitivity; in this case, 0  2308,
1  350, 1  477, R1  85, and R2  79. In other words, 85% reduction in
sensitivity is achieved when the nonlinear controller is used, and 79% reduction in
sensitivity is achieved with the nominal linear controller, C  , utilized.
5.4 Example: Single-Variable Case: Servo System 81

Fig. 5.15 Results of describing function (DF) inversion for controller parameter A0

Fig. 5.16 Results of describing function (DF) inversion for controller parameter A1
82 5 Multirange Nonlinear Controller Design

Fig. 5.17 Results of describing function (DF) inversion for controller parameter B0

Fig. 5.18 Results of describing function (DF) inversion for controller parameter B1

5.5 Example: Multivariable Case: Bank Angle Control

of a UAV

The UAV that is studied in Sect. 4.5 is considered (see Fig. 4.3). It is desired
to design a robust controller for this UAV. As was mentioned in Sect. 4.5, the
effectiveness of a single-range linear controller is investigated by Nassirharand
5.5 Example: Multivariable Case: Bank Angle Control of a UAV 83

Fig. 5.19 Stability test of the closed-loop system with the nonlinear H1 controller

Fig. 5.20 Normalized step responses of the designed nonlinear closed-loop system

and Mousavi Firdeh (2009b). In this case, a linear H1 controller (based on one
operating regime) is also designed; the corresponding step-response plots are shown
in Fig. 4.10. It is apparent that this is not a satisfactory design. The nonlinear
controller synthesis procedure is executed as follows.
In step 1, the desired linear model has a settling time of about 2 s with no over-
shoot. In step 2, the SIDF models for the amplitude set a D f0:75; 1; 2; 4; 8; 16; 32g
and a frequency range from 0.1 to 1,000 rad/s are generated as outlined in the
previous section. The results are depicted in Fig. 4.4. In step 3, arbitrarily, the
model at a D 1 is selected as the nominal model, and a linear proportional
controller is designed. The multivariable gains are Kp D Œ4384; 1912. In
84 5 Multirange Nonlinear Controller Design

Fig. 5.21 Normalized step-response plots of the closed-loop system with a nonlinear controller
and the nonlinear UAV. Source: (Teh 2010) © S.H. Teh

Fig. 5.22 The normalized roll-rate plots of the closed-loop system with a nonlinear controller and
the nonlinear UAV. Source: (Teh 2010) © S.H. Teh

order to characterize the controller gains at small, medium, and large signals, the
following amplitudes are selected: E D f0:05; 0:1; 0:3; 0:5; 0:7; 1; 2g. In step 4, with
E D Œ0:7; 0:7 and the frequency range from 0.1 to 50 rad/s, the desired open-loop
®
DF model is generated. In step 5, the fminsearch function of the MATLAB that
uses a Simplex search technique (Reklaitis et al. 1983) is used, and the Objective
Function (5.5) at various operating regimes is minimized; the data for frequencies
below 4 rad/s are weighted ten times more than the others to improve on the steady-
state conditions. The following controller gains are obtained:
5.5 Example: Multivariable Case: Bank Angle Control of a UAV 85

 
f5:26; 5:33; 4:84; 4:69; 4:44; 4:40; 4:38g
Kp D 1000
f1:51; 1:55; 1:68; 1:79; 1:94; 2:06; 2:12g

In step 6, the SIDF models of the controller gains obtained in the previous step
are inverted, and nonlinear functions of the type shown in Fig. 5.2 are identified.
The results of inversion follow:

ı1 D Œ0:1349; 1:19e5 ; ı2 D Œ51:08; 3:63e6 ;

d1 D Œ0:0; 20:53 ; d2 D Œ0:0; 74:77 ;

d3 D Œ0:0; 1:72e5 ; m1 D Œ5:35e3; 1:95e3 ;

m2 D Œ4:36e3; 1:18e6 ; m3 D Œ2:48e3; 4:06e6 :

In step 7, the normalized step responses of the closed-loop system with the
synthesized nonlinear multivariable controller are depicted in Figs. 5.21 and 5.22.
The settling time is about 1.5 s with no overshoot. A fourth-order nonlinear H1
controller is designed by Colgren (2004); in that design a settling time of 40 s with
no overshoot is achieved (see Fig. 4.12).
Chapter 6
Experimental Study: Unstable SISO Systems

6.1 Background

The presented methods developed so far are limited to stable nonlinear systems.
This is mainly due to the fact that Fourier integrals are not able to converge if
the outputs are not bounded; hence, the corresponding describing function models
are not obtainable. In this chapter, an approach is outlined that would allow the
techniques developed so far to be applied to unstable systems (Li and Nassirharand
2011). An experimental rig is designed and fabricated to demonstrate the approach.

6.2 Experimental Rig

The experimental rig is shown in Figs. 6.1, 6.2, and 6.3.

This setup is comprised of an AC motor with a rated output of 2 KW, rated torque
of 19.1 Nm, maximum instantaneous torque of 57.3 Nm, rated speed of 1,000 rpm,
maximum speed of 2,000 rpm, rotor inertia of 98.4e-4 kgm2, rated power rate of
37.0 KW/s, and a total weight of 27.4 kg. The encoder has a shaft diameter of
8 mm with a resolution of 2,500 pulses/revolution, and it requires a power supply of
5VDC. The gear ratio is 5:1. An additional pulley with mass configuration is added
to increase the degree of instability of the plant by constantly pulling down on the
pendulum from its equilibrium position as shown in Fig. 6.4.
The pendulum is an aluminum bar with an outer diameter of 22 mm, an inner
diameter of 15 mm, and a length of 250 mm. The electronic instrumentation
corresponds to National Instruments products with LabView® software. There are
also mechanisms for adjustable Coulomb friction and adjustable backlash that are
not utilized in this research.

A. Nassirharand, Computer-aided Nonlinear Control System Design: Using Describing 87

Function Models, DOI 10.1007/978-1-4471-2149-7 6,
© Springer-Verlag London Limited 2012
88 6 Experimental Study: Unstable SISO Systems

Fig. 6.1 The overall picture

of the experimental rig.
Source: (Li and Nassirharand
2011) © Sage

6.3 Nonlinear PID Synthesis

The problem statement is to develop an approach for synthesis of nonlinear PID

controllers for highly nonlinear unstable systems with no restrictions on the number
of nonlinear terms, nonlinearity type (could be continuous and/or discontinuous),
arrangement (could be in forward path and/or in the feedback path), and system
order; the approach is to be demonstrated by applying the procedure to a highly
nonlinear inverted pendulum experimental setup.
The developed synthesis procedure is composed of eight primary steps as follows
(Li and Nassirharand 2011):
1. Stabilize the unstable nonlinear system around the operating regimes of interest.
For problems that a mathematical model is available, one could use a variety
of techniques such as sliding mode techniques (Utkin et al. 1999). For cases
that a mathematical model is not available, intuition may be used. For example,
6.3 Nonlinear PID Synthesis 89

Fig. 6.2 Servomotor and pendulum assembly. Source: (Li and Nassirharand 2011) © Sage

Fig. 6.3 Coulomb friction adjuster and encoder. Source: (Li and Nassirharand 2011) © Sage

for the experimental setup considered here, it is found that a linear proportional
controller with a gain of ten may be used to stabilize the system.
2. Generate the sinusoidal-input describing function models of the stabilized
closed-loop system around the operating regimes of interest as outlined in
Chap. 2. The user may choose to fit a linear model to each of the identified
closed-loop frequency domain models. For example, for the example problem
90 6 Experimental Study: Unstable SISO Systems

Fig. 6.4 Pulley system to increase the degree of instability. Source: (Li and Nassirharand 2011) ©
Sage

under study at nominal conditions (i.e., a reference amplitude of 20 degrees), the

following closed-loop transfer function may be identified by using the invfreqs
function of MATLAB® .
2:055s C 7:065
Gcl D (6.1)
s 2 C 3:304s C 6:385

With the known stabilizing controller of step 1, and the closed-loop frequency
domain model of the previous step, use algebra and determine the corresponding
unstable open-loop frequency domain models denoted G .j!I ai /. For the
inverted pendulum problem, the corresponding unstable open-loop frequency
domain model at the nominal condition is determined to be as follows:
0:2055s C 0:7065
Gol D (6.2)
s2 C 1:249s  0:6803

The set of unstable open-loop models are determined using algebra, and they
are depicted in Fig. 6.5. The spread in the pseudo-Bode plots indicate that system
is highly nonlinear. If this were a linear plant, then there would be only one
pseudo-Bode plot for all excitation amplitudes.
3. Select one of the models of the previous step as the nominal model, and design
a PID controller. Then use the MATLAB® function developed in Appendix A
6.3 Nonlinear PID Synthesis 91

Fig. 6.5 Frequency domain models of the unstable plant (a corresponds to the amplitude of the
sinusoid at the input to the plant). Source: (Li and Nassirharand 2011) © Sage

or Appendix B, and design a linear PID controller denoted C  . In this case,

the model corresponding to Eq. 6.2 is selected, and the MATLAB® function
cad controller developed in Appendix B is used, and the following PD (integral
gain turns out to be negligible, i.e., KI  0) controller is designed.

C  .s/ D 0:09272s C 20:01 (6.3)

The primary inputs to the cad controller function are the plant given by
Eq. 6.2 and a desired closed-loop transfer function; in this case, a second-order
transfer function with a damping ratio of 0.9 and a settling time of 0.5 s is
specified to be the desired closed-loop transfer function. The output of the above-
mentioned function is the PD controller given by Eq. 6.3.
4. Generate the closed-loop sinusoidal-input describing function model of the
system comprised of the controller of the previous step and the nonlinear plant
similar to what is done in step 2. Then, with the known desired closed-loop SIDF
model and the known controller of step 3, use algebra, and determine the desired
open-loop frequency domain model denoted C  G  . For the inverted pendulum
problem, this is shown in Fig. 6.6.
5. Design a set of linear PID controllers at various operating regimes of interest
that mimic the desired open-loop behavior obtained in the previous step. This is
accomplished by minimizing the following objective function:

E .j!/ D jC  G   Ci .j!/ G .j!; ei jCi .j!/j/j (6.4)

G .j!; ei jCi .j!/j/ is evaluated by using the G .j!; ai / data of step 1

coupled with interpolation. This will allow one to execute this step with standard
available personal computers. The output of this step is a set of values for the
92 6 Experimental Study: Unstable SISO Systems

Fig. 6.6 The desired open-loop frequency domain model (e is the amplitude of the sinusoid at the
input to the controller). Source: (Li and Nassirharand 2011) © Sage

Table 6.1 The amplitude- e, error KP KD

dependent gains of the PD
controller. Source: (Li and 0.1338 21.62308 0:08978
Nassirharand 2011) © Sage 0.1453 19.95128 0:08963
0.1604 20.48715 0:028779
0.1624 20.23243 0:027646
0.1654 19.86186 0:026103
0.1679 19.56322 0:024987
0.1704 19.27333 0:024041
0.1829 15.85274 0:100826

parameters of the PID compensator at various operating regimes of interest. For

the inverted pendulum problem, the optimized values of the PD controller (the
gain for integral action is determined to be negligible) are given in Table 6.1.
These amplitude-dependent parameters are treated as describing function models
of the desired nonlinear gains of the nonlinear controller.
The corresponding open-loop system comprised of the controllers given in
Table 6.1 and the unstable plants derived in step 2 are depicted in Fig. 6.7.
6. Determine a set of nonlinear functions whose SIDF models mimic the SIDF
models of the compensator parameters that were determined in the previous step.
A typical nonlinear function that a user may select is shown in Fig. 5.4. The
parameters of this function are determined in such a manner that the SIDF model
approximates the amplitude-dependent gains of the compensator parameters at
various operating regimes of interest. Such a function will allow for high gains at
low amplitudes, constant gains at specific amplitude ranges, and gain reductions
or jumps at specific amplitudes. Note that this step is the inverse of activities of
step 2. In step 2, the model is available and the SIDF models are to be obtained;
6.3 Nonlinear PID Synthesis 93

Fig. 6.7 The frequency domain models of the open-loop system comprised of the amplitude-
dependent controllers and the unstable nonlinear plant. Source: (Li and Nassirharand 2011)
© Sage

Table 6.2 Results of SIDF ı1 ı2 d1 d2 d3

inversion. Source: (Li and
Nassirharand 2011) © Sage KP 2.14 2.15 2.19 0.444 6:55
KD 6.72 9.27 0.077 0.621 4:98
m1 m2 m3
KP 2.32 2,640.3 15.83
KD 0.621 107.2 1.938

in this step, the SIDF models are known, and the nonlinearity model must be
obtained. In order to determine the parameters of the nonlinearity function shown
in Fig. 5.2, optimization is used, and the fminsearch function of the MATLAB®
software is utilized. The details of this inversion process are (see Chap. 5): (a) use
the describing function generator MATLAB® function from Chap. 2 and generate
the describing function model of the nonlinearity, (b) if satisfactory results are
obtained, then stop; otherwise, use fminsearch function of the MATLAB® in
order to modify the independent variables and go to (a).
With reference to Fig. 5.2, for the inverted pendulum example, the results of
SIDF inversion using the software outlined in Chap. 5 are given by Table 6.2.
The quality of fit is shown in Figs. 6.8 and 6.9.
Finally, the design is verified using experiments. The normalized step-response
plots of the closed-loop system at various operating regimes of interest are depicted
in Fig. 6.10. This figure demonstrates the results of design verification step. The
normalized step responses at various amplitudes of reference signals indicate that
closed-loop system behavior is fairly insensitive to the amplitude level of the
command signal. It is observed that at lower step magnitudes, there exist high-
frequency oscillations. The main cause of this is because at the connection point of
the pendulum to the shaft of the motor, there is a two degrees free play. The stability
94 6 Experimental Study: Unstable SISO Systems

Fig. 6.8 The quality of fit after SIDF inversion for the proportional nonlinear gain. Source:
(Li and Nassirharand 2011) © Sage

Fig. 6.9 The quality of fit after SIDF inversion for the derivative nonlinear gain. Source: (Li and
Nassirharand 2011) © Sage

test is shown in Fig. 6.11. In this figure, the pseudofrequency response plots of the
closed-loop system comprised of the designed nonlinear PD compensator and the
nonlinear process are successfully generated; hence, stability at operating regimes
of interest is demonstrated.
A linear controller based on one frequency domain model is designed for
comparison purposes; the corresponding normalized step-response plots are shown
in Fig. 6.12.
It is apparent that the single-range linear controller design is not satisfactory
(because of sensitivity and large steady-state errors), and the nonlinear PD controller
is preferred. Another linear controller based on two frequency domain models is also
designed (see Chap. 4); the corresponding normalized step-response plots are shown
in Fig. 6.13.
6.4 Conclusions 95

Fig. 6.10 Normalized step-response plots of the synthesized nonlinear closed-loop system.
Source: (Li and Nassirharand 2011) © Sage

Fig. 6.11 Stability test (r corresponds to the amplitude of the sinusoid for the reference signal).
Source: (Li and Nassirharand 2011) © Sage

These results are competitive with the nonlinear PD controller results; these
results are consistent with prior theoretical work in previous chapters.

6.4 Conclusions

In short, the synthesis procedure involves stabilization of the nonlinear plant

followed by determination of the SIDF models of the stabilized closed-loop system.
Algebra is used, and the unstable frequency domain open-loop models are obtained
followed by a linear controller design at nominal conditions. The desired open-loop
96 6 Experimental Study: Unstable SISO Systems

Fig. 6.12 The normalized step-response plots with the single-range linear controller. Source:
(Li and Nassirharand 2011) © Sage

Fig. 6.13 The normalized step-response plots with the dual-range linear controller. Source:
(Li and Nassirharand 2011) © Sage

frequency domain model is determined, and a set of linear controllers that force the
open-loop frequency domain models at all operating regimes that mimic that of the
desired is determined by optimization. The amplitude-dependent gains are inverted
to obtain the nonlinear functions describing the nonlinear gains of the controller.
Finally, design is verified by obtaining the normalized step-response plots as well
as the corresponding closed-loop frequency domain models for stability test. An
experimental setup consisting of an AC servomotor and an inverted pendulum is
used to demonstrate the procedure. Based on the results of the experimental work,
the following conclusions are made: (1) the nonlinear PID synthesis procedure is
effective and viable to be applied to highly nonlinear SISO systems, and (2) a stable
robust design may be obtained.
Appendix A
Closed-Form Solution to Linear Classical
Controllers

A.1 Linear Lead-Lag Controller Design

Consider the linear controller design problem as depicted in Fig. A.1.

Closed-form solution for the design of lead-lag controllers based on an exact
model-matching criterion is developed (Nassirharand and Karimi 2004c). The
problem statement is to design a linear lead-lag controller for a linear plant in a
unity feedback configuration that would satisfy a set of user-defined performance
measures. The linear plant is expressed as follows:

g1 C g2 j
Gp .j!/ D (A.1*)
g3 C g4 j

The lead-lag controller model is given below:

a0 C a1 s C a2 s 2
Gc .s/ D ;s D j! (A.2*)
1 C b1 s C b2 s 2

The desired closed-loop system behavior, or the desired reference linear model,
is denoted F :
h1 C h2 j Y1
F .j!/ D D (A.3*)
h3 C h4 j Y2

The designed closed-loop system is denoted Gcl :

Gc Gp Z1
Gcl D D (A.4*)
1 C Gc Gp Z2

*Source: (Nassirharand and Karimi 2004c) © Manchester University Press

A. Nassirharand, Computer-aided Nonlinear Control System Design: Using Describing 97

Function Models, DOI 10.1007/978-1-4471-2149-7,
© Springer-Verlag London Limited 2012
98 Appendix A Closed-Form Solution to Linear Classical Controllers

Fig. A.1 The structure of the

considered feedback system. Gc Gp
Source: (Nassirharand and
Karimi 2004c) © Manchester
University Press

Then, the exact model matching objective function is formed:

Z!2 ˇ ˇ
ˇ Z1 Y1 ˇ2
0
J D ˇ ˇ  ˇ d! (A.5*)
Z2 Y2 ˇ
!1

Minimizing the above function, with respect to the compensator parameters,

results in a set of nonlinear algebraic equations that must be solved numerically.
However, the goal is to develop a closed-form solution. Notice that under optimality
conditions Z 1
Z2
D YY12 ) Z1 Y2  Z2 Y1 D 0. Therefore, assuming that a solution
exists, the following alternate objective function may be selected:

Z!2
J D jZ1 Y2  Z2 Y1 j2 d! (A.6*)
!1

By substituting Eqs. A.1–A.4 into Eq. A.6, the following is obtained:

2 2 3
Z!2 a0 A C a1 B!  a2 A! 2 C b1 D! C b2 E! 2  E C
J D 4 5
 2 d! (A.7*)
!1 a0 B C a1 A! C a2 B! 2  b1 E! C b2 D! 2  D

where

A D h3 g1  h4 g2  h1 g1 C h2 g2 (A.8*)

B D h3 g2  h4 g1 C h1 g2 C h2 g1 (A.9*)

D D h1 g4 C h2 g3 (A.10*)

E D h1 g3  h2 g4 (A.11*)

Now by taking the derivative of the objective function with respect to the
compensator parameters, and by setting the result equal to zero, a set of linear

*Source: (Nassirharand and Karimi 2004c) © Manchester University Press

A.1 Linear Lead-Lag Controller Design 99

simultaneous algebraic equations is obtained as follows:

2 3 2 3 2 3
T0 0 T2 R1 S2 a0 S0
6 0 S2 R3 7 6 7 6 7
6 T2 0 7 6 a1 7 6 R1 7
6 7 6 7 6 7
6 T2 0 T4 R3 S4 7 : 6 a2 7 D 6 S2 7 (A.12*)
6 7 6 7 6 7
4 R1 S2 R3 Q2 0 5 4 b1 5 4 0 5
S2 R3 S4 0 Q4 b2 Q2

where
Z!2
 
Tk D A2 C B 2 ! k d! (A.13*)
!1

Z!2
Rk D .AD C BE/ ! k d! (A.14*)
!1

Z!2
Sk D .AE  BD/ ! k d! (A.15*)
!1

Z!2
 2 
Qk D D C E 2 ! k d! (A.16*)
!1

Therefore, by supplying the linear plant, the desired closed-loop transfer func-
tion, and the frequency range of interest, one could use Eq. A.12 and directly solve
for the compensator parameters.

A.1.1 Software
®
A MATLAB function is available that automates the above design process. The
®
MATLAB function is of the following form:

Œgc D leadlag.gp; f; w1; w2/

The printout of the software is given in Fig. A.2. The function inputs are (1) the
linear plant transfer function, gp, for which a compensator is to be designed; (2) the
desired closed-loop transfer function,f ; (3) the lower bound of the frequency range
of interest, w1; and (4) the upper bound of the frequency range of interest, w2. The
function output is the designed lead-lag compensator, gc.

*Source: (Nassirharand and Karimi 2004c) © Manchester University Press

100 Appendix A Closed-Form Solution to Linear Classical Controllers

A.1.2 Demonstration Example Problem

Consider the example problem studied by Ogata (1990). The linear plant is of the
following form:

Fig. A.2 Listing of the leadlag MATLAB® function. Source: (Nassirharand and Karimi 2004c) ©
Manchester University Press
A.1 Linear Lead-Lag Controller Design 101

Fig. A.2 (continued)

4
Gp D (A.17)
s .s C 0:5/

It is desired to design a lead-lag controller that results in a closed-loop system that

would have the damping ratio of the dominant closed-loop poles equal 0.5 with an
undamped natural frequency of 5 rad/s; the static velocity error constant is to be 80
s–1 . The corresponding desired closed-loop transfer function is obtained as follows
Chen and Shieh (1970). The desired transfer function is of the following form:
102 Appendix A Closed-Form Solution to Linear Classical Controllers

Fig. A.3 Design comparison for linear lead-lag design example. Source: (Nassirharand and
Karimi 2004c) © Manchester University Press

xs C !n2
F D (A.18)
s2 C 2&!n s C !n2

where from the problem statement, !n D 5and & D 0:5. Based on the definition of
velocity error constant, kv , the following relation holds:
 
!n2
2&!n  x D0 ) x D 4:6875 (A.19)
kv

Hence,

4:6875s C 25
F D (A.20)
s 2 C 5s C 25
®
The developed MATLAB function is executed with w1 D 0 and w2 D 5:

gc D leadlag .Gp; F; 0; 5/

The designed compensator is of the following form:

3:75s 2 C 21:875s C 10
Gc D :Let"D 0:01: (A.21)
"s 2 C 3:2s C 1

The design given by Ogata 1990 using a classical technique is of the following
form:

10 .2s C 1/ .5s C 1/
Cogata .s/ D (A.22)
.0:1992s C 1/ .80:19s C 1/

The performance of the two compensators may be compared by studying

Fig. A.3.
A.2 Linear PID Controller Design 103

A.2 Linear PID Controller Design

With a similar formulation for the linear lead-lag controller design that is presented
in the previous section, the gains of a PID controller may be determined to be as
follows (Nassirharand et al. 2003):

Q
KP D (A.23*)
T2
    
T0 T2 KI R
D (A.24*)
T2 T4 KD S

where
Z!2
 
Tl D ! l A2 C B 2 d!; l D 0; 2 (A.25*)
!1

Z!2
RD .AC C BD/ d! (A.26*)
!1

Z!2
SD .AC C BD/ ! 2 d! (A.27*)
!1

Z!2
QD .BC  AD/ ! d! (A.28*)
!1

A D g1 h3  g2 h4  g1 h1 C g2 h2 (A.29*)

B D g2 h3 C g1 h4  g2 h1  g1 h2 (A.30*)

C D ! .g4 h1 C g3 h2 / (A.31*)

D D ! .g4 h2  g3 h1 / (A.32*)

Gc D KP C KI =s C KD s (A.33*)

and the rest of the variables are defined as the previous section (see Eqs. A.1
and A.3).

*Source: (Nassirharand et al. 2003) © Elsevier

104 Appendix A Closed-Form Solution to Linear Classical Controllers

Therefore, by supplying the linear plant, the desired closed-loop transfer func-
tion, and the frequency range of interest, one could use Eqs. A.23 and A.24 and
directly solve for the controller parameters.

A.2.1 Software
®
The above described method is automated via the development of a new MATLAB
®
function. The developed MATLAB function is of the following form:

Œkp; ki; kd  D getpid .w1; w2; ss; g; hd /

Where in the above command, Œkp; ki; kd  are the controller gains; the inputs
are (1) w1; w2 the frequency range of interest, (2) ss the step size for integration of
Eqs. A.25–A.28, (3) g the transfer function of the process, and (4) hd the transfer
function of the reference linear model. The software listing is shown in Fig. A.4.

A.2.2 Demonstration Example Problem

Consider the following first-order plus delay model studied by Wang et al. (1995):

e 5s
Gp .s/ D (A.34*)
10s C 1

The desired transfer function is of the following form:

e 5s
F .s/ D (A.35*)
20s C 1

Notice that the developed procedure must have all transfer functions as rational
functions in s; for this reason, we approximate the e 5s term as below (Kuo 1982):

1
e T s  (A.36*)
Œ1 C T s=nn

where T D 5, and we arbitrarily select the order of approximation as n D 4. By

substituting Eq. A.36 into Eqs. A.34 and A.35, the developed MATLAB® function
is executed with w1 D 0, w2 D 0.025, and h D 0.001. Using Wang et al. (1995), the
results are compared with four other techniques in Fig. A.5, and the corresponding
values of PID controller gains are tabulated in Table A.1. It is apparent that, in
this case, the developed approach as well as method given by Wang et al. (1995)
produces better results than other considered techniques.

*Source: (Nassirharand et al. 2003) © Elsevier

A.3 Proportional Plus Rate Feedback Controller Design 105

A.3 Proportional Plus Rate Feedback Controller Design

Consider the feedback system depicted in Fig. A.6. In this Appendix, the for-
mulation for design of proportional plus rate feedback controllers is presented
(Nassirharand 2011). The linear plant is expressed as follows:

Gp .j!/ D A .!/ C jB .!/ (A.37*)

Fig. A.4 Listing for PID design function. Source: (Nassirharand et al. 2003) © Elsevier

*Source: (Nassirharand 2011) © Elsevier

106 Appendix A Closed-Form Solution to Linear Classical Controllers

Fig. A.4 (continued)

A.3 Proportional Plus Rate Feedback Controller Design 107

Fig. A.5 Design comparison for PID controller design example. Source: (Nassirharand et al.
2003) © Elsevier

Table A.1 Comparison of PID gains with different design methods – (1)
Wang et al. (1995), (2) Rivera et al. (1986), (3) Ziegler and Nichols (1942),
and (4) Zhuang and Atherton (1993)
Present (1) (2) (3) (4)
Kp 0.42 0.42 1.00 2.28 1.99
KI 0.04 0.04 0.08 0.26 0.13
KD 0.30 0.17 2.00 4.89 3.84

Ref. Error Gp(s)

A2
Output
+

Kts

Fig. A.6 The structure of the proportional and rate feedback controller. Source: (Nassirharand
2011) © Elsevier

The desired closed-loop system behavior, or the desired reference linear model,
is denoted F :
F .j!/ D C .!/ C jD .!/ (A.38*)

The designed closed-loop system is denoted Gcl :

A .!/ A2 C jB .!/ A2 X1 .j!/

Gcl D D
.1 C Kt !B .!/ C A .!/ A2 / C j .Kt A .!/ ! C A2 B .!// X2 .j!/
(A.39*)

*Source: (Nassirharand 2011) © Elsevier

108 Appendix A Closed-Form Solution to Linear Classical Controllers

Then, the exact model matching objective function is formed:

Z!2 ˇ ˇ2
ˇ X1 .j!/ ˇ
0
J D ˇ ˇ  C .!/  jD .!/ˇˇ d! (A.40*)
X2 .j!/
!1

Minimizing the above function, with respect to the compensator parameters,

results in a set of nonlinear algebraic equations that must be solved numerically.
However, the goal is to develop a closed-form solution. Notice that under optimality
conditions.

X1 .j!/
D C .!/ C jD .!/ ) X1 .j!/  X2 .j!/ C  j X2 .j!/ D .!/ D 0
X2 .j!/

Therefore, assuming that a solution exists, the following alternate objective

function may be selected:
Z!2
J D jX1 .j!/  X2 .j!/ C  j X2 .j!/ D .!/j2 d! (A.41*)
!1

By comparing Eqs. A.35 and A.37, the following is obtained:

20 12 3
A .!/ A2  C .!/  C .!/ Kt !B .!/  C .!/ A .!/ A2
Z!2 6 B C 7
6 B D .!/ Kt A .!/ !  D .!/ B .!/ A2 C B .!/ A2 C 7
J D 6 6BB
C
C
7 d!
7
4 @ CKt A .!/ !C .!/  A2 B .!/ C .!/  D .!/
!1
A 5
D .!/ Kt !B .!/  D .!/ A .!/ A2
(A.42*)

Now by taking the derivative of the objective function with respect to the
controller parameters, and by setting the result equal to zero, a set of linear
simultaneous algebraic equations is obtained as follows:
    
˛1 ˇ1 Kt 1
D (A.43*)
˛2 ˇ2 A2 2

where
Z!2
  
˛1 D D .!/ ! A2 .!/ C B 2 .!/ d! (A.44*)
!1

*Source: (Nassirharand 2011) © Elsevier

A.3 Proportional Plus Rate Feedback Controller Design 109

Z!2 n
 2 h io
ˇ1 D A .!/ C B 2 .!/ .1  C .!//2 C D 2 .!/ d! (A.45*)
!1

Z!2
   
˛2 D ! 2 A2 .!/ C B 2 .!/ C 2 .!/ C D 2 .!/ d! (A.46*)
!1

ˇ2 D ˛1 (A.47*)

Z!2
  
1 D A .!/ C 2 .!/ C D 2 .!/  A .!/ C .!/  B .!/ D .!/ d! (A.48*)
!1
Z!2
  
2 D !B .!/ C 2 .!/ C D 2 .!/ d! (A.49*)
!1

Therefore, by supplying the linear plant, the desired closed-loop transfer func-
tion, and the frequency range of interest, one could use Eq. A.43 and directly solve
for the controller parameters.

A.3.1 Software
®
The design process is automated via development of a new MATLAB function.
®
The MATLAB function is of the following form:

ŒKt; A2 D get L rate.w1; w2; h; g; f /

The function inputs are (1) the lower bound of the frequency range of interest,
w1; (2) the upper bound of the frequency range of interest, w2; (3) the integration
step size; (4) the linear plant transfer function, g, for which a compensator is to be
designed; and (5) the desired closed-loop transfer function,f . The function outputs
are the tachometer gain and the amplifier gain. The corresponding software listing
is given in Fig. A.7 (Nassirharand 2011).
®
The design process is automated via development of a new MATLAB function.
®
The MATLAB function is of the following form:

ŒKt; A2 D get L rate.w1; w2; h; g; f /

*Source: (Nassirharand 2011) © Elsevier

110 Appendix A Closed-Form Solution to Linear Classical Controllers

The function inputs are (1) the lower bound of the frequency range of interest,
w1; (2) the upper bound of the frequency range of interest, w2; (3) the integration
step size; (4) the linear plant transfer function, g, for which a compensator is
to be designed; and (5) the desired closed-loop transfer function,f . The function
outputs are the tachometer gain and the amplifier gain. The corresponding software
listing follows.

Fig. A.7 The software listing for proportional and rate feedback design. Source: (Nassirharand
2011) © Elsevier
A.3 Proportional Plus Rate Feedback Controller Design 111

Fig. A.7 (continued)

A.3.2 Demonstration Example Problem

Consider the example problem studied by D’Azzo and Houpis (1966). The linear
plant is of the following form:
0:88
Gp D (A.50*)
s.s C 1/ .0:2s C 1/

It is desired to design an amplifier and rate feedback controller that results in a

closed-loop system that would have Mm D 1:25 and !m D 3. These performance
measures correspond to a desired second-order transfer function with & D 0:4472
and !n D 3:873; i.e.:
15
F D (A.51*)
s 2 C 3:464s C 15

In arriving at the above desired transfer function, the following formulas were
utilized:
1
Mm D p (A.52*)
2& 1  & 2

*Source: (Nassirharand 2011) © Elsevier

112 Appendix A Closed-Form Solution to Linear Classical Controllers

Fig. A.8 Design comparison for the proportional and rate feedback controller example. Source:
(Nassirharand 2011) © Elsevier

p
!m D !n 1  2& 2 (A.53*)

The developed MATLAB function is executed with w1 D 2:5 and w2 D 10 as

follows:

ŒKt; A2 D get L rate .2:5; 10; 0:01; gp; f /

The designed gains are Kt D 6:3557 and A2 D 20:4545.

The designed gains using a classical technique are Kt D 4 and A2 D 11:4
(D’Azzo and Houpis 1966).
The performance of the two compensators may be compared by studying
Fig. A.8. It should be noted that by selecting F the same as the closed-loop transfer
function with gains of D’Azzo and Houpis (1966), then the present approach also
gives the same exact gains.

A.4 Proportional-Integral Plus Rate Feedback Controller

Design

The structure of the considered feedback system is depicted in Fig. A.9.

The problem statement is to design a PI plus rate feedback compensator for
a linear plant (as shown in Fig. A.9) that would satisfy a set of user-defined
performance measures. The formulation for PI plus rate feedback follows. The linear
plant is expressed as follows (Nassirharand 2011):

*Source: (Nassirharand 2011) © Elsevier

A.4 Proportional-Integral Plus Rate Feedback Controller Design 113

KI / s
Error Gp(s)
Ref.

Kp

Kts

Fig. A.9 Structure of the proportional-integral plus rate feedback controller. Source:
(Nassirharand 2011) © Elsevier

Gp .j!/ D A .!/ C jB .!/ (A.54*)

The desired closed-loop system behavior, or the desired reference linear model,
is denoted F :

F .j!/ D C .!/ C jD .!/ (A.55*)

The designed closed-loop system is denoted Gcl :

.AKI  KP B!/ C j .BKI C AKP !/

Gcl D
.AKt ! 2 C AKI  KP B!/ C j .! C BKt ! 2 C BKI C AKP !/
X1 .j!/
D (A.56*)
X2 .j!/

Then, the exact model matching objective function is formed:

Z!2 ˇˇ ˇ2
X1 .j!/ ˇ
0
J D ˇ ˇ  C .!/  jD .!/ˇˇ d! (A.57*)
X2 .j!/
!1

Minimizing the above function, with respect to the compensator parameters,

results in a set of nonlinear algebraic equations that must be solved numerically.
However, the goal is to develop a closed-form solution. Notice that under optimality
conditions

*Source: (Nassirharand 2011) © Elsevier

114 Appendix A Closed-Form Solution to Linear Classical Controllers

X1 .j!/
D C .!/ C jD .!/ ) X1 .j!/  X2 .j!/ C  j X2 .j!/ D .!/ D 0:
X2 .j!/

Therefore, assuming that a solution exists, the following alternate objective

function may be selected:

Z!2
J D jX1 .j!/  X2 .j!/ C  j X2 .j!/ D .!/j2 d! (A.58*)
!1

By comparing Eqs.A.50 and A.54, the following is obtained:

2  2 3
AKI  KP B!  AC Kt ! 2  AC KI C KP CB!
Z!2 6 7
C!D C BDKt ! 2 C BDKI C ADKP !
6 7
J D 6 6 7 d! (A.59*)
 7
4 BKI C AKP !  DAKt !  ADKI C KP DB! 5
2 2
!1
C
!C  BC Kt ! 2  BC KI  AC KP !

Now by taking the derivative of the objective function with respect to the
compensator parameters, and by setting the result equal to zero, a set of linear
simultaneous algebraic equations is obtained as follows:
XY D Z (A.60*)
where
2 3
˛1 ˇ1 1
X D 4 ˛2 ˇ2 2 5 (A.61*)
˛3 ˇ3 3
2 3
KI
Y D 4 KP 5 (A.62*)
Kt
2 3
1
Z D 4 2 5 (A.63*)
3
Z!2 n
 2 h io
˛1 D A C B 2 .C  1/2 C D 2 d! (A.64*)
!1

ˇ1 D 0 (A.65*)

*Source: (Nassirharand 2011) © Elsevier

A.4 Proportional-Integral Plus Rate Feedback Controller Design 115

Z!2
˚ 2 2  

1 D ! A C B 2 C 2 C D 2  C d! (A.66*)
!1
Z!2
˚  

1 D B! C 2 C D 2 C ! .AD  BC / d! (A.67*)
!1

˛2 D 0 (A.68*)

Z!2 n
 h io
ˇ2 D ! 2 A2 C B 2 .C  1/2 C D 2 d! (A.69*)
!1

Z!2
˚ 3  2 

2 D ! D A C B 2 d! (A.70*)
!1

Z!2
˚ 2  2 

2 D ! A C C D 2  .AC C BD/ d! (A.71*)

!1

Z!2
˚ 2  

˛3 D ! C .C  1/ C D 2 A2 C B 2 d! (A.72*)
!1

Z!2
˚ 3  2 

ˇ3 D ! D A C B 2 d! (A.73*)
!1

Z!2
˚ 4 2  

3 D ! A C B 2 C 2 C D 2 d! (A.74*)
!1

Z!2
˚  

3 D ! 3 B C 2 C D 2 d! (A.75*)
!1

Therefore, by supplying the linear plant, the desired closed-loop transfer func-
tion, and the frequency range of interest, one could use Eq. A.60 and directly solve
for the compensator gains.

*Source: (Nassirharand 2011) © Elsevier

116 Appendix A Closed-Form Solution to Linear Classical Controllers

A.4.1 Software

The design process is automated via development of a new MATLAB® function.

The MATLAB® function is of the following form:

ŒKI; KP; Kt D get PIL rate.w1; w2; h; g; f /

The function inputs are (1) the lower bound of the frequency range of interest,
w1; (2) the upper bound of the frequency range of interest, w2; (3) the integration
step size, h; (4) the linear plant transfer function, g, for which a compensator is to be
designed; and (5) the desired closed-loop transfer function, f . The function outputs
are the integral, proportional, and tachometer gains. The corresponding software
listing is given in Fig. A.10 (Nassirharand 2011).

A.4.2 Demonstration Example Problem

Consider the example problem studied by D’Azzo and Houpis (1966). The linear
plant is of the following form:

0:88
Gp D (A.76*)
s .s C 1/ .0:2s C 1/

It is desired to design a PI plus rate feedback compensator that results in a closed-

loop system that would have Mm D 1:25 and !m D 3. These performance measures
correspond to a desired second-order transfer function with & D 0:4472 and !n D
3:873; i.e.:

15
F D (A.77*)
s 2 C 3:464s C 15

The developed MATLAB® function is executed with w1 D 2:5 and w2 D 10 as

follows:

ŒKI; KP; Kt D get PIL rate .0:001; 6; 0:01; Gp; F /

The designed gains are KI D 0; KP D 20:45; and Kt D 5:74.

The gains designed by D’Azzo and Houpis (1966) using a classical technique for
only proportional and rate feedback are Kt D 4 and Kp D 11:4.
The performance of the two compensators may be compared by studying
Fig. A.11.

*Source: (Nassirharand 2011) © Elsevier

A.4 Proportional-Integral Plus Rate Feedback Controller Design 117

Fig. A.10 Software listing for proportional-integral plus rate feedback design. Source:
(Nassirharand 2011) © Elsevier
118 Appendix A Closed-Form Solution to Linear Classical Controllers

Fig. A.10 (continued)

®
Remark The MATLAB functions, which are developed in this appendix, are
®
integrated with the cad controller MATLAB function as described in Appendix B.
With the cad controller function, the specific values of w1 and w2 are determined
by optimization; if f requires enhancing (see Appendix B), then that function will
automatically perform enhancing.
A.4 Proportional-Integral Plus Rate Feedback Controller Design 119

Fig. A.10 (continued)

Fig. A.11 Design comparison for proportional-integral plus rate feedback controller. Source:
(Nassirharand 2011) © Elsevier
Appendix B
Algebraic Linear Multivariable Controller
Design

B.1 Theory

The design procedure is demonstrated for a typical two-input two-output process

(Mousavi Firdeh et al. 2006). With reference to Fig. B.1, the following relations
hold:
   
yi D ei gi i ci i C gij cj i C ej gi i cij C gij cjj ; fi D 1; 2I j D 1; 2I i ¤ j g
(B.1*)

To achieve decoupling, the following relations must hold:

gij cjj
cij D  ; fi D 1; 2I j D 1; 2I i ¤ j g (B.2*)
gi i

To achieve tracking, compare Eqs. B.1 and B.2 to obtain the following:

yi ci i ki
D ; i D 1; 2 (B.3*)
ri 1 C ci i ki

where
gij gj i
ki D gi i  ; fi D 1; 2I j D 1; 2I i ¤ j g (B.4*)
gjj

In order to design c11 and c22 , a factorization approach may be utilized.

The class of all compensators that stabilize a linear plant G.s/ may be parame-
terized in terms of a stable function parameter r.s/ as follows (Youla et al. 1976;
Vidyasagar 1985):

*Source: (Mousavi Firdeh et al. 2006) © Elsevier

A. Nassirharand, Computer-aided Nonlinear Control System Design: Using Describing 121

Function Models, DOI 10.1007/978-1-4471-2149-7,
© Springer-Verlag London Limited 2012
122 Appendix B Algebraic Linear Multivariable Controller Design

r1 e1
c11 g11

y1
c12

g12
c21

g21

r2 e2
c22 g22

y2

Fig. B.1 A linear two-input two-output system

P .s/  r.s/D.s/
C.s/ D (B.5*)
Q.s/ C r.s/N.s/

where C.s/ is the stabilizing compensator, N.s/ and D.s/ are stable coprime
N.s/
transfer function factors of the plant satisfying G.s/ D D.s/ , P .s/ and Q.s/ are
stable coprime transfer functions and satisfy the Bezout’s identity: P .s/N.s/ C
Q.s/D.s/ D 1, and r.s/ is an as yet unknown stable transfer function parameter.
A transfer function is said to be coprime if there are no common pole-zero
cancellations. This transfer function parameter must be selected in such a manner
that when substituted in Eq. B.5, the desired compensator would be obtained.
In continuation, define the needed transfer functions as follows:

a0 1 C b 0 1 j
N.j!/ D (B.6*)
a0 2 C b 0 2 j

*Source: (Mousavi Firdeh et al. 2006) © Elsevier

B.1 Theory 123

c1 C d1 j
D.j!/ D (B.7*)
c2 C d2 j
iP
Dm
ai :.j!/i
i D0
r.j!/ D iP
Dn
(B.8*)
1C bi :.j!/ i
i D1

e1 C f1 j
P .j!/ D (B.9*)
e2 C f2 j
N 1 C N2 j Y1
y;u D
hD D (B.10*)
P1 C P2 j Y2

where N; D; r; and P are defined as before and hD

y;u is the desired reference linear
model. The input–output map in terms of the coprime factors is of the following
form:
Z1
hy;u D N.s/D.s/r.s/ C N.s/P .s/ D (B.11*)
Z2

Then, the exact model matching objective function is formed:

Z!2 ˇ ˇ
ˇ Z1 Y1 ˇ2
J D ˇˇ
0
 ˇˇ d! (B.12*)
Z2 Y2
!1

Minimizing the above function, with respect to the compensator parameters,

results in a set of nonlinear algebraic equations that must be solved numerically.
However, the goal is to develop a closed-form solution. Notice that under optimality
conditions Z Z2 D Y2 ) Z1 Y2  Z2 Y1 D 0. Therefore, assuming that a solution
1 Y1

exists, the following alternate objective function may be selected:

Z!2
J D jZ1 Y2  Z2 Y1 j2 d! (B.13*)
!1

By substituting Eqs. B.11 and B.6–B.10 into Eq. B.13, the following relation is
obtained:
Z!2 " .AA C BA C B C  B D/2 C#
1 2 1 2
0
E D d! (B.14*)
2
!
.AA 2  BA 1 C B 1 D C B2 C /
1

*Source: (Mousavi Firdeh et al. 2006) © Elsevier

124 Appendix B Algebraic Linear Multivariable Controller Design

where

A1 D a0  a2 ! 2 C a4 ! 4     (B.15*)

A2 D a1 !  a3 ! 3 C a5 ! 5     (B.16*)

B 1 D 1  b 2 ! 2 C b4 ! 4     (B.17*)
   
A D a0 1 c1  b 0 1 d1 .e2 P1  f2 P2 /  a0 1 d1 C b 0 1 c1 .f2 P1 C e2 P2 / (B.18*)

   
B D a0 1 c1  b 0 1 d1 .f2 P1 C e2 P2 / C a0 1 d1 C b 0 1 c1 .e2 P1  f2 P2 / (B.19*)

C DEG (B.20*)

D DF H (B.21*)
   
E D a0 1 e1  b 0 1 f1 .c2 P1  d2 P2 /  a0 1 f1 C b 0 1 e1 .c2 P2 C P1 d2 / (B.22*)
   
F D a0 1 e1  b 0 1 f1 .c2 P2 C P1 d2 / C a0 1 f1 C b 0 1 e1 .c2 P1  d2 P2 / (B.23*)

   
G D a0 2 c2  b 0 2 d2 .e2 N1  f2 N2 /  a0 2 d2 C b 0 2 c2 .e2 N2 C f2 N1 / (B.24*)

   
H D a0 2 c2  b 0 2 d2 .e2 N2 C f2 N1 / C a0 2 d2 C b 0 2 c2 .c2 N1  f2 N2 / (B.25*)

By imposing the necessary optimality condition, i.e., ra;b D 0, a set of

simultaneous linear algebraic equations is obtained as follows:

XY D Z (B.26*)

where
2 3
T0 0 T2    R1 S2 R3 
6 0  S2 R3 S4  7
6 0 T2 7
6 T2 T4    R3 S4 R5  7
6 0 7
6 :: :: :: :: :: :: :: :: 7
6 7
6 : : : : : : : : 7
X D6 7 (B.27*)
6 R1 S2 R3    Q2 0 Q4  7
6 7
6 S2 R3 S4    0 Q4 0  7
6 7
6 R3 S4 R5    Q4 0 Q8  7
4 5
:: :: :: :: :: :: :: ::
: : : : : : : :

*Source: (Mousavi Firdeh et al. 2006) © Elsevier

B.1 Theory 125

 T
Y D a0 a1 a2    b1 b2 b3    (B.28*)

 T
Z D R1 S2 R3    0 Q2 0    (B.29*)

Z!2
Tl D ! l .A2 C B 2 / d! (B.30*)
!1

Z!2
Rl D ! l .AD  BC / d! (B.31*)
!1

Z!2
Sl D ! l .AC C BD/ d! (B.32*)
!1

Z!2
Ql D ! l .C 2 C D 2 / d! (B.33*)
!1

Therefore, by supplying the linear plant, the desired closed-loop transfer func-
tion, and the frequency range of interest, one could use Eq. B.26 and directly solve
for the unknown function parameter r.s/.

B.1.1 Discussion (Nassirharand and Karimi 2004c)**

1. “Desired frequency range” – Based on the experience of the authors, the

frequency range of interest is not of major concern. The user may determine
the bandwidth of the desired closed-loop system, and select the lower and upper
bounds of the desired frequency range. Then, the developed approach may be
executed; if the transient response is not satisfactory, then the upper and/or the
lower frequency limits may be multiplied by an arbitrary factor such as 2˛ (where
˛ is a trial number) until the desired transient response is obtained. Similarly, if
the steady-state response is not satisfactory, then the lower and/or the upper limits
may be divided by an arbitrary factor such as 2˛ until the desired steady-state
response is obtained. The procedure is not very sensitive to the selected frequency
range. Based on application of the software to a number of different problems, it
is recommended that w1 D 0 and w2 D 10 are to be used as a starting point; this
recommended range may have to be modified as outlined earlier.

*Source: (Mousavi Firdeh et al. 2006) © Elsevier

**Source: (Nassirharand and Karimi 2004c) © Manchester University Press
126 Appendix B Algebraic Linear Multivariable Controller Design

2. “Bad frequency range” – If a bad frequency range is used, the developed

procedure will not find the desired solution. The question is when the user can tell
if a bad frequency range is used? The user may be assured that a bad frequency
range is not used if he has gone through the exercise of item 1 above without
any useful results. In this case, the desired performance measures may have to be
modified or a different feedback structure may have to be selected.
3. “Desired transfer function” – For most problems the desired closed-loop transfer
function would be a second order transfer function that possesses the desired
dominant poles, and satisfies the desired steady-state error conditions. The user
may translate the time and/or frequency performance measures to the desired
natural frequency and the desired damping ratio by considering the performance
specification equations in standard automatic control textbooks (Rowland 1986);
then the steady-state error specifications determine if a zero is also required.
The zero may be determined from the relations for definition of a steady-state
error. Also, in some cases, additional high order dynamic terms may have to be
introduced to balance the matching of polynomial coefficients (see Eq. B.13);
this is called enhancing the desired transfer function (Pujara 1988).
4. “Meaningless designs” – If the designed parameters of the controller are such
that results in an unstable closed-loop feedback system, then the designer must
explore the possibility of modifying the required performance measures or a
different feedback structure may have to be assumed.
5. “Improper designs” – It is possible that the model-matching solution is an
improper transfer function; as is known, improper transfer functions are not
realizable. In such cases, one approach to accommodate such designs is to
include adequate high-order dynamic terms until the order of the denominator
polynomial equals that of the numerator. Alternatively, one could add "s n terms
(n D 1; 2; :::) to the denominator until the order of the denominator equals
that of the numerator. This is a standard practice in realizing such improper
transfer functions as a pure derivative. The numerical value of " must be selected
such that this does not require unreasonably small integration step size, and
it also must not be so large that it alters the behavior of the original transfer
function.
6. “Pole-zero shifting” – It should be noted that as the order of the plant increases,
the traditional techniques, which give insight into how the pole-zero locations
of the controller affect the response, lose effectiveness. As was mentioned
earlier, the presented approach complements the traditional techniques (such
as pole-zero shifting) for design of classical controllers; for simple plants, the
developed tool may be used to verify results, and for complicated plants, where
application of the pole-zero shifting approach becomes difficult, the tool may be
used to arrive at a solution which may then be analyzed in order to determine
where the poles and zeros of the controller ought to be placed for optimum
results.
B.2 Software 127

B.1.2 Systematizing the Design Approach

In order to complete the design procedure above, the user must take into account
the remarks made in the above Discussion section. This results in a trial-and-error
scheme by which the determination of the desired frequency range, determination
of the enhancing requirements, and determination of the orders of numerator
and denominator of the transfer function parameter (in case of n-th order linear
controller design) would be possible. This trial-and-error scheme becomes tedious
and time consuming, and at the end the user is not sure if all the possibilities have
been considered.
In order to alleviate this problem, the following algorithm is developed:

Algorithm B.1

1. Let the user set the lower and upper limits of desired frequency range of
interest; the default values are 0 and 10.
2. Solve design equation (e.g., Eq. B.26) for controller parameters.
3. Validate design.
4. If satisfactory results are obtained, then stop; otherwise, vary the lower
and upper limits of the desired frequency range of interest using numerical
optimization.
5. If the variation of lower and upper limits of integration is exhausted with
no solution, then go to Step 6; otherwise go to Step 2.
6. If an n-th order controller is being designed, increase the degrees of
numerator and denominator of the function parameter, and go to Step 2.
If increasing the degrees of numerator and denominator of the function
parameter is exhausted with no solution, then go to Step 7; otherwise stop.
7. Provide plant transfer function enhancing (by introducing high-order
dynamic terms to increase the order of the desired transfer function), and
go to Step 2. If enhancing does not produce the desired result, conclude
that a solution does not exist; consider relaxing the performance measures
or consider a different feedback structure.

B.2 Software

A MATLAB® function based on the above versatile controller design approach

is available (Nassirharand 2009b). This function allows design of lead-lag, PID,
proportional plus rate feedback, proportional-integral plus rate feedback, and n-th
order linear compensators. The primary inputs to this are the plant transfer function
and the desired closed-loop behavior in terms of an n-th order transfer function.
128 Appendix B Algebraic Linear Multivariable Controller Design

The output is the desired controller parameters. The unified design approach that is
implemented is based on minimizing a mean-squared error type objective function
as given by Eq. B.13. This minimization gives the controller parameters in terms
of a set of linear simultaneous algebraic equations. This solution method requires
the user to input such parameters as the lower and upper limits of frequency
range of interest as well as those discussed in Sect. B.1.1; such parameters could
be automatically determined by the use of numerical optimization as outlined in
Algorithm B.1.
A MATLAB® function is developed that automates the design of classical
controllers. This function is of the following form:

gc D cad controller .gp; hd; params/

where
gp is the transfer function of the plant that must be controlled
hd is the transfer function of the desired closed-loop system
params is an array of size 6 that controls the execution of the function
params(1): this input is between numbers 1 and 5. A 1 implies design of a lead-
lag compensator, a 2 implies design of a PID compensator, a 3 implies
design of an n-th order compensator using the factorization approach,
a 4 implies design of a proportional and rate-feedback compensator,
and a 5 implies design of a proportional-integral plus rate-feedback
compensator.
params(2): the final time to be used for simulation of the final feedback system;
if set equal to 1, then the final simulation time value is automatically
determined from MATLAB® .
params(3): the desired error between step response of hd and the designed closed-
loop system; if set equal to 1, then a value of 1 is used.
params(4): the integration step size; if set equal to 1, then a value of 0.005 is used.
params(5): the lower limit of the frequency range of interest; if set equal to 1, then
a value of 0.01 rad/s is used.
params(6): the upper limit of the frequency range of interest; if set equal to 1, then
a value of 10 rad/s is used. The maximum allowed value is arbitrarily
set to 20 rad/s.
gc is the output of the function. In case of lead-lag, PID, and nth order compen-
sators, gc is the transfer function of the compensator; in case of proportional plus
rate-feedback design, gc is an array whose first element is the gain of the rate-
feedback term, and its second element is the gain of the proportional term; in case
of proportional-integral plus rate-feedback, gc is an array whose first element is the
gain of the integral action, its second element is the gain of the proportional term,
and its third element is the gain of the rate-feedback term.
Appendix C
Additional Example Problems for Multivariable
Single-Range Controller Design

C.1 Example Problem: One Combustion Chamber Liquid

Propellant Engine

The design Algorithm 3.1 is applied to a problem of the sort encountered in

propulsion control engineering. The specific propulsion system is a liquid propellant
engine of the same class whose simplified schematic drawing is shown in Fig. 3.1. A
brief engine operation description along with sample propulsion equations of motion
is given in Chap. 3. The primary difference between the present process and that
studied in Chap. 3 is due to specific constants and input data which make this a
coupled nonlinear multivariable system.
The nonlinear and dynamic computer model of this liquid propellant engine is
given, and a controller must be designed to control the combustion chamber pres-
sure; it must also regulate the difference in exit pressure of the oxidizer shut-down
valve and that of the stabilizer in order to control the mixture ratio. In short, this plant
has two actuating signals (the regulator valve piston travel and the stabilizer valve
piston travel). It also has two outputs (the combustion chamber pressure and the
difference between the pressure at the exit of the shut-down valve (not shown) and
the exit pressure of the stabilizer valve. The shut-down valve is in the oxidizer path,
and it is placed between the oxidizer pump and the combustion chamber. Also, it is
required that speed of response of both outputs to be as fast as possible without any
overshoot. Algorithm 3.1 is executed as follows (Nassirharand and Karimi 2004a).
Step 1 – Trial and error is used, and the following multivariable plant is defined as
the reference linear model to be used during the design process.
2 3
252
6 2 0 7
Gref D6 s C 50s C 252 7 (C.1*)
4 25 2 5
0
s C 50s C 252
2

*Source: (Nassirharand and Karimi 2004a) © Sharif University of Technology

A. Nassirharand, Computer-aided Nonlinear Control System Design: Using Describing 129

Function Models, DOI 10.1007/978-1-4471-2149-7,
© Springer-Verlag London Limited 2012
130 Appendix C Additional Example Problems for Multivariable Single-Range...

Fig. C.1 Matrix magnitude plot of pseudofrequency response

The above reference linear model is chosen so the closed-loop system would be
decoupled, and the diagonal dynamics correspond to a linear second-order system
whose natural frequency and damping ratio are 25 and 1, respectively. Use of a
higher value for the natural frequency would result in nonlinear system response to
have an overshoot, and the use of a lower value for the damping ratio would result
in nonlinear system response to be very sluggish.
Step 2 – For the nominal operating regime of interest, the input/output behavior
of the process is characterized as outlined in Chap. 2. The nominal value of the
regulator control valve piston position and the nominal value of the stabilizer control
valve piston position (along with their limits) determine the DC part of excitation,
and the maximum travel values of these pistons are used to determine the amplitudes
of excitation. From the knowledge of the physical process, the frequency range
of interest is determined to be between 1 and 150 rad=s; as was mentioned, the
selected frequency pair must not be identical, and they are chosen as nearly equal
and related rationally. Initially, the phase shifts are set equal to zero; if satisfactory
results are not obtained, then effects of phase shift may also be considered. In
continuation, the plant is excited and the outputs (combustion chamber pressure and
the difference in shut-down valve exit pressure and the stabilizer control valve exit
pressure) are determined. Then, Fourier integrals are evaluated when the outputs are
at steady state. Finally, the matrix SIDF model is calculated. The resulting matrix
SIDF model and the fitted linear model of the next step are depicted in Figs. C.1
and C.2.
Step 3 – In this step, by providing the real and imaginary parts of the pseud-
ofrequency response obtained in the previous step as a function of frequency, the
®
invfreqs function of the MATLAB software is used to identify the individual
elements of the matrix linear approximation one at a time. This function allows
C.1 Example Problem: One Combustion Chamber Liquid Propellant Engine 131

Fig. C.2 Matrix phase plot of pseudofrequency response

introduction of weights to weigh the data at different frequency ranges differently.

In this example, no weights are required to obtain a reasonable fit. It should also
be noted that the numerator and denominator degrees of the transfer function
polynomials are also an input that is determined by trial and error. For example, the
degrees of the numerator and denominator polynomials of the transfer function to
be identified are set to 1 initially, and they are increased until no improvement in the
quality of fit is observed. By examining Fig. C.1, it is apparent that the contribution
of g12 (where g12 corresponds to the first row and second column of the plant matrix
SIDF model) is very small when compared to contributions of the g11 element, and
it may be set equal to zero.
The identified linear model is of the following form:
2 3
13:99
6 s C 5:617 0 7
G.s/ D 6
4 2:254 0:6848s C 3:8493
7
5 (C.2*)
 0:6853
s C 5:617 s C 5:617

Step 3 – The diagonal elements of the controller are determined by solving for
r.s/ using Eq. B.26 followed by application of Eq. B.5. Determination of c11 .s/ is
described, and determination of c22 .s/ is similar. In order to apply Eq. B.26, the co-
prime factor pairs (ŒN.s/;D.s/ and ŒP .s/; Q.s/) must be determined. Since g11 .s/
is stable, one could choose these factors as follows: N.s/ D g11 .s/; D.s/ D 1;
252
P .s/ D 0; and Q.s/ D 1. From Eq. B.1 it is concluded that hD y;u D s 2 C50sC252 .
The only remaining parameters that need to be specified in order to solve for r.s/

*Source: (Nassirharand and Karimi 2004a) © Sharif University of Technology

132 Appendix C Additional Example Problems for Multivariable Single-Range...

are !1 ; !2 in Eqs. B.30–B.33 and m; n in Eq. B.8. Some trial and error must be
employed to determine these parameters; the procedure is computer aided, and
therefore, the cycle time for determination of these parameters is not very long. For
starting point, set m D n D 0 and !1 D 0; !2 D !BW , where !BW is the bandwidth.
To improve on the quality of behavior at high frequencies, increase !2 and/or
increase !1 , and in order to improve the quality of behavior at low frequencies,
decrease !2 . If satisfactory results are not obtained, then increase m; n. Repeat
this cycle until a design is obtained. For this example, !1 D 0; !2 D 10; m D
2; and n D 2. Then Eq. B.26 is applied, and r.s/ is determined. Upon substitution
of r.s/ in Eq. B.5, the desired c11 .s/ is obtained as follows:

44:67s C 250:9
c11 .s/ D (C.3*)
s 2 C 50s

The off-diagonal elements are determined from application of Eq. B.2:

146:9s  824:9
c12 .s/ D 0; c21 .s/ D (C.4*)
s3 C 55:62s 2 C 280:9s

Hence, the synthesized controller is of the following form:

2 3
44:67s C 250:9
6 0 7
s 2 C 50s
C.s/ D 6
4 146:9s  824:9
7 (C.5*)
0:01979s C 911:3 5
s 3 C 55:62s 2 C 280:9s s 2 C 49:91s

For comparison purposes, this step is also solved by application of a H1 design

®
procedure (Safonov et al. 1989) that is also automated by MATLAB function hinf.
The considered feedback structure is shown in Fig. C.3. In this figure, R1 is the
set point, Z1 is the weighted control signal, Z2 is the weighted disturbance output
signal, Z3 is the weighted error signal, U is the control signal, Error is the error
signal, and W is a weighting function which may be subscripted for clarity. The
following remarks are in order (Nassirharand and Mousavi Firdeh 2009c):
1. In order to compensate for erosion and high levels of friction in mechanical parts,
the following control signal weighting function is selected:
2 0:2 .s C 40/ 3
0
6 s C 200 7
Wcntrl D 4
0:2 .s C 40/ 5
(C.6**)
0
s C 120

*Source: (Nassirharand and Karimi 2004a) © Sharif University of Technology

**Source: (Nassirharand and Mousavi Firdeh 2009a) © Elsevier
C.1 Example Problem: One Combustion Chamber Liquid Propellant Engine 133

Fig. C.3 Structure of the H1 feedback system. Source: (Nassirharand and Mousavi Firdeh
2009a) © Elsevier

The factor 0.2 is a scalar weight, the zero at 40 is selected so the signals
with frequencies less than the required bandwidth would not be penalized, and
the poles at 200 and 120 are used to severely penalize the control signals
corresponding to the two actuating signals with frequencies higher than 200 and
120 rad=s, respectively.
2. Considering a multiplicative uncertainty, the following disturbance weighting
function is selected to achieve robust stability:
2 0:2 .s C 1/ 3
0
6 s C 20 7
Wdist D 4
0:2 .s C 1/ 5
(C.7*)
0
s C 20

This choice reflects a 5% uncertainty for low frequencies (here taken to be 1

rad=s) and a 100% uncertainty for higher frequencies (here taken to be 20 rad=s).
3. Since the natural frequency is 25 rad=s and since it is desired to have a small
steady-state error, the following weight for Wcntrl is selected:
2 0:2 .s C 30/ 3
0
6 7
Wprfrmnc D 4 s C 0:001
0:2 .s C 30/ 5
(C.8*)
0
s C 0:001

*Source: (Nassirharand and Mousavi Firdeh 2009a) © Elsevier

134 Appendix C Additional Example Problems for Multivariable Single-Range...

Fig. C.4 Bode plots of the algebraic and H1 controllers. Source: (Nassirharand and Mousavi
Firdeh 2009a) © Elsevier

A reduced order model of the H1 controller, which is designed using the

®
mentioned MATLAB function, follows:
2 3
26:91s 2 C 5536s C 3:025e4 0:0004415s 3  0:5571s 2  125:7s  685:5
6 s 3 C 161s 2 C 5235s C 5:235 s 3 C 170s 2 C 7018s C 7:018 7
6 7
C1 .s/ D 6 7
4 0:7272s C 109:5 0:1711s C 34 5
s 2 C 38:67s C 0:03867 s C 0:001
(C.9*)

The frequency responses of the two designed controllers are compared in

Fig. C.4. The following observations are made:
1. Although the gain of the (1,1) element of C1 is slightly higher than that of the
algebraic controller, the dynamic behavior of the (1,1) elements is similar. It is
clear that the phases of (1,1) elements are almost identical.
2. The gain of the (1,2) element of C1 rapidly decreases as a function of frequency,
and in effect one could set this element to zero similar to the algebraic controller.
3. The gains of the (2,1) elements are almost identical. However, there is a large
shift in phase; since the contributions of this element are drastically less than
those of the (2,2) element, the effect of difference in the phase shifts on the
stabilizer axis is very small. For example, at 2.83 rad=s the gain of the (2,1)
element is 0.374 dB while the gain of the (2,2) element is 16.5 dB.
4. The gains and phases of the (2,2) elements up to a frequency of approximately
20 rad=s are identical; beyond this frequency, the gain values are small.

*Source: (Nassirharand and Mousavi Firdeh 2009a) © Elsevier

C.1 Example Problem: One Combustion Chamber Liquid Propellant Engine 135

Fig. C.5 Time history of the combustion chamber pressure. Source: (Nassirharand and Mousavi
Firdeh 2009a) © Elsevier

Fig. C.6 Time history of the stabilizer pressure difference. Source: (Nassirharand and Mousavi
Firdeh 2009a) © Elsevier

From the above observations, it may be concluded that the dynamic behaviors
of the two controllers are similar; however, the algebraic controller is simpler; and
hence, it is easier to implement.
Step 4 – Finally, the design is verified. The combustion chamber pressure and the
difference in input and output stabilizer pressures for the two designs in cascade
with the nonlinear plant along with desired responses are depicted in Figs. C.5 and
C.6. The corresponding time histories of the control signals are shown in Figs. C.7
and C.8. It is apparent that control system behavior is acceptable. To check the
robustness of the designed control system, the pumps’ heads were reduced by 10%.
The corresponding time responses are depicted in Figs. C.9 and C.10. Again, it is
136 Appendix C Additional Example Problems for Multivariable Single-Range...

Fig. C.7 Time history of the regulator control signal. Source: (Nassirharand and Mousavi Firdeh
2009a) © Elsevier

Fig. C.8 Time history of the stabilizer control signal. Source: (Nassirharand and Mousavi Firdeh
2009a) © Elsevier

apparent that system is responding as desired. With reference to Figs. C.5, C.6, C.7,
C.8, C.9, and C.10, the following observations are made:
1. The combustion chamber pressure response with the H1 controller includes
a higher level of undershoot than that of the algebraic controller, which may
not be desirable. Qualitatively, the remaining parts of the two responses may be
considered to be equal.
2. The stabilizer pressure difference response with the H1 controller has a faster
rise time than that of the response with the algebraic controller, and it has a
small amount of overshoot. However, the settling times may be considered to be
identical. There is no overshoot with the algebraic controller.
C.2 Example Problem: Five Combustion Chamber Liquid Propellant Engine 137

Fig. C.9 Robustness test – combustion chamber pressure. Source: (Nassirharand and Mousavi
Firdeh 2009a) © Elsevier

Fig. C.10 Robustness test – stabilizer pressure difference. Source: (Nassirharand and Mousavi
Firdeh 2009a) © Elsevier

3. Control signals of both the regulator and stabilizer axis are smooth and
acceptable.
4. The time histories of the outputs under a drop in pumps’ pressures are
competitive.

C.2 Example Problem: Five Combustion Chamber Liquid

Propellant Engine

The controller design Algorithm 3.1 is applied to the propulsion control of a five
combustion chamber liquid propellant engine (LPE). The schematic drawing of the
considered liquid propellant engine is shown in Fig. C.11. A brief engine operation
138 Appendix C Additional Example Problems for Multivariable Single-Range...

8 7

16

2
OXIDIZER FUEL
14

6 5
4

15

18 3

17

11
10 12
13

Fig. C.11 Schematic diagram of the LPE (see Table C.1 for key)

description is given next, and the corresponding mathematical model is similar to

that presented in Chap. 3.
This engine is composed of the following elements (not all are shown in
Fig. C.11): (1) combustion chamber, (2) gas generator, (3) starter, (4) turbine, (5)
fuel pump, (6) oxidizer pump, (7) fuel diaphragm valve, (8) stabilizer control valve,
(9) oxidizer shut-down valve of the gas generator, (10) oxidizer diaphragm valve,
(11) oxidizer shut-down valve of the combustion chamber, (12) regulator control
valve, (13) fuel discharge valve, (14) fuel shut-down valve of the gas generator, (15)
exhaust, (16)–(19) thrusters (or verniers), (20) fuel shut-down valve of thrusters,
C.2 Example Problem: Five Combustion Chamber Liquid Propellant Engine 139

Table C.1 Key for Fig. C.11 No. Description

1 Combustion chamber
2 Gas generator
3 Starter
4 Turbine
5 Fuel pump
6 Oxidizer pump
7 Regulator valve
8 Gas generator stabilizer valve
9 Combustion chamber stabilizer valve
10–13 Thruster
14 Gas generator combustion products
15 Starter combustion products
16, 17 Control signal

(21) oxidizer shut-down valve of the thrusters, (22) staged start valve, (23) valve for
simultaneous discharge of the tanks, (24)–(35) orifices, (36)–(49) pipes, and (50)
nonreturn valve.
The LPE operates in the following manner. First, the pressures inside the fuel and
oxidizer tanks reach to specified values. Then, electric commands to starter, fuel, and
oxidizer diaphragm valves (not shown) are issued to start the turbine and allow fuel
and oxidizer to enter the system. As the exit pressurized gas of the starter turns the
turbine, the shaft of the turbopump starts rotating, and the pumps are commissioned.
As a result, fuel and oxidizer would be pumped.
In the fuel path, after liquid modulation, the fuel path is divided into three
branches. In the first branch, some fuel enters the stabilizer of the combustion
chamber. In the second branch, some fuel is guided toward the regulator, and
finally, in the third branch, some fuel heads toward the thrusters. In the first branch,
the combustion chamber stabilizer guides a controlled amount of fuel to the fuel
discharge valve (not shown). Under the condition that fuel discharge valve is not
activated, the fuel that enters this valve travels toward the combustion chamber after
going through appropriate connections. In the second branch, the amount of fuel that
is guided toward the regulator valve passes through appropriate connections, and it
enters the regulator valve. The regulator control valve passes a controlled amount of
fuel to the gas generator. Finally, in the third branch, some fuel enters the thrusters.
In the oxidizer path, the oxidizer exiting the oxidizer pump is divided into two
branches. In one branch, the oxidizer enters the combustion chamber. This amount
of oxidizer and the amount of fuel that enters the combustion chamber from the fuel
path are combined, and due to the resulting combustion, the engine is turned on. The
second branch itself is divided into two other subbranches or parts.
In the first subbranch, some oxidizer enters the gas-generator stabilizer control
valve; this stabilizer sends a controlled amount of oxidizer into the gas generator.
This amount of oxidizer and the amount of fuel that enters the gas generator from
the fuel path are combined, and combustion is resulted. The exit pressurized gas
140 Appendix C Additional Example Problems for Multivariable Single-Range...

Fig. C.12 Gain plots of SIDF models

of the gas generator keeps the turbine rotating, and the starter gradually leaves the
circuit. In the second subbranch, some oxidizer enters the thrusters. This amount of
oxidizer that enters the thrusters is combined with the amount of fuel that comes
from the fuel path, and combustion takes place.
The nonlinear and dynamic computer model (Karimi and Nassirharand 2006;
Nassirharand and Karimi 2004b) of the liquid propellant engine is utilized to design
a controller for the system. It is desired that the final nonlinear closed-system input–
output behavior would mimic the following reference linear model (Abbassi 2003):
2 9:35s C 28:06 3
0
6 7
D 4 s C 12:15s C 28:06
2
56 5
Gref (C.10)
0
s C 56

The controller synthesis procedure steps are executed as described below:

Step 1 – The desired reference linear model is given by Eq. C.10.
Step 2 – The SIDF models of the engine are obtained as outlined in Chap. 2
by simultaneously exciting both input channels with sinusoids of different but
nearly equal frequencies, obtaining the output data, and finally postprocessing the
output data using the describing function generator software. The corresponding
pseudofrequency response plots are depicted in Figs. C.12 and C.13. From the
examination of these plots, it is observed that the spread in the SIDF models is not
significant. Therefore, it may be concluded that a single-range linear controller
might suffice.
C.2 Example Problem: Five Combustion Chamber Liquid Propellant Engine 141

Fig. C.13 Phase plots of SIDF models

Step 3 – In continuation, MATLAB® is used, and the following linear model is

identified:
2 3
28:13s C 35850 4:718s 2  1671s C 6923
6 s 2 C 1283s C 3534 s 2 C 1283s C 3534 7
Gf .s/ D 6
4
7
5 (C.11)
0:0 0:275

Step 4 – In this step, two different feedback structures are considered for controller
design (Abbassi 2003). The first structure is similar to that shown in Fig. B.1
with g21 D 0; the corresponding feedback structure would be given by Fig. C.14.
Substitute appropriate terms into Eq. B.3; then compare the main return ratios for
loops 1 and 2 with the main diagonal terms to obtain c11 .s/ and c22 .s/. Finally,
use Eq. B.2, and solve for c12 .s/ and c21 .s/. The designed controller (with the
feedback structure given by Fig. C.14) is given below:
2 3
s C 3 8:386s 2 C 2970:14s  12305:368
6  7
6 3s s 2 C 1274:44s 7
6
C.s/ D 6 7 (C.12)
7
4 200 5
0
s

With the second structure (see Fig. C.15), the controller matrix would not
be presentable in a standard fashion as in the case with the first structure. With
142 Appendix C Additional Example Problems for Multivariable Single-Range...

Fig. C.14 Structure of the r1 e1

first considered feedback
system

u1
c11(s) g11(s)

y1

c12(s)
g12(s)

r2 y2
c22(s) g22(s)
e2 u2

reference to Fig. C.15, c11 and c22 would be as designed in the first structure. By
inspecting the block diagram in Fig. C.15, the following equation is obtained:

y1 D g11 c11 e1 C g11 M r2 C g12 u2 (C.13)

To achieve decoupling, the following equation must hold:

g11 M r2 C g12 u2 D 0 (C.14)

Note that
u2 22 1 203
D gref g22 D (C.15)
r2 s C 56
22
where gref is the (2,2) element of matrix Gref given by Eq. C.10. Compare
Eqs. C.15 and C.14, and substitute for known quantities to obtain M as follows:

33:63s 3 C 11820s 2  15970s  135900

M D (C.16)
s 3 C 1339s 2 C 75233s C 189560
C.2 Example Problem: Five Combustion Chamber Liquid Propellant Engine 143

Fig. C.15 Structure of the r1

second considered feedback
system

u1
c11(s) g11(s)

y1

M
g12(s)

y2
c22(s) g22(s)
r2 e2 u2

Another controller is designed by minimizing an exact model-matching

objective function as follows (Nassirharand and Karimi 2004b):
2 3
3:6677s 2 C 12:2743s C 52:2724
0:4309s 2 C 0:734s C 0:062 
6
6 0:9248s 3 C 0:2823s 2 C 2:1193s C 0:3302 7
7
C.s/ D 6 s2 7
4 0:0538s 2 C 0:070044s C 0:1853 75:430s C 0:4989 5
s2 s 2 C 0:0133
(C.17)

Step 5 – In this step, design is verified by nonlinear simulation. The results of

design verification are depicted in Figs. C.16, C.17, C.18, and C.19. In these
simulations, step disturbance signals of four units at the inputs of control valves
are also included. The depicted simulation results correspond to the two designs
developed above and one prior work (Nassirharand and Karimi 2004b). From the
examination of the response plots, it is fair to conclude that the second feedback
structure would be preferred to the other two designs. Furthermore, unlike the
work by Nassirharand and Karimi 2004b, the new designs have disturbance
rejection properties. The control signals corresponding to the present designs
have a more acceptable behavior when compared to the prior work.
144 Appendix C Additional Example Problems for Multivariable Single-Range...

Fig. C.16 Design verification for the combustion chamber pressure (Abbassi 2003)

Fig. C.17 Design verification for mixture ratio (Abbassi 2003)

Fig. C.18 Behavior of control signal no. 1 (Abbassi 2003)

C.2 Example Problem: Five Combustion Chamber Liquid Propellant Engine 145

Fig. C.19 Behavior of control signal no. 2 (Abbassi 2003)

Fig. C.20 Magnitude plot for stability test (structure 1)

As is demonstrated in Figs. C.20 and C.21, the SIDF models of the final closed-
loop system at various operating regimes of interest are successfully obtained;
hence, one may conclude that closed-loop system is stable when operating at the
desired operating regimes.
146 Appendix C Additional Example Problems for Multivariable Single-Range...

Fig. C.21 Phase plot for stability test (structure 1)

C.3 Example Problem: An Unstable Multivariable

Robot Arm

It is emphasized that the basic SRLCD procedure that is outlined in Sect. 3.1 is
limited to stable plants. For unstable plants, one approach is to stabilize the plant,
and then apply the outlined procedure to the stabilized plant. This approach is
demonstrated by the following example problem.
A two degree-of-freedom robot arm is considered (see Fig. C.22; Teh and
Nassirharand 2010b). Let Li , Mi , and qi be the length, the lumped mass, and angle
of link i ; for this example problem: M1 D 3Kg, M2 D 1Kg, L1 D 1m, and
L2 D 1m (Utkin et al. 1999). The equations of motion of the robot as given by
Utkin et al. 1999 follow:
      
1 m11 m12 qR 1 n
D C 1 (C.18)
2 m21 m22 qR2 n2

where

m22 D L22 M2 (C.19)

m12 D m21 D m22 C L1 L2 M2 cos .q2 / (C.20)

m11 D L21 .M1 C M2 / C 2m12  m22 (C.21)

C.3 Example Problem: An Unstable Multivariable Robot Arm 147

Fig. C.22 A two degree-of-freedom robot arm (Teh 2010)

n2 D L1 L2 M2 qP 12 sin .q2 / (C.22)

 
n1 D L1 L2 M2 2qP 1 qP 2  qP 22 sin .q2 / (C.23)

First, velocity feedback is used to stabilize the system. The gains of the
velocity feedback for the first and second links are arbitrarily set to be 50 and 40,
respectively. In continuation, Algorithm 3.1 is executed for the stabilized system as
follows:
Step 1 – The desired reference linear model is a diagonal matrix transfer function
whose (1,1) element is a second-order transfer function with a damping ratio of
0.5 and a natural frequency of 100 rad=s, and its (2,2) element is another second-
order transfer function with a damping ratio of 0.5 and a natural frequency of 150
rad=s.
Step 2 – The stabilized system is simulated in the Simulink environment, and the
outputs as functions of time are obtained. Then, the MATLAB® software outlined
in Chap. 2 is utilized, and the describing function models are generated. These
models along with linear system identification results of the next step are shown
in Figs. C.23 and C.24 (these figures originally appeared in Nassirharand 2008c).
Step 3 – As was mentioned, the invfreqs function of MATLAB® is used, and the
following linear model is identified (Govindan 2011):
2 10:78 37:99 3
6
GD4 s .s C 7:84/ s 2 C 129:8s C 667:9 7
0:003457s C 0:0994 5
(C.24)
38:08
s 2 C 129:7s C 665:6 s
148 Appendix C Additional Example Problems for Multivariable Single-Range...

Fig. C.23 The multivariable gain matrix of the robot. Source: (Nassirharand 2008c) © Emeralds

Step 4 – The design equations developed in Appendix B are used to design the
controller. For the return ratio corresponding to both loops, the following design
parameters are used: m D n D 6I !1 D 0I !2 D 10: The designed controller is
of the following form (Govindan 2011):
2 2:56 3
c12 .s/
6 s 2 C 3:2s C 2:56 7
C D4 5 (C.25)
4
c21 .s/
s 2 C 4s C 4

where

992s 3 C 1:219e5s 2 C 8:988e5s

c12 .s/ D (C.26)
s 3 C 289:7s 2 C 3:362e4s C 1:723e5

8:526e4s 2 C 5:132e5s
c21 .s/ D (C.27)
s 3 C 664:5s 2 C 4:318e4s C 2:213e5
C.3 Example Problem: An Unstable Multivariable Robot Arm 149

Fig. C.24 The multivariable phase matrix of the robot. Source: (Nassirharand 2008c) © Emeralds

Step 5 – Finally, the design is verified via a nonlinear simulation. The results are
compared with an alternate design (Nassirharand 2008c) in Figs. C.25, C.26,
C.27, C.28, C.29, and C.30. The following remarks are made:
1. There is less coupling (see Figs. C.26 and C.29) with the new design at the
expense of increase in settling time (see Figs. C.25 and C.28).
2. The robustness of both designs is competitive (see Figs. C.25, C.26, C.27, C.28,
C.29, and C.30).
3. The control signals of both designs are acceptable. The new design has less
control effort (see Figs. C.27 and C.30).
150 Appendix C Additional Example Problems for Multivariable Single-Range...

Fig. C.25 Tracking quality of the first axis

Fig. C.26 Decoupling from the second axis

C.3 Example Problem: An Unstable Multivariable Robot Arm 151

Fig. C.27 Quality of the control signal for the first axis

Fig. C.28 Tracking quality of the second axis

152 Appendix C Additional Example Problems for Multivariable Single-Range...

Fig. C.29 Decoupling from the first axis

Fig. C.30 Quality of the control signal for the second axis
Appendix D
Dual-Range Controllers and Simultaneous
Stabilization Theory

D.1 Bézout’s Identity

Application of the factorization theory of linear systems requires one to solve for
the coprime factors satisfying the Bezout’s identity, i.e.:

PN C QD D I (D.1)

N P C DQ D I (D.2)

where
D D a right coprime factor of the plant G and a transfer function matrix
D D a left coprime factor of the plant G and a transfer function matrix
1
G D the plant transfer function matrix, G D N D 1 D D N
I D the identity matrix
N D a right coprime factor of the plant G and a transfer function matrix
N D a left coprime factor of the plant G and a transfer function matrix
P D a right coprime factor and a transfer function matrix
P D a left coprime factor and a transfer function matrix
Q D a right coprime factor and a transfer function matrix
Q D a left coprime factor and a transfer function matrix

A. Nassirharand, Computer-aided Nonlinear Control System Design: Using Describing 153

Function Models, DOI 10.1007/978-1-4471-2149-7,
© Springer-Verlag London Limited 2012
154 Appendix D Dual-Range Controllers and Simultaneous Stabilization Theory

D.1.1 Multivariable Case

For multivariable systems, consider the following state-space representation:

P
x.t/ D Ax.t/ C Bu.t/ (D.3)

y.t/ D C x.t/ C E (D.4)

where
A D system matrix, 2 Rnn
B D input matrix, 2 Rnp
C D output matrix, 2 Rqn
E D transmission matrix, 2 Rqp
t D time variable
u D vector of inputs of size p
x D vector of state-variables of size n
y D vector of outputs of size q
The coprime factors may be obtained from the following relations (Nett et al.
1984):

N D .C  EK/ .sI  A0 /1 B C E (D.5)

D D I  K.sI  A0 /1 B (D.6)

 1
P D K sI  A00 F (D.7)

 1
Q D I C K sI  A00 .B  FE/ (D.8)

 1
N D C sI  A00 .B  FE/ C E (D.9)

 1
D D I  C sI  A00 F (D.10)

P D K.sI  A0 /1 F (D.11)

Q D I C .C  EK/ .sI  A0 /1 F (D.12)

where

A0 D A  BK (D.13)
D.1 Bézout’s Identity 155

A00 D A  F C (D.14)

Matrices F and K are determined such that eigenvalues of A0 and A00 have
negative real parts; if such F and K exist, then the pairs .A; B/ and .A; C / are said
to be stabilizable (i.e., all unstable poles are state controllable) and detectable (i.e.,
all unstable poles are observable), respectively (Skogestad and Postlethwaite 2005).

D.1.2 Single-Variable Case

For single-input single-output systems, it has been shown that the application of the
above relations to obtain the coprime factors results in high-order coprime factors
(Nassirharand and Patwardhan 1988; Patwardhan 1988; Nassirharand 1993, 2003).
Given the plant transfer function, G, a lucid factorization of G is given by (see
Vidyasagar 1985):
8 n0 9
< G D N D 1 ; G 2 R.s/; N 2 H; D 2 H; N D ; >
ˆ =
m
(D.15*)
:̂ n0 2 ; m 2 h ; D D d0 ; d0 2 >
;
m
where
d0 D the denominator polynomial of G
H D a ring; this ring is the set of all rational functions whose poles lie on the left
half of the s-plane
m D Hurwitz polynomial whose poles are in a user-defined algebraic ring H
n0 D the numerator polynomial of G
R.s/ D field of rational functions in s
D the set of all polynomials in s
h D the set of all Hurwitz polynomials in

Theorem 1 (Nassirharand 2003). Lucid factors of Bezout’s identity are acquired

by a lucid factorization of the plant, a user-defined algebraic ring, and a set of linear
simultaneous algebraic equations whose coefficients form a Sylvester matrix.
Proof. Without loss of generality, bring in a lucid factorization of plant G.s/ with
degree of m.s/, ım , equal to max fın0 ; ıd 0 g; i.e.:
   
N.s/ n0 .s/ d0 .s/ 1
G.s/ D D : (D.16*)
D.s/ m.s/ m.s/

*Source: (Nassirharand 2003) © Elsevier

156 Appendix D Dual-Range Controllers and Simultaneous Stabilization Theory

Now let

p0 .s/
P .s/ D (D.17*)
p1 .s/
q0 .s/
Q.s/ D (D.18*)
q1 .s/

where p0 .s/ 2 , p1 .s/ 2 h , q0 .s/ 2 , q1 .s/ 2 h , n0 .s/ 2 ,

d0 .s/ 2 , and m.s/ 2 h . By substituting Eqs. D.15–D.18 into Bezout’s identity
fP .s/N.s/ C Q.s/D.s/ D 1g, we obtain the following equation:

p0 .s/n0 .s/q1 .s/ C q0 .s/d0 .s/p1 .s/ D m.s/q1 .s/p1 .s/ (D.19*)

let p1 .s/ D q1 .s/ D d.s/ with d.s/ 2 h. Then, the Bezout’s identity becomes:

p0 .s/n0 .s/ C q0 .s/d0 .s/ D m.s/d.s/ (D.20*)

Note that coprime factors P and Q must at most be proper. For this reason we let
ıp0 D ıq0 D ıd , where the symbol ı is used to denote the degree of a polynomial,
and it may be subscripted for clarity. The following relation gives the number of
unknown coefficients on the left-hand side of Eq. D.20, denoted Tl :

Tl D 2ıd C 2 (D.21*)

And the number of unknown coefficients on the right-hand side of Eq. D.20,
denoted Tr , is given by:

Tr D ım C ıd C 1 (D.22*)

In order for equality (D.20) to hold, we must have:

T l D Tr ) ı d D ı m  1  l (D.23*)

Let

n0 .s/ D n00 C n10 s C n20 s 2 C    (D.24*)

d0 .s/ D d00 C d01 s C d02 s 2 C    (D.25*)

p0 .s/ D p00 C p01 s C p02 s 2 C    (D.26*)

*Source: (Nassirharand 2003) © Elsevier

D.1 Bézout’s Identity 157

q0 .s/ D q00 C q01 s C q02 s 2 C    (D.27*)

m.s/ D m0 C m1 s C m2 s 2 C    (D.28*)

d.s/ D d 0 C d 1 s C d 2 s 2 C    (D.29*)

Select m.s/ 2 h and d.s/ 2 h such that coprime pairs ŒP; Q 2 H and
ŒN; D 2 H ; then, substitute the above polynomials of known degrees into Bezout’s
identity (or Eq. D.20); expand identity terms and equate terms of equal power in
complex frequency. The following equations would be obtained:

n00 p00 C d00 q00 D m0 d 0 (D.30*)

n10 p00 C n00 p01 C d01 q00 C d00 q01 D m0 d 1 C m1 d 0 (D.31*)

nl0 p00 C nl1

0 p0 C    n0 p0 C d0 q0 C d0 q0 C    C d0 q0
1 0 l l 0 l1 1 0 l

D m0 d l C m1 d l1 C    C ml d 0 (D.32*)

n0lC1 p00 C nl0 p01 C    C n10 p0l C d0lC1 q00 C d0l q01 C    C d01 q0l
D m0 d lC1 C m1 d l C    C mlC1 d 0 (D.33*)

n02lC1 p00 C n2l

0 p0 C    C n0 p0 C d0
1 lC1 l
q0 C d02l q01 C    2 Cd0lC1 q0l
2lC1 0

D m0 d 2lC1 C m1 d 2l C    C m2lC1 d 0 (D.34*)

˚
˚
By noting that nk0 D 0; k 2 i WD .l C 2; l C 3; :::; 2l C 1/ , d0k D 0; k 2 i W D
˚ k

.l C 2; l C 3; :::; 2l C 1/g, m D 0; k 2
i WD .l C 2; l C 3; :::; 2l C 1/ , and
˚ k
d D 0; k 2 i WD .l C 1; l C 2; :::; 2l C 1/ , the above set of equations in matrix
form follows.

SX D B 0 (D.35*)

where
 
S11 S12
SD (D.36*)
S21 S22

*Source: (Nassirharand 2003) © Elsevier

158 Appendix D Dual-Range Controllers and Simultaneous Stabilization Theory

2 3 2 3
n00 0  0 0 d00 0   0
6 1 : :: :: 7 6 1 : 7
6n
6 0 n00 : : : : 7
7
6d
6 0 d00 : : 0 0 7
7
6 7 6 7
S11 D 6 ::: :
n10 : : 0 07 S12 D 6 ::: :: ::
: : 0 0 7
6 7 6 7
6 : :: : : 7 6 : :: :: 7
4 :: : : n0 0 5
0 4 :: : : d 0 5
0
0

nl0 nl1
0    n10 n00 d0l d0   
l1
d01 d00

2 3 2 3
n0lC1 nl0       n10 d0lC1 d0l     d01
6 7 6 7
6 0 n0lC1 nl0 : : : n20 7 6 0 d0lC1 : : : ::
: d02 7
6 7 6 7
6 7 6 7
S21 D 6 ::: :
0 :: ::
: ::
: 7 S22 D 6 ::: :
0 :: d0l
::
: 7
6 7 6 7
6 : : : : : lC1 7 6 : :: :: 7
4 :: : : n0 nl0 5 4 :: : : d0lC1 d0l 5
0 0  0 n0lC1 0 0  0 d0lC1

2 3
m0 d 0
6 m0 d 1 C m 1 d 0 7
6 7
6 : 7
6 :: 7
6 7
6 :: 7
6 7
6 : 7
6 0 l l 0 7
6 m d C m 1 l1
d C ::: C m d 7
B0 D 6 7 (D.37*)
6 m1 d l C m2 d l1 C ::: C mlC1 d 0 7
6 7
6 :: 7
6 : 7
6 7
6 :: 7
6 : 7
6 7
6 :: 7
4 : 5
mlC1 d l

In the above equation, S is known, B 0 will be chosen by the user, and X is

unknown and will be determined from Eq. D.35. Notice that we started with a
lucid factorization of the plant that was followed by selection of the coprime pairs
ŒN; D and ŒP; Q in a user-defined algebraic ring, and finally, it was shown that
remaining unknowns may be solved via a linear matrix equation of the form given
by Eq. D.35 whose coefficient matrix, S , is a Sylvester matrix. Notice that Sylvester
matrix is always nonsingular provided that n0 .s/ 2 and d0 .s/ 2 do not have
any common factors.

*Source: (Nassirharand 2003) © Elsevier

D.2 DRLCD Using Simultaneous Stabilization Theory 159

Remark 1. In order to obtain the lower-order coprime factors that are solutions to the
Bezout’s identity, we must select m.s/ to enforce either N.s/ 2 H or D.s/ 2 H to
have left-half plane pole/zero cancellations.

D.1.3 Software
®
A MATLAB function for computation of coprime factors given by Eqs. D.5–D.12
is developed in Teh (2010). With reference to the work presented in Nassirharand
®
and Patwardhan (1988), a much simpler software in the MATLAB environment for
single-variable systems is developed in Nassirharand (2003). The m-file listings are
given in Fig. D.1.
®
The corresponding MATLAB function is of the following form:

ŒN; D; P; Q D Lucid f actor.G; mpoles; dpoles/

where N; D; P; Q; and G are defined as described in Sect. 3; mpoles is an array

of size ım D max.ın0 ; ıd0 / containing the desired poles of plant lucid factors, and
dpoles is an array of size ıd D ım  1 containing the desired poles of the coprime
pair ŒP; Q 2 H .
The software first obtains the lucid factorization of the plant followed by
determination of the Sylvester coefficient matrix, S ; then, the input matrix B 0 is
formed, and finally the unknown vector X is solved for. The last step of the software
is to enforce any left-half plane pole-zero cancellations that may exist.

D.2 DRLCD Using Simultaneous Stabilization Theory

D.2.1 Theory and Algorithm

The original idea for design of dual-range controllers for nonlinear systems (Taylor
1983) calls for the application of simultaneous stabilization theory of Vidyasagar
and Viswanadham (1982). In this section, all controllers that simultaneously
stabilize the closed-loop systems corresponding to the two linear system models
of the nonlinear plant are parameterized. This involves developing a mathematical
expression for all stabilizing controllers in terms of the coprime factors of linear
system models and another as yet unknown stable nth-order transfer function
denoted r.s/.
The term parameterization arises from the fact that the class of all stabilizing
controllers is expressed in terms of the parameter r.s/. Such a parameterization
provides a logical means for obtaining the desired minimum-sensitivity controller.
The following is single-input/single-output version of the simultaneous stabilization
theory.
160 Appendix D Dual-Range Controllers and Simultaneous Stabilization Theory

Fig. D.1 Listing of the m-files for lucid factorization and solution to the Bezout’s identity. Source:
(Nassirharand 2003) © Elsevier
D.2 DRLCD Using Simultaneous Stabilization Theory 161

Fig. D.1 (continued)

162 Appendix D Dual-Range Controllers and Simultaneous Stabilization Theory

Fig. D.1 (continued)

D.2 DRLCD Using Simultaneous Stabilization Theory 163

Theorem 2. The class of all controllers that stabilize a plant G.s/ may be parame-
terized in terms of a stable linear transfer function, r.s/, as follows:

P .s/ C r.s/D.s/
C.s/ D (D.38)
Q.s/  r.s/N.s/

where
C.s/ D the stabilizing controller
r.s/ D any stable transfer function parameter to be selected by the designer; may
be selected to achieve certain performance criteria,
and the rest of the symbols are defined as in Eq. D.1. See Youla et al. (1976) for
proof.
The parameterization of all controllers that stabilize two stable linear system
models is given next. Let the pair ŒNi .s/; Di .s/ correspond to the stable coprime
factors of Gi .s/, i D 0; 1 and designate the corresponding factors that satisfy the
Bezout’s identify by the stable coprime pair ŒPi .s/; Qi .s/.
Theorem 3. The set of all controllers that simultaneously stabilize G0 .s/ and G1 .s/
is given by:

P0 .s/ C r .s/D0 .s/

Css .s/ D ; (D.39)
Q0 .s/  r.s/N0 .s/

where r.s/ belongs to the set of all stable controllers that stabilize Ga .s/

N0 .s/D1 .s/ C D0 .s/N1 .s/

Ga .s/ D (D.40)
Q0 .s/D1 .s/ C P0 .s/N1 .s/

The set of all controllers that stabilize Ga .s/ is given by Theorem 2; the condition
that r.s/ must itself be stable (it is said that the plant G a must be strongly
stabilizable) must be imposed in addition. See Vidyasagar and Viswanadham (1982)
for proofs.
The following algorithm (Patwardhan 1988; Nassirharand 1991; Ujile and Nas-
sirharand 2010), which is based on the simultaneous stabilization theory presented
above, may be applied to arrive at a dual-range controller.

Algorithm D.2
1. Define the desired performance measures for the final closed-loop system.
For example, one could specify the desired percentage overshoot, rise time,
settling time, undershoot, gain margin, and/or phase margin.
2. Identify a reference linear model whose characteristics are similar to those
of the performance measured defined in the previous step.
164 Appendix D Dual-Range Controllers and Simultaneous Stabilization Theory

3. Set the denominator of the design parameter equal to that of the reference
linear model identified in Step 2. This choice for the design parameter
denominator is based on dominant pole placement ideas.
4. Use the Lucid factor MATLAB® function (see Fig.D.1) and compute the
coprime factors of the plant transfer functions as well as the coprime
factors that are solutions to the Bezout’s identity.
5. Compute the auxiliary plant, Ga .s/, given by Eq. D.40. Then, classify the
set of all controllers that stabilize the auxiliary plant in terms of a function
parameter, Ra .s/, as below (see Theorem 2):

Pa .s/ C Ra .s/Da .s/

Ca .s/ D (D.41)
Qa .s/  Ra .s/Na .s/
where
Da D a coprime factor of the plant Ga and a transfer function
N D a coprime factor of the plant Ga and a transfer function
Pa D a right coprime factor that is a solution to the Bezout’s identity and
a transfer function
Qa D a right coprime factor that is a solution to the Bezout’s identity and
a transfer function
The simultaneously stabilizing controller, Css .s/, is given by Eq. D.39.
InEq. D.39, r is given by Ca .s/ for which Ga .s/ is strongly stabilizable.
Substituting Ca .s/ from Eq. D.41 into Eq. D.39 as r solves the simultane-
ous stabilization problem.
6. Search for the optimum coefficients of the numerator polynomial of Ra .s/
using dominant pole placement ideas coupled with optimization to yield
the desired simultaneously stabilizing controller that forces the closed-
loop system to behave in a desired way. To achieve this, the following
objective function is minimized using a simplex optimization routine as
implemented by fminsearch function of MATLAB (see Nelder and Mead
1965 for a Simplex Algorithm):

F D ˇF0 C .1  ˇ/ F1 (D.42)

where
h˚
2 i X 2
j j
Fj .x/ D ˛ j
ymax  ymax
d
C < ymin >2 C yi  yid ; j D 0; 1
i
(D.43)
8 
< j 2 j
j ymin if ymin < 0
< ymin > D
2
(D.44)
: 0 if y j > 0
min
D.2 DRLCD Using Simultaneous Stabilization Theory 165

d D a superscript denoting the desired value(s)

y D the unit step response; may be subscripted and/or superscripted for
clarity
˛ D a weighting factor
ˇ D a weighting factor
The terms in brackets of Eq. D.43 represent the errors in the over-
shoot and undershoot specifications, respectively. The remaining term in
Eq. D.43 represents the squared error between the actual and desired
closed-loop response. The objective function has two weighting factors, ˛
and ˇ. Weighting factor ˛ is a trade-off between the overshoot and settling
time conditions, while ˇ is a trade-off between the closed-loop responses
of the two different plants.
The vector of independent variables is comprised of the coefficients of
the numerator polynomial of Ra .s/. Initial values for these coefficients
must be assumed.
9. Validate design using a digital simulation.

D.2.2 Software

There exist two pieces of software based on Algorithm D.2. One is in the FORmula
TRANslation language whose listing is given in Patwardhan (1988). The other is
in terms of MATLAB® functions developed in Ujile (2010). The flow chart of
the MATLAB® version is given in Figs. D.2 and D.3 (Ujile and Nassirharand
2010).

D.2.3 Example Problems

D.2.3.1 A Linear System with Integrity

Consider a linear system with integrity whose dynamic behavior at two different
conditions is given by G0 .s/ and G1 .s/ as follows:

1
G0 .s/ D (D.45)
4 .s C 2/
3:0
G1 .s/ D (D.46)
4 .s  2/
166 Appendix D Dual-Range Controllers and Simultaneous Stabilization Theory

Fig. D.2 Flowchart for the

main function of DRLCD.
Source: (Ujile and
Nassirharand 2010) ©
Hypersciences

This example problem was originally studied in Nassirharand (1986), and later
in Patwardhan (1988), Nassirharand (1991), Ujile and Nassirharand (2010), Teh
(2010), and Nassirharand (2010b). It is desired to design a dual-range controller
so the closed-loop system comprised of the designed controller and the two linear
models above would mimic a second-order linear system with a percent overshoot
of 37 and a settling time of 0.3 s, i.e.:

1867:9
y;u D
hD (D.47)
s 2 C 26:5s C 1867:9

The plant G0 .s/ is stable. Hence, we choose N0 .s/ D G0 .s/ and D0 .s/ D 1: The
auxiliary plant Ga .s/ is given by the following equation:

sC4
Ga .s/ D (D.48)
2 .s 2  4/
D.2 DRLCD Using Simultaneous Stabilization Theory 167

Fig. D.3 Flowchart for

computation of the objective
function for DRLCD. Source:
(Ujile and Nassirharand
2010) © Hypersciences

Since Ga .s/ has the parity interlacing property (PIP), a compensator that strongly
stabilizes Ga .s/ exists (a plant has PIP if the number of poles of the plant between
every pair of real zeros on the right half plane is even). To find the auxiliary
compensator Ca .s/, first the coprime factors of Ga .s/ are formed as given below:

sC4
Na .s/ D (D.49)
s 2 C 4s C 4
2 .s  2/
Da .s/ D (D.50)
.s C 2/

For the above coprime factors, the solution to the Bezout’s identity is obtained
as follows:
Pa .s/ D 2:6667 (D.51)

s C 3:33
Qa .s/ D (D.52)
2 .S C 2/
168 Appendix D Dual-Range Controllers and Simultaneous Stabilization Theory

Hence, the strongly stabilizing controller is given by:

   
rad 21:3s 3 C 127:9s 2 C 255:9s C 170:6 C ran 16:0s 3 C 32:0s 2  64:0s  128:0
Ca .s/D    
rad 4:0s 3 C 23:9s 2 C 47:9s C 31:9  ran 12:0s 2 C 48:0s C 48:0
(D.53)

where
rad D the denominator polynomial of the transfer function parameter Ra .s/ in
Eq. D.41
ran D the numerator polynomial of the transfer function parameter Ra .s/ in Eq. D.41
To maximize the number of the independent design variables for optimization,
we set the degree of the numerator polynomial of the parameter Ra .s/ equal to that
of its denominator. Hence,

ran .s/ D a0 s 2 C b 0 s C c 0 (D.54)

The design variable c0 in Eq. D.54 is selected to have a zero steady-state error due
to a step input, i.e., c 0 D 1243:36. Optimization is used to select the design variables
a0 and b0 in the following fashion. The value of ˇ D 0:5 is chosen to give the same
weight to the responses due to both plants. Keeping ˇ fixed at 0.5 with initial values
a0 D b 0 D 1:0, and ˛ D 1000, the two software mentioned in Sect. D.2.2 produce
the following compensators:

39:4s 3 C 2518s 2 C 1:13e4s C 1:055e4

Css1 .s/ D (D.55)
s 3 C 3:297s 2 C 102:6s

15:06s 2 C 2723s C 3834

Css2 .s/ D (D.56*)
s 2 C 19:3s

where
Css1 D the controller obtained using the FORTRAN software after simplification and
model reduction
Css2 D the controller obtained using the MATLAB® software after simplification
and model reduction
Unlike the FORTRAN solution, in MATLAB® solution the value c 0 is also
determined by optimization; its initial value is set to 1; also, poles of the coprime
factors are set to 2. For comparison purposes, the software based on direct
optimization approach is also applied; the following controller is obtained:

*Source: (Ujile and Nassirharand 2010) © Hypersciences

D.2 DRLCD Using Simultaneous Stabilization Theory 169

Fig. D.4 Comparison of frequency response of the three simultaneous stabilizing controllers

Fig. D.5 Step-response comparison of the performance of different simultaneously stabilizing

controllers with plant G0 (linear system with integrity)

1:024s 3 C 54:98s 2 C 7395s C 1:416e4

Css3 .s/ D (D.57)
s 2 C 24:26s

where
Css3 D the controller obtained using the MATLAB® software using direct optimiza-
tion (Teh 2010) after simplification
The frequency responses of all three controllers may be compared by examining
Fig. D.4. The unit step responses of the closed-loop systems comprised of both
plants (see Eqs. D.45 and D.46) with all three controllers along with that of the
desired are depicted in Figs. D.5 and D.6.
170 Appendix D Dual-Range Controllers and Simultaneous Stabilization Theory

Fig. D.6 Step-response comparison of the performance of different simultaneously stabilizing

controllers with plant G1 (linear system with integrity)

D.2.3.2 A Nonlinear Servo System

The application of the simultaneous stabilization theory to a nonlinear plant is

investigated in this section. Consider the nonlinear plant that is also studied in Sects.
2.2.1.1, 4.4, 5.4.1, 5.4.2, and 5.4.3. The describing function models of this nonlinear
plant are depicted in Figs. 2.5, 2.6, 2.7, and 2.8. The two describing function models
whose gain characteristics bound those of all the others in class are identified to be
as follows (Nassirharand et al. 1988; Nassirharand 1988):

1:231 C 0:003989s
G0 .s/ D (D.58)
s .0:05475 C 0:003761s/
1:000 C 0:04031s
G1 .s/ D (D.59)
s .0:22521 C 0:23641s/

The desired reference linear model is given as follows:

1953
y;u D
hD (D.60)
s2 C 26:67s C 1953

The MATLAB® implementation of Algorithm D.2 with the same starting

solution as the previous example and the same location of the coprime factor poles
with ˇ D 0:1 is executed to obtain the following controller:

228s 4 C 3441s 3 C 1:051e4s 2 C 1:424e4s C 7386

Css4 D (D.61)
s 4  5:166s 3 C 643:7s 2  668:1s C 1484
D.2 DRLCD Using Simultaneous Stabilization Theory 171

Fig. D.7 Step-response comparison of the performance of different simultaneously stabilizing

controllers with plant G0 (nonlinear system)

The particular reason for selecting ˇ D 0:1 is because the nonlinear effects (i.e.,
Coulomb friction) are more pronounced at small signals, and hence, it is decided to
pay more attention to the model given by G1 .s/. This value for ˇ is also selected in
Nassirharand (1991).
The controller obtained by direct optimization (Teh 2010) is given as below (see
Chap. 4):

3:1671s C 1:1469
Css5 .s/ D (D.62)
0:0229s C 0:1089

The unit step-response plots of both controllers with both linear approximations
are depicted in Figs. D.7 and D.8.
The unit step-response plots of Css5 .s/ with the nonlinear plant are depicted in
Fig. D.9. In this example, although Css4 .s/ stabilizes both G0 .s/ and G1 .s/, it does
not stabilize the actual nonlinear system (see Fig. D.10). Hence, the original call
to use the simultaneous stabilization theory of linear systems for nonlinear control
system design becomes questionable.

D.2.4 Discussion

The primary objectives of this appendix were to demonstrate the following two
major problems with the application of the simultaneous stabilization theory of
linear systems to nonlinear systems:
172 Appendix D Dual-Range Controllers and Simultaneous Stabilization Theory

Fig. D.8 Step-response comparison of the performance of different controllers with plant G1
(nonlinear system)

Fig. D.9 Normalized step-response plots of the closed-loop system with the nonlinear plant and
the simultaneously stabilizing controller (D.62)

1. The simultaneously stabilizing controllers tend to be of a high order.

2. The simultaneously stabilizing dual-range controllers may not stabilize the actual
nonlinear system.
These goals have been met. In Sect. D.2.3.1, a linear system with integrity is
considered. In Sect. D.2.3, a nonlinear example problem is considered, and for this
problem it is demonstrated that simultaneously stabilizing dual-range controller
obtained from the application of the simultaneous stabilization theory of linear
systems is of a much higher order than that obtained using a direct optimization
approach. It is also demonstrated that, unlike the latter dual-range controller, the
former dual-range controller does not stabilize the actual nonlinear plant.
D.2 DRLCD Using Simultaneous Stabilization Theory 173

Fig. D.10 Normalized step-response plots of the closed-loop system with the nonlinear plant and
the simultaneously stabilizing controller (D.61) – unstable!

It could certainly be argued that the direct optimization approach (Teh 2010; Nas-
sirharand et al. 1988; Nassirharand 1991; Patwardhan 1988) described in Chap. 4
could also result in controllers that would stabilize the two linear models but not
the actual nonlinear plant. In such cases, one should consider design of a nonlinear
controller as described in Chap. 5. However, application of the direct optimization
approach to various nonlinear problems in aerospace and servomechanism has not
shown such a characteristic (see Teh 2010; Teh and Nassirharand 2010b, c; Chap. 4;
and Sect. D.2.3.1).
References

Abbassi N (2003) Controller design for simultaneous control of combustion chamber pressure and
mixture ratio of a specific liquid propellant engine. M.S. Thesis Department of Electrical and
Computer Engineering, Amir Kabir University, Tehran
Al-Kindi SSA, Nassirharand A (2010) State-variable control of quadrotor helicopter. In: Proceed-
ings of international conference on advances in mechanical engineering, Shah Alam
Atherton DP (1970) Comment on describing function inversion. Electron Lett 6:779–780
Atherton DP (1975) Nonlinear control engineering. van Nostrand Reinhold, London
Atherton DP, Benouarets M, Nanka-Bruce O (1993) Design of nonlinear PID controllers for
nonlinear plants. Proc IFAC World Congr Sydney 3:355–358
Bengea SC, Li X, Decarlo RA (2004) Combined controller-observer design for uncertain time
delay systems with application to engine idle speed control. ASME Trans J Dyn Syst Meas
Control 126:772–780
Besancon-Voda A, Blaha P (2002) Describing function approximation of a two-relay system
configuration to Coulomb friction identification. Control Eng Pract 10:655–668
Bongard JC, Lipson H (2005) Nonlinear system identification using coevolution of models and
tests. IEEE Trans Evol Comput 9:361–384
Chen CT (1984) Linear system theory and design. Holt, Rinehart, and Winston, New York
Chen CF, Shieh LS (1970) An algebraic method for control systems design. Int J Control 1:717–
739
Colgren RD (2004) Applications of robust control to nonlinear systems. American Institute of
Aeronautics and Astronautics, Reston
D’Azzo JJ, Houpis CH (1966) Feedback control system analysis and synthesis, 2nd edn. McGraw-
Hill Book Company, New York
Dhupia JS, Ulsoy AG, Katz R et al (2008) Experimental identification of the nonlinear parameters
of an industrial translational guide for machine performance evaluation. J Vib Control 14:645–
668
Doyle JC, Glover K, Khargonekar PP, Francis BA (1989) State-space solutions to standard H2 and
H1 control problems. IEEE Trans Auto Control 34:831–847
Garvin C, Mathew A (1996) The application of the method of simultaneous stabilization to the
control of a nonlinear servo valve. IEEE Trans Control Syst Technol 4:654–664
Ge HW, Qian F et al (2008) Identification and control of nonlinear systems by a dissimilation
particle swarm optimization-based Elman neural-network. Nonlin Anal Real World Appl
9:1345–1360
Gelb A, Vander Velde WE (1968) Multiple-input describing functions and nonlinear system design.
McGraw-Hill, New York

A. Nassirharand, Computer-aided Nonlinear Control System Design: Using Describing 175

Function Models, DOI 10.1007/978-1-4471-2149-7,
© Springer-Verlag London Limited 2012
176 References

Govindan P A/L R (2011) Controller design for a 2 degree-of-freedom robot arm. BEng Thesis,
Faculty of Engineering, The University of Nottingham, Semenyih, Selangor
Haber R, Keviczky L (1999) Nonlinear system identification-input–output modeling approach.
Kluwer Acaemic Publishers, Dordrecht
Horowitz IM (1976) Synthesis of feedback systems with nonlinear time-varying uncertain plant to
satisfy quantitative performance specifications. Proc IEEE 64:123–130
Hunt LR, Su R, Meyer G (1987) Global transformation of nonlinear systems. IEEE Trans Auto
Control 28:24–30
John J (2001) Nonlinear continuous system identification by means of multiple integration II. Acta
Polytechnica 41:64–67
Karimi H, Nassirharand A (2006) Application of a simulation algorithm for dynamic and nonlinear
analysis of a liquid propellant engine. J Aerosp Sci Technol 3:23–30
Karimi H, Nassirharand A, Beheshti M (2003) Dynamic and nonlinear simulation of liquid
propellant engines. AIAA J Propuls Power 19:938–944
Kerschen G, Worden K, Vakakis AF et al (2006) Past, present and future of nonlinear system
identification in structural dynamics. Mech Syst Signal Process 20:505–592
Kim MS, Chung SC (2006) Friction identification of ball-screw driven servomechanisms through
the limit cycle analysis. Mechatronics 16:131–140
Kuo BC (1982) Automatic control systems, 4th edn. Prentice-Hall, Inc., Englewood Cliffs
Li SC, Nassirharand A (2010) A single-range controller design procedure for one degree-of-
freedom servo systems with experimental results. In: Proceedings of international conference
on advances in mechanical engineering, Shah Alam
Li SC, Nassirharand A (2011) Non-linear proportional-integral-derivative synthesis for unstable
non-linear systems using describing function inversion with experimental verification. Proc
Inst Mech Eng Part I J Systems and Control Eng (in press)
Li L, Yang Y, Peng H (2010) Fuzzy system identification via chaotic ant swarm. Chaos Solitons
Fractiles 41:401–409
Lyashevskiy S, Chen Y (1997) Identification of nonlinear system parameters using describing
function approach. Proc Am Control Conf 3:1944–1945
Mousavi Firdeh SR, Nassirharand A, Abbassi N et al (2006) A systematic single-range controller
synthesis procedure for nonlinear and multivariable liquid propellant engines. Aerosp Sci
Technol 10:392–401
Nanka-Bruce O, Atherton DP (1990) Design of nonlinear controllers for nonlinear plants. Proc
IFAC Congr Tallinn 6:75–80
Nassirharand A (1986) Controller design for nonlinear systems based on simultaneous stabilization
theory and describing function models. PhD Thesis, School of Mechanical and Aerospace
Engineering, Oklahoma State University, Stillwater, Oklahoma
Nassirharand A (1987) Input/output characterization of highly nonlinear systems. Adv Eng Softw
9:129–133
Nassirharand A (1988) Identification of frequency domain models for nonlinear systems. Adv Eng
Softw 10:195–201
Nassirharand A (1991) Design of dual-range linear controllers for nonlinear systems. ASME Trans
J Dyn Syst Meas Control 113:590–596
Nassirharand A (1993) Factorization approach to control system synthesis. AIAA J Guid Control
Dyn 16:402–405
Nassirharand A (2003) Computation of lucid factors for Bezout identity. Adv Eng Softw 34:527–
531
Nassirharand A (2008a) A new software tool for design of linear compensators. Adv Eng Softw
39:132–136
Nassirharand A (2008b) Controller synthesis procedure for linear unstable plants. Int J Automat
Control 2:137–149
Nassirharand A (2008c) Control system design of multivariable non-linear space robots. Aircr Eng
Aerosp Technol 80:27–34
References 177

Nassirharand A (2009a) MATLAB software for inversion of describing functions. Adv Eng Softw
40:600–606
Nassirharand A (2009b) Computer-aided design of classical linear compensators with application
to aerospace. J Syst Eng Electron 20:1373–1377
Nassirharand A (2010a) Nonlinear controller synthesis for SIMO systems with application to cruise
missile. ASCE J Aerosp Eng 23:105–110
Nassirharand A (2010b) Controller design for nonlinear systems. LAP Lambert Academic
Publishing, Saarbrucken
Nassirharand A (2011) Computer-aided design of P or PI plus rate feedback compensators for
linear systems. Adv Eng Software 42:1035–1040
Nassirharand A, Housseini SA (2009) Tracking and decoupling of multivariable nonlinear systems.
Int J Model Identif Control 6:341–348
Nassirharand A, Karimi H (2002) Input/output characterization of highly nonlinear multivariable
systems. Adv Eng Softw 33:825–830
Nassirharand A, Karimi H (2004a) Design of a single-range controller for the pressure control of
a combustion chamber. Scientia Iranica 11:153–158
Nassirharand A, Karimi H (2004b) Controller synthesis methodology for multivariable nonlinear
systems with application to aerospace. ASME Trans J Dyn Syst Meas Control 126:598–607
Nassirharand A, Karimi H (2004c) Closed-form solution for design of lead-lag compensators. Int
J Elect Eng Educ 41:172–180
Nassirharand A, Karimi H (2006) Nonlinear controller synthesis based on inverse describing
function technique in the MATLAB environment. Adv Eng Softw 37:370–374
Nassirharand A, Mousavi Firdeh SR (2008) Design of nonlinear lead and/or lag compensators. Int
J Control Automat Syst 6:394–400
Nassirharand A, Mousavi Firdeh SR (2009a) Combustion chamber pressure control based on
mixed H1 -describing function approach. Acta Astronaut 64:206–211
Nassirharand A, Mousavi Firdeh SR (2009b) Single-range controller design for a cruise missile.
Aircr Eng Aerosp Technol 81:283–287
Nassirharand A, Mousavi Firdeh SR (2009c) Design of nonlinear H1 controllers using describing
functions with application to servomechanism. Asian J Control 11:446–450
Nassirharand A, Patwardhan A (1988) A fast algorithm for obtaining low-order coprime fractional
representations of SISO systems. Control Theory Adv Technol 4:277–281
Nassirharand A, Taylor JH (1990) Synthesis of linear PID controllers for nonlinear multivariable
systems. In: Proceedings of the American control conference, San Diego, California, pp 2223–
2228
Nassirharand A, Taylor JH (1991) Frequency domain modeling of nonlinear multivariable systems.
Control Theory Adv Technol 7:201–214
Nassirharand A, Taylor JH, Reid KN (1988) Controller design for nonlinear systems based on
simultaneous stabilization theory and describing function models. ASME Trans J Dyn Syst
Meas Control 110:277–281
Nassirharand A, Karimi H, Dadfarnia M (2003) A new software tool for synthesis of linear PID
controllers. Adv Eng Softw 34:551–557
Nelder JA, Mead R (1965) A simplex method for function minimization. J Comput 7:308–313
Nett CN, Jacobson CA, Balas MJ (1984) A connection between state-space and doubly coprime
fractional representations. IEEE Trans Automat Control 29:268–280
Ogata K (1990) Modern control engineering, 2nd edn. Prentice Hall, Englewood Cliffs
Ozer MB, Ozguven HN, Royston TJ (2009) Identification of structural non-linearities using
describing functions and the Sherman-Morrison method. Mech Syst Signal Process 23:30–34
Patwardhan A (1988) Compensator synthesis using factorization approach. Masters Thesis,
Department of Mechanical Engineering, University of Kentucky, Lexington, Kentucky
Petrovic I, Baotic M, Perik N (2000) Model structure selection for nonlinear system identification
using feedforward neural networks. In: Proceedings of IEEE-INNS-ENNS international joint
conference on neural networks, Como, 1:1053
178 References

Prazenica RJ, Brenner MJ, Kurdila AJ et al. Nonlinear system identification for aero-elastic
systems. Research and Engineering Education Facility, University of Florida, Shalimar Florida.
http://www.reef.ufl.edu/nonlinear%20System%20Id%20fo..ro-ElasticB.htm
Pujara LR (1988) Computer-aided control systems design technique with applications to aircraft
flying qualities. AIAA J Guid Contr Dynam 11:250–255
Reklaitis GV, Ravindran A, Ragsdell KM (1983) Engineering optimization. Wiley, New York
Rivera DE, Morari M, Skogestad S (1986) Internal mode control. PID Controller design. Ind Eng
Chem Process Des 25:252–265
Rowland JR (1986) Linear control systems. Wiley, New York
Safonov MG, Limebeer DJN, Chiang RY (1989) Simplifying the H1 theory via loop-shaping,
matrix-pencil and descriptor concepts. Int J Control 50:2467–2488
Shi J, Sun HH (1990) Nonlinear system identification for cascaded block model: an application to
electrode polarization impedance. IEEE Trans Biomed Eng 37:574–587
Skogestad S, Postlethwaite I (2005) Multivariable feedback control. Wiley, West Sussex
Slotine JJ, Li W (1991) Applied nonlinear control. Prentice Hall, Englewood Cliffs
Suzuki A, Hedrick JK (1985) Nonlinear controller design by an inverse random-input describing
function methods. In: Proceedings of American control conference, Boston, MA, pp 1236–
1241
Taylor JH (1983) A systematic nonlinear controller design approach based on quasilinear system
models. In: Proceedings of American control conference, San Francisco, CA, pp 141–145
Taylor JH, O’Donnell JR (1990) Synthesis of nonlinear controllers with rate feedback via
sinusoidal input describing function methods. American control conference, San Diego, pp
2217–2222
Taylor JH, Strobel KL (1985) Nonlinear control system design based on quasilinear system models.
In: Proceedings of American control conference, Boston, MA, pp 1242–1247
Teh SH (2010) Design of multivariable dual-range linear controllers for nonlinear systems. BEng
Thesis, Faculty of Engineering, The University of Nottingham, Semenyih, Selangor
Teh SH, Nassirharand A (2010a) Idle speed control using a describing function approach. Proc Inst
Mech Eng Part D J Automob Eng 224:533–545
Teh SH, Nassirharand A (2010b) Software for design of linear controllers for nonlinear systems.
In: Proceedings of international conference on advances in mechanical engineering, Shah Alam
Teh SH, Nassirharand A (2010c) Multivariable temperature control of a reactor. Proceedings of
international conference on advances in mechanical engineering, Shah Alam
Ujile A (2010) MATLAB® software for the design of dual-range linear controllers with application
to robotics. BEng Thesis, School of Electrical and Electronics Engineering, The University of
Nottingham – Malaysia Campus, Semenyih, Selangor
Ujile A, Nassirharand A (2010) MATLAB software for design of SISO dual-range linear
controllers. Int J Syst Control 1:149–155
Utkin V, Guldner J, Shi J (1999) Sliding mode control in electromechanical systems. CRC Press,
Boca Raton
Vaerenbergh SV, Via J, Santamaria I (2007) Nonlinear system identification using a new sliding-
window kernel RLS algorithm. J Commun 2:1–8
Vidyasagar M (1985) Control system synthesis: a factorization approach. MIT Press, Cambridge
Vidyasagar M, Viswanadham N (1982) Algebraic design techniques for reliable stabilization. IEEE
Trans Automat Control 27:1085–1095
Wang L, Barnes TJD, Cluett WR (1995) New frequency-domain design method for PID con-
trollers. IEE Proc: Control Theory Appl 142:265–271
Youla DC, Bongiorno JJ, Jabr HA (1976) Modern Wiener-Hopf design of optimal controllers – pt.
1: the single-input case. IEEE Trans Automat Control AC-21:3–14
Zhuang M, Atherton DP (1993) Automatic tuning of optimum PID controllers. IEE Proc Part D
140:216–224
Ziegler JG, Nichols NB (1942) Optimum settings for automatic controllers. ASME Trans J Dyn
Syst Meas Control 62:759–768
Index

A E
Algebraic design Experimental study, 4, 87–119
algorithm, 127
software, 127–128
F
Feedback structure, 126, 127, 132, 141,
B 143
Bezout identity, 153–159
software, 159
H
H1 design, 31– 34, 55, 57, 58, 78–81, 84, 97,
132, 134, 136
C linear, 57, 58
CACSD. See Computer-aided control system nonlinear-example, 57, 78–81, 84
design
cad controller, 57, 91, 118, 128
Case studies I
idle speed control, 34–41 Inversion
liquid propellant engine, 24–35, 127–143 algorithm, 64–65
robotics, 146–152 software, 69–72
servo, 10–14, 46–48, 72–85, 167–170
UAV, 81–85
Computer-aided control system design L
(CACSD), 1, 3 Lead-lag
linear
example, 100–102
D software, 99
Design platform, 1–3, 9 nonlinear-example, 75–78
Dual-range linear controller design (DRLCD)
algorithm, 45
example, 46–55 M
experiment, 44–46 Multi-range nonlinear controller design
simultaneous stabilization (MRNCD)
algorithm, 159–165 algorithm, 58–60, 64, 65
examples, 165–171 experiment, 60–65
software, 46, 159 software, 66–72

A. Nassirharand, Computer-aided Nonlinear Control System Design: Using Describing 179

Function Models, DOI 10.1007/978-1-4471-2149-7,
© Springer-Verlag London Limited 2012
180 Index

P Single-range linear controller design (SRLCD)

PID algorithm, 24
experiment, 88–95 example, 25–41
linear, 103–104 experiment, 23–25
example, 104 Sinusoidal-input describing function
software, 104 (SIDF)
nonlinear-example advantages, 5, 6
software, 66–72 examples (see Case studies)
PI plus rate feedback generation of, 4
example, 116–119 software, 10
software, 116 SRLCD. See Single-range linear controller
P plus rate feedback design
example, 111–112 State-variable model, 4, 9
software, 109–111

S U
SIDF. See Sinusoidal-input describing function Unstable systems, 87, 88