This action might not be possible to undo. Are you sure you want to continue?
AN INTRODUCTION TO
CONTROL SYSTEMS
ADVANCED SERIES IN ELECTRICAL AND COMPUTER ENGINEERING Editor: W. K. Chen
Published Broadband Matching: Theory and Implementations(Second edition) by W. K. Chen Volume 2: Active Network Analysis by W. K. Chen Volume 3: Linear Networks and Systems: ComputerAidedSolutions and Implementations (Second edition) by W. K. Chen Volume 4: Introductionto Engineering Electromagnetic Fields by K. ~mashankar Signal Processing Volume 6: ln~~oductory by R. Priemer Volume 7: Diagnostic Measurements in LSlNLSl Integrated Circuits Production by A. Jakubowski, W. Marciniak, and H. M. Przewlocki Volume 9: Orthogonal Functions in Systems and Controls by K. B. Datta and B. M. Mohan Volume 10: Introductionto High Power Pulse Technology by S. T. Pai and Q. Zhang n Volume 11: Systems and Control: An l n t r ~ u c t i o to Linear, Sampled and Nonlinear Systems by T.Dougherty Volume 1:
Forthcoming Volume 5: Applied Graph Theory: Graphs and Electrical Networks (Third edition) by W. K. Chen An lnt~oduction Control Systems (Second edition) to byK. Warwick Infrared Characterizationfor Microelectronics by W. S. Lau Identification, Estimation and Statistical Signal Processing by Y.F. Huang
Volume 8:
Advanced Series in €lectrical and C o ~ p u t e Engineer~ng r Vol. 8
AN INTRODUCTION TO
CONTROL SYSTEMS
(Second Edition)
KEVIN WARVSrICK
Department of Cybernetics, University of Reading
Y
$World Scientific
Singapore New Jersey London Hong Kong
Published by
World Scientific Publishing Co. Pte. Ltd. P 0 Box 128, Farrer Road, Singapore 912805
USA ofice: Suite IB, 1060 Main Street, River Edge, NJ 07661
U K ufice: 57 Shelton Street, Covent Garden, London WC2H 9HE
British Library CataloguinginPublication Data A catalogue record for this book is available from the British Library.
First published by World Scientific Publishing Co. Pte. Ltd. 1996 First reprint 1996 (pbk)
A
First printed in 1989 by Prentice Hall International (UK) Ltd division of Simon and Schuster International Group
AN INTRODUCTION TO CONTROL SYSTEMS
Copyright 0 1996 by World Scientific Publishing Co. Pte. Ltd
All rights reserved. This book, or purts thereof; may not be reproduced in unyfurm or by any meuns, electronic or mechunicul, including photocopying, recording or any information sturuge und retrieval system now known or to
be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, M A 01923, USA. In this case permission to photocopy is not required from the publisher.
ISBN 9810225970 (pbk)
Printed in Singapore by U t t r P n n t
To the Staff and Students of the University of Reading, Cybernetics Department
.
2 A concise history of control 1.5 Some examples of control systems 1.5 Electromechanical devices 2.4 Closedloop control 1.Contents xi xiii xiv Preface Preface to the Second Edition About the author 1 Introduction 1.2 Steadystate response 3.1 Introduction 2.3 Block diagrams 2.7 Summary Problems Further reading 1 1 3 5 6 8 10 12 12 14 2 System representations 2.6 Definitions of standard terminology 1.1 Introduction 3.4 Errors 3.1 General introduction 1.5 Servomechanisms 56 56 57 65 73 84 .3 Transient response 3.6 Summary 15 15 16 25 32 42 47 48 54 Problems Further reading 3 System response using differential equations 3.4 Physical system realizations 2.2 Transfer functions 2.3 Openloop control 1.
1 4.7 Compensation using the Nichols chart 5.6 Summary Problems Further reading 87 88 92 4 System stability and performance 4.6 Decomposition 6.7 Summary Problems Further reading 197 197 198 201 205 209 213 216 216 220 7 State variable feedback 7.3 Controllability and observability in transfer functions 7.4 The Nyquist stability criterion 5 .3 4.viii Contents 3.4 The characteristic equation 6.4 4.1 Introduction 7.1 Introduction 6.1 Introduction 5.6 93 93 95 102 115 125 131 131 136 Introduction The RouthHurwitz criterion Root loci construction Application of root loci Compensation based on a root locus Summary Problems Further reading 5 Frequency response analysis 5.2 4.3 Solution to the state equations 6.2 Controllability and observability 7.6 The Nichols chart 5.3 Compensation using the Bode plot 5.4 State variable feedback 7.2 The Bode diagram 5.2 State equations from transfer functions 6.5 Canonical forms 6.5 4.8 Summary Problems Further reading 137 137 139 149 159 170 177 184 191 192 195 6 State equations 6. 5 Compensation using the Nyquist plot 5.5 State estimation 222 222 223 229 232 238 .
7 Introduction Sampled data systems Stability analysis Digital compensation State variable forms Pole placement design Summary Problems Further reading 249 249 249 253 256 26 1 264 269 269 273 9 Computing methods 9.3 8.5 Summary Problems Further reading 306 306 307 312 319 329 330 334 Appendices 1 Matrix algebra 2 Tables of common transforms 3 Nonlinear systems 4 Unity feedback loops index 335 335 343 346 355 359 .5 8.2 8.Contents ix 7.4 8.2 9.6 8.3 9.1 9.1 8.6 Summary Problems Further reading 243 243 248 8 Digital control 8.1 Introduction 10.4 PID control 10.3 Proportional + integral control 10.5 Introduction System simulation Computer control Computeraided design Summary Problems Further reading 274 274 275 279 298 302 303 305 10 Threeterm controllers 10.4 9.2 Proportional control 10.
.
are considered. These two chapters. important aspects of control systems design. This book treats digital control as one of the basic ingredients to an introductory control course. In the 1960s control engineers were predominantly concerned with analog design procedures. detail the necessary mathematical tools used in the analysis of a system and give an indication as to how the system responds to a particular set of . the increasing influence of microcomputers has led to a much greater need for an emphasis to be placed on digital aspects of control. Significantly. Basic handling rules for these models are given and the model’s relationship with everyday engineering tools and apparatus is stressed. This includes a brief history of control and points out many of the standard definitions and terms used. Many control ideologies that were once thought to be unrealizable. in effect. The undergraduate students of engineering taking an introductory control course as part of their overall degree must necessarily become aware of the developments in digital control which have now become standard practice in many industrial applications and are fundamental to any further degree courses or research in control. have now become everyday realities in practical industrial applications due to the transference of complex algorithms to a software base. Differential equations are examined in Chapter 3 and in Chapter 4 system stability and performance. because of hardware limitations. Organization of the book is as follows. With the advent of VLSI design in association with specialpurpose machines. Many new techniques have risen to prominence in the area of control systems during the last decade. the analog computer being a fundamental building block. higher speed and greater memory capacity computers are becoming available. except in a few extreme cases. A general overview of the field called control systems is introduced in Chapter 1. Things have changed considerably since that period. including the more regulady found topics with equal weight. In Chapter 2 it is shown how mathematical models can be formed in order that they may be representative of a system. In common with most engineering disciplines it is a topic which is constantly changing in order to keep pace with modern day application requirements. allowing further control schemes to be considered.Preface Automatic control encroaches upon many of the defined engineering and mathematics university courses in terms of both theory and practice.
Chapter 4 concentrates mainly on the RouthHurwitz criterion in terms of stability and rootlocus techniques for performance by investigating the effect of parameter variations on the roots of the system characteristic equation. Chapter 5 looks into the frequency response methods due to Bode and Nyquist. The book is intended to be used in relation to an introductory course on control systems. computeraided control system design being highlighted. however. John Finch. for which students need a mathematical background such that Laplace transforms. Realization of the state variables when these are not measurable or physically obtainable is also considered. including practical help from my daughter Madeline: I hope the reader concludes that it has been worth it. I would also like to acknowledge the assistance of Sukie Clarke and Karen Smart in preparing the text and correcting my tewible English. whereas Chapter 7 is devoted more to system control using state feedback. are also used for control design in terms of system simulations and modeling. I would particularly like to thank John Westcott. It is considered that Chapters 3. I am immeasurably grateful for the support I have received from my family during the writing of this book. Finally. The same approach is then taken in the discussion based on the Nichols chart which is used in obtaining the system closedloop frequency response. and includes details of its implementation in both continuous and discrete time. and numerous former students who have provided valuable feedback on the material presented. subject to this being possible. Chapters 8 and 9 include the important topics of digital and computer control. . Mohammed Farsi and Keith Godfrey for their technical appraisal. The use and variety of state equations and their relationship to system transfer functions is found in Chapter 6. 4 and 5 cover what is now termed classical control techniques.xii Preface conditions. The final chapter looks in depth at the most widely used control scheme. Computers. This text has been put together with the help of several people. Details are given with emphasis placed on how to design compensation networks using the methods described. solutions to differential equations and complex numbers have been taught. the rest of the book is therefore devoted to more recently devised methods. Design philosophies and system descriptions based on the state space are brought together in Chapters 6 and 7. the PID controller. Digital control uses the z operator and is closely linked with computer control via sampleddata systems.
. I am delighted that World Scientific Publishing Co. have decided to publish this second edition. her husband Rodney for helping with the figures and Barbara Griffin at World Scientific for her patience. following the first edition selling out very quickly. My thanks go to Liz Lucas for all her help in putting together this edition. apart from a brief mention. In the second edition. to be replaced by a completely new section on Fuzzy Logic Controllers. which is very much of interest both from an academic and industrial viewpoint.Preface t o the Second Edition One of the reviewer’s comments on the first edition was that the section on Analog Computers was rather dated. this section has been completely removed.
About the Author Kevin Warwick has been Professor of Cybernetics and Head of the Department of Cybernetics at the University of Reading. UK. neural networks and robotics. He has also published well over 200 papers in the broad areas of his research interests. since 1988. a Fellow of the IEE and an Honorary Editor of the IEE Proceedings on Control Theory and Applications. as well as being employed by British Telecom for six years. including intelligent systems. including Virtual Reality in Engineering and Applied Artificial Intelligence. Imperial College. . Professor Warwick has received Doctor of Science degrees for his research work on Computer Control both from Imperial College London and the Czech Academy of Sciences. Newcastle University and Warwick University. Prague. He previously held appointments at Oxford University. Kevin is a Chartered Engineer. London. a number of books. He has written or edited.
the response itself can be characterized by essentially the same information. systems often exhibit characteristics which are of a similar. signs of good taste and common sense . Certain properties are common between many different fields. although it is more usual for a system to be regarded as something to which the concepts of cause and effect apply. In its most general sense a system can be virtually any part of life one cares to consider. flow rate. engineering. heating a pot of water: this exhibits the same type of exponential response as that witnessed if either a capacitor is charged up or a . which involves the movement of our hands in a coordinated fashion. nature. height. as opposed to other books written by different authors. speed. mechanical. possibly different branches of engineering.. marine. This movement is made in response to signals arising due to information obtained from our senses . This is usually witnessed in terms of a system’s response to certain stimuli. electronic. etc. irrespective of the field in which it lies.both of which are difficult properties to quantify and study. In terms of engineering it therefore follows that the study of control systems is multidisciplinary and is applicable equally well in the fields of chemical. picking up and opening this book. for example. is an example of the human being behaving as an intelligent control system exhibiting. if not identical.in this case sight in particular. Take. it should in no way be considered that the topic is a completely new field with which one has had no prior contact whatsoever.Introduction 1. and it is the mere fact of this commonality that has given rise to the field of control systems. In everyday life we encounter control systems in operation and actually perform many controlling actions ourselves. in this case. etc. voltage. Consider. examples being distance. electrical. for example. nuclear. and although the physical properties of the stimuli themselves can take a different form. A control system can then be thought of as a system for which we manipulate the cause element in order to arrive at a more desirable effect (if at all possible). It is worth mentioning however that the selection of this book.. Although lying within distinctly different fields.1 General introduction When first encountering the subject of control systems.
requiring that a level of water does not vary by more than 2% from a nominal mean value. quantity of gas (if the water is gas heated). voltage and force respectively.g. A model can often take the form of an appropriate mathematical description and must be satisfactory in the way in which it expresses the characteristics of the original system. or models. the response. is still characterized as in Fig. in terms of system performance. is therefore made great use of in the design stage. whether this is to be in terms of a modification to the system itself or in terms of a separate controller block. 1. There are essentially two main features in the analysis of a control system. An underlying theme in the study of automatic control systems is the assumption that any required controlling/corrective action is carried out automatically by means of actuators such as electromechanical devices. however.1. 1 . charge or stored energy. Modeling a system is the means by which our picture of the system is taken from its own discipline into the common control systems arena. The study of control systems in terms of a multidisciplinary basis therefore means that the results of performance and design tests in one discipline can readily be reposed and reevaluated within an alternative discipline. 1 . achieving a particular level of water or a speed of rotation. The concept of a human operative reading a value from a meter and applying what is deemed an appropriate response by pulling a lever or twisting a dial is therefore not really part of the subject matter covered in this . by means of an appropriate design. in which a suitable control strategy is both selected and implemented in order to achieve a desired system performance. 1 . The design of a control system only makes sense if we have some objective. which can be readily dealt with and understood. e. e. Actual performance objectives are very much dependent on the discipline in which the system exists. which we are aiming t o achieve. previously obtained. although the objectives can be stricter.g. say. see Fig. 1 Exponential response of a system Time  spring is compressed. The stimulus is. We are therefore trying to alter the present performance of a system in order t o meet our objectives better.2 Introduction " Fig. The system model. whether it be in terms of temperature. Firstly system modeling. which means expressing the physical system under examination in terms of a model. The second feature of control systems analysis is the design stage.
It must be remembered however that interfacing is required.A concise history of control Input signal System output signal 3 Fig. 1. the speed of rotation of the flyballs. via the piston and flywheel. As the flyballs rotate so they determine.2. It is also a difficult task to point the finger at someone way back in time and accuse them of starting the whole thing. a system can therefore be depicted as a block which operates on an input signal in order to provide an output signal.2 Basic system schematic book. dependent on the particular text one is looking at. it serves a primary lesson in control systems analysis not to rely completely on any measured value. James Watt’s flyball governor is shown in Fig. Although the actual year itself is not particularly important. The human being as an element in a control system is briefly considered later in this first chapter. The study of control systems is certainly no exception in this context. 1. in terms of speed variation.the less steam is supplied. Indeed some historical accounts refer to times well before the birth of Christ for relatively early examples. As far as automatic control (no human element) is concerned. .2 A concise history of control To put forward a chronologically ordered set of events in order to show the logical development of a technical subject is certain to be fraught with wrong conclusions and misleading evidence. can be obtained with such a device. In terms of the system under control. The exact date of the invention/ development seems to vary. how much steam is supplied. The rate of steam supplied then governs. merely to serve as an example. with dates ranging from 1767 to 1788 being quoted. the faster the rotation . a particular problem being posed by much of the development of the mathematical tools which now form the basis of control systems analysis. 1. 1. via the valve. as shown in Fig. where the control objective is to ensure that the speed of rotation is approximately constant. an actuator is used to apply a signal as input to the system whereas a suitable measurement device is employed to witness the response in the form of an output from the system. firstly to feed the measured value into the computer and secondly to convert the calculated response from the computer into a signal which is suitable for a physical actuator. It is in fact becoming more often the case that a digital computer is employed to turn the measured (meter) value into an appropriate responsive action. a historical example cited in many texts is James Watt’s flyball governor. Although tight limits of operation. Schematically. the main advantage of such a technique being the speed with which a computer can evaluate a set of what could be complicated equations.3. The characteristics of that system are then contained within the block.
The Second World War provided an ideal breeding ground for further developments in automatic control. thereby stressing the importance of the phase. turned his attention to the problem of stability in repeater amplifiers. particularly due to finance made available for improvements in the military domain. provided over a century before by Laplace and Fourier. C. which has become widely used as a name to describe many types of feedback control system. He successfully tackled the problem by making use of standard function theory. In the early 1930s Harry Nyquist. many years after Watt’s governor. In 1934 a paper appeared by Hazen entitled ‘Theory of servomechanisms’ and this appears to be the first use of the term ‘servomechanism’. In the first instance (1920s) this was due to analysts such as Heaviside who developed the use of mathematical tools. in particular forming links with the decibel as a logarithmic unit of measurement. Around 1868. . J. in the study of communication systems. a physicist who had studied noise extensively. A common feature with the systems was the employment of information feedback in order to achieve a controlling action. Examples of military projects of that time are radar tracking. Rapid development in the field of automatic control took place in the 1920s and 30s when connections were drawn with the closely related areas of communications and electronics. as well as the gain. characteristics of the amplifier. thereby explaining in mathematical terms reasons for oscillations within the system.4 Introduction Flywheel Steam from __t boiler Piston Fig. an example being piloting (steering) of ships. It was gradually found that Maxwell’s governor equations were more widely applicable and could be used to describe phenomena in other systems.3 James Watt’s flyball governor there are unfortunately several negative features. 1. Maxwell developed a theoretical framework for such as governors by means of a differential equation analysis relating to performance of the overall system. not the least of these being the tendency of the speed to oscillate about a mean (desired) speed value.
An excellent example is an automatic electric toaster in which the control is provided by means of a timer which dictates the length of time for which the bread will be toasted. The need for highspeed control devices has in the 1980s been a contributing factor too and has made great use of hardware techniques.3 Openloop control An openloop control system is one in which the control input to the system is not affected in any way by the output of the system. e. Finally the low cost and ease of availability of personal computers has meant that many control systems are designed and simulated from a software basis.Openloop control 5 antiaircraft gun control and autopilots for aircraft. has provided a stimulus for more recent developments. The 1970s saw further progress on the computer side with the introduction of microprocessors. It is apparent therefore that if the system characteristics change at some time then both the response accuracy and repeatability can be severely impaired. Since that time mechanization in many industries. with a computer completing the feedback loop.) However. Implementation. In almost all cases however the openloop system will present no problems insofar as stability is concerned. It is also necessary however that the system itself is not varied in any way in response to the system output. whereas at the same time ideas from the field of artificial intelligence have been employed in an attempt to cope with increased complexity needs. along with the root locus method proposed by Evans in the late 1950s. 1. in the 1960s the influence of space flight was felt.g. Digital control also became widespread due to computers. the assumption being that once the timer has been set the operator can only wait to examine the end product. . Clearly the response of an openloop system is dependent on the characteristics of the system itself in terms of the relationship between the system input and output signals. is then only carried out when good control is assured. The output from the toasting system is then the brownness or quality of the toast. such as parallel processors. Originally. which may itself make use of a computer. Such a definition indicates that openloop systems are in general relatively simple and therefore often inexpensive. each of which requires tight performance achievements in terms of both speed and accuracy of response. with optimization and statespace techniques gaining in prominence. process and power generation. (Note: This was also known in the UK at one time as Westcott’s method of ?r lines. frequency response techniques such as Bode’s approach and Laplace transform methods were prominent. The use of robot manipulators not simply for automating production lines but as intelligent workstations also considerably changed the requirements made of a control system in terms of speed and complexity. manufacturing. thus allowing for the implementation of relatively complicated control techniques on oneoff systems at low cost. which were particularly relevant in the process control industries in which many variables must be measured and controlled.
where the error = reference input . It is often the case that the signal fed back from the system output is compared with a reference input signal. if an input is applied the output will not shoot off to infinity . A problem with this openloop system is that although today the water may be provided at a nice temperature at the output. 1. Another example of a n openloop system is the water heater shown in Fig. tomorrow this might not be the case. Very often the reference input is directly related to the desired value of system output. By using output information to affect in some way the control input of the system. We have now closed the loop. the result of this comparison (the difference) then being used to obtain the control o r actuating system input.system output.4.it is not much use as a n openloop system if this is the case. This is simply because the purpose of the controller will most likely be . feedback is being applied to that system. and where this is a steady value with respect t o time it is called a set point input. but rather is closedloop. in which the controller is merely a n on/off switch which determines when the heater is supplying heat in order t o provide heated water at a certain specified temperature. a change in the temperature or amount of water input to the storage device or a drop in the voltage used to supply the heater. Merely leaving the system as it is and hoping for the best is clearly not satisfactory in the majority of cases. Such a closedloop system is shown in Fig. It would be much more sensible to measure the temperature of the heated water such that the on/off information can be varied appropriately in order to keep the output at approximately the temperature desired. hence the system is n o longer openloop. Reasons for this could be a change in ambient conditions. 1. 1. By means of the negative feedback loop shown in Fig.5. 1.e.6 Introduction Input Controller Input (On/Off information) (water) Input (electrical energy) *I Fig. 1.4 Closedloop control In a closedloop system the control input is affected by the system output. by providing feedback from the system output to an input.4 Water heating device Heater l z * l storage Output (heated water) i.5 (negative because the system output is subtracted from the reference input) the accuracy of the system output in relation to a desired value can be much improved when compared to the response of an openloop system.
any system variations or effects caused by outside influences can seriously degrade this accuracy. Unfortunately the advantages obtained when feedback is employed are at the expense of system stability (oscillations can often be induced by increasing system gain). the speed of rotation. where oscillations can occur in the system output. The openloop heating system of Fig. 7 System. even if this means swinging the output first one way and then the other. With the system in closedloop however.Closedloop control Reference input ? L Control or Error Controller . 1. i. as are effects due to disturbances or nonlinear behavior. Once feedback is applied.zing 1 i control System output fed back .4. Closedloop systems can achieve much greater accuracy than openloop systems. Effects of system variations or outside influences are much reduced.e. A disadvantage of feedback is related to the flyball governor of Fig. Let us assume that we require the water temperature to be 50°C. Although only common in relatively simple systems.3. The oscillations are due to the attempt to get the error signal to as low a magnitude as possible. output Fig. with feedback the system is less sensitive to variations in conditions. 1. the effect on the system output will be much reduced because of the feedback arrangement.. a variation in system characteristics would merely cause the temperature to drift away from the required value.1 Effects of feedback Openloop systems rely entirely on calibration in order to perform with good accuracy. although they rely entirely on the accuracy of the comparison between desired and actual output values and therefore on the accuracy of the measured output value. 1. With the system in openloop. the system is in closedloop. . which would not occur if the system were connected in openloop mode.5 Closedloop system to minimize the error between the actual system output and the desired (reference input) value.4 can be converted into a closedloop system by measuring the temperature of the heated water (output) and feeding this measurement back such that it affects the controller by means of modifying the switch on/off information. any variation from 50°C will be seen as an error when the actual measured temperature is compared with the required value. and the controller can then modify the control input in order to reduce the error to zero. an important property is found in that an increase in system gain (amplification) does not affect the stability of an openloop system (oscillations cannot be induced in this way). It is then apparent that when variations in performance occur due to system modifications or a change in ambient conditions. 1.
certain automatic electric trains fall into this category along with autonomous guided vehicles (AGVs) which are not restricted t o operation along certain lines. comparing this with the desired value and then modifying the control input accordingly. examples being driving a n automobile and cooking. by means of computational procedures either operating in situ or from a remote station. Many control systems have been automated by replacing the human feedback element with more accurate equipment which also responds more rapidly. where the human can merely form the feedback loop . is used to ensure constant velocity and constant heading. both ship docking and aircruff landing can also be automatically carried out. It would not be sensible t o include the effect of high winds when modeling the act of drinking a cup of coffee because the vast majority of cups of coffee are not drunk in the presence of high winds.5 Introduction Some examples of control systems Many simple control systems in everyday life include a human being. Examples of this are ship steering and flight control. Note that outside influences such as windy weather. Taken further. This is really an example of a biological control system. neglecting the effect of rough weather or the presence of other craft could be fatal in this instance. not only in the feedback loop but also to provide the actuating signals and even as the control system itself.measuring the actual output. This is an important initial point because when modeling a control system the more characteristics that are included in the model.e. conversely lifting it too slowly will result in severe boredom and even death through lack of liquid intake. i. such as rails. These latter examples serve as particular reminders that many different characteristics must be taken into account when modeling a system. Although all of these would affect the performance in question. such as roll detectors and gyros respectively. being part of the human being. Feedback is provided via touch and sight. Performance is measured in terms of how quickly we can lift the cup . which involves lifting a cup through a desired height. An actuating signal is then provided in terms of lifting the cup with our hand at a rate which is dependent on how large the error signal is at any time. With both visual .lifting it too quickly or going past the desired end position will result in coffee being spilt and possibly several broken teeth. they were considered to be outside the scope of the system definition in the first instance. except the cup and coffee. Often the human being merely forms part of a n overall control system. Consider drinking a cup of coffee. a lead weight in the bottom of the cup o r the temperature of the coffee have not been considered. the more complex that model becomes. by our mouth . A tradeoff therefore has to be made in terms of model complexity and the ability of that model t o account for all eventualities. Any change in either of these requirements can be fed in merely as a change in setpoint value.8 1. for which pilots are replaced by autopilots whereby information from sensors. with a comparison being made between where the cup actually is at any time and where we want it to be. Other transit systems also employ automatic control techniques.this is the error signal. everything within the system.
radar tracking and missile control. Further down on the electrical power scale. In manufacturing. Controllers are required in order to minimize the effect of disturbances on the output signal. concentration) in many cases results in millions of dollars being saved. to machine tools which can modify their procedure in response to varying conditions such as differing workpiece requirements. Nuclear reactor control is concerned with the rate at which fuel is fed into the reactor and it is in areas such as this that the effects of hazardous environment must also be taken into account. Rollers are often not perfectly cylindrical and. hence tight limits are necessary. voltage stabilizers are employed to retain the output voltage from a source as close as possible to a reference voltage value.Some examples of control systems 9 (camera) and distance (laser and sonar) information it is now possible for these vehicles to ‘drive’ themselves and it should not be long before greater use is made of such techniques in automobiles. Letting the thickness increase merely wastes money whereas letting it decrease produces an inferior product. but each individual workstation can consist of a robot manipulator which either merely performs a set routine repeatedly with good accuracy and speed or which modifies its actions due to visual and/or tactile information feedback. Uniformity in the thickness of such as paper. controlling output thickness within tight limits is not a simple task. automated production lines are not only supervised by an overall automatic controller. On a similar basis the control of machine tools is gradually increasing in complexity from simple profile copying mechanisms. Stricter controls imposed due to environmental pollution have meant not only increased efficiency requirements in the conversion schemes themselves but also the use of automatic control devices to control effluence. any difference (error) then being used to vary the resistance provided by a transistor which is connected in series with the source. This is done by comparing the actual voltage output with the reference. Indeed other systems concerned with power generation and distribution share many instrumentation difficulties. through numerical control for exact coordination. all of which involve rapid tracking of a target. combining this with possible large variations in the product quality (e. In oil and chemical industries a small percentage improvement in reducing output variations (e.g. pulp obtained from a different factory). Process plants are usually characterized as fairly slowly varying systems with only occasional changes in reference input. with minimum error. flow rate. as indeed it is for future developments in gun positioning. Power generation is also an important area in the application of automatic control systems. The military interest in such vehicles is important for their development. Biochemical control and biomedical control are in most cases closely related to process . Certainly the large number of variables which need to be measured in order to provide a useful picture of the system presents many problems in terms of accuracy and reliability. steel and glass is based on the control of the motor speed of rollers and drawing machines. The control of motors both in terms of position and speed is however a much wider field than mentioned here. large amounts have been spent in providing accurate automatic systems for process control. In terms of financial outlay. the complexity of control being dependent on the accuracy requirements made.g.
A list of definitions is therefore given here in order to help remove any barriers which do exist (see Fig. It can.6). It should be remembered that many of the simple techniques considered remain as simple techniques even when applied to the most complicated plant. is the signal obtained from the plant which we wish to measure and control. an example being the price of shares on the stock market. The system output. and the latter respiratory monitoring and pain relief. 2 . Hopefully a good idea of the wide variety of system types in which automatic control schemes operate has been given. speed. Obvious advantages can be obtained by successfully modeling economic systems in order to obtain suitable automatic control schemes. and are functions of time. The reference input.g. u. 1. Note: s is the Laplace operator and u = 27rf where f is frequency. When one is new to the subject. Apart from the most simple cases however it is extremely difficult to account for many of the spurious events that can occur in practice and which have considerable effect on the economic system under control.10 Introduction control systems. y . or otherwise Laplace transformed quantities. G . chemical plant. electronic amplifiers. however. 1. 5 . mechanical systems and electrical machines. as in the case of u ( t ) = U cos w t . voltage. 3. Social systems are similar to economic systems in terms of modeling difficulties and also suffer from the fact that very often there are both a large number of outputs and also different measures of performance which are often contradictory. be seen as rather vague and in many cases confusing from the outsider’s point of view. Improved automobile engine performance and efficiency has become more and more dependent on complex control schemes. e. also known as the controlled output. Capital letters denote signal magnitudes. is an external signal applied in order to indicate a desired steady value for the plant output. also known as the setpoint or desired output. The system under control is also known as the plant or process.6 Definitions of standard terminology As with most subject areas. .this is indicated by W u ) . Lower case letters refer to signals. terminology is used in the study of control systems which aptly describes phenomena within the field. Automatic control has become increasingly more important in agricultural engineering. 4.g. u = u(t). with power generated by means of the engine itself being used to provide energy for the electronic/microcomputerbased control circuitry. with the former including such examples as drug production and fermentation.Where s = ju. electromechanical systems. 1. with intelligent tractors now able to cope with wheel slip and to provide a constant ploughing level despite surface variations. U = U ( s ) . e. examples encountered are often related to relatively simple applications with which the reader may be familiar.
15. The control input. 14. where a period T (sample period) is defined separately for each system. 16. Note: Often H = 1. 7. Almost all except the most simple systems are nonlinear in practice. e. A continuoustime system is one in which the signals are all functions of time t. also known as the actuating signal. in an analog sense. to an extent at least. A timeinvariant system is one in which the characteristics of that system do not vary with respect to time. it is no longer the error between u and y . 11.Definitions of standard terminology 11 Fig. D . A discretetime system is a system such as a digital system or a sampled data system in which the signals. 8. 1. A disturbance. In many cases it takes the error signal as its input and provides an actuating signal as its output. The error signal. at least over a specific range. 17. amplifier saturation at the extremes. e. i.g. 13. e. 12. u. 1. The controller. and includes D and G. . The operator z is used to define a discretetime signal such that z 3 y ( t )= y ( t + 3 ) means the value of signal y ( t ) at a point in time three periods in the future. The feedback element H provides a multiplying factor on the output y before a comparison is made with the reference input u. The feedback signal is the signal produced by the operation of H on the output y . A nonlinear system is one in which the principles of superposition do not apply. or noise (not shown in Fig. ageing. is the element which ensures that the appropriate control signal is applied to the plant.6 Closedloop system with feedback element H 6. is the difference between the desired system output and the actual system output (when H = 1). or hysteresis effects. is applied to the plant G and is provided by the controller D operating on the error e. Note: See 8.6). The feedback path is the path from the output y . although H can represent the characteristics of the measurement device in which case most likely H # 1. The vast majority of systems can however be dealt with by approximating the system with a linear model.e. is a signal which enters the system at a point other than the reference input and has the effect of undermining the normal system operation. Most systems do vary slowly with respect to time. which consist of pulses. 10. 9.g. The forwardpath is the path from the error signal e to the output y . control action or control signal. When H # 1 the error e is the error between u and Hy. only have values at distinct time instants. through H. however over a short period they can be considered to be timeinvariant.
1. These diagrams are described in much greater detail in the following chapter. e. pressure to voltage. in terms of applications. In this text only single variable systems are considered. A multivariable system is one which consists of several inputs and several outputs. The ideas put forward should not be seen as limiting or restrictive in any way. In the chapters which follow many techniques are shown for the modeling. To do this it has been necessary to employ representative diagrams to show connections between signals and systems. 19. about a setpoint mean value. A transducer converts one form of energy (signal) into another. Positive feedback is obtained when e = u + Hy.7 Summary In this first chapter the purpose has been to gently introduce the subject area of control systems. but rather as an insight to the wide variety of possibilities. 1. Control engineers therefore come from many different backgrounds in terms of subjects studied and an even more diverse set exists in terms of areas in which a control engineer can show a certain amount of understanding. such variations being caused by disturbances. by providing a brief historical account of developments leading to its present state and also by putting forward the idea of feedback control in terms of some elementary forms. 22. If the water is heated from cold. Note: A regulator differs from a servomechanism in which the main purpose is to track a changeable reference input.6. they are generally applicable in many other fields. control and study of system behavior. Note: This is not shown in Fig. A regulator is a control system in which the control objective is to minimize the variations in output signal. Why is a simple openloop system not used instead? Explain the openloop operation of traffic signals at a road crossing. Once the fundamentals of a particular control technique have been studied in one field. 21. How can improved traffic control be achieved by means of a closedloop scheme? .Hy.1 1. with the aid of a sketch of water temperature versus time. Negative feedback is obtained when e = u .12 Introduction 18. ~ ~ Problems 1. the principle of operation of such a device. 20. explain.2 A thermostat is often used in order to control the temperature of a hot water tank. Many examples of the application of control systems were given in an attempt to show the multidisciplinary nature of the subject.g.
in many cases. How can a closedloop technique be devised in order to track a target and what problems can this cause? Consider the potential divider in Fig. 1. Meanwhile. once the missile has been launched. 1.a constant K Z multiplying the error between a number P and F. and use the rabbits as a food source. while the number of newborn young which survive to the following year is twice the number of adult females alive throughout the present year. How could a closedloop system be constructed for the manipulator? In a liquid level system. Demand for the toy will decrease if the price of the toy increases.8 1. By finding the difference between supply and demand. whereas supply of the toy increases if its price increases. Foxes move into the field inhabited by the rabbits of Problem 1. adult females and young rabbits. Consider the simplified economic system describing the price of a toy.6 1.and closedloop system at the same time? The population of rabbits in a particular field is unaffected by events external to the field.4 1. The population consists of adult males. F is the number of foxes at any one time and R is the number of rabbits. this is operated in openloop mode. Explain how.3 1.7.5. show how this can be used as an error signal to effect a controlling action on the price. Each year 40% of the young become adult males the following year and 40% of the young become adult females the following year.7 Consider a missile guidance system which is based on an openloop scheme. Each year 20% of adult males and females die. the remaining 20% of the young die. The rate of change of foxes is equal to a constant K I multiplying the error between R and a number i. including visiting rabbits. what problems could occur if an openloop scheme is employed? Can all of these problems be removed by employment of a . The output voltage is VZand the input voltage is V I . How can the system be both an open.5 1. whereas the rate of change of ? rabbits is equal to . 1. what happens to F and R with respect to time? Consider the end effector of a robot manipulator.9 How does the system operate.7 1.Problems 13 1 A Fig. Find equations to describe each section of the population from one year to the next.
218.. . Enos.).the history of an idea’. Dorf. Peter Peregrinus Ltd.. ‘Stabjlity of a given state of motion’.. New York. ‘Control problems’.. Black. Stroudsberg.). Peter Peregrinus Ltd. Dowden. Adams Prize Essay.. ‘Aircraft flight control’. H o w could t h e vehicle driving scheme b e configured without a h u m a n operator? A r e there a n y particular advantages o r disadvantages t o such a technique? Further reading Bennett. Nyquist. H. 232. close your eyes a n d ears a n d drive (maybe you drive this way already!). Control of manufacturing processes. J. 12647. S. ‘Regeneration theory’. Bell System Technical Journal. 1922. in Selectedpapers on mathematical trends in control theory. J.. Pergamon Press.. Minorsky. (ed. 27083. (ed. pp. Minorsky. 10623. Origins of feedback control.. 1977. Lighthill. 1991. Bode.10 closedloop technique? A r e a n y problems brought a b o u t d u e to t h e closedloop format itself? Consider driving a vehicle without using a n y feedback. Am. 1964. J. 18001930. The dynamics of automatic control systems. 1877. and Malkin. H. J. . L.. E. 1962. Tzafestas. N... 1986. K. H. A history of control engineering. (eds. 1934. Tischler. 11016. M.. G. Naval Eng. 1959. J. S.. 1971. R. C. ‘Feedback . ‘Theory of servomechanisms’. Marcel Dekker. 1978. pp. 1941. p. pp. 1987. P. H. Control (Special Issue). which includes a h u m a n operator? (obvious answers). 1 Popov. Control theory: a guided tour. Routh. Cambridge University Press. Automatic control: classical linear theory. H. p.. R. W h a t advantages a r e obtained if a closedloop vehicle driving scheme is used. ‘A brief history of electrohydraulic servo mechanisms’.g. J. ‘Inventing the negative feedback amplifier’. W. IEEE Spectrum. ‘Directional stability of automatically steered bodies’. N.). W. 1992. Huchinson and Ross Inc. 59.. G. ASME J . J . S .. Mathematical SOC. e. Maskrey. Int. AddisonWesley. R . pp.. Encyclopedia of systems and control.. AddisonWesley. J. 1 . M. 1994. Dover. J.S. Modern control systems. Leigh.14 Introduction 1.. Franklin Institute. of Dynamic System. M. Macmillan. J. 280. 1993. Mayr. 451. Franklin Inst. Maxwell.. MIT Press. Danai. 5560.). ASME. Proceedings of the Royal Society (London). 1992. Pa. M.. Fourier anafysis and generalisedfunctions. pp. Singh.. E. C. O. Thaler. SOC. 6th ed. 1974. 1932. Measurement and Control. and Thayer. 1868. (ed. pp. Applied control. 1993. Dynamics and control of mechanical systems. ‘On governors’. 54380. No. . 16. J. 34. B. 1. Hazen.
it must be remembered that once the equations have been obtained. it is possible in the vast majority of cases to linearize the equations in the first instance such .System representations 2. as will be seen in the proceeding chapters. especially if this means an extremely costly design exercise due to the time taken because of the complexity. will most likely result in a large number of equations. Therefore a certain amount of common sense and practical experience is necessary when forming a mathematical model of a physical system in order to decide which characteristics of the system are important within the confines of the particular setup in which the system must operate. they will be used for analysis purposes and hence a complicated collection of identities and equalities will not generally be widely acceptable. the most likely occurrence being that a physical system is already in existence and it is required that a controller be obtained for this system. no matter how simple the system might at first appear to be. and this can be quite a complicated procedure. In some cases the nonlinear characteristics are so important that they must be dealt with directly. However. although for physical systems the equations will in general be nonlinear to a certain extent. Conversely the model describing a physical system should not be over simple so that important properties of the system are not included. Further. The systems considered in this text are inherently dynamic in nature. Completely describing.1 Introduction In order to analyze a system in terms of controller design a basic assumption is made. the performance and operation of a physical system. that we already have in our possession an accurate mathematical model of the system. especially if it is attempted to account for all possible eventualities. which means that differential equations are an appropriate form in which their characteristics can be described. It is not usually the case however that a mathematical model is employed as a starting point from which a physical system is constructed. in a mathematical sense. something that would lead to an incorrect analysis or an inadequate controller design. Forming a mathematical model which represents the characteristics of a physical system is therefore crucially important as far as the further analysis of that system is concerned.
by thinking of the system's operation in terms of a transfer function. that the input to output transfer function is found to be: (D" + (In. It is important to remember. u ( t ) . both electrical and mechanical systems are considered in much greater depth in order to show the relationship between initial system networks and the transfer functions which are obtained as a result. and hence their transfer function which relates input to output must take on a more complex form. i. as is done in Section 2. If an input is applied to the function. for a given input our model must tell us what output will occur. in general terms. An alternative form of representing system characteristics in terms of a more visual approach is then discussed with the introduction of block diagrams in Section 2.l D"'+ + 61 D + b o ) u ( t ) (2. we must produce a function which operates on an input in such a way as to arrive at the correct output. In forming a model we need to supply an answer to the question 'What will be the system response to the particular stimulus?'. 2. it will transfer us to the corresponding output value such a function is therefore called a transfer function. y(t). so the transfer function of this system is simply 1/50. and system output. how linear approximations can be made.* a = (6. though. This could mean. as long as the input and output are scaled in cents and bars respectively. to nonlinear system description equations. that any design carried out on the basis of a linearized model is only valid within the range of operation over which the linearization was carried out.2 of this chapter. It is often the case that the relationship between system input. and this must correspond with that of the actual system if the same input were applied. of linear differential equations. which may be only one. over a set operating range. If we are modeling the system in a mathematical way.3. In its simplest form a transfer function can be simply a multiplying operator. Finally.e.1) . Unfortunately many systems are rather more complicated than the simple chocolate bar machine example.16 System representations that subsequent analysis is much simplified. If we insert 50 cents (the input) into an appropriate slot on the machine we will be rewarded with one chocolate bar (the output).2 Transfer functions In order to analyze a control system in any reasonable depth it is necessary to form a model of the system either by obtaining response information from a system which is already in existence or by combining the information from individual components to assess the overall system obtained by bringing those components together. It is shown in the same section. consider a chocolate bar machine for example. Once a physical system has been modeled by a finite number.2. Let us assume that from the machine in question each chocolate bar costs 50 cents. D'" + b. by means of two examples.1 D"' + + a1 D + ao)y(t) . can be written in the form of a differential equation. it is possible to invoke the Laplace transform method with a consequent ease of analysis.
1) by the Laplace operator s. .2. are real values which are defined by the system characteristics. and that the coefficients a and bi are timeinvariant.. ..and is therefore not the standard technique employed for the manipulation of transfer functions in order to find a solution to the output response. i .2. as the ratio of the Laplace operator polynomials with initial conditions zero.Transfer functions in which 2 d2 D y ( t ) = 7 ( t ) .. As long as the integers n and m are known. this is not a simple problem. .. where the property of linearity is defined in Section 2.2. along with all of the coefficients ai: i = 0. and many computer packages are available for such cases.1) in order to find an exact solution.1. n . m .2. As far as the rest of this text is concerned a transfer function will be regarded..2. The effect of this is simply to replace the D operator in (2. However.. in its most general form. whereas the differential orders n and m are linked directly with the system structure.. an1 and bo.. with respect to a set of input and input derivative values for the same t > to.then for a given set of initial values at t = to. y dt 17 Also the coefficients ao. etc.. and this results in an input to output transfer function: in which it must be noted that both the input and output signals are also transformed values.3) where it can be seen that G(s). Equation (2.which itself is independent of both the input and output signals. the output response for any t > to can be calculated by finding the solution to the differential equation (2..3. forms the function which transfers the input U ( s ) to the output Y ( s ) .b.2) can be written as Y ( s )= G ( s ) U ( s ) (2. especially for large values of n and m...2. I). Having said that.2. A much simpler approach to the analysis of a system transfer function is to take the Laplace transform of (2. for specific cases it may well be necessary to revert to a differential equation description (2.. such that Also it will be considered that G ( s )is linear. and bi: i = 0.1) and to assume that all initial signal values at time t = to are zero.
C.18 System representations 2.in terms of its transform. we can initially put t o one side the output voltage and find the relationship between v i ( t ) and i ( t ) if it is remembered that by the voltage law of Kirchhoff.1 Series RLC circuit with input v i ( f ) L w R a . the voltage input. to output voltage. and also that until t = 0 the capacitor. both sides of (2.e. Ui(t) = L ~ +$1: i ( t ' ) d t ' + R i ( t ) (2.5) where t' is the time. this becomes CSVi(S) = (LCS2 + RCs + l)I(s) vi ( S ) (2.1 Transfer functions . ( t ) in terms of the Laplace operator.1. u i ( t ) must be equal to the sum of the voltage drops across the inductor. The problem is then to find the transfer function relating input voltage. remains uncharged.6) and on taking Laplace transforms of both sides. 2. In order to remove the integral term. which is also the output voltage. V o ( s ) .2. U i ( t ) .2. considered over the interval 0 < t' < t. 2. In order to find the solution. i ( t ) . u .worked example Consider the series RLC circuit shown in Fig.2. The voltage across the capacitor C .5) can be differentiated to achieve Dui(t)= L D 2 + R D +  ( 7 c i(t) (2. is given S 0T  R Fig.2.7) or I ( s )= LCS* + RCs + 1 c s We have therefore obtained a transfer function which relates the input voltage signal transform Vi(s) to the current transform I(s). assuming that no current. A further step is therefore required to find an equation (transfer function) which relates the current to the output voltage transform.2. for any t 2 to. capacitor and resistor. i. flows until the switch S is closed at time t = to = 0.
7) and (2. Applying an impulse to a system is one of the oldest forms of analysis. o r if a signal lasts for a small time period.2. the output can be found. as will be shown here. in terms of the impulse response and the applied input.10) In this example one of the basic properties of transfer functions has been shown.4) is related t o a time domain solution.2. If any other input signal is applied. In fact the approach taken is to define a unit impulse function as being one which lasts for a + + .e.7).8) D V o ( t )= . As a n example consider the chocolate bar machine mentioned earlier: if the machine was to fail. Actually applying an impulse as introduced. then the output.10).r(t) C with a Laplace transform V o ( s ) . if the input is applied at t = 0. the resultant transfer function from input to output voltage is thus (2. then as 6 f ' 0 so the signal becomes a n impulse.2 The impulse response Having discussed the use of a transfer function to describe the response of a system to an input signal. in this section it is shown how the Laplace transformed function of (2. 6 t ' . we hit the machine as hard as we can. then we apply an impulse as a means of fault diagnosis.2.to a system.2.9) in order to obtain the final form of (2. namely the multiplicative property exhibited by the combination of the transfer functions (2. will also be the impulse response. i.9) and on substitution of I(s) from (2. g ( t ) : where for all causal systems (all physical systems) g(t) = 0 for all t < 0. possibly by not presenting us with a chocolate bar after our money is inserted. is not a realistic idea.Transfer functions 19 as v . The output from a system which is obtained in response to a n impulse applied at the input is then known as the impulse response. Strictly speaking a n impulse is a signal which lasts for a time period equal to approximately zero. u ( t ) . 2.2. y ( t ) .2.2.2. ( f )= or c !' o 1 . with 6 f ' 0 .2.I(s) = cs 1 (2. in particular the response to an impulse input is investigated. i ( f ' )d t ' (2. If a n impulse is applied as the input.
e. etc. But.2.2. 2. the output at time t will in fact be equal to the summation of all of the output parts due to each of the input pulses up to t . thus resulting in the ( t . 2. impulse strength multiplied by the impulse response.20 System representafions time period 61' and has a magnitude equal to unity. the output found in terms solely of the second pulse is: y ( t ) = lim ( u ( 6 t ' ) 6 t ' 61' 0  g(t  where. u ( f ' ) .2 Input signals as a set of impulse functions . in 0 +I 6r'p t kbt ' t f = nbt' t ' =time Fig.e. i. the first pulse.e. The unit impulse function is said to have a n impulse strength equal to unity such that a n impulse function of magnitude 5 and time period 6t' would have a n impulse strength of 5 . for this second pulse the impulse strength is given as the pulse magnitude u ( 6 t ' ) multiplied by its time period 6 1 ' . The continuous time signai u ( t ' ) is in Fig.. for the pulse at t' = 0. as follows. This basic technique can be extended to account for a general input. approximated by a series of pulses. for the signal shown in Fig. I t is not difficult to see therefore that if the impulse response is the output of a linear system in response to a unit impulse function then the system output in response t o a 5unit impulse function will be 5 x the impulse response. for the purpose of our analysis. 2. the output due to the first pulse must be added to the output due to the second pulse.g. the impulse response will only have effect from time 6t' t o time t . so + y ( t ) = lim ( u ( 0 ) 6 t ' 61' 0  g(t)/O < t' < 2 6 t ' j 6 t ' ) / 6 t ' 6 t ' 6 26t'j Similarly. In terms of the pulse approximation to ~ ( t ' )the output y ( t ) can be calculated as the .shown in Fig. in which the fixed time t is a n integer number n of the time periods 6t' and the input is assumed to be zero for all t' < 0. Also in respect of the output at time instant t .2. t = 3 seconds. The impulse response of a linear system is then generally taken t o be the response of the system to a unit impulse function applied at the input. Consider the input signal. i. such that as the pulse width 6t' * 0 so (a) the approximation u ( t ' ) and (b) the pulses become impulses. 2.s t ' ) indexing. The general time scale is given as t ' and the time instant t is considered to be a certain point along this scale.
+ u(t . In terms of a n integral this becomes y ( t )= i‘ 0 u(t‘)g(t .2. with .k 6t’) + * .12).2.2. i. however. In other (2..2.2.6 t ’ ) Y(s)= G ( s ) U ( S ) 6t‘ * g ( 6 t ’ )+ u ( t ) * 6t‘ * g(0)) (2.12) which is known as the convolution integral. and in what way.12) and (2. The problem of approximating a nonlinear system with a linear model will be discussed in the next section.3) If now Laplace transforms are taken on both sides of (2.11) which are: y ( t ) = lim I u ( 0 ) 6t‘ 6/’ 0  * g(t) + u ( 6 t ’ ) .2.2.t ’ ) dt’ (2. A basic assumption was made earlier that the system G ( s )is linear and timeinvariant. written as Y ( t )= u ( t ) * s ( t ) (2. both continuous time signals. The means of dealing with a timevarying transfer function. it is merely another way of writing (2. It can in fact be observed quite simply by reference t o the first two and last two terms in (2.6t’ * g(t .11) where t = n6t‘.14) This does not imply that the input function and the impulse response change roles. So the transform relationship between system input and output has been obtained by starting from an impulse function input and impulse response.14) we have in which G(s) is the Laplace transform of g ( t ) and U ( s )is the Laplace transform of the input signal u ( t ) . and the results of this section have been derived on that understanding.2. I t is however the case for many systems that neither of these assumptions hold.13) mean exactly the same thing. are very much dependent on just how rapidly the transfer function is changing.. In particular the transfer function has been shown t o be the Laplace transform of the impulse response.e. One property of the convolution integral is the interchangeability of the integrands.13) such that (2.Transfer functions order t o find the output at time t due to the input from time 1’ = 0 to I ‘ words the actual output at time t is found t o be: = 21 t . we also have which means that y ( t )= 1‘ 0 g(t’)u(t.2.t ’ ) dt‘ (2.2.
If the system transfer function varies slowly with respect to time then it will most likely be necessary to redesign the controller every so often. can appear t o be large in terms of component and signal values. after a while one of the components would malfunction. but as the time variations become more rapid so a more frequent redesign may well be necessary. However.where b1 and bz are constant multipliers. If an input u 1( t )to a particular system produces a n output yl ( / ) and if a n input u z ( t ) to the same system produces an output y z ( t ) .2. This then leads on to the idea of an adaptive controller which periodically updates its transfer function representation in order to carry out a controller redesign. In the previous section it was shown that a system input to output relationship could be given in terms of the convolution integral (2. where with the switch closed.1. the component might be permanently damaged and it may well not be possible to return to the initial state. i.3 Linear approximations All physical systems can exhibit nonlinear behavior. a technique which is particularly suited to a computer control framework.1 ’ ) = 0 for t < t ’ .12). some examples of which are given in the following sections. We have therefore for a physical system that it will be in some sense nonlinear. it would no longer operate in its normal mode.22 System representations respect to time. Consider for example the series RLC circuit of Fig. for other system types it may well be necessary to consider . i. if the input voltage v i ( t ) is continually increased. Both of these properties are necessary to define a linear system and when combined mean that a n input b l u l ( t ) + b Z u 2 ( f ) .2. ) 2. 2.t ’ ) u ( / ’ ) d r ’ where it has been assumed that (a) u ( t ’ ) = 0 for t‘ < 0 and (b) g ( t .e. This may be due to overheating such that when the input voltage is reduced. One thing to remember is that any controller design carried out which is based on a system transfer function depends primarily on the accuracy of the transfer function coefficients. y(t)= l‘ 0 g(/’ . normal operation may be resumed.e. in most cases this is simply due to a system variable o r signal being increased well beyond its normal scope of operation. but that it will very likely have a certain range of operation over which it can be considered to be linear. Although the linear range of many electrical and mechanical systems. Homogeneity: If an input u l ( t )to a particular system produces an output y l ( / ) . will produce an output b l y l ( t ) + b2yz(t). But what d o we mean when we say a system is linear? A system is said to be linear if it has the t w o following properties: I . where bl is a constant multiplier will produce a n output b l y l ( t ) . then in order for the system to be linear it is necessary that an input blul ( t ) .then in order for the system to be linear it is necessary that an input u l ( t )+ uz(1) will produce a n output y ~ ( t+ y z ( t ) . 2.
2. Consider 0 t o be the system input and dZ8/dt2the system output.4 Linear approximations . depending on the frequency of the input signal.Transfer functions 23 the linear range as existing over only a small set of signal values.2.worked examples The first example of approximating a nonlinear mechanical system by a linear model is the pendulum oscillator system.2. If the sine function can be approximated by a linear function then the equation of motion will be ‘linearized’. 2.2.16): (2. and indeed this is a possibility for small values of 0 about the midposition 8 = 0 . indeed fluid and heating systems fall into this category.3 Pendulum oscillator system . 2. which is usually analyzed in terms of several different linear models.15) in which g is the acceleration due t o gravity. as is true for midfrequencies. It is then apparent that the equation of motion is nonlinear as it breaks the linear property rules set out in the previous section because of the sine function. the nonlinear equation Fig. 2. The equation of motion for the pendulum in terms of the displacement angle is (2.3. which can be seen by truncating the Taylor series expansion t o its first term only. depicted in Fig. An excellent example of a nonlinear system which can be considered t o be linear over a small signal range is the transistor amplifier. The assumption can be made that for small values of 8 we have 8 = sin 8. as shown in (2. whereas for high frequencies a different model is employed. So for a low frequency signal one linear amplifier model is used. Two examples of linearly approximating mechanical systems are detailed in the next section.16) It follows that by using the truncated Taylor series.
4 Massspring system (2. if the mass displacement is restricted to lie within the range I . applied by the spring.5.17) for small values of 0.4.17) dt"= 7 d20 g0 (2.15) becomes the linear equation (2.2.2.18) to be a nonlinear equation. 2. So the nonlinear pendulum oscillator system has been approximated by a linear system over a small range for the input signal 0.2.5 Plot of spring force against displacement . 2.18) where f ( x ) is the force. 2. as a function of displacement.2.24 System representations 0 Fig. The equation of motion for the massspring system in terms of the displacement x is: M7 dt d2X = f(x) (2. As a second example consider the massspring system of Fig 2. Although the general effect of the force f ( x ) is to cause (2.2. as shown in Fig.XI I I * x Fig.
by considering the input signal to exist within a relatively small range the system response can be effectively described by means of a linear equation.3. 2. alternative forms of connection can be considered.2. When other systems and/or controllers are connected on to the original system it is then quite straightforward to consider each separate system in terms of its own block diagram. It is one advantage of the block diagram form of system representation that. diagrams XI 25 < x < XI then the term f ( x ) in (2.2 is represented in block diagram form as Fig. as soon as that system is connected up in some fashion to another system.2 the mathematical description of systems in terms of a transfer function representation was considered.1 Basic building blocks The transfer function system description introduced in Section 2.k x dt (2. Hence the block diagram in Fig. they are just alternative representations. it rapidly becomes a difficult problem to relate various equations to the underlying physical systems that they represent. by means of block manipulation. .2.2. dependent on just how the connections were made.Block. Both of the examples given in this section have shown that. While dealing with a particular system in its simplest form there is really no need for any other alternative representation.6 and the equation Y ( s )= G ( s ) U ( s ) mean exactly the same thing. then the overall transfer function will become that due to the combined functions.6. However. although the system under examination is inherently nonlinear..3 Block diagrams In Section 2.XI < x < X I . When a set of systems are connected it is quite possible that the overall relationship produced could have been achieved either by connecting the same system up in a different way and/or by connecting some slightly different systems.18) can be replaced by kx where k is d2x M y = . 2. In this section the basic ingredients for block diagram system representation are introduced and subsequently methods for block manipulation are discussed..18) is then as long as the displacement lies in the range . The result of this is that it is often best to think of the system concerned in terms of a graphical block diagram representation. So the linearized version of (2. 2.19) some constant value. 2. Although it is perfectIy acceptable to think of the system connections in terms of solely algebraic relationships. or a controlling function is applied to the system.
but a different takeoff point has been employed.8 where the (s) indexing is dropped t o aid explanation.sign to indicate whether they are to be added o r subtracted.8. 2. In terms of an overall block diagram Fig. each being given a + or . 2.7. 2.26 System representations U(S) F]*  Y(S) Fig. 2. can have any number of signals entering.9 can be examined. Laplace functions are assumed. 2. If a takeoff point is introduced into the cascading example shown in Fig.2 Block diagram manipulation If two blocks are connected together in cascade.8 Blocks in cascade . as shown in Fig. which is sometimes shown with the symbol C inside the circle. The summing point.6 Block diagram basic transfer function Takeoff point Summing point Fig. There is only one resultant signal from a summing point. as shown in Fig. The only other two basic elements of a block diagram representation are the summing point and the takeoff point. 2. the block containing the operator G ( s ) . In each of the cases shown in Fig. they can be replaced by a single block.6 that the signals U ( s ) and Y ( s ) are shown in terms of a direction given by the arrows. 2. 2. then the equivalence of the three representations shown in Fig. the function within which is a multiple of the two original block functions.e.9 we have that (a) Y = G IGz V and (b) E = GI V .3. 2. and if this signal must appear in a number of places then an appropriate number of takeoff points can be used.7 Summing and takeoff points It can be seen in Fig. 2. i.
3. 2.9 Moving a takeoff point any of the three representations can be used. when the output is fed back by means of a takeoff point.1 1 . From the original (lefthand) negative feedback block diagram we have two equations. The one remaining basic operation to be considered is the method of dealing with the use of feedback.H Y On substitution for U into the first of these equations we have our final closedloop transfer block given by Y = GV. one which describes the openloop system transfer function Y = GU and the other which describes the system input as consisting of an applied input V summed with a signal fed back from the output U = V.e. the decision being dependent on whichever seems most appropriate for a particular purpose. where in each case the output is found to be Y = GIG2 V + G IG3E.G H Y or y= 1+GH G v (2. to be one of the inputs as well. 2. as shown in Fig.10. along with its equivalent single block representation.1) . whether this be in terms of an algebraic or a physical sense.Block diagrams 27 E d E + Fig. i. 2. A negative feedback loop is shown in Fig. The same is true for a summing point.
3.3 Block diagrams . The first is t o show how the basic ideas of block manipulation can be employed one set at a time to deal with systems which initially appear quite complex. reduce even the most complicated control system either algebraically in terms of the component transfer function or in terms solely of block diagrams.11 Negative feedback loop With the basic elements as described in the previous section and the block manipulations discussed in this section it is possible t o analyze.worked examples Two examples of block diagram manipulation are given in this section. . The second example is included as a reminder that the transfer functions contained within each block can often contain factors which may cancel out in a later part of the analysis.28 System representations El E t E I E Fig. where necessary.10 Moving a summing point Fig. 2. 2. modify and. 2.
2.15. 2. consider the block diagram shown in Fig.13 becomes Fig.12. as described in the previous section.16 in which the G3 block has been cascaded. The feedback loop can be reduced. For the second example in this section. All that remains is to cascade G I and then to add on G in order to obtain the 4 final block diagram as in Fig.Block diagrams 29 H i Consider the block diagram shown in Fig. As a first step towards diagram reduction the Gt block can be moved onto the other side of the leftmost summing junction. It is required to reduce this diagram into a single block with input V and output Y . so Fig.19. 2. The (s + 2) feedback numerator factor can be taken over to the other side of the summing junction to give Fig. 2.18. note that K is some constant gain value.14(b) and on reduction of the positive feedback loop we have Fig. 2. 2. 2.14(a) or Fig. Now we can investigate the negative feedback loop shown in Fig. 2. 2. . that is in terms of a system of the form introduced in Fig. to give the overall diagram as shown in Fig.20.17. 2. 2. which is required to be redrawn in terms of a unity feedback loop. I 1 with H = unity.
30
System representations
X
I
Fig. 2. I 4
Fig. 2. I6
Block diagrams
31
Fig. 2.18 Worked example, final block diagram
(s + 2)
Fig. 2.19 Worked example, initial block diagram
Fig. 2.20
Fig. 2.21
32
System representations
O n cascading the ( s + 2 ) term now in the forward path, this becomes Fig. 2.21, On moving the (s + 1) feedback denominator factor to the other side of the summing junction, the final unity feedback system is as shown in Fig. 2.22.
2.4
Physical system realizations
In this chapter methods of representing physical systems in terms of an algebraic sense by means of a transfer function o r a graphical sense by means of a block diagram have been introduced. Once a system model has been obtained, further analysis and controller design work can then be carried out, employing the model instead of the actual system. Hence testing of a new control scheme can be carried o u t in terms of the model rather than the actual system  so if a closedloop control system was to go wrong because of poor control action, the result would only be in terms of the model output, possibly just a number in a computer, rather than an actual output. To accurately characterize, with a model, the exact nature of any real physical item would require a model which was at least as complicated as the item itself, any simplification meaning that at least some information was lost. However, in a complicated form the model may well be timeconsuming and expensive, not only to obtain but also to comprehend and make use of, causing it to be unsuitable for analytical purposes. So the basic requirement is for a relatively simple model which encapsulates the mainlimportant properties of a physical system. I n terms of a straightforward inputoutput transfer function relationship simplicity is directly related to the order of the system, i.e. the number of coefficients required such that the model well describes the system’s behavior. One advantage in the study of control systems is that results obtained for say an electrical system are also true for a chemical system o r an aeronautical system, etc. In each case the basic components may well be different as will the type of input and output signals; however the mathematical model employed in order to carry out the analysis could have been obtained from any kind of physical system. Once a mathematical model has been arrived at, the physical system from whence it came loses much of its significance for analysis purposes and only really matters when one checks to see how sensible or practically applicable the analysis results are. Indeed the physical system itself could easily consist of several different types, for instance it might be part electrical, part mechanical, etc. In this section the text deals with obtaining transfer functions, as described in Section 2.2, in order t o represent firstly electrical and secondly mechanical systems. Methods are considered in which variables from one system type can be regarded as being analogous to those from another system type, showing how the study of control systems brings together the various engineering and physical disciplines by means of a common mathematical toolbox. Subsequently in Section 2.5 overall transfer functions
Physical system realizations
33
of systems in which at least two physical systems are present are discussed in terms of a motor/generator problem.
2.4.1
Electrical networks
An electrical network has already been used as the worked example in Section 2.2.1, and in a sense this section merely acts as a backup for that previous work. The three basic elements of an electrical circuit are the resistor, the capacitor and the inductor, each of these forming a different relationship between the signal variables with which they are associated. Consider initially a resistor, Fig. 2.23, which relates the voltage across it to the current through it by means of the equation v ( t ) = R i ( t ) , or in terms of the Laplace operator this is

V ( s )= R * /(s)
(2.4.1)
This basic equation (2.4.1) can be viewed in two ways, either we apply the current (input) I(s) and measure the voltage drop (output) V ( s ) or vice versa. So although just one equation (2.4.1) is deemed to characterize the physical response of a resistor, hence relating current to voltage, it can also relate voltage to current. The block diagram representing a resistor can therefore be drawn as either Fig. 2.24 (a) or (b) dependent on which direction is appropriate for signal flow. A capacitor, Fig. 2.25, can similarly be viewed in one of two ways as far as block diagram representation of its transfer function is concerned. We have that i ( f )= C dv(t)/dt, or
s V ( s )=  I(s) (2.4.2) C and this means that either Fig. 2.26 (a) or (b) is suitable for representation purposes.
1
Fig. 2.23 A simple resistance
(a)
(b)
Fig. 2.24
Block diagram for the resistor
34
System representations
Fig. 2.25
A simple capacitor
(bf
Fig. 2.26 Block diagrams for the capacitor
Fig. 2.27 A simple inductor
The final individual electrical element is the inductor, Fig. 2.27, which has a tran Fe function relating current to voltage of u ( t ) = L di(t)/dt, or
V(S)=
Lsl(s)
(2.4.3)
As far as block diagrams are concerned the inductance transfer function can be written as that in either Fig. 2.28 (a) or (b). As an example of the use of block diagrams and transfer functions in the modeling and analysis of an electrical network, consider the series RC circuit with input u i ( t ) and output u , ( t ) shown in Fig. 2.29. A block diagram for the RC circuit can be built up quite easily by recognizing that the voltage dropped across the resistance R is in fact U i f t )  u o ( t ) . The block diagram shown in Fig. 2.30 then follows directly.
Fig. 2.28 Block diagram for the inductor
Fig. 2.29 Series RC circuit with input v i ( f )
1 + RCs
1
Physical system realizations
35
Fig. 2.30 Block diagram representation of the RC circuit
Using the block manipulation techniques described in Section 2.3, once the two forward path blocks have been cascaded, the final transfer function can be obtained by removing the negative feedback loop, resulting in
Vo(s) = ___ Vi(S)
(2.4.4)
2.4.2
Mechanical systems
The elementary components and basic relationships involved in electrical circuits were described in the previous section. Mechanical systems will now be investigated, the overall intention being to characterize the essential physical relationships by means of equations similar to those used for the electrical analysis. The mechanical case is however a different entity in that there are two completely different types of mechanical action, translational motion in a straight line, and rotational motion about a fixed axis. The two types will be dealt with separately, the initial discussion being concerned with translational motion.
Translational
The fundamental signals within the study of translational mechanical systems are force f ( t ) and velocity w ( t ) ;in general the force is applied to a certain mass in order to move
36
System representations
it at a specific velocity, although as will be shown shortly a more direct link is apparent via Newton's law of motion between force and rate of change of velocity, i.e. acceleration. Some of the applied force is however used up in order t o overcome frictional effects, which are unfortunately nonlinear and rather difficult to model. However, by neglecting effects such as static friction, which opposes the commencement of motion, it is possible to model friction simply as viscous friction, which is a linear retarding force dependent o n the relative velocity. The idealized viscous friction resistive component called a dashpot is shown in Fig. 2.31, and has the effect that the applied force f(t) causes a velocity drop between piston and cylinder. If B is the viscous frictional component then the transfer function of a dashpot is f ( f )= B w ( t ) or
F ( s )= B W ( s )
(2.4.5)
Hence two possible ways occur in which this can be described in block diagram format, as shown in Fig. 2.32. A second element present in translational mechanical systems is a spring which in practice can constitute an actual spring, although more likely it is the capacity to store potential energy which is present in a cable or connecting rod. As was the case for a dashpot, the response of a spring is inherently nonlinear, although over a small range it can be approximated fairly closely by the simple linear relationship df(t)/dt = K w ( t ) or
sF(s) = K W ( S )
(2.4.6)
in which K is the spring stiffness, or K  ' is the compliance of the spring. In terms of a block diagram, the response of a spring can be shown as in Fig. 2.33.
41)
Fig. 2.31
Dashpot as a frictional element
Fig. 2.33
Block diagram for the response of a spring
Physical system realizations
37
.rrrrrrrrrmW)
f ([I
*
Fig. 2.34 Spring as an energy storage element
Perhaps the simplest way to consider the physical effect of a spring is that the application of a force produces a related displacement (Fig. 2.34), i.e. f ( t ) = K j w ( t ) dt. If one end of the spring is fixed to an immovable object, then the end of the spring attached to that object will have zero velocity whereas the other end will have a velocity which depends on the force applied, via (2.4.6). The final element for study in the area of translational mechanical systems is mass M , where the weight of an object is equal to its mass multiplied by g , the free acceleration due to gravity (9.81 m/s2). It was stated earlier in this section that one of the basic concepts of a translational mechanical system is the application of force to a mass in order to achieve an acceleration of that mass in terms of the relationship f(t) = M dw(t)/dt, or
F ( s ) = MsW(s)
(2.4.7)
When represented by a block diagram this relationship then takes on one of the two forms shown in Fig. 2.35. Although it is quite possible for a mass to have different forces applied to it at different points, the mass is considered to be solid and therefore will have just one absolute velocity associated with it; this is shown in Fig. 2.36 as the mass having a velocity compared to the zero velocity of an immovable object. As an example of the use of schematic diagrams and transfer functions in the modeling and analysis of a translational mechanical network, consider the mechanical system
(a)
Fig. 2.35
Block diagram for a mass
Fig. 2.36 Mass with absolute velocity
2.4.Wz(s)) (2.W Z ( S= M ~ s W ~ ( S ) ) (2.By eliminating force FI (s) and velocity WZ(S) from the three equations (2. Assuming that FI(s) is the resultant force on the dashpot with frictional resistance B I then from Fig.38 System representations Fig. balancing the forces across mass M I gives F I ( s ). 2.810) we obtain the input impedance transfer function of the MIS   MIS * F(s) . The system can also be described by means of element schematic diagrams as in Fig.4. there being a frictional resistance BI between the two masses.37 Mechanical system with force input f(t) shown in Fig.8) (2.F (s)= MISWl (s) I Further the velocity drop across the BI dashpot is described by FI (s) = BI ( WI (s). 2. 2.37.B 2 [ +1 1 .9) Also. where a force f(t) is applied to a mass M I which sits o n another mass M z .38.4.4.10) where the resultant force on the BZdashpot is equal to B2 W2(s)and the resultant force on the spring equal to K W z ( s ) / s .38 the forces across mass M I balance via the equation F ( s ) .
therefore a different speed on each shaft but with the same torque transmitted through the element.40. is then sQ(s) = K Q ( s ) (2.13) (a) (b) Fig.40 Torsional spring 410 . shown in Fig.11) The above final transfer function was obtained by algebraic manipulations on the three previous equations. Resistance to rotational motion is due largely to friction and is modeled by the damper. indicating a simple braking system. the lefthand shaft of the damper is shown fixed. and (b) two separate shafts.39.12) However.39 Damper depicting resistance to rotational motion a Fig. Rotational The second type of mechanical system considered is that concerned with rotational motion. As with translational mechanical systems there are three basic idealized elements and these are briefly introduced here. 2. 2. 2. it could however have equally well been obtained by means of a block diagram approach. 2.39 indicate (a) one shaft. is depicted by the torsional spring shown in Fig.Physical system realizations mechanical system as: 39 F(s) WI (s)  Mis(Mzs2 [ B I+ B Z ] S K ) + BI(Mzs’ + K ) + + + &S M2s2+ [ B I+ &IS + K (2. The compliance equation in which K is the torsional spring constant.4.4. the two different diagrams in Fig. 2.4. therefore one speed but a difference in torque along the shaft. The compliance of a shaft. In many cases with a damper of type (b). when a torque is applied. In general the resistance equation is given as: Q ( s )= W s ) (2. in which the fundamental variables are torque g ( t ) and angular velocity w ( f ) .
which is a measure of that body’s ability to store kinetic energy. the input impedance for this network is found to be: Q ( s ). 2. 2. 2. 2.B I ( Js’ + B2s + K ) + K ( JS+ B2) Js‘ + B2s + K fl(s) (2.4. By block manipulation or algebraic transfer function manipulation. 2. The inertia equation is Q ( s )= J s ~ ( s ) (2.42. Consider as an example the rotational mechanical network shown in Fig.14) in which J is dependent on the density of the body and its rotational axis. 2.43 Block diagram for the rotational mechanical example . where a torque d r o p occurs over each of the dampers and an angular velocity drop over the torsional spring of constant K .43.4.42 Rotational mechanical network Fig. for example its radius.41. as shown in Fig.15) 611 K  B Z example Fig.41 Inertia of rotating bodies The final element in the study of rotational systems is the inertia Jof a body. given that that body has an absolute velocity associated with i t . A block diagram for this system is given in Fig.40 System representations Fig.
44.4.s2 + B2s + *(s) K) . The dashpot Bz however has a velocity drop w z ( t ) . as shown in Table 2.u l ( t ) across it. 2. consider the network shown in Fig. this therefore becomes a resistance with a voltage drop u 2 ( t ) across it. In forming an analogous circuit.(M2s + B.1. There are in fact two possibilities: either a link can be formed between force in a translational mechanical system and voltage in an electrical system or. The input impedance of the network considered can be found by means of either the mechanical or the electrical setup to be: W2(S) .B i s from the balance equations: and . is directly mapped to become current i ( t ) . expensive and potentially dangerous whereas its electrical analog will most likely be a cheap benchtop exercise with the inherent safety that that affords. force and current.wl(t)across it and this becomes a resistance with a voltage drop u 2 ( t ) . As an example of forming an electrical analog from a translational mechanical system.4 thus far the relationships of basic electrical and mechanical systems with transfer functions and block diagrams have been discussed.16) + B2) . Once the initial link has been made the remaining connections between the different network types follow automatically. Also the masses M I and MZ have absolute velocities associated with them.45. The dashpot B I also has a velocity drop wz(t)across it.Physical system realizations 2. in the analog.3 Analogous circuits 41 In Section 2. It is readily observable that many similarities exist between electrical and mechanical systems in terms of the equations considered and this makes possible the investigation of a mechanical system by means of an electrical analogy or vice versa.s2 + B s + K (M. One big advantage of employing the mechanical to electrical analogy is that the mechanical network under consideration may well be large. these are mapped as capacitances with absolute voltages across them u l ( t ) and u z ( t ) respectively. The input to this mechanical network is in the form of a force f(t) which. The overall electrical circuit diagram is then as shown in Fig.= M.4. 2. (2. as will be considered here. the latter connection being the most widely encountered. an analogy is drawn not only between the different network elements concerned but also between the fundamental signals.
Q ( s ) I Conductance.45 Electrical analog of the mechanical network 2. K Capacitance. B Torsional compliance. Q ( s ) Velocity.Dashpot. J . Damper.42 System representations Table 2.5 Electromechanical devices Having considered separately mechanical and electrical networks. L Spring compliance. I(s) Voltage. forcecurrent Mechanical (translational) Mechanical (rotational) Electrical Current. M . an important aspect also arises in devices . . Fig. C Mass.44 Mechanical network. V ( s ) Force. B R Inductance. and discussed their similarities in terms of analog relationships.1 Analogies. 2.2. analog example Fig. K Inertia. W ( S ) Angular velocity. F(s) Torque.
There are in fact two electrical inputs to a separately excited DC motor. 2. by far the most commonly encountered motor being of the direct current type. making them ideal for use in highspeed environments such as disk drives and printers. one to the armature circuit and one to the field circuit. nowadays they are found only occasionally. is in fact only used for special lowspeed. as shown in Fig.m. A basic assumption is made that the motor is of the more usual separately excited field type which means that the field and armature windings are separate. hightorque applications. Although many years ago A C motors proliferated. ( t ) when motoring operation takes place. 2. usually when energy conversion takes place.Electromechanical devices 43 which involve both types of energy. but a detailed analysis of such motors is considered to be beyond the scope of this text. The mechanical output circuit which contains an inertia J and frictional resistance B. In this section the discussion is concerned primarily with the modeling of D C motors. Due to design improvements D C motors are made with very low inertia rotors and hence their time constant of operation is generally small. in series with a n inductance La and a n idealized back e. Such electromechanical devices can exhibit very nonlinear characteristics. The series field motor.1 DC motors The DC motor is used to convert a n electrical input signal into a mechanical output signal. usually when a special feature such as high noise rejection is required. the reverse process of electrical generation from a mechanical input will also be considered briefly from a control point of view. an example being the twophase induction motor which requires a n A C electrical input in order to provide its nonlinear torquespeed output relationship.46. 2. The armature circuit consists of a resistance R .5.46 Separately excited DC servomotor . For completeness. Fig.f. The field circuit however is simply a resistance Rfin series with a n inductance Lf. u . which is not discussed further here. thereby resulting in a wide control signal range. has a generated torque 4 e ( t )and a resultant load torque qi(t).
Two basic physical equations then describe the electromechanical characteristics of the motor. although it is quite straightforward t o employ each of them in a linear fashion by holding one of the terms constant.5. the air gap flux 4 ( t )is considered t o be proportional t o the field current.6) T h e mechanical output circuit means that load speed QI(s) is obtained from the load torque by means of the equation (2.7) A block diagram for the complete armature control motor is shown in Fig.1) (2. in the case of (2. which presents us with a linear relationship between if(t) and qe(t). and these are: (2.5. The load torque is therefore given by QI(~) = Qe(s)  Qd(S) (2. Because they both involve products of two variables the two basic equations (2.2) where K . So. however in some cases a torque due t o external disturbances Qd(S) must be overcome before any movement is possible.4) The armature current is related t o the input armature voltage by means of the armature circuit equation (2. such that where K m f = K.5. if(t).5.44 System representations On the assumption that various nonlinear characteristics such as saturation are to be neglected.o r we can hold the armature current i a ( t ) constant ( = l a ) . which presents us with a linear relationship between i a ( t ) and qe(t).5.2) are inherently nonlinear. which are termed armature control and field control respectively. Both o f these cases. The disturbance torque is included here for completeness although in many cases it will not be present and hence can be neglected from any analysis. An armature controlled motor is one in which the field current is held constant.1) we can either hold the field current i f ( t ) constant ( = If).47.5.5.5.5. 2.5) Also the load torque is made u p simply of the torque provided by the motor Qe(s). The load position can be obtained directly from the load speed by employing the . are discussed in this section. starting with a n armature controlled motor.1) and (2.If = constant (2. is the electromechanical coupling constant.
is A field controlled motor is one in which the armature current is held constant such that Qe(s) = Krnazf(s) where (2.48.5.5.1 2) A block diagram of the field controlled DC motor is shown in Fig.5.9) (2.Electromechanical devices 45 ' Kmr I 4 . 2. Fig.1 1) which results in a transfer function from input voltage Vf(s) to output speed of (2.10) Kma = KmIa = constant The field current is related to the input field voltage by means of the field circuit equation (2.' nl(s) S 1 (2.8) where el(s) the output position transform.48 Block diagram of a field controlled DC motor .5.5. 2. equation 6 l ( S ) = .
I ) and (2. Mechanical energy is supplied by means of the rotor and this is converted to electrical energy which appears in the form of a generated armature current.2): (2. o r the field current can be held constant while varying the rotor speed.5.14) By referring t o equation (2. one slight alteration being that the motor output speed 01 and torque ql are now replaced by the generator input speed url. So. 5 ) (2. A Jield controlled generator is one in which the rotor speed is held constant. such that. the usual method of obtaining a direct current being to rectify a generated alternating current.. 2.5.49 Armature output circuit for a DC generator . respectively. we have for the armature output circuit that (2.2). there are two possibilities for the linear control of a DC generator. either the speed will can be held constant and a varying field current if applied. It must be emphasized from a practical point of view. and torque q.46 and basic equations (2. as this procedure is both much cheaper and more robust. 2. the basic mode of operation is though essentially different and can be described as follows. that DC generators d o not commonly occur. from (2. Fig.5.46 2.2.5.5.2) are employed again to discuss the operation of a generator. A DC generator can be dealt with in roughly the same way as that used for the DC motor. however.5. if Rb and Lb are the load resistance and inductance respectively.49.13) (2.5. ( t ) the generated voltage as shown in Fig.5.2 System representations Generator In order to carry out a relatively simple analysis the motor diagram Fig.5. In the case of the DC generator the armature circuit is thus the output with u .
in practice it is often the case that the system is quite complex. thermodynamic. 2. indeed this is shown in Fig.4) It is fairly straightforward to obtain a closedloop block diagram of the speed controlled DC generator by referring to q n .50. 2.5. and hence an overall block diagram can be built up for the field controlled generator as shown in Fig. 2..5. energy is converted between electrical and rotational mechanical systems.46. see Fig. a physical system is a consideration of the relationship governing the variables. such that certain of the variables may well relate one physical type to another. both signals and parameters. y means of the field circuit equation (2. It must be pointed out that in this look at electromechanical systems in terms of DC motors and generators.6 Summary Step number one in the analysis of. 2. mechanical and electrical. such that (2.r + R b ) + (La + Lb)S The field current is related to the input field voltage. or the design of a controller for. The overall system may well in fact consist of a number of different system types such as fluid.. resulting in at the least a set of nonlinear . ( t )as the applied input torque and neglecting any torque due t o external disturbances. 2.50 Block diagram of a field controlled DC generator Fig. within the system in order t o obtain a mathematical model which adequately describes their interactions.1 l). Since the systems to be controlled are dynamic.17) where Kmfwas defined in (2.51. rules of the same kind apply and a similar analysis can be carried out for the conversion of energy between electrical and translational mechanical systems. differential equations are the most usual form employed for describing operating relationships.51 Block diagram of a speed controlled DC generator A = (Ra wQe (S ) K.5. A speed controlled generator is one in which the field current is held constant. However.Summary 47 Fig. 2.
i. as was shown in Section 2. the signal which is applied. where Y ( s ) = Y I(s) + Gs Yz(s). It is therefore necessary to make certain assumptions with respect to system operation. Fundamental analysis of the system variables can then be carried out either by means of a n algebraic transfer function description or by means of the equivalent block diagrams.e. This leads to the idea of analogous circuits in which a physical system of one type is considered within the framework of a physical system of another type. (b) The output Y z ( s )as a function of the inputs. What is the effect of making G I G ~ = ( G z G ~ ) . Y(S) 0 2. Y z ( S ) .2 s i 2 4 The twoinput/twooutput system shown (Fig. i. Although in this chapter the links between mechanical and electrical systems were dwelt upon.48 System representations differential equations.(s).' Assuming ? that GIG^ # (GzG3)'. such as range of signal values possible.I .2.4 it was shown how system transfer functions can be obtained from electrical and mechanical system types.53) has inputs VI(S). * . Find an expression for: (a) The output Y I(s) as a function of the inputs. in order to linearize the system equations and ease further analysis. as was shown in Section 2. In Section 2. and another signal as the system output. Once a set of linear differential equations have been obtained. and how once a transfer function has been obtained it can in fact be regarded in the form of any desired appropriate physical type. in order to define a system transfer function. 2.Ks.e. similar analogies can be made with other physical system types. or indeed a combination of the two approaches can be used. . the Laplace transform can be made use of. the two systems being related by a common transfer function. find the value of Gs such that Y ( s )= G I VI(S). The transfer function approach to system representation highlights the importance of labeling one signal as the system input. namely block diagrams. the signal which results. VZ(S) and outputs Y . This technique also leads o n to a graphical form of system representation.3.
Problems 49 Fig.1 pF 2.54. 2.4 Find the transfer function which relates input to output in Fig.5 p F Y(s) 0.53 2.55. 2. 2. . T T II == 0.3 Find the transfer function for the electrical circuit shown in Fig.
57 2. at a rate __t 4P Fig. 2. Water flows out of tank 1. 2.7 Two tanks of water are connected by a pipe (Fig.7 11 I + Fig.56 2.6 R I == c Y(S) R c == Fig.58 .5 System representations Find the transfer function which relates input to output in Fig. 2. which is of unit crosssectional area. .50 2. and into tank 2. 2. has a head of water equal to height r. 2.56.58). Water flows into tank 1 at a rate u and this tank.
61.7 r + 7 p Tank 2 : p = 7r . 2. Consider the system in Fig. if no unit feedback loop was present? Find the relationship between the input and output voltages in Fig. where p is the head of water in the unit crosssectional area tank 2. Fig. 2.Probferns 51 7(r . 2.9 What is the effect o n the input transfer function if K is reduced to K/4? What would have happened if K was part of an openloop system.e.10 Find the relationships between the input and the outputs Y I ( s ) .60 2. Water flows out of tank 2 at a rate 4 p . Show that the equations which govern the system are: Tank 1: i = u .1 l p Until time t = 0 the tanks are empty.60.8 YI!2l Fig. 2.59 2. i. u = 1. 2. From that time onwards.p ) . 2. Yz(s) and Y3(s) in Fig.61 . Find values for r and p from time I E 0 onwards and show that they tend to steady values. Fig.59 where K % I . Show the solution in the form of a block diagram. 2.
1 1 System representations The input to a system is designated as V ( s ) .62 2. Determine the transfer function relating the input applied force F ( s ) to the resultant velocity W ( s )in Fig. 2.13 Determine the transfer function relating the input applied force F ( s ) to the resultant velocity W ( S )in Fig. Fig. 2.52 2. 2. 2. JI BI J2 B 2 J3 Fig.whereas the output is X ( s ) .64 2.63.15 Find a rotational mechanical system which is an analog of the electrical system shown. using the forcevoltage analogy. That output is in turn the input to a second system whose output is Y ( s ) .64. Fig. construct an electrical circuit analog.65.62. .63 2. If the systems are characterized by the equations (a) (Dz + 7D + 12)x(t) = 3(D + l)u(t) and (b) (Dz 4D + 3 ) y ( l ) = 5(D + 4)x(t) + 2. 2. by using a torquecurrent analogy in Fig. 2.12 find the overall transfer function which relates input V ( s )to output Y ( s ) . Thereby.14 Find the transfer function relating the output angular velocity Q(s) t o the input torque Q(s) in Fig. 2.
The rotor has a resistance BrNm/rad/s and inertia J. show that the transfer function which relates input armature voltage to output load angular velocity is given by: K G ( s )= S+CY where 2. capacitance C F load. Find the transfer function which relates the desired to the actual load voltages. kgm2 and Kmf = K m I f= constant.19 2. . = J. find the transfer function which relates input field voltage to output load voltage when Kn. = 0 and Kmr = 3 Nm/A. in terms of the transfer function relating armature current to armature voltage (the input impedance). drives a load with inertia J = 4 kgm' and negligible resistance. The generator output of Problem 2. The field voltage is found as the output from an amplifier.16 2. Given that R.20 A field voltage controlled generator. If the armature resistance is Ra Q and the armature inductance La H .drives a purely capacitive load C F. Find an electrical analog for this circuit. An armature voltage controlled motor. with gain K . drives a load with resistance B Nm/rad/s and inertia J kgm2. An armature voltage controlled motor. with armature resistance Ra = 5 and negligible inductance. which has as its input the error between desired load voltage and actual load voltage.. Assume the effect of external disturbances Qd = 0.17 2. = KmOm = constant. and the rotor has a resistance B. Find the transfer function which relates the input armature current to output load angular velocity.Problems 53 V(S) L c 2 == 0  2. with field resistance R f Q and field inductance Lr 0. Nm/rad/s and inertia Jr kgm2.18 An armature current controlled motor drives a load with inertia Jkgm' through a shaft with compliance 1 / K rad/Nm. The coupling constant Kmr = KmIr = constant.19 is now connected to a series resistance R a. with armature resistance Ra Q and inductance L a H.
1975. H.. J... E. NewnesButterworths. J. Johnson. McGrawHill (Schaum). Mechanical technology... L. C. Close. Doebelin.. 1991.66 Further reading Bacon. . Electrical circuits. Karnopp. D. Principles o mathematical modelling. Gould. obtain t h e transfer function relating VI (s) t o F ( s ) (Fig. 2. New York. Signal analysis in linear systems. O.. 1978. . John Wiley. Process dynamics. H. C. Macmillan. A. 1977. I. R. DiStefano. E. f Boston. R. A .21 If Vl(s) and V z ( s ) are absolute velocities associated with masses MI a n d M 2 respectively. 1972. AddisonWesley. R. Introduction to signals and systems. J. 1969. Houghton Mifflin. A. Macmillan. Beachley. D. E. Introduction to dynamic systems analysis. and Harrison. D. Bender. Draper. and Stephens. L. J. Academic Press. Electrical technology. and Rosenberg. 2nd ed.. construct an analogous electrical circuit. A. C. 1984. 1990.. Houts. Feedback and control systems. t F(s) Fig. 1985. An introduction to mathematical modelling. and Williams.B. System dynamics. D. Wiley. Chemical process control. Electrical machines. A. System modeling and response. Longman. and Frederick. Modeling and analysis o dynamic systems. 1986.. E.. Wiley. K.. Using the forcecurrent analogy... C. Hughes. R. Peter Peregrinus Ltd. Kamen. C.. Longman. 1975. Dym. 2. estimation and control. 1980. C. Saunders. 1978. W. 1978. M. 1976. N. Harper and Row.. H.66). Stubberud. f Edwards. L.54 Sysrem representations 2.
J.. 1974. H. Modelling of dynamical systems . (eds. 1981. 1981.). 1976. H. AddisonWesley. (ed. Peter Peregrinus Ltd. E. Academic Press. Automatic control engineering. Nicholson. Dynamic modelling and control of engineering systems. Physical structure in systems theory.Vol. 2. 1990. 2nd ed. 1990. J. 1990. Van Dixhoorn. Harrap.. J. Systems modelling and optimisation. P. Ziegler.. Macmillan. Peter Peregrinus Ltd. McGrawHill. Nash. F. f . B. Truxal. M. I. 1980. 1979.. 1955. Theory o modelling and simulation. H. 1978. F. Power control of AC Motors..). G. G... and Evans.Further reading 55 MacFarlane. J. Raven. Dynamical system models.. Pergamon Press. Wellstead. Automatic feedback control system synthesis. Modern control system theory and application. Introduction to physical system modelling.(ed. Peter Peregrinus Ltd. Nicholson. J. Shearer. Academic Press. Murphy.. John Wiley. P. Modelling of dynamical systems . A. S. J..).Vol. Shinners. 1980. McGrawHill. J.
‘Does the performance of the present system satisfy our requirements?’ But this raises a further problem in that we must have a clear definition of what we mean by performance. for a car. The total output response witnessed is therefore considered to . and it can be seen through the remainder of this text that this operator is a major mathematical tool in the formulation and design of continuoustime control systems. it is shown here that sdomain system representations are an ideal form and that n o inverse transformation is necessary when investigating time domain specifications. the actual controller design is then of only secondary importance. In this chapter some of the fundamental performance criteria used for system analysis are introduced by considering the output response of a system in terms of its dependence on one of a standard set of basic test input signals. So the initial question to be answered is. will usually require some financial outlay. the subsequent analysis is generally easier than that required for a complex test input and it is most likely the case that the input signals which are in practice applied to that particular system can be thought of as combinations of the basic forms of input. and although it may not be initially obvious. It is important therefore to decide. perhaps more obviously. As a n example. In assessing the performance of a system. whether or not it is necessary to employ a controller in a particular case. The application of a controller to any system will result in an increase in overall complexity and.1 Introduction In Chapter 2 various ways were considered in which a system to be controlled could be modeled in terms of a representative transfer function relating the system’s input to its output. reliability. we could be interested in speed. following a particular test input. o r the reaction observed once things have settled down to a steady form of response. Not only is it much simpler and cheaper to apply a fairly straightforward test input. It is noticeable that the output response is viewed in terms of its time domain features. analysis can be concentrated on either the initial reaction of the system output. The Laplace operator was used to great effect. in the first instance.System response using differential equations 3. comfort and/or fuel consumption in terms of one criterion in particular or more likely in some combination of all of them.
when in normal operation. when they are asked to respond.Steadystate response 57 consist of two parts (a) the initial reaction. may well have to deal with a multitude of different input signals or conversely only have to face one specific type of input. carried out in Section 3.4 the possibibty of an error existing in the steadystate between the actual output and the desired output/reference input is shown to be directly related to the application of specific input signals to certain types of systems. contribution of this chapter is to introduce some of the positive aspects gained with the application of feedback t o form a closedloop control system. Whatever the case it is usual to investigate + + . it is possible to think of the steadystate response as being the response which remains when the transient response has died down. It is wrong to think that the steadystate response of a system refers in every case to a steady system output value although. or more simply as the transient response subtracted from the total response. A variety of standard input signals are available for employment in order to determine the response characteristics of a system. So the steadystate response of a system can only be considered by making reference to the type of input applied. as is shown in this section this might be the case. In most cases the system.2 Steadystate response Because the total response of a system is made up of two parts. Indeed this latter definition is that employed to investigate response errors in Section 3. and (b) the steady reaction. All physical systems. such as reduction of the effect of noise on the output signal and reduction in output sensitivity to plant variations. the transient and steadystate responses. In some cases it might be that the response is rather sluggish and slowly reaches its final value.5. A further. Some of the main points made in the chapter are then brought together in the extended servomechanism example of Section 3. do not do so instantaneously by jumping to a final steady value. whereas in other cases it is possible for the output to very rapidly overshoot its final value and then to slowly home in on it by decreasing its overshoots as time goes on.3.4. If a time invariant signal is applied at the input then a different steady output response would be expected when compared with the response obtained by applying a sinusoidal input to the same system. which is known as the steadystate response and is essentially that part of the response which is left when the transient response dies down to zero. These different transient response possibilities are considered in Section 3. and as the transient response will by definition tend to zero at time t 00. In the detailed study of the steadystate response of a system. but vital. so the steadystate response to an input will be that apparent as t m. in which the use of feedback to improve both transient and steadystate system performance is discussed and this results in a basic controller design problem.2. 3. it is not necessarily so. it is shown how the observed output is dependent on the input signal applied and in Section 3. a common element between these inputs being their relatively simple implementation.
with the more popular types discussed in greater detail. be considered as random or at least pseudorandom. The unit step input is a primary test signal and can be described by u ( t )= I 1: 0: t>O t<o (3. the most commonly encountered standard input test signals are described. one example being systems whose input is affected by a large random disturbance.1. the same is true for many complicated systems which are inherently of high order.58 System response using differential equations the system response to only one or two of the standard input forms. If it is assumed that the step is of unit magnitude and is applied at t = 0. Perhaps the simplest of all inputs to employ is a step input which can be applied. I Fig. It may then be necessary to apply a pseudorandom input in order to obtain detailed information as to the characteristics of the system. the test signals applied as inputs are generally of a fairly simple standard form. during normal operation. Standardizing test input signals is important when assessing the relative merits of several systems by comparing the output responses to a common input signal.1 Control test inputs The input to some systems can. In practice it will most likely be sensible to apply a step input of some magnitude other than unity. by switching a voltage in an electrical circuit or a sudden force which is retained in a mechanical network. this being Time.2. for example.1 Unit step input signal .2. Subsequently the steadystate response of systems to certain of the inputs is considered with examples being given of some low order. However for essentially simple systems. then this is shown in Fig. fairly straightforward system transfer functions.1) where the input magnitude is shown to be unity after time t = 0. In the section which follows directly. and in order to obtain important fundamental information on more complicated systems. 3. 3. because in general this will provide sufficient information for further analysis. 3.
3) when applied as a control input signal. will be much slower than the step transition time.1) is given by U ( S )= I/s (3. The ramp function. logic gates for example. governed by their time constants. which can therefore be neglected. the assumption being that at t = 0 the input changes from 0 to 1 . The unit ramp input can be written as u ( t )= I t: 0: t>O t<O (3. a more complicated description of the step input would. for example. since a finite but very small time is actually necessary to carry out such a transition. 3. If the integral is taken of the step input this results in a function which increases linearly with respect t o time from a zero value at time t = 0. In general though the response of systems considered in this text. For the study of very high speed systems. It can also be noticed that the step input is undefined for t = 0. however. and is a simple derivative of the step input. As the practical application of such a signal poses obvious problems. most likely be necessary.2. 3.4) 0 1 Time.2. it therefore has a Laplace transform of U ( s )= 1 + (3. In its strictest form an impulse function lasts for a time interval St as 61 0. The Laplace transform of the ideal unit step input (3.2. A second test signal is the impulse function. it is usual to approximate the impulse with a signal of unit magnitude which lasts for a very short time period At. and is of infinite magnitude. is shown in terms of a unit magnitude ramp in Fig.2.2.2 Unit ramp input signal .2) and it is worth noting that a 4unit step input. has a Laplace transform U ( s )= 41s. I Fig. an idealistic view. which can also be regarded as a velocity function. and this is in fact discussed at length in Chapter 2.Steadystate response 59 particuIarly true for a low gain system when the presence of noise on the output could be problematic unless a high magnitude step is used. This was considered to an extent in the previous chapter.
resulting in a power series or polynomial . is achieved.5) By integrating the ramp function. * + + 2 P! although a more general expression is obtained by considering each separate term to have a magnitude equal to a constant di.7) by means of the Laplace transform. this can be written as u(t)=l+t+2 t2 It is quite possible.2.+ .3. This can also be considered as u ( ~ 1/s3 = ) (3. a square law signal known as a parabolic function. ramp and parabolic inputs to form one single input signal. for all t > 0.2. however.60 System response using dqferenlial equations Fig.2.2. to consider further integrated terms in order to arrive at an overall input t2 t p u(t)=1 + t + . 3. hence obtaining an acceleration function.2. It is apparent that by combining the step.6) where the division by 2 in (3.6) is included in order to think of the function as a step acceleration term with respect to the original step input (3. 3.3 Unit parabolic input signal or in terms of the Laplace transform U ( s )= l/s2 (3. The unit parabolic input can be written as (3.1). This is shown in terms of a unit parabolic signal in Fig.
2 Sinusoidal and polynomial inputs Analyzing the steadystate response of a system to a sinusoidal input is not in general possible by means of the Laplace transform final value theorem. This polynomial function can though be written more concisely as u(t)= i=O 2 djti/i!: t > 0 where d are scalar constant values. as long as the system to which the input is applied does not contain any instabilities (the stability of a system is discussed in Chapter 4).2. 3. The output signal in the steadystate will then also be a sinusoid.2.9) This type of input signal is considered in the following section in terms of the steadystate response which results from its application.Steadystate response function 61 t2 ~ ( t )do + dl t + d2 .9) we have that for all t > 0. This theorem requires that there exists a limiting value to which a signal converges as time tends to infinity. and for a sine wave no such value exists. u ( t )= cos w t . 3. (3. remembering that if the input is of magnitude U. So if the steadystate system output. i The final standard test signal introduced here is the sinusoidal function which can be written in unit magnitude form as (3. * * = 2! t + dp.P P.+ .8) for t > 0. For the sinusoidal input defined in (3. y ( t ) . is obtained in Fig.in terms of a phasor diagram this can be represented as in Fig. u ( t )= 0 for t < 0.4 Phasor diagram representation of u ( t ) = U cos w t .2. 3.2.4.
where y ( t ) is the response to a sinusoidal input acted upon by a system whose transfer function may be described by means of a differential equation as defined in Chapter 2 . then the output will be a sinusoid of magnitude Y .62 System response using differenfiat equations Imag. as shown on the phasor diagram in Fig.2. A general polynomial test input was defined in (3. k +Ck t!k then in order to obtain equality in the example equation we must have that y ( f )= co + clt + . It is assumed that the system to which the input is applied is defined by a differential equation.5.8). 3. 2! czt2 i. then we have that in which the ai and bi coefficients determine the characteristics of the system transfer function.2.e. k =2 . As an example consider n = 2 and m = 1 in the above equation with a control input defined by U(t)=do+dlt+We have then that: d2f2 2 (a2D2+ atD + a o ) y ( f ) (blD + bo) = If in general the steadystate output is described by y ( t ) =co+ Clf c2tZ ++ 2! **. 3.5 Phasor diagram representation of y ( r ) = Y cos steadystate output (wt + #). as described in Chapter 2.8) will result in a steadystate output signal which will also contain enough terms in powers of time t to cause equality in the system differential equation. The application of a test input (3. Fig. with a phase shift denoted by the angle $.
ramp and parabolic inputs The step input was defined in Section 3. then the system output is obtained from Y ( s )= G ( s )* U ( s ) to be K 1 Y ( s )= . or l/ao is the time taken to reach the final value if the initial rate of increase is continued.2. so k=p 3.e.* exp( . clr c2 can then be found by solving the three equations obtained. The output is then Y ( s )= Klao s K/ao s+ao or in terms of its inverse transform K K y ( t ) = .1..10) and ~ 2 Notice that the power index k is dependent solely on the complexity of the input signal. 3. UOC= ~ ~ 1 bo dz + ~ O C2I = 6 d2 + bo dl U 1 +~G I C I + ~ O C O bi di + bo do 2 = (3. the index p . where a0 > 0.e. i. If such an input is applied to a firstorder system transfer function of the form K G ( s )= s + a0 where K is the system gain. It can be considered either that l / a o seconds is the time taken for the output to reach within exp( .Steadystate response from which it follows that 63 (a2D2+ U I D+ ~ o ) u ( t ) 0 2 ~ 2+ = UICI + uoco+ ( ~ 1 ~ U 2 + aoc2t2 OCI)~  + 2 The steadystate output coefficients co. The final value of the output signal.a t ) a 0 a0 the time response of which is shown in Fig. i.6.s+ao s in which l/s is the unit step Laplace transform. can be found ..e. the time to reach 63. that is the steadystate output value.2.2% of the final value.2.1) of its final value. i.3 Step.
which states that if the limit of a function f ( t ) as t 00 exists.2.. A similar result in which the output tends t o infinity is witnessed as the firstorder system response to a unit parabolic input. lirn s Y ( s ) = s .s0 So. application of a unit ramp input will result in the system output tending to infinity as t + 00.* s0 s + a0 s3 s0 in which I/s3is the unit parabolic input... shown in Fig.6 Firstorder system response to a unit step input by means of the final value theorem of the Laplace transform. then + lirn sF(s) = lirn f ( t ) s+o 1m (3. . This can also be seen by finding the inverse Laplace transform which is 1 such that the output tends to infinity as t 00. the output transform is now given by K 1 Y ( s )= . By applying a unit ramp input to the same system transfer function. .6.s + a sz 0  and on applying the final value theorem we find that lim s Y ( s ) = s * . given from the final value theorem as + lim s Y ( s )= s * .e inverse Laplace transform as t + 00.11) So for the firstorder example just discussed. 3. for a firstorder system as shown.64 System response using differential equations 0 l/ao Time. s + a s ls0 0 l s+o K =a0 and this agrees with the result found from t. 3. I  Fig.
when set to zero. By analyzing the output of this system in response to a unit step input signal. the output value will always approach its final value from one direction only. All of the analysis in this section is based on the assumptions that (a) the control test input is a unit step function. For the firstorder system considered in this section it is not possible to obtain any overshoot. and (b) the system is represented exactly by a secondorder + . 3. a topic discussed in Chapter 4.4. The value of K merely part defines the actual final output value for a firstorder system whereas the value of ao.2).it is a type 1 system. the steadystate response of a system was considered to be that part of the total response which remains as time t 00.12) Note: This is second order because the highest power of s = 2. This is the same as requiring that the system is ‘stable’. and this factor s. i. the type of response obtained being dependent on whether the system denominator roots (the system poles) are real or complex. in fact with a0 = 0 the system denominator reduces to s(s+ a l ) . is considered at length in order to introduce the possibility of an output response overshooting its final value. see Section 3. characterizes how rapidly that final value is approached and is therefore an important measure of system performance. as well as partly defining the actual final output value. The concept that a system has a transient response which tends to zero implies that the system has certain properties which prevent the output from constantly increasing in magnitude towards infinity.3 Transient response In the previous section (3. to a step input. and this diagram holds for all a0 > 0.Transient response 65 In the following section the response of a secondorder system. It is worth noting that if a0 = 0 the steadystate output signal tends to infinity. In a similar fashion the transient response of a system can be regarded as that part of the total response which tends to zero as time t + 00. 3. in particular to a unit step input. The total response therefore consists of the steadystate response summed with the transient response.2. it contains a common s factor . causes the infinite output.1. The time response.6. different values of a0 merely affecting the time and magnitude axis scaling. of a firstorder system was shown in Fig. the final value theorem reveals lim sY(s) = s * s0 K s 2 + als+ a. s s+o = a0 which is an identical result to that found for a firstorder system.e. and this is the subject of the following section. A general secondorder system transfer function can be written as G ( s )= s2 + als + a0 K (3. The step input time response of a secondorder system is rather more complicated. the figure l/ao being called the time constant.
66 System response using dgerential equations transfer function. they can in many cases be modeled adequately by a secondorder system.3.2) It can be noted from the step response that by setting t to infinity in (3.3. /a K K lim y ( t ) = . This result agrees with that obtained by application of the final value theorem in Section 3. a0 = a@. from (3. this is exp( .= QP a0 where.1 Overdamped response If the system transfer function has two real denominator roots.3. then it follows that: (3. and for the majority of systems it will be a step response which is most relevant in practice. A further point t o note is that at t = 0: and ..3.c or in terms of its timedomain representation. the total response obtained being dominated by that due to a secondorder model. Restricting the analysis to secondorder systems only by no means limits the applicability of the results.exp( . Although certain systems are of order greater than two. and assuming a > 0.1). P > 0.1) By applying a unit step input to such a system.2.2).P t ) P 11 (3. however no further important aspects are revealed by such a procedure when compared with a step input response. 3. Ramp or parabolic inputs could be applied as well as or instead of the step input.3.a t ) .3. the output signal transform is found to be: K 1 Y ( s )= (s + a ) ( s + (3) s 1 i +.
as shown here it increases from zero to its final value with no overshoot. If the two denominator roots of the system transfer function are identical. In common with the response of a firstorder system to a step input.at exp( .a t ) ) .3) When a unit step input is applied to this system the output signal transform is Y ( s )= 2 * (s+a) s K 1 .! L a2 s (s+a) (s+a)2 “1  and in terms of its timedomain representation: K y ( t ) = a [ 1 .Transient response 67 Fig. It has very little effect on the time response other than that. e. 3 . is to cause the differential d y ( t ) / d t at time t = 0 to be nonzero.exp( .3. then G(s) =  (s + a)2 K (3. where the initial rate of change of output signal is identically zero.g. The effect of having a numerator root in the system transfer function.K [ .2 Critically damped response A critically damped response can in a certain sense be regarded as a special or extreme case of the overdamped response.a t ) .7. 3. 3.3. the output signal always approaches its final value from one side. 7 Overdamped response to a step input The overall time response for the overdamped case is shown in Fig.
in which it would be extremely dangerous to have any overshoot and yet taking a long time to reach the desired position.e. and such a response is described in the next section. 3. The critically damped response represents the most rapid type of response which does not result in any overshoot. For many systems however it is desirable for an overshoot to occur before the output settles on its steadystate value.3. . the output signal is in Fig.4) Assuming that a > 0. in which the initial rate of change of output signal is zero.3. would mean a waste of revenue due to less productivity. we have that K lim y ( t ) = 1W a2  Qo K where a0 = LY 2 Initial conditions can be found by setting f to zero in (3.8. This property is used to good effect in some measuring instruments and is particularly useful for example in certain machine tool applications or robot welders.4) which gives Y ( t )= 0 and also The overall time response for the critically damped case is shown in Fig.68 System response using differential equations K ~ ( t ) =1 e xp(a t )[l + a t ] ) ( = ~ a (3. 3. i.8 never greater than its final value. as would be the case for an overdamped response.
3. This is by far the most widely studied type of transient response.5) or G ( s )= which is also K s2 G ( s )= K (S such that when compared with the standard polynomial denominator form given in (3..at)sin P t  a P or . then we have G(s) = (s +a + jP)(S +a . More parameters are used to describe the response than in the previous two cases and hence in an academic environment it is a much more fruitful topic as far as exam questions are concerned.5) the output signal transform is Y ( s )= K ( S + a ) 2 + p2 * s 1 In terms of its timedomain representation this is 1 .a t ) cos P t .3.3.j P ) + 2as + (a2 + p 2 ) + a)2 + p2 and al=2a K (3.exp( .exp( . If the two denominator roots of the system transfer function are complex (complex conjugates in fact) then the output signal will exhibit oscillations when a step input signal is applied. Just how quickly these oscillations die down is dependent on the degree of damping. However.3 Underdamped response The underdamped response is characterized principally by the output overshooting its final steadystate value following the application of a step input. as the amount of damping is increased so the oscillations die down more quickly until we reach a critically damped response in which the degree of damping is so great that the oscillations do not even start up! Assuming that the system transfer function contains complex roots.1) it follows that in this case ao=a2+P2 By applying a unit step input to the system (3. such that if the response is not damped at all the oscillations will in theory carry on for ever.Transient response 69 3.3.
5) is in fact in the form s2 + ttw.see (3. i. The most usual way to write the system denominator of (3. is defined as the time taken for the step + .6) the time response can be written as (3. the setfling time. The initial part of the response is most often described by means of two terms. which is the time taken for the output step response to reach 50% of its final value. If the damping ratio is zero then = 0 = 01 and an= b. O n the other hand if a 0.7) So the response is a sine wave of frequency 0 with an exponential damping governed by the factor CY and a steadystate value given by K / ( a Z+ b 2 ) .3.3. which is the time taken for the output step response to rise from 10 to 90% of its final value.8) in which { is the damping ratio and because: We have from (3. but this will only be equal to the natural frequency anin the limit as the damping tends to zero.. At a later stage in the time response. and secondly the rise time. shown as T. Hence for a n underdamped response 0 < f < 1 . s2 + 2as + (aZ B Z ) = + W. hence for a n undamped system the response will be dictated from time t = 0 by a cosine waveform.0. Note that for a particular response. as shown in Fig. . These definitions arise and ofi = a’ + 0’ such that i f P 0.5) and (3.1 in which the system denominator consists of two nonidentical real roots. where the damping is heavier as t tends to unity.70 System response using differential equations and if we define I $ = tan IP a (3.8) that: ( = a/wn + the natural frequency. in Fig. something which causes the roots of the denominator to tend t o those of the critically damped system .then the damping ratio j 1. 3. such that wn is the frequency at which oscillations will occur when there is n o damping and in such a case the oscillations would be sustained indefinitely. where sz + 2 f w n s + a:.s + a$.3) .3.3. the performance of that system for comparison or design purposes.7).3.e.3.see (3.9 indicates several basic characteristics of a system which are used t o describe. 3. . In fact for > 1 this corresponds to the overdamped case discussed in Section 3.3. + + r r r T h e response shown in Fig. 3.9.3.3. in relatively simple terms. so the exponential damping tends to zero and we approach the case of a purely oscillatory response . firstly the delay time. (3. In the initial part of the response it must be remembered however that the phase shift 4 will have a certain effect such that 4 90” as the damping tends to zero.9. /3 is the actual frequency of the oscillations.
7) to zero. Tirne. 1.Transient response 71 I I I I I I I I I 0 T. are termed overshoots.3.9. A commonly encountered figure is 5 % .2. etc. The maximum overshoot is .9) It can be seen that the maximum and minimum values of the differential are dictated by the sine term only. the initial condition being y ( i ) = 0 for i = 0 .3. i. which results in (3 * 3.3. or more usually maximum overshoot. Another term used to describe the transient performance of a system is overshoot.9 Underdamped response to a step input response to reach and stay within a specified percentage of its final value. such that they occur when t = h / P .2. 3. This means that on substitution back into (3.e.2. 1. At the time instants denoted by the index i .10) in which i = 0.1 . which is the maximum difference between the transient and steadystate step response values that occurs after the response has first passed its final value. maximum overshoot x 100% steadystate value The maximum and minimum values of the output response can be found by equating the differential form of (3.7) the magnitude of each maximum and minimum value is given by y ( i ) = . the output response therefore has values of Relative to its final value the output response values for i = 1.exp( . 3. etc. etc.ina/P) cos in) I K a2+p  (3.t * Fig. where i = 0.. The maximum overshoot is frequently written as a percentage of the steadystate output response value. and this was employed for the response shown in Fig.
m / P ) x 100% (3. overshoot is t = T / P . when the exponential term is at its largest. I t is often written as a logarithmic value. which is successive overshoot ratio = exp( .m / P ) This is then also an indication of how quickly the oscillations die away in magnitude. I0 Variation of transient response with damping ratio . and this particular relationship between 01 and P results in what is known as ideal damping.3.11) and the time taken to reach this maximum. The maximum overshoot value is in fact also the absolute value of the ratio of successive overshoots.3. and this is found from the actual magnitude at i = 1 to be . with a transfer function G ( s )= (S + a)’ + K (3. in which case we observe a loss of 27.29 ((YIP)dB/overshoot A special case of the logarithmic decrement is arrived at when (Y = 0. that is logarithmic decrement = 20 loglo exp( .72 System response using differential equations then the first overshoot.12) 012 Fig.e. Consider for example the magnitude of the second overshoot divided by the magnitude of the first overshoot.exp( . 3. In terms of a percentage this is maximum overshoot (070) = exp( .29 dB per overshoot. i.m / P ) +P where the steadystate value K / ( a 2+ 0’) has been subtracted from the actual response maximum overshoot = Q‘ K magnitude in order to obtain the magnitude of overshoot at this time instant.m / P ) =  27.4 2a01’P) 3 = exp( . first.
given in Fig. 3. although it is quite possible that even though the model was a good one when obtained. the system has aged or has been modified since then and the representation that we have of the system is no longer suitable. A control system design can be investigated in order to find out how sensitive it is to variations in the system description. Any error which does occur can be caused by one or more of a number of reasons. caused by disturbances. and this leads to the idea of acceptable damping.4.Errors 73 However in the design of a system it might.4 Errors The performance of a system is usually assessed in terms of the error between the demanded or required output signal and the actual output signal that is obtained. about a reference value. The first classification of error to be considered here is that caused simply by the effect of certain system types on straightforward input signals. 3. The problem of noise is quite a serious one in many systems. in terms of disturbance error reduction. not be possible to achieve an exact equality between (Y and p.e. When applying a specific desired reference input to a system it is generally hoped that the system output will settle down to the desired level as steadystate conditions are approached. are discussed.1 System type When discussing a system type it is important to note that reference is always made to the openloop transfer function of a feedback system. by how much is the control design affected? The sensitivity of a control system is also discussed later in this section. 3. on the assumption that unity feedback is auulied as shown in Fig.1 1. namely step. . for which 3 < cY/p < 3 To conclude this section. if the actual system is not quite the same as we have pictured it to be. for instance the output measurement could be incorrect because of poor calibration or noise interference. The effect of noise signals on a system is considered later in this section and the advantages gained by the employment of a feedback control scheme. 3. It is however shown here that this is by no means necessarily the case and is strongly dependent on the system type.10 in order to show the difference in signal decay due to the damping factor in each case. i. It might be that a poor model was calculated in the first instance. a set of step responses is. and the objective of some controllers is simply to regulate the variation in output level. Another major source of errors is the employment of an incorrect system model for controller design purposes. in practice. In the first instance therefore a definition of system type is given. ramp and parabolic inputs.
1) where q. bo # 0 and ( q . Note that the order of the system.74 System response using dgferential equations Fig. denotes the type of the system. these roots being termed poles and zeros of the transfer function.. and this is the case considered here. thirdorder system. e..4. m. the openloop transfer function . I 1 Unity feedback system If the original block diagram representation of a system contains elements other than unity in its feedback path. m and n are integers. respectively. . is then n + q . in order to obtain Go(s).1 This is a type 1.g. a type 1 system is achieved when q = 1. ao. 1.this is shown in detail in the following section.2. Example 3. by means of the techniques introduced in Chapter 2. It is worth noting that it is usual for n 2 m.p This is a type 0. usually given by the order of the denominator.4. thirdorder system. by defining Go(s) as: (3. The transfer function definition given in the previous chapter is now further generalized. . 3. this would signify a thirdorder system. e. K is a gain constant and q = 0.4. if n = 2 and q = 1.n 2 01.g. the representation must be manipulated into a unity feedback system.2 'E . Example 3. In many cases it is possible that the denominator and numerator polynomials of order n and m respectively can be factorized.
this is given by (3. however: Y ( s )= 1 G + G ((ss))H ( s )V ( s ) (3.3) Go(s)= 1 + G ( s ) [ H ( s . as shown in Fig. 1q p p Fig. 3. the transfer function Go(s)= G(s).Y ( s ) . the error is considered to be the difference between the desired output V ( s )and the value H ( s ) Y ( s ) .2) and (3. Only when H ( s ) is unity are the two approaches identical.12. the desired reference output signal.2 Steadystate errors Consider the unity feedback closedloop system in Fig. Example 3.4.2. is important in that it categorizes the error considered in this section always to be the difference between the desired and actual outputs E ( s ) = V ( s ). Kuo (1982). obtained by connecting (3.e.g.it is imperative that the reader is aware of this point.2) For a general feedback system. In certain other texts. 3.11. the unity feedback case.4. 3. in which the error E ( s ) .5) such that the error between desired and actual output signals depends on the input signal V ( s ) .4. and also on the openloop transfer function G&) whether this is obtained directly or via equation (3. i.Errors 75 The second of these examples.1 1. 3.4) where it can be noted that if H ( s ) = 1 .12 General feedback system .3) The equivalent openloop transfer function Go(s)can therefore be found from G ( s )and H ( s ) by means of the equation.4.4. is found to be (3. Consider the closedloop equation obtained from the unity feedback system shown in Fig. 3. e.4.4.4).4.11 ) G(s) (3. the closedloop equation is.4. i.e.
4. for q > 0. for a type > 0 system.6) in which sE(s) has n o denominator root whose real part is positive.6). it will therefore only be considered briefly here: Transient error = e ( t ).76 System response using differential equations The total error can be regarded as consisting of two parts. The remainder of the total error is then the transient error and this in fact links directly with the transient system response discussed in Section 3.4.4. and where (3.e.7) Steadystate step error I f a unit step input is applied to a unity feedback system.4. the term [ Go(s)] Is4o = K .3 Consider the system 1 The steadybtate step error is given as s E ( s ) 1s0 = 1 1 5= 0.5) and (3. however. a constant Example 3.4.e. = sE(s) Is+ (3.e ( t ) (3. firstly the steadystate error can be written as e ( t ) 1. I f Go(s)has any common s factors in its denominator the steadystate error to a step input will be zero. For a type 0 transfer function.3. i. we have that 1 V ( s )= S The steadystate error following a unit step input is then which is obtained from (3. .6) is merely the Laplace transform final value theorem result. i.4 +1.4. .
a smaller steadystate errar has been achieved.4. For a type 0 system.g.. the steadystate step error can be made smaller by increasing K.e.. = 4 and therefore SE(S) JS+O = 0. we have that 1 V ( S )= S2 The steadystate error following a unit ramp input is therefore For a type 0 system it is observed that [sGo(s)l IS+O =0 such that the steadystate ramp error is 00.e. which can be done by increasing Ihe gain K. a constant such that the steadystate ramp error is l/K. we have [sGo(s)] ISo = K. Steadystate ramp error If a unit ramp input is applied to a unity feedback system.3. .Errors where 77 It is important to note that if the step input is of a magnitude other than unity. For a type 1 system however.. but with This would make K. consider the system i n Example 3. the magnitude of the type 0 steadystate step error will be different from that calculated for a unit step input.2 i.
o = such that the steadystate parabolic error is 0. In the following section the remaining part of the total error. we have [sGo(s)] ISo = Kar a constant 00. a halt will be called here by summarizing the results obtained in Table 3. such that the steadystate parabolic error is l/Ka. It can be noted that if the ramp input is of a magnitude other than unity. If the parabolic input is of a magnitude other than unity. The analysis carried out for the three classes of input can in fact be continued further if desired. is considered.78 System response using differential equations For a type > 1 system though [SCO(S)l I s 4 = 00 such that the steadystate ramp error is 0. the magnitude of the type 1 steadystate ramp error will be different from that calculated for a unit ramp input. For a type > 2 system though [sGo(s)l I s . Steadystate parabolic (acceleration)function If a unit parabolic input is applied t o a unity feedback system.1. In this section we have been concerned primarily with the steadystate error obtained from different system types in response to certain test inputs. namely the transient error. the magnitude of the type 2 steadystate parabolic error will be different from that calculated for a unit parabolic input. we have that 1 V ( S )= j S The steadystate error following a unit parabolic input is therefore For a type < 2 system it is observed that such that the steadystate parabolic error is For a type 2 system however. . however.
It follows then that the essential time . as shown in (3.[sE(s) (3. in the sdomain. The steadystate error 1/(1 + K ._ 1 1+Ks s in which the transform of the steadystate error has been taken. and as the results in this section are closely related. The transient error is the error which remains when the steadystate error is subtracted from the total error. 10 0 m 3 0 0 0 3. the analysis is not carried out to such a great depth here.4.4. as ISol (s) transient error = E ( s ) . I Steadystate errors System Unit step Unit ramp Unit parabolic input input input type 0  79 I + K.9) Transient step error For a type 0 system. this means that the denominator characteristics of the transient error transfer function are identical to those of the closedloop transient response._. The transient error of a type 0 system in response to a step input is therefore transient error = 1 1 1 + Go(s) s 1 .3.Errors Table 3.4. If the total error between desired and actual outputs is d t )= u(r) . the steadystate error following a unit step input is given by with the step input transform V ( s ) = l/sapplied. ) is simply a scalar value.9).4.Y ( t ) then the transient error can be written. see Problem 3.3 Transient errors The transient response of a system was considered at length in Section 3.
e._ . a certain amount of interference caused by disturbances will affect any control system. In conclusion. For a type > 1 system the steadystate error following a unit ramp input is zero.4. 0 for a type 0 system. although it can be regarded as a special case of the following. transient error = 1 + I Go(s) 1 ' s ' A similar analysis can be carried out for a parabolic input.9) transient error + = 1 + Go(s) 1 1 1 . For a type 1 system the steadystate error following a unit ramp input is l/Kr. i.4 Disturbances N o matter how well the system design and construction phases are carried out and how much protection is introduced. In some cases it is negligible with respect to the noisefree . by means of (3.4. the same input signal being assumed for both cases. _ . The transient error is therefore also the total error.80 System response using differential equations domain characteristics of the type 0 system transient error following a step input will also be those of the transient response to a step input. For a type > 0 system. 1 s2 Kr s where K. The transient error is then. Transient ramp error For a type 0 system the steadystate error following a unit ramp input is Q). transient error = 1 I + CO(S) * s 1 and as the closedloop response to a step input is given by Y ( s )= 1 + Go(s) Go(s) s it follows that as the denominator of the transient error is also that of the output response. i. it must be remembered that the transient error behavior of a system depends on a denominator which is the same as that for a closedloop transform of the same system.e. the characteristics of the two responses will be essentially the same. the steadystate error following a unit step input is zero. The transient error is therefore also the total error. 3. transient The error in such a case is therefore only of academic interest and is not really defined well in any useful sense.
Consider a feedback control system. The closedloop transfer function for this system is Y ( s )= K (s+cY+ K ) V ( s ) (s + +a+K ) 1 * NO) (3. as shown in Fig. N ( s ) (deterministic) part of the system operation and need not be considered further. the time constant of which is 7 = 1/(a K). a steady nonzero value .or (b).13. which is affected by the disturbance term N ( s ) . with a selectable gain K. It is usual for disturbances to be regarded as either (a).Errors 81 Fig. In practice such a + rapid response may not be possible and the resulting large signal at the output of the gain K may not be achievable. The total output value is therefore a summation of the desirable signal with the undesirable noise. 3.We would hope that in the steady state the output signal would be equal to the desired value V. the steadystate output value is in fact given by: Hence. For a large number of systems.10) which in fact consists of two separate transfer functions. One of the main reasons for employing a feedback loop in a control system is to reduce the effects of noise on the output signal. Disturbances are however considered in a fairly general way in this section and designs particular to either (a) or (b) are not discussed.4. By the Laplace transform final value theorem. Assume the desired input V ( s )is a step of magnitude V and the disturbance N ( s ) is a load disturbance of magnitude N. though.13 Unity feedback system with disturbance. it presents a major problem and can result in the total objective of the control design being to reduce the effect of noise on the output signal to an acceptable level. one which relates the signal V ( s )to the output Y ( s )and the other which relates the noise N ( s ) to the output Y ( s ) . 3. a random noise signal.these are called load disturbances or offsets . which may be corrupted. if we make K a very large value this will satisfy both the steadystate final value y 2: V and noise reduction problems. However the (s + a + K ) denominator term means that in the time domain the output signal will tend to its final value = V by means of an exponential increase. Increasing the gain can . and it has been shown here how this is possible merely through the selection of an appropriate gain value. which at the same time reduces the steadystate step error in the example given.
4. it is by no means always the case that a n adaptive controller is necessary o r even particularly desirable. Y ( s ) . For a n openloop system. U ( s ) . It is shown here how the use of feedback in the control of a system. for which a n increase in K would result in oscillatory behavior such that for a very high value of K output signal oscillations would be wild and would take some time to settle down. reduces the sensitivity of that system t o variations in the parameters which represent the system’s characteristics. something which is easily seen if the firstorder system example in Fig. since the system was modeled. For simplicity.1 2) . to the output signal transform.82 System response using differential equations though cause problems.1 1) which is simply the fractional change in output divided by the fractional change in system transfer function. then it is sensible t o redesign the controller in order t o reduce the dependence of output performance on the accuracy of that parameter.4. exact representations of their realworld physical counterparts. However. When investigating the sensitivity of a system though. SE. if it is found that when one particular parameter varies by a few percent from its nominal value the output performance is degraded considerably. results in an equation Y ( s )= G ( s ) U ( S ) from which it follows that for a n invariant input signal U ( s ) (3. One of the main reasons behind the employment of adaptive control schemes is the need to vary the control action in accordance with variations in the system characteristics. 3. It is important to discover how sensitive system performance is t o certain variations o r inaccuracies in the parameters used t o describe the system. the following discussion restricts itself to the sensitivity of the output signal in respect of variations in the system transfer function. the particular system performance that is being measured must be specified as must the parameter(s) which the sensitivity is considered in terms of. This assumption ignores not only the errors inherent in the modeling and construction stages but also measurement problems. For instance. 3.4.5 Sensitivity When designing a controller for a system it is generally assumed that both the system model itself and the resulting controller description are accurate. we have that the transfer function relating the input signal transform. and the fact that either the system itself o r its environment will have altered to some extent.13 is replaced by a secondorder system. This can be summarized by means of the sensitivity function. defined as (3.
4.4.4. looked at in another way.1.4. The original closedloop transfer function is given as Y ( s )= such that 1 G + G ((ss))H ( s ) V ( s ) (3. dependent on variations in H ( s ) rather than G ( s ) . it is usually a desirable feature in that precision and modifications can be much more readily built in t o H ( s ) rather than G ( s ) . is reduced from its openloop value by a factor I/(l + G ( s ) H ( s ) ) The selection of feedback H ( s ) . Although this places much more importance on the accuracy of the feedback path.4.13) which means that the output varies in a directly proportional relationship to variations in the openloop transfer function . 3. as shown in Fig.14) and the sensitivity function becomes or (3.12. if the system transfer function is only accurate to 2 10% then so is the system output. (3. it will be only weakly linked with the accuracy of the system transfer function. we have that (3. when in closedloop. Consider now a closedloop system. is therefore very important in determining the sensitivity of a closedloop system to variations in the system transfer function.4. .16) with a sensitivity function. but the feedback function H ( s ) varies.Errors such that sensitivity function becomes 83 (3. If it is the case that the system transfer function is invariant.17) This means that when I G ( s ) H ( s ) S. while for the same conditions.15).4. with the desired reference input transform V ( s ) and feedback path H ( s ) considered as invariant.or.15) Hence (3. the sensitivity function magnitude is equal to 1 1 and thus the accuracy of the output will be directly dependent on the accuracy of the feedback path.15) shows that the sensitivity of a system. from (3.
will be discussed in detail. Also the load impedance is ZI(S)= BI + JIS. 3. The latter signal is subtracted from the former at a summing junction in order to obtain the error.15 to introduce the idea of a gear ratio n . 3. 3. commonly referred to as a servomechanism. E ( s ) . for an armature controlled motor.14. and K y respectively.e. This error is then itself amplified in order to provide the armature current. the effect of the gearbox. The motor is assumed to have a constant field current and also to contain a gearbox placed between the motor and the load.1 Gears The bIock described in Fig. as shown in Fig.14 as the 'motor' is expanded in Fig. the output of which is the output position Y ( s ) . a position control system. in which B m is the resistance of the motor and Jm is the motor inertia. between the amplified desired and actual output signals.5. O m = nol.84 System response using differential equations 3.15 Servomotor in openloop . Fig. 3.14 Position control servomechanism Fig. 3. The reference input signal V ( s ) and sensed output position Y ( s ) are provided by rotational potentiometers with gains K . in which BI is the resistance of the load and JI is the load inertia. which means that n revolutions of the motor produce 1 revolution of the load. Before considering the overall inputoutput servomechanism transfer function. rather in the form of an extended example. is considered here. + J m S . Za(s). 3. 3.5 Servomechanisms In order to draw together the motorlgenerator analysis carried out in the previous chapter with the differential system equation properties of damping and steadystate error introduced in this chapter. The motor impedance is Zm(S) = B. regarded as being part of the motor in Fig.14. i.
3) or by substitution (3. Now the torque required to accelerate the load.7) . when differentiated gives and this must be equal to zero for a minimum of Qa(S).5. a criterion often employed is that of obtaining n such that Qa(s) is a minimum.5. + Qe(s) = z m ( s ) Qm(s) + Qg(s) (3. Qe(s)is in fact split up between that used on the motor and that input to the gearbox. hence (3. the smaller the reflections. i.e.4) The effect of the gearbox is thus to reflect back the load impedance to combine with that of the motor.5. Hence. whereas Qs(s) is that required in order to keep the speed of the load constant.5.1) or (3. the larger the gear ratio. and this value of n is then referred to as the ideal gear ratio. also Ql(s) is the load torque output from the gearbox. subject to the square of the gear ratio.5.5. that is Qg(s)Qm(s> = QI(S)QI(S) (3.2) The total torque obtained from the motor.Servomechanisms Qm(s) 85 is the motor speed input to the gearbox and Ql(s) is the load speed output from the gearbox.5) or Qe(s) = Q a ( s ) + Qs(s) in which Q~(s)n J m + = [ 3 SQI(S) (3. Qa(s).6) is the torque required in order to accelerate the load. power in = power out.5. This can also be regarded as (3. lossless gearbox. Therefore the torque input to the gearbox Qe(s) must be such that for an ideal. When selecting the gear ratio n for a particular servomechanism.
= B. to minimize Qa(S).5.7) results in a minimum value of Qa(s) rather than a maximum. + f r s ) Kmf * By cascading the blocks shown in Fig. such that B.16. 3.4.5.3 it is possible to set the transient response of the servomechanism considered.4) and (3.10) this results in a n overall closedloop equation relating desired output value (reference input) V ( s ) to actual output Y ( s ) of or (3.Ky Y(s)l (3. that the value of n given in (3.16.5.5 8) J. By referring to the transient response forms considered in Section 3.5.. 3. the torque QS(s) is disregarded completely. = Jm + J1/n2 ns(B. we have that Y ( s )= la(s) shown in Fig. by taking the second differential of Qa(s) with respect to n. by an appropriate choice of . 3.5) which reduce the servomotor block diagram to that in Fig. la(S) = K [ K v V ( S ).14 to be (3.1 1) J. 3.9) and on substitution for la(S).15.5. and + Bl/n2 (3. where the impedance of the load is referred to that of the motor.2 Servomechanism transient response The relationship between torque Qe(s) and load speed QI(s) can be viewed as that in the servomotor block diagram.86 System response using differential equalions It can be noted that: (a) (b) In the selection of n . Fig. 3.5. or in terms of the relationships (3. It is easy t o show.
as t 00. The analysis of system steadystate and transient responses carried out made use of the + .5.Summary 87 the potentiometer/amplifier gains..can be investigated by means of the final value theorem. V ( s )= I/s. the steadystate response which remains after any initial response fluctuations have died down. will be those selected.1 1) K” lim sU(s) IS+o =.5.11) and Assuming B. The denominator of (3.12) such that the output y ( l ) will follow its required value. 3.3. the system response can be considered to consist of two separate parts. u ( t ) .1 1) should also be of this form. When V ( s )is a step input. K. it is considered that they die down to zero as I . is equal to that of the feedback potentiometer. the transient conditions which occur. and the transient response which is the name given to those initial response fluctuations and although these can persist for quite some time after the input signal is applied.6 Summary It was shown in this chapter how some fairly simple input signals can be used to test the response of a system and hence to investigate some of the fundamental properties of that system.3. i. Following the application of an input signal. i. n. as long as the input potentiometer gain.5. i. 00. in the steadystate.12) is of the form (s + a)2 + a 2 + a)’ + a 2= s2 + 2as + 2 a 2 and if the servomechanism control parameters are to be chosen with this in mind it is required that the denominator of (3. Consider for example the case of ideal damping discussed in Section 3. whereby from (3. we therefore must set the feedback potentiometer gain K y to Steadystate conditions following a unit step input.e.e. (3 such that from (3. when a desired output position value i s asked for. Jr.e.= lim y ( t ) Ky Im (3. Kmf and K to have been previously set.3. Ky.5..
and these are made use of in the chapters which follow. as shown in Fig. Secondly. 3. I7 (a) (b) (c) 3. also considered in terms of its sdomain representation. Both the steadystate response and transient response can be discussed with reference to sdomain operators only.2 Find the value of K necessary to provide a critically damped closedloop system.88 System response using dvferential equations fact that once a system transfer function is obtained in terms of the Laplace operator s. Fig. via the final value theorem and in terms of the degree of damping respectively. some of the useful properties obtained from the employment of feedback in a control system were highlighted by showing how disturbance corruption and performance sensitivity can be reduced to a n acceptable level simply by a suitabIe selection of feedback terms. . Firstly. 3. Find the range of values of K which will provide acceptable damping for the closedloop system. Problems 3. it is not necessary to take inverse transforms in order to investigate the time domain properties of the response. when K=0. o n the application of a n input signal. where K is an adjustable gain. Many of the fundamental performance requirements in the design of a control system have been introduced in this chapter.1? A control system with unity feedback has a n openloop transfer function: G ( s )= K ( s + 12) s2 + 25 where K is a scalar gain. What is the magnitude of overshoot which follows a unit step input. Some of the main reasons for control system error were considered in Section 3.4.1 A unity feedback closedloop system contains an openloop transfer function 9 G ( s ) = ___ s(s + 5) A factor D ( s ) = Ks + I is cascaded with G(s). the importance of system type and the input applied were shown in terms of their effect on the steadystate error between the desired and actual output signals.17.
Find the value of K for which the transient response is ideal and the steadystate and transient step errors in this case. Also. . show that the steadystate ramp error which is produced by such a choice is 10%.3 Calculate the closedloop denominator and find the range of values for K over which the closedloop poles are real. this will result in a transient step error with damping ratio [ = 0. Show how this function is modified if a feedback gain of H ( s ) = 2 0 0 is employed. and (b) a scalar multiple in the denominator does not affect the answer. Fig. consider how derivative feedback.5 show that (a) the steadystate error to a step input is K . With K = 24.5 and natural frequency wn = 10. rather than H ( s ) = 1 in the unity feedback case. 3.6 show that if K = 100 and CY = 10. = K / 2 . type 1 system: G ( s )= K ~ s(s + (Y) 3. with gain K . can be employed to improve the transient performance. I8 3. Given an openloop.Problems 89 3.18. Consider the unity feedback closedloop system with an openloop transfer function: G ( s )=  K s+3 Find the sensitivity function S i which relates the sensitivity of the output to variations in the gain K . find the step response of the closedloop system and evaluate the steadystate error which follows a step response. 3. For the control system shown in Fig. By using the openIoop transfer function: C(S)= K ( s + 1)(s+ 2) 3.4 Show that the denominator of a type 0 system’s transient error following a unit step input is also the denominator of the system’s closedloop transient response.
find the gain value K and amount of derivative feedback K I for which the system has ideal transient performance and a ramp error of 5 % .7 System response using differential equations For the feedback control system with openloop transfer function G ( s )= K s(s + 20) 3.19. For a unity feedback system with openloop transfer function show that for the closedloop system.11 Find the steadystate step and ramp errors for the unity feedback closedloop system whose openloop transfer function is s+4 s(s 5 ) G ( s )= 3.9 A unity feedback closedloop system contains an openloop transfer function: G ( s )= K s(s + 25) 3.90 3. For this value of K obtain (a) the steadystate error to a unit step input.10 Find the value of K which will provide a steadystate ramp error of 5 % . 3. and (c) a unit ramp input.e. where N ( s ) is an unwanted noise signal and s(s+ 1) Determine values for K and G ( s )= 1 ~ K ( s + 1) D ( s ) = ___ (s + Q ) Q which will achieve a critically damped output . = 20. K . and the transient errors produced by (b) a unit step input. the damping ratio is {= 1 2SKP and the natural frequency 3.12 ~ + Consider the system shown in Fig. i.5 s2(2s+ 1) 3.8 and feedback H ( s ) = Kts + 1. Find the steadystate parabolic error for the unity feedback closedloop system whose openloop transfer function is G ( s )= 2.
15 3. in order to provide a damping ratio l' = 0. Consider a unity feedback closedloop system in which the openloop transfer function is G(s)= 4(s s(s2 + 3)K + 4s . step magnitude. Find the range of values for K over which acceptable damping is obtained. 3.25 radls. (b) Ideally damped transient step error?. in which G(s) = ] / h a n d N ( s ) is a ramp signal of magnitude 0.3.16 For a particular armature current controlled servomechanism.02?.12.14 D ( s )= K ( s + a) S find values for K. By reflecting back the load impedance to combine with that of the motor. the total inertia is J kgm' and the total resistance R Nm/rad/s with a gearbox ratio n : 1. find the steadystate error to a ramp input of magnitude 0.75 and a natural frequency w = 1.13 response to a unit step input V ( s ) . It is required that the steadystate signal value is equal to 25 units 4 2%.03 0. a and input signal.5.03K What value of K will result in: (a) Steadystate step error = 0. 3.Problems 91 I Fig.35s+ 0. find a compensator D ( s ) which when connected in cascade with G(s) will produce a damping ratio { = 1. A unity feedback closedloop system contains an openloop transfer function: G(s)= s2 + 0. . what are the steadystate and transient errors if a unit step input signal is applied? Consider the system diagram shown in Problem 3.and a disturbance effect on the output signal equal to 1/10 that of a step signal N ( s ) . For K = 20. V ( s ) .5 units. the error between the reference position and sensed output position is amplified by a gain of K Nm/rad to achieve the total torque obtained from the motor. If K = 0.5.4) If K = 10. If 3.19 I 3.
1979. and Mayer. Introduction to the design of servomechanisms.. R.. H. 1958. C. McGrawHill. Linear control system analysis and design. Wiley Eastern. B. John Wiley & Sons. P. Longman. and Schultheiss.... System dynamics. A field current controlled servomechanism is subject to a total inertia of 15 kgm’ and a total resistance of 2 Nm/rad/s (referred to the motor) with a gearbox ratio 10:I . Pergamon Press.. Healey. H. f Chestnut. McGrawHill (Schaum). C. L.P. The motor armature has a resistance equal t o 3 Q and an inductance equal to 6 H. when the ratio of the gearbox is ideal. PrenticeHall. Principles of automatic control. 1990. Brown. McGrawHill. 1978.. M.92 3. and Willsky. Automatic control systems. PrenticeHall. A field voltage controlled servomechanism is subject to a field resistance of 2 fl and a field inductance of 4 H. and Piasco. Control systems. 1983. Find a value for the amplifier gain K Nm/rad which results in ideal damping. J. K. Richards.. Modern control systems. An armature voltage controlled servomechanism is subject to a motor resistance of 0.4 kgrn’. Feedback and control systems.19 3. Introduction to system sensitivity theory.18 3. Dorf.5 A and the electromechanical coupling constant is 25 Nm/A2. 1978. J. F. R. Sinha. 1981. Frank. N. M. V. Kuo. and the load has a resistance of 10 Nm/rad/s with an inertia 15 kgm2. D.. M. J. J. Infroducfion to dynamics and confrof. with a motor resistance of 1 Nmfradfs and a motor inertia of 2 kgrn’. . The constant relating torque to field current is 25 Nm/A. with a load resistance of 10 Nm/rad/s and a load inertia of 120 kgrn’.. 1983. R. 1976. Signals and systems. Hodder and Stoughton. 1951.17 System response using differential equations An armature current controlled servomechanism is subject to a total inertia of 4 kgm’ and a total resistance of 3 Nm/rad/s with a gearbox ratio 1 : l .S. 6th ed. 1981. Chapman and Hall. Find the transfer function which relates output load speed to input armature voltage. Wiley (Vols 1 and 2). 2nd ed. P. H. and Campbell. R. Further reading Bower. Find a value for K which results in a suitable damping.2 Nm/rad/s and a motor inertia of 0. 1985.. Elloy.. 1978. 1992. DiStefano. J... S .. Control systemprinciples and design. C. AddisonWesley. Stubberud. and Houpis. Oppenheim. I. R.. E. A.20 between desired position and sensed output position is amplified by a gain K Nm/rad. the field current is a constant 2. PrenticeHall. McGrawHill..M. O.. Raven. Doebelin. Classical and modern control. Automatic control engineering. 1950.. The error 3. 3rd ed. Chapman and Hall. An introduction to dynamics and control. Ogata. and Williams. H. where 0 < K c 10.. Academic Press. W. D’Azzo. J. G. 1994. J. Principles o servomechanisms. A. 1990. and Simpson. Servomechanisms and regulating system design. J. A. . P. Power. Find the output speed following a 10 V step input. K. M. J.
Consider a system transfer function in terms of numerator roots (zeros) and denominator roots (poles).System stability and performance 4. This means that if a bounded input or disturbance (i..(s+ P") where K is a scalar gain and the poles are either real or occur in complexconjugate pairs. in this case the system is said to be critically or marginally stable..e. Certain bounded inputs when applied to a critically stable system can produce an unbounded output signal and therefore critically stable systems should not be regarded in the same way as stable systems. although it must be remembered that they should also not be regarded as unstable . if it is a stable system then its output will also be bounded.the oscillations neither increasing nor decreasing with respect to time. The vast majority of objectives are plant dependent. such that G ( s )= K ( s + z ~. system. timeinvariant. One of the most important requirements for a control system is that it be stable.in this case the system output (speed) must be able to reach a specified level (150 km/h). conversely if its response tends to infinity magnitude as t + 00 then it is unstable. There is. a third possibility in that a system response to an impulse might tend to some finite but nonzero value as I + 00. System stability can be looked at in another way by considering the application of an impulse function at the system input. they can usually be posed. Many different objectives exist. the same definition applying were the system output to oscillate between two finite values . even in the general sense. however. those commonly encountered being dealt with in this book.merely as a special case. .1 Introduction The problem of control system design is one of specifying a number of performance objectives which it is required that the controlled system must meet. If the response of a system to an impulse input approaches zero as time t + 00 then the system is stable. e... a certain automobile is required to travel at 150 km/h.g. (S+Zm) ) (s+ P I ) . however. no bigger than some finite value) is applied to a linear. in a way which is more generally applicable .
system zeros are irrelevant (as long as no zeros are exactly equal t o and therefore cancel with any righthalf plane poles). U ( s )= I . It is only the poles of a system transfer function which are important as far as stability is concerned. s = . as shown in Fig. if all pole real parts are positive then the output will tend to zero as t co. we have that: y ( t )= n C Kie”‘ I = I further terms being necessary where multiple poles exist.1.) 3. In Fig. if one or more lie in the right half then that system is unstable. The poles of a system are the polynomial roots obtained when the system denomi .1 is an example. + Nor es: 1. If the poles of a system transfer function all lie in the left half of the splane then that system is stable. A system is critically stable if one or more poles lie o n the imaginary axis of the splane. It follows that if any one pule real part is negative. It does not matter how many poles of a system lie in the left half of the splane. A positive real part for a particular pole pi means that the denominator root s = pi will lie in the left half of the splane.p4 in Fig. 4. if a pole existed at either s = . Further. 4. I splane stability region (shaded) If the system output Y ( s ) is related to the system input U ( s ) by means of Y ( s )= G ( s ) * U ( s ) then when an impulse is applied at the input. 4. 2. subject t o no poles lying in the right half of that plane. 4. 4.p l .1 s = .p5 or s = . however. the output will tend lo infinity irrespective of the remaining poles. (The pole at s = .94 System stability and performance Fig.p3 are all within the stability region.p6 then the system which contained such a pole would be unstable.pz and s = .
It is shown in Section 4.and closedloop systems. furthermore if it is openloop unstable this does not necessarily mean that it is also closedloop unstable.i. In this chapter it is shown how the loci of characteristic equation roots can be plotted simply by means of a set of readily applicable rules to follow. However. The movement of each root is called the locus of that root and gives a graphical representation when plotted in the splane. The system denominator is known as the characteristic polynomial and when it is equated with zero it is known as the characteristic equation. in most cases. this means that if the feedback H ( s ) is other than unity then the equivalent forward transfer function within a unity feedback system must be obtained by algebraic means or block manipulations before proceeding. can be investigated by inspection of the transfer function denominator roots (poles). be difficult for other than loworder systems. Further. 4. 5 .2 The RouthHurwitr crlterion The stability of a system. this can be done by finding the roots of the system’s characteristic equation. however for values of n greater than five no .The RouthHurwitz criterion 95 nator is equated with zero. as it is only the closedloop denominator which is of importance. If this is so and n is a fairly low value then the poles can be found without too much computational effort. it is assumed that the system G(s) is employed in a unity feedback system. whether it be openloop or closedloop.e. although this will. The concept of stability is applicable to both open. this means that it is only necessary to cascade H ( s )with G ( s ) . if it is openloop stable.2 that it is in fact not necessary to find the characteristic equation roots when testing for stability and that a technique exists for testing a polynomial to ascertain whether or not there are any roots which do not lie in the left half of the splane. The location of system poles in the splane gives an indication not only of the stability and relative stability of that system but also of its transient performance.a. The RouthHurwitz testing criterion merely requires the formation of a table from the characteristic polynomial coefficients. and this procedure can also be used when a system is stable to find out just how stable it is by testing to see whether or not any roots which lie in the left half of the splane are to the right of a specified real value s = . However. obtain G ( s ) H ( s ) .this concept is known as relative stability. the roots may well not be directly available. this does not necessarily mean that it is also closedloop stable. the denominator being represented simply in terms of an nth order polynomial. and to proceed as though this were the original openloop system. If a system is to be tested to find out whether or not it is stable. For a particular system. It is assumed that the gain K is the coefficient which can be varied and hence it is the variation of K which causes the loci. If it is intended to control the system in a closedloop form then it is useful to see how the closedloop characteristic equation roots move around in the splane as certain controller parameters are varied.
l). . n .2. then both sides of the characteristic equation must be divided through by this factor in order to obtain the normalized form given in (4. it merely means that we must also carry out the full RouthHurwitz stability test as described in the next section.11 are constant real numbers. . The procedure is in fact applicable to any general polynomial which contains real coefficients and it merely tests for roots which lie in the right half of the splane. timeinvariant system may be investigated without the need to calculate the transfer function poles. and/or at least one coefficient is negative. the denominator is equated to zero. however. . the method consists entirely of an algebraic technique and hence no inexact graphical stages are required. Another point worth making is that the transfer function for which the polynomial in (4. it is zero.e. . (ai>O: i = O . . 4. in terms of a transfer function denominator this means that the procedure tests to see if the system is unstable. there are assumed t o be n o exact common factors between the numerator and denominator. and this is that all coefficients of the polynomial must be positive definite. i. Thus we have that s"+ a . An assumption has been made in (4. i. The RouthHurwitz criterion presents a method by which the stability of a linear. If the polynomial of (4..96 System stability and performance direct factoring methods exist and obtaining even approximate roots can be a n extremely timeconsuming process. associated with s".1) is found to pass the first test then this does not mean that there are n o roots which lie in the righthalf plane.2.I ) If this test fails and either at least one coefficient is nonexistent.. * .e. hence realizing the characteristic equation.. i.1) in which { ai: i = 0.2.. but by definition must be nonzero.l s " .1): note that the coefficient an could be positive or negative.1) in that if initially there is a constant real term a. n .1) are rewritten t o form the . + a 2 s 2 + a1s+ ao=O (4. and there is therefore n o need to apply the full test.2.1 RouthHurwitz stability test The coefficients of the characteristic equation shown in (4. this means that roots of the polynomial exist either in the right half of the splane or on the imaginary axis and the system is then considered not t o be stable. Where a polynomial has been found to fail the initial test.2.2.l + . it has been shown that the system is not stable (although this could mean it is critically stable).1) is the denominator is assumed to be proper. The point is only important if there are any transfer functions zeros which lie in the right half of the splane o r on the imaginary axis. Further.2. In order to find the poles of a transfer function.2. A n initial test by inspection can be carried out on the polynomial of (4.e.
. The remaining rows of the array are constructed by bl = an2 ..I 1 b2 = an4 ..an.3 .... (a) Note: The most important aspect of the RouthHurwitz criterion is that the system is stable if there are no sign changes in the first column of the Routh array. etc.. .. . the coefficients of each row depend only on the coefficients in the two rows immediately above.bz. (b) If there are any negative elements in the first column..I sn2 97 1 Qn2 an4 an5 an1 bi CI $13 an3 bz c2 b3 c3 .I  an5.2. 1 an. etc. The RouthHurwitz criterion can be stated as follows: The roots of the characteristic equation all lie in the lefthand side of the splane if all the elements in the first column of the Routh array are positive definite. = an3 .. ... It is of no consequence whether the a0 term actually appears on the top or next to top row. in the vertical direction the array should then contain ( n + 1) rows in total. .1 Consider the polynomial equation s4 + 6s3 + 1 ls2 + 6s + 15 = 0 The Routh array is s4 s3 S2 S1 SO 3 15 I .* ~ ~ .The RouthHurwitz criterion first two rows of the ROUTH ARRAY. c .. which is defined in full as follows: S" Sn. the number of sign changes indicates the number of roots which lie in the righthand side of the splane. then if we start at the top of the array and work to the bottom by comparing successive elements in the first column. Example 4.. bi  So.b3. .. S ' SO I * ..1 cz = an5 . Elements in the array are calculated until only zeros appear.l bi an.a .
when a zero appears in the first column there are two possibilities. 4.2 Consider the polynomial equation s3+6s2+ l l s + 6 = 0 The Routh array is s3 S2 1 10 I S ' 6 lI ' 11 6 A system with such a denominator polynomial is therefore a stable system because there are n o sign changes in the first column of the Routh array. either the remaining terms in that particular row will all be zero or at least one of them will be nonzero. However. .2. Further.The new polynomial.3 . + 15.l). The original characteristic polynomial should then be multiplied by a factor (s + a)in which the real constant a > 0. If a zero does appear it is then not possible t o continue forming elements of the array because all elements of the next row are necessarily infinity by means of the definitions given.2 Special cases In the analysis carried out in the previous section the problem of a zero term appearing in the first column was completely avoided.e. + + Example 4. something which is very easy to test. 10 . although the remainder of the row in which the zero appears does not consist entirely of zeros. with a factor (s + a) should then be tested for stability via the Routh array. although one must be certain that the original polynomial does not then contain a factor (s. however it must be checked that for any particular selection of a the original polynomial does not contain a factor (s.98 System stability and performance A system with such a denominator polynomial is therefore a n unstable system because of the negative sign in the first column of the Routh array. i. Consider a polynomial for which a zero appears in the first column of the Routh array.2. there are two roots which lie in the righthand side of the splane. Normally a selection such as a = 1 is perfectly suitable.a). The latter case will be dealt with first. as there are two sign changes.
in the lefthand side of the splane.3 99 Consider the polynomial equation: s4 + 3s3+ 2s2 + 6 ~ 2+= 0 The first three rows of the Routh array are s3 s4 S2 and so a zero has appeared in the first column. as there are two sign changes. there are two roots which lie in the righthand side of the splane. i. and so on until the final coefficient in the row is associated with so. + 3 . .1.10.e.e.2 + 10. formed from the new polynomial.2. in which the real part is not necessarily nonzero. It follows that (s4 + 3s3+ 2s2 + 6s + 2 ) ( ~ 1) = sS + 4s4 + 5s3 + 8s2 + 8s + 2 = 0 + The full Routh array. and/or by pairs of real roots with opposite signs.The RouthHurwitz criterion Example 4. this will always be even.5 2 I 2 S' So A system with such a denominator polynomial is therefore an unstable system because of the negative sign in the first column of the Routh array. i. The second coefficient is then associated with the power of s two less than that denoting the row. + + A second special case exists when all the elements in a row of the Routh array are zero. The multiplying factor merely consisted of a single root at s = . after first checking in the equation above that s = 1 is not a factor of the original polynomial. When a row of zeros occurs. Further.5. Let us choose the multiplying factor s + a = s + 1. is then: S2 s3 s4 SS p 7 . both of the roots in the righthand side of the splane are therefore roots of the original polynomial. The auxiliary polynomial is in fact a factor of the original polynomial and its order indicates the number of root pairs. This is caused by pairs of complexconjugate roots. The coefficient in the first column is associated with the power of s which denotes that particular row. an auxiliary polynomial must be formed with the coefficients in the row of the array immediately above the row of zeros.
Example 4.100 System stability and performance The procedure to be carried out when a row of zeros occurs is (a) form the auxiliary polynomial.2. Note: When forming the remainder polynomial.in example 4.4 Consider the polynomial equation s6 + 2s’ + 4s4 + 2s3 + 7s’ + 8s + 12 = 0 1 2 3 0 The first four rows of the Routh array are s6 SS 1 j ’ 1 4 s4 s3 2 3 0 7 8 12 12 +  + Auxiliary polynomial coefficients Row of zeros S2 s’ So 1 1 3 2 3 which shows the remainder polynomial to be stable with roots which all lie in the lefthand side of the splane.2. (b) divide the original polynomial by the auxiliary polynomial and (c) test the remaining polynomial by means of the Routh array.4 there was a factor 3 can be neglected. . any common multiplying constant in the auxiliary polynomial .
4.2. the concept of relative stability can be viewed as one in which the characteristic equation is tested to find out whether or not any roots lie to the right of an imaginary axis which passes through the point s = . however. Example 4. without actually solving for the roots. whether the system. for which the polynomial is its transfer function denominator. It is often useful.5 Consider the polynomial equation: S ' +5 . with a stable system. where 6 2 0. as shown in Fig. such that the equation becomes one in factors of r rather than s.5 3 The Routh array formed from this polynomial is s3 SZ S' So 1 5.3 Relative stability 101 Employment of the RouthHurwitz criterion as considered thus far merely tells us whether or not a polynomial has any roots which d o not lie in the lefthand side of the splane. 3 s= 0 Fig.5 175/22 3 I . The method of testing for relative stability by employing the Routh array is shown by the following illustrative example. 5 ~ 8.e.p is made. The point s = . i.p.e. to find out how close the poles are to the imaginary axis.5s ~ + +3 =0 I I 8.The RouthHurwitz criterion 4.2. is stable or not. to find out how far away the system is from being unstable. i. Introducing the constant real value p. I . 4.p then occurs when r = 0.2 Imaginary axis shifting Testing of the characteristic equation by means of the Routh array can in fact be carried out if the substitution s = r .2.
1 and s = 0 points. the new polynomial equation is r3 + 2 S r 2 + 0. the system transient response analysis carried out in Chapter 3 also concentrated specifically on the transfer function denominator.  1 1 From the Routh array it is apparent that as only one change of sign occurs in the first column. 4. The importance of this polynomial is highlighted again in the remaining sections of this chapter. a particular response pattern will always be obtained.5 0. If. or more generally roots of a polynomial. Further.5r . such as those due to ageing or temperature drift. r3 r2 r‘ 1 ! 2.1) point. at least one parameter such as a n amplifier gain can be varied. although in a different vein from the previous section. by considering the resulting movement of poles.1. from the previous Routh array it was shown that no roots exist t o the right of the s = 0 point. The root locus method involves plotting the splane loci of polynomial roots caused by .3 Root loci construction The RouthHurwitz criterion is a method for testing the stability of a system by investigating merely the system transfer function denominator. Looked at in a slightly different way one can investigate the effect of parameter variations. The imaginary axis is now shifted from s = O to s = . and hence how the performance of the system is affected by certain values of that parameter. so only one root of the polynomial lies to the right of the r = 0 (s = .1 = 0 and although it is now apparent by the negative sign in the polynomial that at least one root of this polynomial lies to the right of the point r = 0 . The movement of poles. about the splane can be shown in a graphical sense by means of a locus drawn for each root. Also.102 System stability and performance A system with such a denominator polynomial is therefore a stable system because there are n o changes of sign in the first column of the Routh array. as is often the case. it is of great interest to see how the poles of the transfer function vary as the parameter varies. it is the position in the splane of the individual roots that is central to the discussion. When all system and controller parameters are fixed.1 by making the substitution s = r . thus the root in question must lie between the s = .9 I 05 . the Routh array will be formed for completeness. however. It was shown in the previous chapter how the transient performance of a system is dependent on the poles of the system’s transfer function.
It is also apparent from (4. A plot of the locus of each root as K varies then forms the basis of the root locus analysis method. in that for a particular closedloop system the root loci will commence at the openloop poles and end at the openloop zeros. The closedloop transfer function obtained when G ( s ) is placed within a unity feedback system is given by (4.3. in this chapter we will restrict ourselves to the case of the denominator roots belonging to a unity feedback closedloop system. such that the characteristic equation of this feedback system can be written as (a) A ( s ) + K B ( s ) = 0 or (b) 1 (4.3. It is readily observed from (4.3) that as + it follows that (4. B ( s ) .3. Although the root locus procedure is appropriate for the investigation of root positions when one parameter of any general polynomial is varying.3.5) . A ( s ) . whereas they will tend to those of the openloop numerator. as K varies from zero to infinity.as K 0.cs"' + b m .1) where K is an adjustable gain value.3) + G ( s )= 0 The position in the splane of the roots obtained as a solution to (4.Root loci construction 103 variations in a particular parameter. The technique gives important information with respect to system stability and performance without the need to continually solve the characteristic equation to find the root positions for a large number of values of the varying parameter. (4.4) / G ( s )= (2k + I)T (4.l ~ ' " .+ 01s + a0 *.3..3.3(a))that the roots of the characteristic polynomial will tend to those of the openloop denominator.l + + IS + bo) + '. This initial result in general forms a starting point for all root loci plots.3.3.3)will therefore vary as the gain K is varied.as K + 00. Consider then such a feedback system with an openloop transfer function: G ( s )= K B ( s ) = A (s) K(s" s" + all.2) Hence the closedloop denominator is A ( s ) + K B ( s ) .
n ) are the openloop i poles.. the definitions (4.6) In terms of the poles and zeros of the transfer function. Although it might be appropriate to perform careful plotting and multiple calculations around certain points of interest. the magnitude of the transfer function.3.3..7) we have that (4.4).8) i.. By means of the definition (4. ( s + Zm) (s+ PI) .. all of which are either real or occur in complexconjugate pairs. .3.e. Also.. (4.6) can be reconsidered by letting G ( s )= K ( s + ZI) .7) in which (zi: = 1. In general though it is not necessary to carry out the laborious exercise of plotting each locus with great accuracy by searching for all points which satisfy the angle and magnitude criteria. whereas if K < 0 / G ( s ) = 2k1r (4. and it is root locus construction by this means that is considered here.. So any point in the splane which is on a root locus for a particular system must satisfy (4.. it is possible t o construct the required root loci. .104 System stability and performance in which k = 0 .. which must equal unity.3. 2 2 . The magnitude condition (4. This is known as the magnitude criterion. the sum of the angles from the zeros to that point minus the sum of the angles from the poles to that point must satisfy (4. if there exists a point in the splane which lies on a root locus for a particular system.8). . then the value of K which results in a root at the point can be found from (4.. This is known as the angle criterion. for any point in the splane to lie on a root locus of a system.9) Hence.4)can now be written as (4. . m ) are the openloop zeros and (pi: i = 1. (s+ Pn) (4.. is proportional t o the product of the individual zero term magnitudes divided by the product of the individual pole term magnitudes for any value s. Armed with the magnitude and angle criteria. root loci are usually constructed by making use of cardinal points and asymptotic properties in forming a sketch of the loci.3.8) for that system. given a system transfer function. along with loci start and finish points.3.9).3.3.3.3.3...3.5) and (4. 2 1.
m by definition. Property 2 The total number of loci for a particular system is equal to the number of poles n of the openloop transfer function.3. Property 3 If K > 0. assuming that n >. If K e 0. ~~ Example 4. a similar result being true for K < 0 if even is substituted for odd.1 In order to illustrate some of the properties discussed thus far.1 Root loci: some basic properties 105 Property 1 Because the openloop poles and zeros are all rea. m by definition. consider the following example: G ( s )= (s K ( s + 2) + I)($ + 3) .Root loci construction 4. This property follows from the angle criterion. then for a point on the real axis to lie on a root locus there must exist an even number of finite poles and zeros to the right of the point. then if K > 0.3. and can be regarded slightly differently in that if there are no points which exist on the real axis to the left of an odd number of finite poles and zeros. the polynomial A (s)+ K B ( s ) can only have roots which are either real or occur in complex pairs. or occur in complex pairs. no part of a root locus exists on the real axis. The root loci will therefore be symmetrical about the real axis. assuming that n >. then for a point on the real axis to lie on a root locus there must exist an odd number of finite poles and zeros to the right of the point. Property 4 The number of loci which tend to infinity is given by the number of openloop poles n minus the number of openloop zeros m .
It follows from (4. the loci only exist on the real axis. 4 .1. Consider the point S = . i.4 from the zero at s = . i I K=w Fig.1.I is T .e. whereas the angle from the pole at s = . The latter end point can be confirmed by making reference to Property 3. and 0.3.10) in which 0.4 is the magnitude from the pole at s = .1.3 is O".3.6 is the magnitude to the point at s = .3. to the left of an odd number of poles and zeros.6 is the magnitude from the pole at s = .6) G(S)'=(0.10) that a value of K = 32/3 will result in a root on the locus at s = .3 . Notice that.1 and . .4 The angle criterion is satisfied because the angle to that point from both the zero at s = . then the two loci will end at s = . s = .1 The characteristic equation for this openloop system is given by: (S + l)(s + 3) + K ( s + 2) = 0 For K = 0 the two loci will therefore start at the openloop poles.9) that and from the magnitude criterion ' K(0. 4. as directed by Property 3.3.106 System stability and performance Imag.2 and the pole at s = .6)= (4. We then have from (4.3 .3.2 and s = . The direction of increasing K is shown by the arrows on the root loci of Fig.4)(1.1 whereas 1. 3 Root loci for Example 4.2.4. If we are just considering K > 0.
i.1 only one branch tends to infinity. states that n . The asymptotes do not emanate from the splane origin.Root loci construction 107 4.m 2 k ~ The total number of asymptotes is therefore equal to n . at an angle T .for either K > 0 or K < 0.3. (2k+ 1 ) ~ (4. the angle of the asymptote being given. and this does so along the negative real axis.3. by + 4k+l = n .( n .3.1).m = 2 so two branches (loci) will tend to infinity magnitude following asymptotes with angles 41 =z 7r and 42 = 7 3T .11) . but rather are centered on a point on the real axis given by (4. therefore n . in which real part is required. In Example 4.3.12) 4k+l =n .l and an] defined in (4.m .m in which k = 0.3.2 Branches and asymptotes Each separate root locus is also referred to as a branch.e. For complex poles/zeros only the Example 4..m of these branches will tend to infinity magnitude.7).1 = i= I 2 pi (sum of poles) respectively. given in the previous section. Whereas if K < 0. we have (4. and Property 4.3..2. 1. when K > 0.1) are also equal to bml = CI l i i= Ill (sum of zeros) and an..3.13) where b m . In fact for n > m each of the branches involved will tend to infinity magnitude by means of a straight line asymptote.3.m . as K 00.2 Consider the system transfer function G ( s )= (s K ( s + 2) + l)(s + 3)(s + 4)' K>Q such that n = 3 and m = 1. zi and pi are defined in (4.
9) a condition is found which determines the angle from any one specific pole or zero to the locus as it departs from or arrives at that pole or zero respectively.6. the angle of arrival of a locus to a zero at s = . The angle of departure of a locus from a pole at s = ..9). This can be interpreted as..(1 + 3 + 4) = .ZIis given by /(s + ZI = (2k + l)r + /s ( j= 1 + pj) .108 System stability and performance Fig.i = 2 /(s + zi) 2 (4. 4.4. in these two equations is not really of any great significance here. for K > 0.15) The index k = 0.3.3.3 2 A complete root loci diagram.3. 4.3. in the same fashion. is equal to (2k + 1 ) ~ . indicating the calculated asymptotes. the sum of the angles from each zero to a particular point on a root locus minus the sum of the angles from each pole to that same point.3.P I is given by /(S+pI)=k /(s+zi).14) I j=2 and.3 and these asymptotes are centered on a point at s=rJ= 2 .3 Angles of departure and arrival Previously the angle criterion was defined in (4. other than for completeness of the equations. 2 1.3. Simply by rearranging (4. .4 Angles of departure and arrival for Example 4.?/(s+pj)(2k+1)* I= (4.. is shown in Fig. It is probably easier in practice to .
4. Referring to Fig.3.14) be: e3 = ( . 4. = (0"+ 63. due to symmetry of the loci around the real axis.15)to be: .180' i.1. along the negative real axis.j ) ( s + 1 + j ) ' )s K>O To find the angles of departure and arrival. 82 is the angle of arrival of the locus at the zero at s = this is found from (4.360" or 8 . Example 4.43" Similarily.180" so e3 = .Root loci construction 109 assume k = 0 initially and then to take away the required multiples of 2a when an answer is obtained.3 34' . It is usually the case that breakaway points occur on the real axis as shown in Fig.43") (90"+ 90").3.2 + j .180" =6.e. = 63. only one complex pole and one complex zero need to be considered.4.j ) ( s + 2 + j ) (s + l( + 1 . and this to is found from (4.5.4 Breakaway points Breakaway points are points on root loci at which multiple order roots occur.45" + 45").1 + j .90" + 90") .14)to be: e.57" Finally 133 is the angle of departure of the locus from the pole at s = .( .3. and e2 = 180" or + (180" + 135" + 116.57" e2 = 161.3.3. 81 is that angle of departure of the locus from the pole at s = . they can therefore be considered as points at which two or more branches either join or split. .3 Consider the system transfer function: G(s) = K(s+ 2 .57") 90" = 360" + 161. 4. and this is found from (4.
noting that in order to be a breakaway point a solution to (4. and are The angle at which branches approach o r leave a breakaway point is dependent on the root multiplicity.3.1 10 System stability and performance Fig.remembering that symmetry about the real axis must be preserved.3). Example 4.3. is necessary to remember this latter point because not all solutions to (4. such that if r is the number of roots which exist at a breakaway point.5 Breakaway points on the real axis although this is not necessarily true.3. the branches approach with an angle between them of 360°/rand leave with an angle between them of 360°/r. it being possible for breakaway points to exist in complex conjugate pairs. breakaway points must satisfy all dK d ds=z [m] b)l =ds =O A(s) d B(s) (4.4 Consider the system transfer function: G(s) = (s K ( s + 2) + l)(s + 3)(s + 4) where B ( s ) = (s + 2) and A (s) = (s + l)(s + 3)(s + 4) Breakaway points for the root loci obtained from this system can be found by means of the solution t o ds (s + 2) .16) breakaway points.3.3).3.3. 4.16) It satisfy (4.3.16)must also comply with (4.16) as this gives the maximum and minimum values of K for which the roots are real.3) therefore not all solutions to (4. By reference to (4.3.
by reference to Example 4. 3 7 9 ) ( ~ 1. The fairly simple example considered here highlights the problem of finding a breakaway point by a purely analytical method. 4.(s + 2)(3s2 + 16s + 19) = 0 or s3+ 6s2 + 13s + 14 = 0 and it follows that: ( ~ + 3 .311 .379 back into (4.3.3.3. the problem becomes much more difficult and it may well not be possible to obtain a solution.6 Complete root loci for Examples 4.j 1 . However if the complex solution is substituted back into (4.311 + + j 1 .Root loci construction 11 1 v Real Fig. For a higherorder system therefore.3.3.3) a solution cannot be found. 5 5 8 ) = 0 On substitution of the solution s = . 5 5 8 ) ( ~ + 1. in that it is not an easy problem to find the solution to the thirdorder equation which results.6. 4.3. and therefore only the point s = 3.379 is a breakaway point Having calculated the breakaway point for this example.2 the complete root loci for the system considered are shown in Fig. In practice it is most likely that an approximate solution can be obtained by graphical means from the remainder of the loci. 4. an exact answer can then be found if necessary by employment of for example a NewtonRaphson iteration.406.2/4 such that 2(s3 + 8s2+ 19s + 12) .3) it is found that K = 0.5 Worked example In order to explain more fully how the root loci are plotted for a particular .
which are shown plotted in Fig. 5n 4 ’ 4 ’ 4 and 7n 4 . an extended example is considered in detail here.e.17) and this means that: The total number of loci is equal to n = 4.3 + = a 3a .7 Root loci for the worked Example 4.1 12 System stability and performance Fig.1 has an odd number of poles to the right. = (2k+1)a 4 for k=0. and as the number of zeros m =O. Only the real axis between the poles at s = 0 and s = . 4. The openloop transfer function is given by: G ( s )= K * K>O s(s + l ) ( s 2 + 4s + 8) ’ (4.5 (poles at s = 0.3. The number of loci which tend to infinity is given by n .1. the angle of each one being given by: 4k+l i.2 f j 2 ) system transfer function. Branches will tend t o infinity by means of asymptotes.2. and these will start from the four openloop poles.1 and s = . and therefore this is the only part of the real axis on which loci exist.7.+ 00. s = . 4.3. this means that all four loci tend to infinity as K .m .
4402.(2k + I)* = .2 . in which 2 is the number of roots which exist at s = . found to be s=(I= bm1 . The asymptotes are shown by broken lines in Fig.243.116. this means that they must leave at angles of n/2 and .(s4 d ds + l)(s' + 4s + 8)) = 0 + 5s3+ 12s' + 8s) = 0 +2 4 ~ +=0 8 4s3 + 15s' and the result of importance. is calculated to be s = .161. 4.57" .(2k + I)* = 738..j 2 ) . However.(2k + I)* 341.270" .43"= + 161.57' Similarly. the angle of departure of the locus from the pole at s = given by /(s . angles of departure of the loci from the complex poles at s = .5 . Since there are no finite zeros./ ( s + 2 .and as they arrive along the real axis this property is certainly true.0. Breakaway points for the root loci are obtained as the solution to: . which is expected from the loci already plotted.2 2 j 2 need to be calculated.Root loci construction 1 13 Also. where the latter is found from the openloop denominator s4 + 5s3 + 12s' + 8s.57" .43" .nrn 4 in which bmI = 0 and an. This means that the loci will arrive at the breakaway point with an angle 'IT between them .180" = .n/2 to the real axis in order to retain symmetry of the loci about the real axis. the asymptotes are centered on a point on the real axis. .4402 The angle at which the branches approach or leave this breakaway point is given by 360'/2.s(S that is d ds such that . there are no angles of arrival to calculate.90" ./(s + 1) . 5.j2 is + 2 + j2) = /(s + 0) .[ .I = 5 .2 + j 2 is given bY /(s+ 2 j2) = /(s+ = 0) /(s+ 1) /(s + 2 + j 2 ) .43" . Because the loci also leave with an angle 'IT between them.180" = .an1 . The angle of departure of the locus from the pole at s = .7.(2k + I)a = 225" .198.0.57' 6.135" .I . to lie on the real axis between s = 0 and s = .
3.5K110. having obtained the critical points/angles in (1)(7).17) is found to be: s4+5s3+ 12s2+8s+K=0 For which the Routh array is: s4 s3 S2 10.8).4 A system with such a denominator is therefore stable as long as K > 0 and 5K 8>0 10. 16. divided by the product of the magnitude from each zero t o that point. The magnitude criterion then means that for any point on a root locus.e.1 14 System stability and performance 7.18) i. Consider then the case when the two righthand complex roots are looked . Imaginary axis crossing: the points at which the loci cross the imaginary axis and the value(s) of K at those points can be found from the RouthHurwitz criterion as follows: the characteristic polynomial equation (closedloop denominator) for the transfer function (4.3. the gain K is equal to the product of the magnitude from each pole to that point.64 into the characteristic polynomial equation. the remainder of the loci are drawn with good accuracy.265 j 5 8.e. and these are the points we are looking for. results in s4 + 5s' + 12s' + 8s + 16.64 > K Hence when K = 16. see (4.3. by following the angle criterion. Substituting K = 16.4 I I S' So K where X = 8 . o n condition that.4 i. the value of gain K can in fact be calculated at any point o n the root loci.64 = 0 and this has roots on the imaginary axis of s = tm = ? j 1. As a n extension to the previous calculation. (4.64 roots will exist on the imaginary axis.
which means that these latter roots will tend to represent the system response as a whole. the effect of modifying or adding system poles and zeros can be witnessed by considering the effect the change will have on the root loci. without the need to follow all of the rules or to make accurate measurements. i. 4. t c o s .only a secondorder equation needing to be solved The real part of each of the roots denotes. how quickly their effect dies away. as K=4. These are fairly simply obtained. from this value of K . Because of this.153 fj 1.e. .Application of roof loci 1 15 at in terms of natural frequency and damping.973.7.' shown in Fig. 4. For example. which gives roots at and these lie at angles of t 60" (i. In the following section the application of root loci plots to specific problem areas is discussed. Hence. then +2 l ~ +nw ~ = S i ' + WnS + u'. let the damping ratio f S' = 0. the technique of root locus plotting is a very useful tool with which to investigate system transfer functions without the requirement of detailed algebraic calculations.e. in this case. Also the two other complex root positions are found. it is in fact more likely that an experienced control system designer (certainly one who has read this book) will be able very quickly to plot approximate loci for most lower order systems. and for this example the lefthand root pair real part values are much larger than those for the righthand roots.3.18) results in: r ) to the negative real axis. s2 + 2 f w n s + w . A calculation of (4. to be s = 2.695. The example looked at in this section highlights the main points to be considered when plotting the root loci for a system transfer function and provides a set of rules (1) to (8) which need to be investigated for root loci in general. from the closedloop denominator . in the time domain.114 with wn = 0.5. In this section the use of root loci in finding an approximate solution to the roots of any .4 Application of root loci Although a strict set of rules was outlined in the previous section for the construction of root loci.
2) which is in the same form as (4. The root locus procedure itself becomes quite difficult to carry out on high order transfer functions and this means that it is not really suitable in the analysis of high order polynomials.3. The roots for K = 20 are in fact s = .By reference to Section 4. it is shown how the root locus method is of great use in the study of system parameter variations and their effect on system performance in the sense of control system sensitivity.e.rule 5 . the advantage of the technique being to greatly simplify the procedure necessary.2. The root loci plot for the transfer function (4. Also the branches which tend t o infinity will d o so at a n angle of asymptotes given by ?7~/2.1) + 7 1 (4. it has certainly not been chosen t o show the root locus method at its best.4. Certainly it is fairly simple.1 S*(S + 7) (4.4.1 16 System stability and performance polynomial is discussed. the occurrence of a system pure time delay. s = 0 and s = 7.7)/(3 . i.and a center point s = (1.1. The explanation is given by means of an example polynomial.2. the figure 20 in the numerator of (4. If.3.9.1. is investigated in terms of its effect on the root loci of a transfer function.8. to obtain a very rough idea as to the complex/real nature of the roots and their approximate position.2 and s = .2) = . for which angles of departure .3 and + .4. Later in the section. The root loci plot shows that one of the three roots is real and lies between s = .4.2) is replaced by a general K > 0.2) sZ(s 7) (4.7.2.have also been calculated ( 27r/2). a root loci plot can be made for which the required polynomial solution exists when K = 20.3) with the gain K = 20. rule 3 shows that the only part of the real axis on which a locus exists is that between the pole at s = .4.9. with the remaining two loci tending to m as K a. 4. Further. which although it has been selected because of its exact integer root solution. Finally. Consider the polynomial equation s3 + 7s2 + 20s + 24 = 0 This can also be written as 20(s + 1.3) and this has a total of n = 3 loci which start at the poles s = 0.2. 4.1) = . ending at the zero s = .1. as is shown here.5. Let the transfer function under consideration be: G ( s )= K ( s + 1. therefore.7 and the zero at s = .3) is shown in Fig. a transport delay. the remaining two roots can be seen to form a complex pair with a real part lying between s = 0 and s = .2 .4.1 Roots of a polynomial The technique of root loci plotting is considered here merely as a method by which approximate solutions can be found t o the roots of a polynomial.
4.4. So.1) rewritten as: s= s(s2 + 7s + 20) . consider (4.. If T seconds is the transport delay. It would have been equally correct to set K to be any one of the polynomial parameters.2 k j 2 and it would be expected that with the approximate root values obtained from the root loci in Fig. the point is made that when obtaining approximate solutions to polynomials by means of root loci. 4. the use of (4. each one of the loci in this case tending to infinity magnitude. the parameter associated with the sterm.8. the exact solution can be found without too much difficulty.g.1) the gain K was arbitrarily chosen to be equal to 20. the lag can be represented in the s domain by means of the . and to obtain a different root loci from the alternative starting point. this time delay is also known as dead time or transport lag. the equation (4. From the starting point of the original polynomial (4. 4.1 24  (4.4.4. it is worthwhile selecting a K which simplifies the procedure as much as possible. to conclude this section.3) .4) rather than (4.2 Time delay systems A large number of systems exist for which there is a time delay between an input signal variation and the witnessing of an effect on the system output signal in response to that input variation. The three new loci then start at the poles s = 0 and s = . e.4.5 2 jd31/2.then a plot can be made of the resulting root loci.4. In the example shown.8 Root loci for the transfer function (4.4.3.4.4.2) was therefore preferred.4) such that if the numerator 24 is substituted by a gain K.Application of root loci 1 17 I \I I I Fig.2) would have necessitated the calculation and plotting of complex poles as starting points for two of the loci.
resulting in a n infinite . 2 1. ?2. the magnitude condition for this system becomes: (4.3s) U(s). as G(s) = G(s)exp( . expl .~ s ) A (s) ~ (4. s = u + j w . consider the problem of plotting the root loci for the openloop system characterized by the transfer function: Y ( s )= K expl ..3.4.. therefore the angle condition for this system becomes: (4.e.4. The term exp( .9) for K > 0 and in which k = 0.Ts).12) 1.4.Tsl = K B ( s ) expl . the final angle condition can now be written as / G ( s ) = (2k + I)* +wT (4. T = 3 s (4. Hence.10) Armed with the versions of angle and magnitude criteria applicable for systems with a time delay. Also the phase component of the term exp( .3s) = 0 (4. i.1 1) The characteristic equation for this system is: s(s + 2) + K exp( .4).4. hence the characteristic equation has a n infinite number of roots .7) Since I exp( .1 18 System stability and performnnce exponential function.1). it follows that (4.j w T ) I = 1.4.Ts) = 0 (4. in a modified form of (4.Tu) is zero.3. The closedloop denominator.6) or Because s consists of a real and imaginary part.4.e. s(s + 2) where K > 0. . resulting from this system then produces the characteristic equation: A(s) + KB(s)exp( .4.8) or I G ( s )I = exp( + T u ) which compares with (4. the characteristic polynomial..TS]can be written in terms of an infinite series in powers of s.4.5) where it must be remembered that it is assumed that unity feedback is to be applied and K is a n adjustable gain. i. The overall openloop transfer function of a system with a transport delay can therefore be written.
4..4.12) give s = 0 and s = . the angle of each one being R (when u = + O D ) . Fig. so for this problem loci exist between the poles at s = 0 and s = .4. Breakaway points for the root loci are obtained as the solution to + + =dK ds ds (s(s + 2)expI + Ts)] = 0 (4. and therefore an infinite number of branches tend to infinity magnitude.6). 4. the asymptotes are not in this case centered at points on the real axis + due to the asymptote angles being R.2 . Angles of departure and arrival do not need to be calculated here . but u = .Ts) written as an infinite series. + ~ R / T . as shown in Fig. the openloop poles. number of loci. there are an infinite number of openloop zeros. 6.9 Root loci for the transfer function (4. Branches will tend to infinity by means of asymptotes.4. 4. When K = 0.. m > n . 4.the only applicable poles being those on the real axis..10) with (2k + 1)n = w T . Existence of loci on the real axis is dependent on identical conditions to those given previously (w = 0 on the real axis). .4.1 I ) with T = 3 seconds . as this is the only part of the real axis for which an odd number of poles lie to the right.from (4. + 3 a / T . 5.13) as can be seen with the help of (4. when a = ?TIT. this is kSrr/T.4. the roots of (4.8) K 00 as u 00: from (4. With exp( .9.Application of root loci 1 19 2.00 also gives K = 0 and therefore further starting points when w = ? T I T . From (4.4.10) with (2k+ 1 ) R = OT.2. 3.. . as loci starting points.. .
4.j w we obtain s2 (4. . This can only be assured when 0 < K < 0.2. 7 . since the characteristic equation (4.8 5 J40 6 and this gives the required solution of a breakaway point at: s=  0. (c) It might be expected that the root loci would cross the imaginary axis a t values of w equal to the halfway point between associated asymptote values. 24.in order to retain symmetry. 4.680 K = 42.14) and we know from Fig. because at these points the characteristic equation must be: + w 2 = s2 + 2st K exp( 3s) = o and by substituting s = + j w or s = .384.333.4.925 stability properties are therefore given by the value of K corresponding to the imaginary axis crossing nearest the origin.16). For the closedloop system to be stable the characteristic equation roots must lie in the left half of the splane. etc.9 that the desired solution is expected to lie between the limits 0 > s > .279 The angle at which the branches approach or leave this breakaway point is given by 360"/2.13) that T e x p { + Tsj (4.451. K = '7.4.693.14) is s= .16) w 2= K cos 3w from which the following solution pairs occur: w= w= w= 20. With T = 3 seconds the solution to the quadratic part (4. It is worth noting from these results: For any one value of K there are an infinite number of possible solutions. (b) Solutions for K < 0 can also be obtained from the equation set (4. 8. 22.331. Points at which the loci cross the imaginary axis cannot be calculated by means of the RouthHurwitz criterion.164 K = 20.925 w = k6.4. This is (a) .4.15) (a) and (b) 2 0 = K sin 3w (4. K=0. The loci therefore leave with a n angle T between them which means they must leave at angles r / 2 and .4. For this example however it is fairly straightforward t o obtain both w and K values where the loci cross the imaginary axis.4.12) is not algebraic in terms of powers of s.r / 2 to the real axis .120 System stability and performance We have from (4.
then let us consider in the first instance its sensitivity to variations in gain K. usually a gain K.17) The root loci plotting procedure examines the variation in closedloop roots (poles) as one specified parameter.a .a to the gain K is then (4. 4.189].5 the sensitivity of a system output Y ( s ) to variations in the system’s transfer function G(s) was defined as: (4.e. the crossings occurring at w = k2..4.333. is varied. not necessarily the gain K. without too much difficulty.4.3 Sensitivity t o parameter variations In Section 3. etc.19) Note that here we are not interested in the fractional change in closedloop pole position. but merely the change in its position. If the root position under investigation is that at s = .4. although K at the imaginary axis crossing points can be calculated algebraically. It follows then that .18) can be written as (S + P I ).. and hence more accurately.Application of root loci 121 not the case however. the characteristic equation of the system (4.)] = (S + a ) X ( s )= 0 (4..331.4. where the unity feedback system openloop transfer function is (4.0941.18) The sensitivity of the root at s = . s + a is a root of the closedloop denominator.(S + z.4. i. as was done previously.20) in which X ( s ) is the remainder of the closedloop denominator. it is convenient to investigate the sensitivity of a particular root location (position) to variations in one system parameter or other. (d) The value of K can be calculated at any point on the loci by graphical means.. as was shown in the example discussed.4. and not w = f2?r/3 [2. Combining the ideas of root loci plotting and sensitivity analysis. etc. once the factor (s + a ) has been removed. 24.(S + P n ) + K [ ( s+ z l ) .4. Further. ?47r/3 [4.
G(s) (s aKIKX(s) I + G(s) .23) in which K is nominally equal to 4.4. and also the distance from each of the other closedloop pole positions (with openloop gain K ) to the root at s = a.22) can also be written as Hence the sensitivity function can be calculated directly from the root loci plot by finding the distance from each zero to the root located at s = .a . (s + z m ) l .22) s=..a .21) which results in SE = (s 1 + a)G(s) + G(s) (4. The root sensitivity function is thus redefined as: (4. It is necessary however to make a slight modification to the root sensitivity function before continuing.4. + where (s + a ) is also a denominator factor. The root sensitivity is then found as the gain K multiplied by the product of distances from each zero to s = .122 System stability and performance from which aK X(s) aa ( s + z I ).ol which is a complex measure whose magnitude and direction gives an indication of how that root position will change for a small change in gain K . hence the root sensitivity function calculated is only valid in the neighborhood of s = . Note that (4..a ... in that the differentiation procedure assumes that only a small change occurs in the root position (aa). on the . As an example. It follows then that the closedloop expression is I + ~ ( s ) .01 and divided by the product of distances from each of the other closedloop poles to s = . consider the openloop system transfer function: (4.4.( s + 2)(s+ 1 + j ) ( s + G(s)  4 I j) I f we wish to investigate the effect of a small change in gain K from its value 4.4. ( s + z m ) or  aa K [ ( s+ z l ) .
10 Use of sensitivity function on pole locations .a. that the variation in a is caused by the movement of a closedloop zero.4.10 in terms of an increase in gain K. This sensitivity function therefore gives an indication in the direction of / 4 ' terms of magnitude and phase of the increase in parameter value Q caused by an increase in gain K.e. the sensitivity function is merely negated when aK is in the negative direction. 4.24) s= 2 Fig. Because the closedloop pole is given by s = a. The effect on the pole position s = . i. the magnitude of the sensitivity function is the same as that for a small increase in gain. i. however. the sensitivity function shows the decrease in root magnitude and an angle equal to ?T + the phase change in root position caused by an increase in gain K. The discussion thus far has restricted itself to changes in a due to changes in the gain K.where s + a is a closedloop denominator factor.e. noting that the closedloop zeros are also openloop zeros.1 . in . It might be. however the angle becomes ?T + that for a small increase in K.will then vary slightly in the direction of .l ) .Q = .j. which is originally located at s = .Application of root loci closedloop pole located at s = .1 . The sensitivity of the parameter a. The corresponding closedloop root.j in the example is shown in Fig. to changes in a zero value Zi is given by (4.then we have S+' = CJ 123 or From this we can conclude that if K increases slightly then the parameter Q = 1 + j will vary slightly in the direction of /135". 4. assuming no polezero cancellations take place. For a small decrease in gain K.5 .(j.
a in response to variations in the openloop pole s = pi.1.4.Zi. Root sensitivity analysis is a very useful tool with which to investigate the effects on a closedloop root of variations in various system parameters.4. but to be near the imaginary axis. thereby affecting the stability properties of the system.. Finally in this section. one root is shown to lie in the left half of the splane. It is then possible to see how slight variations in certain of the system parameters could cause that root to migrate towards or even over the imaginary axis.4.q = Zi .(s  + P I ) .' .a in response to variations in the openloop s = .25) or St: s.. if the parameter pi increases from pi = 2 to pi = 2.a = 1.e. the sensitivity of Q to changes in the openloop pole value pi is considered by means of the sensitivity function s. i.1 then this means that an openloop pole moves from s = .2. = St: Pi Q noting that G ( s ) . The sensitivity function SFi denotes variations in the parameter a in response to variations in the parameter pi. It has particular application when from a root loci plot.25) It must be noted that the sensitivity function SFi denotes variations in the parameter Q in response to variations in the parameter zi.Ly (4.27) or s .2 to s = .4..a and this is also  s = a G(s) 1 1 + G(s) (4. which is = (4. Both parameters must be negated in order to measure the variations in closedloop pole s = . . (s + Pn) (S + pi)X(s) (s+Q) s= .26) .124 System stability and performance and this is which can also be written as (4. Both parameters must be negated in order to measure the variations in closedloop pole s = . Is= .
an appropriate openloop gain K. and the result of this is that a controller block must be connected to the original openloop system such that the overall closedloop transfer function can be tailored by means of choosing appropriate controller parameters.2) then by using Fig.36 in order to satisfy the desired maximum overshoot value. 8=53. such that when t = 0 . 3.5 Compensation based on a root locus 125 In this chapter it has been shown how the root loci of a system transfer function can be plotted in the splane by means of following a simple procedure in the form of a set of rules. The steadystate error following a unit ramp input is l/K.167 and K < 17.1) It is required that the steadystate error to a unit ramp input is no more than (a) 33% or (b) 5%. in many cases the performance specification cannot be met by only a choice of gain K. such that overall closedloop denominator root loci are shifted to ensure that they follow predefined paths at specific points. fi=3.1).5. = lim sC(s) = K/5 s0 such that for condition (a) K. Consider the openloop transfer function: G ( s )= s(s 5 ) K + (4.Compensation based on a root locus 4. When the two roots are complex.5 k j f i .33 and K=17.5. Because { = 2. The root loci for the system (4.167. It is then a fairly straightforward task to employ a root loci plot in the desig of a desired control system in order to meet a performance specification.3. where fi2 = . Further there must be no more than a 10% overshoot of its final value in the unity feedback closedloop system’s response to a step input. For any {>0.10 as a guide. It is shown directly here how it is sometimes possible to achieve the required design objective merely by selecting. and Wn cos B = 2. where + K .6 then a n < 4. (4. K.2) are drawn in Fig. It can be noted that if the closedloop function is written as (4.5.2 S 2 w’. 6 .= K/5 > 3 or K > 15 .5. Later in this section it is shown how controller parameters can be selected by employing information obtained from a root loci plot. s = 2.13”. wn=4.6 in order to achieve a 10% = overshoot at most. A controller design is carried out here to show initially that it is sometimes possible to meet certain closedloop system performance objectives simply by a selection of system openloop gain K. 4.5/wn9 it follows that { = cos 8 or B = COS’ l. also we have that 2{wn = 5 and a’.11 by following the rules presented in Section 4.36: K must therefore lie in the range 0 < K < 17. from the closedloop denominator root loci.36. However. it follows that { 2 0.5.
4.36. = K / 5 > 20. 4. Fortunately. as shown in Fig. as long as 15 < K < 17. The problem is made worse if further performance objectives are stipulated and/or the openloop system is of higher order. a cascade controller consisting of one pole and one zero is suitable for many control purposes.12.1 Firstorder compensation Attempting to meet two o r more control system performance criteria with only one controller parameter. and to replace it with a pole at s = .11 Root loci plot for G ( s ) of (4. as far as at least one of the criteria is concerned. it is often the case that when a system is of a higher order its response can be approximated with reasonable accuracy by means of a low (first.1) I . the gain K . For the steadystate ramp error requirements made in condition (b) it is necessary that K . which means that K > 100. In theory it is then possible t o simply cancel a n undesired openloop pole of G ( s )with the zero at s = .a.b. As the majority of systems to be controlled can be adequately represented by a model of order n o higher than three. in the form of other selectable parameters must be used. 4. and hence controller complexity is not necessarily increased by any great extent .5. second or third) order model.126 System stability and performance (Km jw K=O L + L l s=o a. it being assumed in general that a and b a r e positive definite real scalar values.+K Fig. will in many cases result in an unsatisfactory final selection. It is therefore not possible for both the maximum overshoot requirements and the maximum ramp error requirement (b) t o be met simply by a selection of openloop gain K . the first being that there is not always a n obvious .5. I Hence a K can be selected which meets both the maximum overshoot requirements and provides a maximum ramp error of less than 33%. it is a consequence that in addition t o a selection of forward gain K . There are however two drawbacks to this seemingly sensible idea. and therefore in this case further controller flexibility.
Note that for all s” with w < 0 we have /(s+ a ) >/(s+ b ) and thus the phase of s” is shifted in the negative phase angle direction . points s’ and s” are considered to be two points on an originally uncompensated root loci plot. The cascading of a compensator of the form (s b ) / ( s+ a ) can therefore be viewed in terms of how the compensator affects the original.13.a result which can be deduced from the symmetrical nature of the loci.5. in which case any cancellation of an openloop pole with the controller zero will itself not be an exact cancellation.2 Phase and magnitude compensation The compensator introduced in the previous section can be employed effectively to alter the system phase and/or magnitude characteristics by means of a root loci plot. This can best be seen by making reference to the angle and magnitude root locus criteria considered in (4. uncompensated. 4.Compensation based on a root locus 121 Compensator Openloop system Fig. I2 Lagllead cascade compensator relationship between the openloop system/controller polezeros and the specified performance objectives made in terms of the closedloop system./(s + a ) . Although this may not be serious when the openloop pole in question lies in the left half of the splane.9) and (4. 4.13.5. 4. The compensator is known as a lead compensator because of this phase shifting property. Hence for the compensator with a > b it follows that for all s’ with w > 0 we have /(s+ b ) >/(s+ a ) .8) respectively. Regarding Fig. and thus the phase of s’ is shifted in the positive phase angle direction as shown by the arrow in Fig. if the pole lies in the right half of the splane a controller zero will be introduced which also lies in the right half of the splane and this would result in an unstable mode being present in the closedloop system. The second drawback is a practical one in that it is not often the case that an ‘exact’ representation of the openloop system is available.3. An example of the effect of a lead compensator is shown in terms of the openloop system transfer function K (4. and the compensator has been designed such that + a>b>O. loci in terms of magnitude and phase.3. 4.3) G ( s )= /sl= /s‘ (s + 3)(s + 5) . In terms of the angle criterion the phase of the point s’ is affected by the compensator polezero pair by means of the relationship original +/(s + b ) . the effect that the compensator has being dependent on the magnitude of the coefficients a and b.
when w > 0 . 4. 4. This is due to the fact that when b > a.14(b). we have / s + a ) > /(s + 6 ) and ( original +/(s + b ) /(s + a ) . points s' and s" are two points on an originally uncompensated root loci plot.13 Effect of a lead compensator with a lead compensator whose transfer function is s+ =.14(a) whereas those with the compensator applied are shown in Fig. The compensator is known as a lag compensator because of this phase shifting property. 4.128 System stability and performance a Fig. An example of the effect of a lag /sl= /s' S Fig. Regarding Fig. Note that for all s" with w < 0 we have /(s+ b) >/(s + a ) and thus the phase of s'' is shifted in the positive phase angle direction.14 Phase shifting with a lead compensator . is shifted in the negative phase angle direction as shown by the arrow in Fig. and the compensator has been designed such that b > a 2 0. 4. The angle criterion then means that the phase of the point s'. IS.b s+a s+l s+2 The uncompensated closedloop system root loci are shown in Fig. 4.15. 4.
4. but well to the right of any real axis breakpoints.Compensation based on a root locus 129 / S= b \ " f ' i n s= .16(b). The effect of a compensator selected for steadystate error magnitude conditions can be seen by reference to the example initially considered at the start of Section 4.5.14(a)) whereas those with the compensator applied are shown in Fig. that given in 1Kw I jw / s= 4 s= 3 I s = 3. 4. Hence both the compensator pole and zero must lie to the left of the splane imaginary axis for stability. i.e.5 U Fig. 4.a \ \ I I Fig. 4.b s+a s+2 s+l The uncompensated closedloop system root loci are shown in Fig.3) with a lag compensator whose transfer function is s+ =.16 Phase shifting with lag compensator .15 Effect of a lag compensator compensator is shown in terms of the openloop transfer function (4. 4. Magnitude compensation by means of root loci plots is generally not a complicated procedure and is usually a case of selecting a polezero pair such that steadystate error conditions are achieved without seriously affecting the system's closedloop transient response.5.16(a) (also in Fig.
but nevertheless it can readily be seen that the complex loci are very little affected and hence the transient response is not significantly altered. 4. 4. 5 . In particular Fig.5 (see Fig. the original breakaway point was at s = .130 System stability and performance (4.Ol is shown in Fig.2 lie well t o the right of s = . 4. where or If it is remembered that we needed K < 17. then selecting K = 10.0. so let us select b t o be 0.36 t o satisfy the overshoot conditions (less than lo%). =  a or b > 100 Further.1).2 s+O.17 in order to show the loci movement. we have that 2b 20 < K . 4.17.01 and zero at s = .1 l ) .17 Compensated root loci plot for G(s) = K / s ( s +5 ) . The new root loci plot with a lag compensator s+ bs+ a s+O.01 such that the pole at s = . The effect of the compensator polezero pair is magnified in Fig.2 and a = 0.2.0.5. A requirement is that the steadystate error to a unit ramp input is no more than 5 % and it was shown that use of a simple gain K could not satisfactorily achieve this objective whilst also trying to meet maximum overshoot conditions. The steadystate error following a unit ramp input is l/K.2 .
by means of the RouthHurwitz criterion. which is used to test for system stability.187s + 8. that if the system is connected .233 and hence 131 s2+ 2 f w n ~ 02.997 and f = 0. By employing a computer it is not too difficult to obtain a complete root loci plot for a given transfer function simply by plotting the root positions for a large number of K values. it may well be necessary. A graphical display of relative stability properties is given by plotting the root loci of a closedloop system transfer function denominator. Often it is not in fact necessary to resort to a full root loci plot.and closedloop system transfer functions. I For an openloop system transfer function G ( s )= s(s3 + 6s2 + 11s 1 + 6) show.2 does not increase as the order of the polynomial increases. whose characteristic equation factors (roots) are not readily obtainable. as it is usually the case that a sketch of the loci can be rapidly drawn. it merely means that more calculations are necessary. Problems 4. and although for many systems the simple leadllag type of compensator is of sufficient complexity. 4.Problems when K = 10 a root lies at s = 0. to employ a multipole/multizero compensator. In this chapter it has been shown how the root locus technique can readily incorporate systems which exhibit a transport delay and how the roots of a system are sensitive to parameter variations. This is a major selling point for the method because higherorder systems. can be tested in a relatively straightforward manner to show their stability and even relative stability properties. although problems can exist for higherorder systems.6 value required for a 10% overshoot maximum.984 Thus a n = 2. with only one or two points being required to any great accuracy.6 Summary The topics included in this chapter have been concerned primarily with the study of closedloop denominator characteristics. = S’ + + 4. as the forward gain K is varied. Also the possibility of designing a controller by making use of root loci plots was introduced in Section 4. It must be said however that the RouthHurwitz criterion.799 which is certainly greater than the 0. This means that with very little effort the steadystate ramp error requirement has been satisfied and this has been possible while also satisfying the maximum transient overshoot condition in response to a step input. The complexity of the RouthHurwitz stability testing method described in Section 4.5. in order to tailor the final root loci to meet a particular performance objective. is merely testing the roots of a polynomial and is therefore applicable to both open.
l 0. Determine the conditions for which the feedback control system.4 Find the range of values for K over which the closedloop system is stable.5(3s+ 1)  Fig. of gain KI. A unity feedback control system has an openloop transfer function: G ( s )= s(s K ( s + 2) + l)(s2 + 2s + 2) 4.132 System stability and performance Fig.in which K I is a scalar gain value. 4.18 4.5 Determine whether or not the systems with the transfer functions given are stable.l ) ( s + 2) 4. 4. where G ( s )= 1 (s + 3)(s .2 up with an amplifier. with the block diagram shown in Fig. is stable.19.02 . 1 KI 20 s+O. in the unity feedback loop (Fig. 4. K2  1. Find the closedloop transfer function and find the ranges of values for KI over which the closedloop system is stable. 4.18) then the closedloop system will be stable for 10 > KI > 0. (a) (b) s+ 1 ss + 3s4+ 4s3 + 4s’ 1 + 3s + 1 ss + 2s4 + 6s3+ 10s’ + 8s + 12 s5 + 2s4 + 3s3+ 6s’ + 10s + 15 1 . Determine whether or not the openloop system described by the transfer function G(s) is stable.3 A negative feedback loop is applied with transfer function H ( s ) = K I .19 4.
for what values of K will the closedloop system be stable? . 4. 4.1 .6 Find the number of roots of the polynomial equation s3 133 + 2s2 + i i s + 20= o 4.8 Find the range of values of the gain K for which the unity feedback closedloop system is stable when the openloop transfer function is: (b) 2(1 .10 Determine the range of values for K over which the closedloop system will be stable. Fig.20.Problems 4.7 which lie between s = 0 and s = . Find the conditions for which the feedback control system.9 The openloop transfer function of a unity feedback system is: G ( s )= ~ ( 0 . 5 s 1)(2s + 1 ) + K 4. then the stability range will be increased. with the block diagram as shown in Fig.s)K G ( s ) = 2 s + s2 + 8s+ 1 3 4. The openloop transfer function of a system is given by: G ( s )= 1 s(s + l)(s + 3) If a variable gain controller whose transfer function is is cascaded with G ( s ) and a unity negative feedback loop applied.20 4. Show that if the feedback is made equal to (s + 3) rather than unity. is stable.
15 and find the value of K which gives the complex closedloop pole pair a damping ratio of { = 0. Draw the root locus diagram for the system G ( s )= ~ K s(s + 4) How is the diagram affected. what happens when the zeros lie at s = .l( + I) )s 4.9.13 Draw the root locus diagram for this system and find the range of values for K over which the closedloop system is stable.14 and show the exact breakaway point values.5. find values for K and b which .9? A unity feedback system has an openloop transfer function G ( s )= 2.9 what effect does this have on the diagram? Similarly. Draw the root locus diagram for the system G ( s )= ~ K(s + 2) s(s+ 1) 4.11 System stability and performance Draw the root locus diagram for a unity feedback system whose openloop transfer function is G ( s )= K ( s + 9)2 (s+ 113 4.5K(s + 2) (s . when a cascade compensator of the form s+2 D ( s )= s+6 4. by means of a root locus diagram. Draw the root locus diagram for the system G ( s )= K s(s+ l)(s+ 2) 4.16 is employed within a unity feedback loop? Consider the openloop system transfer function G ( s )= lOK s(s + l )(s + 4) If a cascade compensator of the form s+b D ( s ) = __ s+4 is employed.134 4.8.12 If the openloop zeros lie at s = .1 rather than s = .
Draw the root locus diagram for the openloop system < a < 1.20 Draw the root locus diagram for this system and show how. with particular attention paid to stability properties. by a suitable choice of cascade Compensation.19 s+ 1 s + 10‘ By gain compensation only. (iv) G ( s )= s(s2 + 4s + 8) K when employed within a unity feedback loop. (iii) b a < b < 1. (ii) 1 < b < a.Problems produce the dominant closedloop pole pair s= 1 k f i 135 4. Consider also the cases: (v) a < 1 < b and (vi) b < 1 < a. a faster system step response can be achieved. what happens if the compensator is altered to D ( s )= 7 4. What is the effect on the root loci when a cascade compensator of the following form is applied D ( s )= s+b s+ a a.b>O? 4.17 Draw a root locus diagram for the openloop system 1 G ( s )= s(s+ 1) when employed within a unity feedback loop. What effect does the employment of such a compensator have on the system’s stability properties? Consider the openloop system transfer function G ( s )= s(s + 10) ’ K K>O Plot the root locus for this system when the system is placed within a unity .18 Consider four cases: (i) 1 < a < b . it is desired to control the position of a radio telescope whose openloop transfer function is G ( s )= (s+ 7)(s + loo) * K K>O 4. What is the effect on the loci when a cascade compensator of the form s + 10 D ( s )= s+ 1 is applied? Conversely. within a unity feedback loop.
136 System stability and performance 4. Electronics Con$.. 1970. 1991. Automatic and Manual Control. ‘Application of the electrolytic tank to servomechanisms design’.. A. f Nelson. .. J. W. L.. Feedback control systems. Dynamics o a system o rigid bodies. ‘Control systems synthesis by the root locus method’. Analysis and synthesis o linear active networks. D.. W. McGrawHill. 6th ed. Trans.. Richards. R. W.. IEEE Trans. and Boothroyd.. L. W... Flight stability and control.14 with regard to a n increase in gain K (when the damping ratio is f = 0.21 feedback loop. Process control instrumentation technology. 3rd ed.5). Longman.. 21. pp. Systems analysis and design. 1964. 1957. ‘On conditions under which an equation has only roots with negative real parts’. L. R. Wiley. Dover. Automatic control systems. pp. Reid. AIEE. Wiley. C. 1892. AddisonWesley. 1964. An introduction to dynamics and control. ‘A generalized root locus following technique’. Macmillan. and McDaniel Jr. ‘Saddle points of a complete root locus and an algorithm for their easy location in the complex frequency plane’. Introduction to control theory. Johnson. W. J. on Automatic Control.. R. 1990. R. 1991. 67. J. W . Find a suitable cascade compensator which will ensure that both (a) the closedloop step response has a maximum overshoot of no more than 20% of the final output value and (b) the steadystate error to a unit ramp input is less than 5 % . 69. ASME J. Proc. Jacobs. Modern control systems. AIEE. A. R. 1954. PrenticeHall.. 1969. PrenticeHall. Natl. J. J. McGrawHill. Westcott. 1992. Ser. C. Dickenson. and Harbor R. McGrawHill. Remec.. 1974. Dorf. f f Routh. C. 1948.. M. G. C. C. Mitchell.. 1991. Butterworth. Trans.. Basic Engineering. H. Control system dynamics.. Find the sensitivity function for the dominant poles in the closedloop system of Problem 4. D. K. R... 0. AC15. Kuo. 6058.. pp. Evans. S. 1950. New York. Modern control theory.. p. R. B. Mitra. E. R. Evans. Phillips. PrenticeHall.. 54751. 1989. J. Oxford University Press. Evans. 1979. R. 4835. C. ‘Graphical analysis of control systems’. PrenticeHall.. 1977. Linear systemfundamentals. Wojcik. Hurwitz. 66. 1983. 1965. B. ‘Analytical representation of root locus’. Trans. 86. in Selected papers on mathematical trends in control theory. How does the result compare with the sensitivity function obtained with regard to a decrease in gain K? Further reading Brogan.
e. Each of these. the steadystate output signal will also be a sinusoid of the same frequency w. Frequency response data values for system gain and phase. see Problem 5.Frequency response analysis 5.180". may be obtained either by direct analytical means when an exact transfer function is known.7. by setting s = j w .180" is still critical as far as stability is concerned as is thoroughly discussed in this chapter. and the phase shift 6 will be the phase / G ( j w ) of the transfer function at the frequency w . Once the data is obtained however. i. for the most part a variation in the openloop gain K. So. A short example is given at the end of this section: see Example 5. This is termed the frequency response of the system G ( j w ) when values for I G ( j w )1 and / G ( j w ) are found for all frequencies w . In Section 3. which is of magnitude U . s = j w . of magnitude Y. The output magnitude Y will be the result of the input magnitude U multiplied by the gain I G ( j w ) I of the transfer function at the frequency w .1 Introduction In Chapter 4. the point at which the system has unity magnitude and a phase shift of . then for a stable system the steadystate output signal is also a sinusoid. it is found that Y = I G ( j w ) I U and 6 = / G ( j w ) . s = u + j w . In fact. techniques makes use . there are several available techniques which can be employed for system analysis and subsequently for compensator design.2 it was shown that for a sinusoidal input signal u ( t ) = U cos w t . an important case arises when u = 0. i. essentially graphical. Remembering that each location in the splane consists of a real and imaginary part. i.e. An alternative approach is to set the gain K as a constant value and to consider the magnitude and phase of G ( s )as they vary in response to variations in the roots of G ( s ) .2. y ( t ) = Y cos(ot + 6 ) where 6 denotes a phase shift. in response to a coefficient variation. at many frequency points. where I G ( j w ) 1 is the magnitude of G ( s ) when s = j w . if a sinusoidal input signal of frequency w rad/s is applied at the input of a stable system. A fundamental assumption is made that the system openloop transfer function G(s) must be of unity magnitude and must have a phase shift of . control system analysis and design is carried out by investigating the movement of closedloop roots about the splane.1. or by taking experimental results. In terms of stability analysis though.e.1.
However. each of them has properties which make them useful under certain circumstances. such that 3 G ( j o )= jw + 2 then which produces the data in Table 5. such that the forward path becomes D ( j w ) G ( j w )for which a frequency response plot can be obtained.061 10 m I G(jw)l /G(jw) 0 45' 0. 1 . Where a compensator D ( j o ) is also employed. Nyquist and Nichols. even though for a particular case study one of the approaches will probably be much more appropriate than the others.5 2 1. Example 5. Table 5. merely represent different ways of looking at the same system characteristics. in that the frequency response o f t h e system openloop transfer function G ( j w ) is plotted.7O 0 90" .294 78. namely Bode. but presents the results in a different way. the three design techniques described in this chapter.1. Stability results then apply to the unity feedback closedloop system with D ( j w ) G ( j o )in its forward path. In a certain sense. and therefore it is worthwhile becoming familiar with the different approaches. I Frequency response data for Example 5 . Stability results obtained from each technique then apply to the unity feedback closedloop system which contains G ( j w ) in its forward path.1. therefore.138 Frequency response analysis of the same data values. 1 w 0 1.1 Consider the transfer function: 3 G ( s )= s+2 The frequency response of this system is obtained by setting s = jw. it is considered here that this is connected in cascade with the system openloop transfer function. All three frequency response methods discussed in this chapter have a common element.
if the exact nature of the modification is not known. the output of the system will itself be a sinusoid in the steadystate. Even so. A system transfer function can therefore be characterized in terms of its frequency response considered in the form of two plots. as can be seen in this section. a much simpler process to carry out graphically when compared to multiplication. whose magnitude varies sinusoidally with respect to time. Bode employing the logarithmic value of gain plotted vertically on a linear scale against a horizontal nonlinear logarithmic frequency scale. rather than output magnitude which is.The Bode diagram 5. modification of the system by changing the steadystate gain Kor adding a further pole. a similar procedure being carried out to form a separate plot of the corresponding phase shifts. although its magnitude will be some multiple of the input signal and there may be a phase shift. the latter course of action may well be necessary. If the magnitude of the input signal is held constant and a different frequency of signal is chosen.1 Logarithmic plots For both Bode gain and phasevfrequency plots. Indeed. which is not dependent on input signal magnitude. will mean that all of the values piotted must be multiplied by a relevant factor in order to avoid having to take a further series of measurements. A much better approach is to use logarithmic values of gain such that any modification can be regarded as an addition. when combined with corresponding plots of phase shift on a linear axis against logarithmic frequency. for the design of feedback amplifiers much work was carried out by H. the vertical gain or phase axis is linearly scaled whereas the horizontal axis is drawn to a logarithmic frequency scale. in response to a sinusoidal input signal of frequency w radls can be written as (5. W. 5.2 139 The Bode diagram When an input signal. The system transfer function. For a constant magnitude input signal the resultant output magnitude or gain can therefore be found for a wide range of frequency values and a plot made of the different values. If a complete gainvfrequency plot has been obtained over a wide range of frequencies for a particular system. Hence the frequency response can be presented over a much wider range of frequencies than would be possible over a linear scale. are therefore referred to as Bode plots. Such frequency response plots. is applied to a linear system which is stable. whether by a series of measurements or by calculations from a transfer function model. then another output magnitude and corresponding phase shift will be found. noting the employment of gain.2.2.1) . one of gainvfrequency and the other of phasevfrequency.
2. 5.3. Gain (dB) 20 log1oK * 0 0" 1 2 c Fig.2.3) Example 5. the negative sign indicating a gain of less than unity.02 dB.2 Consider the system transfer function: G ( j w )= Kjw (5. consider I G ( j w )I = 2 which results in a gain = 6.2.140 Frequency response analysis 0.2.1 Bode plots for the system (5.2) gain (dB) = 20 loglo I C ( j w ) I where the gain is written in terms of its standard unit the decibel (dB).01 dB. which can also be regarded as a loss or attenuation. or if 1 G ( j w )I = l / & then the gain = .4) . such that for any frequency the gain can also be expressed in logarithmic form as: (5. As a simple example.
Similarly. where the slope of the gain plot is given as + 20 dB/decade.2. . w=1 w = 10 . phase shift is the + 180" over all frequencies.2. 5.7) +90" r oo . I + 1 0 2 loglow .1.6) it follows that the phase will be [+go"] x p over all frequencies and the gain: gain (dB) = 20 loglo K + 20p loglo w (5.2. For a system transfer function of the form G ( j w )= K[jwI2. for K > 0. and whose (5.The Bode diagram The phase of which will be logarithmic gain is: 141 + 90" over all frequencies.20 gain (dB) = 20 loglo K + 0 gain (dB) = 20 loglo K + 20 The Bode plots for the system (5. but the straight line will have a slope equal to + 40 dB/decade.4) are shown in Fig. gain (dB) = 20 loglo K . when K > 0.2. that is the gain increases by 20 dB every time the frequency w is increased by a multiplying factor of 10.5) gain (dB) = 20 loglo I Kw 1 = 20 loglo K + 20 loglo w such that when w = 0.2. for a general system transfer function of the form G ( j w )= K[jwIP (5. with gain (dB) = 20 logloK + 40 loglo w Hence the gain plot will still pass through the point 20 loglo K when w = 1 rad/s.
2.8) are shown in Fig.3 Consider the system transfer function: G ( j w ) = K [ j w ] .2.8) The phase of which will be . which means that for p negative this will result in a negative phase shift and a straightline plot with negative slope. 5. Example 5. .6) in fact hold for all integers ( p = 0.3 Bode plots for the system ( 5 . 5.20 loglo o (5. as is shown in the following example. ? 2. when K > 0.2. 2 1.).20 dB/decade.3.142 Frequency response analysis such that it passes through the point 20 loglo K when w = 1 rad/s and the straightline plot has a slope of + 20p dB/decade.2. and whose logarithmic gain is: gain (dB) = 20 loglo K .90" over all frequencies..' =K / j w (5. 8 ) . 2 . where the slope of the gain plot is . The phase and gain results for the system (5. Gain (dB) t I 0 1 t I 2 loglow Phase t Fig..2.9) The Bode plots for the system (5.
The two gain values are in fact both equal to 20 loglo(K/a).2.2.01 dB Hence the true response is approximately 3 dB below the asymptotic sketch at the point of largest error. the phase shift exhibited by the system is 90' and the gain is: gain (dB) = 20 loglo [ K / a l .11). at which point G(jw) = Kla j+1: w=a The phase shift exhibited by this transfer function when w = a is 45". is shown in Fig. known as the 3 dB point. so the effect of each individual pole or zero can be found separately. Consider the system transfer function: G(jw) = (jw + a) K (5. although it is occasionally also referred to as the corner frequency.10). The frequency at which this occurs. when w 4 a. An important feature to be aware of is that because logarithmic gain plots are being obtained.2 Asymptotic approximations 143 Thus far we have merely considered Bode plots for systems consisting of a gain K and either poles or zeros which lie at the origin of the splane. the effects simply being added in order to obtain the total system frequency response.10) To simplify the analysis of this system.12) where the a term has been used as a denominator of both K and w in order to link (5. with an associated gain of: gain (dB) = 20 loglo [K/al . the worst error occurring when w = a.The Bode diagram 5.10).2. its response will be considered at high and low frequencies in the first instance. such that an asymptotic approximation of the complete frequency response for the system (5. when w a a. for very high frequencies. is known as the break frequency.12) with (5. when w = a.4.12).2. In order to generalize the results it is necessary also to consider firstly systems with poles and/or zeros which lie along the splane real axis and secondly systems with complex roots.and lowfrequency asymptotes it is then only necessary to ensure that the actual plot passes through the 3 dB point.2. .2. For very low frequencies. The true response is in fact very close to the asymptotic approximation. 5.1 1) However. the phase shift exhibited by the system is 0" and the gain is: gain (dB) = 20 loglo [ K l a ] : w4 a (5.3.20 loglo [w/al: w9a (5. can be employed such that in order to quickly sketch the actual frequency response gain plot for a system such as (5.2.20 loglo[&] = 20 loglo [ K / u ].when w = a in (5. This point.2. in this case w = a.2. having obtained the high.
14) : oQ a However. and is sufficient for many purposes.lu w=u w = IOU w = IOOU b loglow system phase curve 90"  Fig. noting that when p is negative this will result in a positive phase shift. .2.10) can be generalized by considering a system of the form: G(jw)= K [ j w + alp Q a.90" x p . + 2 .response curve dB Phase A w=O.e.2..a. 2 1. This straight line approximation is never more than 6" away from the actual phase plot.. with a gain of: (5.is known for both low and high frequencies and also when w = a . 5. when w %.2.2. the break frequency. for very high frequencies.10). when w gain (dB) = 20 loglo F P I the phase shift will be 0".. i.. 5. the line therefore also passes through the point w = a.144 Frequency response analysis 20 login(K/a) 1 L w=O.20p loglo ["aI : w 9a (5. The gain is then: gain (dB) = 20 loglo . also shown in Fig. It will be seen later though that it is sometimes necessary to find a more accurate plot of the phase characteristics and in certain cases exact calculations employing several values of w are required.10) The phasevfrequency plot. For very low frequencies. the phase shift is .I.2.4 for the system (5..15) . (5. The results given for the system (5.4 Bode plots with asymptotes for the system (5.13) where ( p = 0.lu w=u w = IOU w= IOOU 1 * logiow t3 .2. A simple approximation to the actual plot can be obtained by joining together the plot at w = a/10 with that at o = 10a.
of those given in Fig. Example 5. * I 0 ' 1 2 loglow 20 .1 6) which can also be written as G (j w ) = 3 (1 10. is therefore unaltered by the integer p. w = a..2.The Bode diagram 145 The break frequency. (1 +j43) w 4 2.4. although the slope of the gainvfrequency plot will be .20p dB/decade for all w > a. Hence for p = .b .2. the phase shift exhibited by the system is O". Gain 20 10 . 5.1 the Bode gain and phase plots will be simply a mirror image about the horizontal axis.4 Consider the system transfer function: G (j w ) = 5( j w + 2) + jw/2) t j w + 3) (5.3 For very low frequencies the phase shift exhibited by the system is 0 and the ' gain is: gain (dB) = 20 log10(10/3): However for very high frequencies. I I ] 20 log105 0 10 5 .20 log10  .
A special case must be treated in a slightly different way. and that is the effect of complex roots. The Bode gain plot for such a system is shown in s Fig. in which 0 2 > 01. Consider for example a system transfer function with unity numerator and a denominator equal to (s + @ ~ ) (+ Pz).6 Individual term and overall Bode phase plots for system (5.2. the individual gain and phase frequency responses can be obtained separately and simply added to give the overall response as shown in Figs 5.20 dB/decade over the frequency range between the .3 Complex roots If a system exists in which two roots are real. 5.16) . 5. 5. The slope of the Bode plot is . as was done in the example.5 and 5.6.2.146 Frequency response analysis I I I 1 I Phase t I ! Fig. This procedure of summing the individual pole and zero plots can be carried out for a system with any number of poles and zeros. the effect of these roots can be treated separately and merely added to find the overall Bode plot.7.20 lOglo(w/3) Armed with these checks on the low and high frequency values. although the gain must be calculated at high and low frequencies.asymptotes only shown: note change of scale while the gain is: gain (dB) = 20 log10(10/3)+ 20 log10(w/2).
= 0 K (5. 20 l o g l o ( K / w : ) .e. a simple but rough approximation being to consider the plot to be that pertaining to the case when two real and equal roots exist. From ( 5 .40 dB/decade.18) +2 l w n+ ~ Asymptotic approximations of the frequency response for this system cannot be produced with any great deal of accuracy for all 0 < { < 0.asymptotes shown for PI two break frequencies. In particular consider a system transfer (5.7 Bode gain plot for the system G(jw) = (.20 dB/decade Fig. such that the denominator becomes (s + W n ) 2 . W . w = PI and w = P 2 . 2 Not all roots are simply real values.2. i. gets larger as 0. and which has 0" phase shift.19) and hence: The first part of this expression. 1 8 ) we have that: + G (j w ) = (1 $)+j2iz W KW n / L (5.40 dB/decade directly w = PI = B2. 5. the error incurred. l = 1. they can occur in complex pairs as can be seen from the polynomial equation: s2 + 2 f W " S + a'. 5. 2 . Hence for a plot in which PI = 0 the slope will become .17) which produces complex roots when 0 < f function of the form G ( s )= S* < 1.e. by employing the suggested simpIe approximation.The Bode diagram Gain ( d B ) t 20 147 I oh w[ ) 0 w=pI I 2 w=p2 I 3 Slope = .8.2.a + P l N i W 02 +Pz) . is merely a gain value providing a figure which is constant for all frequencies. which show the gain and phasevfrequency plots for various values of damping factor {. after which it becomes . and .5. i.2. As can be seen from the Bode plots in Fig.
5 Phase angle 1 (degrees) O ° K .ws w =wn w = Sw.3 r = 0.148 Frequency response analysis Gain 10 I { = 0.8 Gain and phase versus frequency plots for a quadratic denominator with varying damping factor.180' 1 w= I ' p log low w = low" w = . . l t= r = 0. 10 2 { Fig.O .3 r = 0. 5..5 r = $2/2 90'  .
4) in which k = 0 . ..22) is 0 dB and this is therefore the frequency at which the two asymptotes cross. i. at all frequencies.where (5.22) If w = con then the gain of (5.3 Compensation using the Bode plot In the root locus technique described principally in Section 4. if the transfer function of (5. 5.2. as was demonstrated in Example 5. f l .2.2. It must be remembered that the plot shown in Fig.4.19) is cascaded with further denominator and/or numerator terms.3) + l)n (5. then the final Bode plots can be obtained.20).2. 5.3. we obtain the characteristic equation 1 + G ( s )= 0 (5.e..3. Similarly.2. the magnitude of (5.3. by adding the plots obtained separately for each of the terms.20).Compensation using the Bode plot 149 therefore which can be simply added to the remaining plot obtained from the second part of the expression in (5. and this means that for any point in the splane to exist on a root locus.20) can be obtained by considering firstly low frequencies for which w 4 anwhereby 2Olog10[(12) 2 w + (Itz) ] 2 I/2 =OdB/O" Wn (5. Asymptotes for the remainder of (5.2. for all K > 0 / G ( s ) = (2k (5. f 2 . and the gainvfrequency plot must be modified by adding the constant gain value caused by K / w i . it must satisfy both I G(s)I = 1 and..2) by setting the closedloop denominator to zero..3. w = lown.2.22) becomes 40 dB and hence the high frequency asymptote has a slope equal to 40 dBfdecade. the loci of closedloop denominator roots are plotted in the splane. as the gain K varies in the system openloop transfer function G ( s ) .1) in which B ( s ) and A ( s ) are the openloop numerator and denominator polynomials respectively.21) and secondly high frequencies for which w 2010g10[(l * wn whereby 1/2 <)'+ (2jZ) ] 2 O n Wn = 4Olog10*dB/i80" On (5.2. It can be seen that if w is increased through a decade.8 is merely that for the second part of (5.2. When G ( s ) is configured within a unity feedback system.3.
5) is also satisfied. We can write down. 3 . Y ( j w ) . a point which is hopefully made a little clearer in Section 5. and due to the symmetrical properties exhibited by a system transfer function to positive and negative frequencies.7) (5. This condition of critical stability only applies for the particular gain value K and frequency w which result in a solution to both (5.3. as is the case when obtaining a Bode plot.6). due to its lack of relevance to this particular topic. rather than calculating the equivalent unity feedback openloop function.e.9) Fig.6) then that system is critically stable. With reference to the unity feedback system shown in Fig. For the case 1 G ( j w ) I = 1 the closedloop system is critically stable and (5. I f in fact the feedback loop is other than unity.3. A basic assumption is therefore made. from Fig.3. it is sufficient to consider the frequency response of G ( j w ) H ( j o ) .3. as long as I G ( j w ) I < 1.3.3 for closedloop stability purposes.3. and actual. V ( j w ) . the closedloop system is stable.8) Neglecting. that a unity feedback loop is applied. resulting in an increase in the error signal. By setting s = jo in (5. E ( j w ) denotes the error between the desired.where H ( j w ) is the feedback transfer function. the error equation: E ( j w )= .1 [Note: In fact for the analysis carried out in Sections 5.3. conversely if 1 G ( j w ) I > 1 the closedloop system is unstable. output signal values and G ( j o )= Go(jo) is the openloop transfer function. the equivalent openloop transfer function within a unity feedback loop must first be obtained.3.4.9.3.1.5) and (5. the only part of the splane considered is the imaginary axis.1. 5. it follows that: E ( j o ) = I G ( j o )I W w ) + U j o ) (5.150 Frequency response analysis If the system frequency response is of concern. because similar stability results hold: see Appendix 41.3. only positive frequencies need be taken into account.3. as with the root locus technique.9 Unity feedback loop . 2 and 5.5) + I)a (5. for all K > 0 / G ( j w ) = (2k (5.3. Assume a particular openloop system to be of the form (5. 5. 5 .9.3. i.2) we are imposing a condition such that when a system satisfies both I G ( b )I = 1 and. the input signal V ( j w ) .G ( j w ) E ( j o ) V ( j w ) + and if G ( j o ) is such that (5.6) is satisfied. any existing error will tend to decrease. as shown in Section 3. 5.
at a phase angle of 180". with o = w ' . and because 20 loglo I G(jw) I = 0 dB when I G ( j w ) I = 1.e. a transfer function could satisfy the magnitude condition (5.6).6) hold for K = K ' > 0 and w = w r . If then. 5. An example is now given to show how the gain and phase margins are measured.5) but the resultant phase angle at that frequency may well be something other than 180".3.e. the system is now closedloop stable because on condition that / G ( j w ' ) = 180". The phase margin can simply be taken as the difference between the Bode phase of G ( j w )and 180" at the frequency for which 1 G ( j w ) 1 = 0 dB. so the gain margin is defined as the amount (in dB) by which the magnitude of G ( j w ) must be increased in order to be equal to 0 dB.1 Gain and phase margins An openloop transfer function.5) and (5.3. both the gain and phase margins can be used to describe the relative instability of a system by considering for the gain margin the amount by which the magnitude of G ( j w ) must be decreased in order to be equal to 0 dB. For an unstable closedloop system both the gain and phase margins are therefore given as negative values. I G ( j w ) I < 1. the further I G ( j w )I is from unity. gives an indication of how stable the system is.3. the more K can be increased before the closedloop system becomes unstable.3. + .6) and exhibits a gain K such that it is closedloop stable. If however the transfer function satisfies (5. As can be seen in the following example. will be closedloop critically stable if it satisfies both (5.e.3.3. Because 20 loglo I G(jw) I = 0 dB when I G ( j w ) I = 1. it is the extra phase lag which would be necessary to make the system unstable. i. remembering that for good accuracy the asymptotic approximation to both Bode gain and phase plots must be corrected as much as possible before measurements are taken. for a stable system the phase angle will be less negative than . i. then the difference between I G ( j w ) I and unity.9) is configured in the same way. so the phase margin is defined as the additional phase lag required to make / G ( j w ) = k 180" at the frequency for which the magnitude of C ( j w ) is equal to 0 dB. whereas for a stable closedloop system they are both positive.Compensation using the Bode plot 15 1 This system is closedloop critically stable when employed within a unity feedback loop.3. the system (5. i. subject to the fact that both (5. it follows that if K = K .3.5) and (5. and for the phase margin the additional phase lead (negative phase lag) required to make/G(jw) = ? 180" at the frequency for which the magnitude of G ( j w ) is 0 dB.180" (this is the same as more positive than + 180"). It can therefore simply be taken as the inverse Bode magnitude of G(jw) at the frequency for which the phase angle of G ( j w ) is 180". when connected within a unity feedback loop. where K' > I? > 0. At a particular frequency. when /G(jw)= +180". Although strictly intended to describe the relative stability of a unity feedback closedloop system.
5.10(b) it can be seen that the phase angle of the response at this frequency is .10 (a) and (b) respectively.5 rad/s. 5. 5.1 Consider the openloop transfer function: G (jo)= 20 jo(j0 + 2 ) ( j o + 5) (5. this will have the effect of shifting the whole gainvfrequency plot up by the same increase in gain for all frequencies.3.11 dB.I w=l w2 WI w= 10 w (rad/s) [log scale] Fig.144" = 36". 5.152 Frequency response analysis Example 5. the gain margin is therefore 11 dB. the phase margin is therefore 180" . 5.3.IO(a) Gain versus frequency plot for system (5.10) is made greater than 20.144". and by reference t o Fig. From Fig.9) . If the numerator gain of G ( j w ) in (5. The phasevfrequency plot is however unaffected by a change of gain. From Fig. resulting in a n unstable closedloop system when the gain is increased to such a n extent that 02 > 01.lO(b) it can be seen that the phase angle of the response is equal to . hence wz will approach W I as the openloop gain is increased.3.10(a) it can be seen that the gain of the response at this frequency is . hence the 0 dB crossover point will occur at a higher frequency.180" at a frequency of W I = 3. S.2 rad/s.10) The Bode (a) gain and (b) phase versus frequency plots for this system are shown in Fig. and by reference to Fig. Gain 0 20  40  I I 60 I I I I I I I I ! I I I I w=O. (02 = W I giving critical stability).10(a) it can be seen that the gain of the response is equal to 0 dB a t a frequency of wz = 1.3.
5.1 w=l w2 WI w= 10 w (radls) [log scale] Fig. approximation Asymptotic I 6 I w=o.20 6'  . Consider again the system described in (5. the phase shift is .2 Gain compensation It is usually the case that certain requirements exist concerning the performance of a closedloop control system. resulting in an amplification of the error signal between the desired and actual output signals. has no effect whatsoever on the phasevfrequency plot.loo"  . With reference to Fig.134" at a frequency of . such that the appropriate specifications can be satisfied. Where performance specifications are relatively simple it is in many cases possible to meet the requirements merely by adjusting the system openloop gain K . often these arise due to specifications made in terms of stability properties such as the gain and phase margin. whereas a decrease in gain K will result in the gainvfrequency plot being shifted down by an appropriate amount over all frequencies.10(b). It is therefore necessary that at the frequency for which the phase shift is 180" + 46" = . In practice this might actually mean that an amplifier is placed in cascade with the openloop system. for a system of the type G ( j w ) = KB( j w ) / A ( j w ) .3.10) 5.Compensation using the Bode plot Phase shift 153 I . Initially the system has a phase margin of 36". obtained simply from the openloop transfer function.134". So. 5.3.10) and employed in Example 5. the specification is made though that a phase margin of 46" is required. the gain must be I G ( j w ) 1 = 0 dB.3.140" . within a unity feedback loop.220" . A variation in the gain K however.180" .3.10(6) Phase angle versus frequency plot for system (5. In so far as the Bode gain and phase versus frequency plots are concerned the result of this is that a controller selection is made in order to reshape the original Bode plot.1. an increase in gain K will result in the gainvfrequency plot being shifted up by an appropriate amount over all frequencies.
only being apparent in the midfrequency range. to saturation effects when a large amplification factor is necessary. when the gain of the openloop system is reduced to K = 13. in practice such a solution might in fact not be possible due. or the lead effect . when connected in a unity feedback loop.11 that the phase shift caused by the compensator itself will be 0" at both high and low frequencies. 3.37 = 22.496.e. for instance. the phase angle will be negative in the midfrequency range. if more than one specification is made it will almost always be the case that with only one control parameter to select. with K = 20.52 dB.10). the phase margin is then 46".37. 5. 5 dB corresponds. I Frequency response analysis + 3. To summarize these results: in the first instance the openloop system defined in (5.a negative phase contribution. So whereas in the first instance 20 loglo 20 = 26. It can also be seen from the compensator in Fig. rad/s. The response of the compensator itself is in fact fairly straightforward to analyze if it is noted that at low frequencies (w 0) the compensator gain. Now 3 .3 Lagllead compensation Although it is sometimes possible to meet one particular specification simply by adjustment of the system openloop gain.5 dB down on its previous value.37. results in a phase margin = 36".11. in the compensator. 5. 5.02 dB. the lag effect . 5. I t follows therefore that if the gain is reduced by 3.e. whether Iag o r lead. i. which means that the effect of the denominator j w + a .5 dB. For a lag compensator. 5. whereas at high frequencies (w + 00) the gain is unity. I I Lag/lead cascade compensator .3. as with very little effort a slightly more powerful controller can be employed in cascade with the openloop system as shown in Fig.5 dB over all frequencies.10(a) the gain at this frequency is found to be approximately then at a frequency of o = 1 . 1 rad/s the magnitude will be 0 dB. now 20 loglo 13. Such an eventuality is neither a desirable nor a necessary feature. the Bode frequency response plots for the compensator can be simply added to those of the openloop system in order to obtain the response of the total plant. is b / a . and hence the system openloop gain of 20 must be divided by 1. then either only one specification can be met or a compromise must be made between the specifications.a positive phase contribution.3. is witnessed + Compensator Openloop system Fig. to a gain of 1. Because it is connected in cascade with the openloop system. thus achieving the desired objective of a phase margin = 46O. i. by taking antilogs. once a unity feedback loop is employed.496 to result in a new openloop gain of 13. and on Fig. Further.154 w = 1. the openloop gain.
5. hence a < b for a lag cornpensator.Compensation using the Bode plot Gain (d6) 155 I 0 ~ 0 0. It is often the case that as well as the essential components of a lag compensator shown by Fig. The effect of a lag compensator is therefore firstly to introduce a negative phase component in the midfrequency range and secondly to increase the importance of the lower frequency response when compared to the higher frequency response. The result of employing a lag compensator in terms of gain and phase margin values is dependent on the choice of break frequency values. 5. I2 Gain and phase versus frequency asymptote plots for a lag compensator with (i) a = I .48 I (w = 3) 0 0.8 I (ii) b loslow (i) Fig. once an appropriate ratio of b : a has been chosen. The actual gain value at any frequency is though stiI1 something which can be selected by means of K. is selected by the ratio b/a. as indeed is the relative increase in gain between low and high frequencies.48 I 4. This means that the phase versus frequency plot. which has the same resultant effect. In general though the lag compensator tends to slow down the . and the extent to which a negative phase shift is introduced. w = a and w = b: in particular their relationship with the frequency at which the system gain is 0 dB and that at which the system phase is . b = 10 and (ii) a = 3 . it can therefore be thought of as a lowpass filter.12(a) and (b) respectively. the system openloop gain K can be adjusted or the compensator has itself an adjustable gain.180'. b = 10 (a) Gain versus frequency (b) Phase versus frequency at a lower frequency than the numerator j w + b. The Bode gain and phasevfrequency asymptote plots for a lag compensator are shown in Fig. 5.12.
b and K . The Bode gain and phasevfrequency asymptote plots for a lead compensator are shown in Fig. + .ow 0. hence a > b for a lead compensator. remembering that now a > 6.8 loglow Fig.48 1 4. It is therefore quite possible to either increase or decrease both gain and phase margin values. compared to the lower frequency response. For a lead compensator.2. b = 3 (a) Gain versus frequency (b) Phase versus frequency response of a system due to an emphasis being placed on low frequency (slower) modes at the expense of high frequency (faster) modes and this tends to improve relative stability properties. is selected by the ratio b / a . and the extent to which a positive phase shift is introduced by a lead compensator. by employing a lag compensator. b = 1 and (ii) a = 10. K. The phasevfrequency plot. The effect of a lead compensator is therefore firstly to introduce a positive phase component into the midfrequency range (the phase shift being zero as both w . 5. 5. the actual gain value being adjustable by means of the openloop system gain. as can be seen in Example 5.0 and w m) and secondly to increase the higher frequency response when .13(a) and (b) respectively.Indeed this ratio also determines the relative increase in gain between high and low frequencies.13 Gain and phase versus frequency asymptote plots for a lead compensator with (i) a = 10. in the compensator.48 ( w = 3) 1 Phase f (ii) I I 0 0.3.156 Frequency response analysis Gain (dB) t I 0 (a) 1 1 * log. the end result being dependent on the values a. it can therefore be thought of as a highpass filter. is witnessed at a lower frequency than the denominator jo + a. the phase angle will be positive in the midfrequency range which means that the effect of the numerator j w + b.
the end result being dependent on the values a.18". and their relationship with both the frequency at which the system gain is 0 dB and that at which the system phase is . where it can be seen that within an uncompensated unity feedback system the gain margin is approximately . hence the 60 40 .180". Consider the openloop transfer function: G (j w ) = jw(jw + l ) ( j o+ 2) 10K (5.3.2 . 5.14(a) and (b) respectively.Compensation using the Bode plot 151 The result of employing a lead compensator in terms of gain and phase margin values is dependent on the choice of break frequency values w = Q and w = 6 .3.40 loglow Fig.2 The objective in the use of lagllead cascade compensation is to reshape the Bode gain and phasevfrequency plots of an openloop system in order that the unity feedback closedloop system meets desired performance requirements in terms of gain and phase margins and low frequency gain.11) where K is an adjustable gain.14 Bode gain and phase plots for the uncompensated system of Example 5.20 . by employing a lead compensator. In general though the lead compensator tends to speed up the response of a system due to an emphasis being placed on high frequency (faster) modes at the expense of low frequency (slower) modes. b and K. are shown in Fig. It is required that a unity feedback system be designed to produce a minimum 0 ' gain margin value of 10 dB and a minimum phase margin of 3. Example 5.3.8 dB and the phase margin approximately . It is possible to either increase or decrease both gain and phase margin values. 5. The uncompensated system's Bode gain and phasevfrequency responses with K = 1.
S.1 and if it is ensured that b / a = 10.48 rad/s. and it can be seen that the frequency at which the gain is O d B is now 0. At this frequency the phase of the uncompensated system is 42" more positive than 180".2 system is unstable. A choice of 10a = b < 0. there will be no change in gain at low frequencies.G o . a phase lag compensator can be used such that (a) the overall gain is reduced at higher frequencies (w > 0. In order to meet the performance specifications. Such a choice of gain K will also mean that the total gain will now be 0 d B at a frequency of w = 0. a gain margin of + 12 d B would be provided. and (b) the negative phase shift introduced by the compensator occurs at a lower frequency.4 rad/s when the phase is .40 20 logluw Fig.1. I5 Bode gain and phase plots for the phase lag compensated system of Example 5.180". 2 rad/s for the uncompensated system). Such a decrease in gain is achieved by selecting K = 0.01) along with K = 0. although as we are well over the required 30" a fairly small negative phase shift at w = 0.158 Frequency response analysis 60 r 40 s E .20 .48 rad/s (the gain was previously + 20 dB at this frequency). Note that the results are found here simply in terms of asymptotic approximations to both the gain and phase versus frequency plots.IS(a) and (b). I)/(jo 0. i.048 would ensure that almost n o phase change occurs at w = 0. that is the frequency at which the gain is 0 dB.1 is shown in + Fig.e.3.48 rad/s is acceptable and this is provided by the selection 10a = b = 0. I ) . Initially the gain margin is measured at approximately w = 1.48 rad/s (as opposed to w = 2 . a gain margin of + 12 dB and a phase margin of 38" are obtained. So employment of a lag compensator has resulted in a lower gain crossover frequency. If the gain is lowered at this frequency by 20 d B with no resultant change in phase. 5 . . a phase margin of 42" (well over the required minimum of 30") would be obtained if n o alteration was made t o the phase characteristics. Further. The resultant effect obtained by employing a lag compensator of the form ( j w + 0. thus meeting the performance requirements.
3. resulting in an increased bandwidth.e. which means that the response speed of the system was slowed down.42 rad/s.6 rad/s. Once gain and phase versus frequency values have been obtained for a system response.3. where necessary.3 dB) point.3 dB. the bandwidth is reduced to w = 0. By ensuring that the . an alternative form in which the results could be presented is by means of plotting gain versus phase over the range of frequencies involved. It is also possible to employ a laglead compensator. Slowing the response down had. by cascading lag and lead compensators. and hence the bandwidth is simply given by the high frequency cutoff ( . so that system’s response is faster. The gain and phase characteristics can then be plotted separately against frequency. will then increase from its original value. i. in which the lead compensator can be effected by ensuring that the peak of the phase lead characteristics occurs in the frequency neighborhood of the uncompensated system’s gain crossover frequency. by means of applying a phase lead (rather than lag) compensator to the uncompensated system in Example 5. The effect of this is to retain the gain and phase characteristics of the uncompensated system at both high and low frequencies. This is in fact considered in Problem 5. in terms of system gain and phase characteristics. 5. The gain crossover frequency. over a wide range of frequency values.2.3 dB at all in the low frequency end of the spectrum. For the phase lag compensator in Example 5.3 dB. It is worth noting that. in the example.2.2 the bandwidth of the uncompensated system is w = 2.2 the bandwidth was reduced. The bandwidth of a system gives an indication of that system’s speed of response. In Example 5. For many control systems the gain does not fall below . whereas once the phase lag compensator is applied. while for a band of midrange frequencies the system’s gain is either increased (band pass filter) or decreased (band limiting filter).The Nyquist stability criterion 159 The bandwidth of a control system is the range of frequencies for which the gain of that system does not fall below . The Bode characteristics of the laglead compensator can be found straightforwardly by adding the separate lag and lead Bode plots.4 The Nyquist stability criterion In Section 5. that is the frequency at which the gain is equal to . once the system has been phase lead compensated.3.2 it was explained how the steadystate response of a stable. several lag or lead compensators can be placed in cascade. once logarithmic values of gain and frequency are employed. such that the logarithmic gain effects and phase modifications of the individual compensators can be added to achieve much larger values of gain and phase shift in the important crossover region. The introduction of a phase lag compensator has therefore resulted in a reduction in bandwidth. such that the larger the bandwidth is. It is however possible to increase the gain and phase margins as well as increasing the bandwidth. the effect of improving the relative stability properties by increasing both the gain and phase margins. linear system to a sinusoidal input of constant magnitude can be found. hence forming Bode plots.
individual factors can be dealt with separately both in terms of gain and phase. the gain and phase of the system transfer function is found. in order to simplify Nyquist plots.1 Frequency response plots There are four important points to be considered when plotting the frequency response of a system. The transfer function. then a plot of system gain versus phase is known as the polar or Nyqiiist plot. i.160 Frequency response analysis response of the system is considered over all frequencies from 0 < w < 00. points will produce an adequate representation. and secondly in terms of Cartesian coordinates.1 Consider the system transfer function: G ( s )= s(s 2qs + 0. such that: G (j w ) = 2O(jo t 5 ) j o ( j w 3)(jw + + 1) . simple to evaluate. Example 5. G(s). the response being termed the frequency response. however. The Cartesian coordinate technique is particularly useful as far as finding real and imaginary axis crossing points is concerned. In fact for the vast majority of system transfer functions it is not necessary to solve for a large number of frequency values. Crossing of the imaginary axis So for each of these points.e. However. A procedure applicable. even when carried out by means of a computer. The points are: 1. and no more. result in a long drawnout plotting process. set s = j w . If the transfer function. real and imaginary parts. Start of plot when w = 0 2. a few important. Crossing of the real axis 4. by setting the imaginary and real part t o zero respectively. of a system is known.4. End of plot when w = OD 3. can be regarded in two ways. as can be seen in the following example.4. by viewing G ( j w ) in terms of polar coordinates. Firstly in terms of polar coordinates.e.5) + 3)(s + 1) In order to investigate the system’s frequency response. magnitude and phase. i. G ( j w ) . over a large number of frequency values. in terms of gain and phase. is given in the following section. Once the system transfer function has been evaluated in terms of gain and phase at these points. the complete plot can subsequently be sketched by joining the points. then its frequency response can be found analytically simply by solving for G ( j w ) . This can. 5.
this is: real [ G ( j w ) ]= and imaginary [ G ( j w ) J = 2 0 ( w 2 .16 Magnitude and phase for an individual factor then and in which the individual factors have been dealt with separately by applying Pythagoras's theorem for the gain and by taking the tangent of the angle.The Nyquist stability criterion hag. Note that when a factor of G ( j w ) is a zero.0 2 + 15)] so in terms of real and imaginary parts. its magnitude appears as a denominator term and its phase contribution is negative. 5. Alternatively the transfer function G ( j w ) can be considered in terms of real and imaginary parts. Thus G (j w ) = 2Oj(jw + 5)(3 w(02 + 9)(w2 + 1) . its magnitude appears as a numerator term and its phase contribution is positive. whereas when a factor is a pole. 161 T P=(d+ 0 = tan' Fig.16.2 q w 2 + 17) (02 + 9 ) ( J + 1) + 9)(02+ 1) .jw)(l . as shown in Fig.jo) then G(j w ) = . 5.2 0 [ 0 ( w ' + w(w2 + 9)(w2 + 1) 1 7 ) + j ( .e.IS) o(w2 . i. which are found by multiplying both the numerator and denominator of G (j w ) with complex conjugate denominator roots.
5 (15 9)(15 1) . imaginary [ G ( j o ) ]= 0 A solution is found when w 2 = 15.513 . w = +JT5 as 0 6 w < 00. I I I I I I I l I . 340 9 imaginary [ G ( j w ) ] = . real [ G ( j w ) ] =  / G ( j o ) = 90" .162 Frequency response analysis The frequency plot can now be sketched by means of the four important points: 1. Real axis crossing / G ( j w ) = 0" o r ? 180".e. 3 . w = o 1 G ( j o ) l = 00. 5.340/9 along the real axis. / C ( j o ) = .4. 4. + + + 5 Note: The solution is found most easily here by setting the imaginary part of G ( j w ) to zero. w = m 1 G ( j w ) 1 = 0. 2. Imaginary axis crossing / G ( j w ) = k90". (not to scale) Fig. but also from an asymptote .17 Nyquist plot for Example 5.00 The plot therefore starts at infinity magnitude. with a phase angle of . real [ G ( j w ) ] = 0.180" imaginary [ G ( j w ) ] = 0 which gives information showing how the plot approaches the origin.340 I r Jw=m b 9 Real . real [ G ( j w ) ] = 0 hag. Then real [ G ( j w ) ] = 20(15 17) . i.1 .90".
as shown in Section 3. Bode).The Nyquist stability criterion 163 There are no real solutions for w. it must first be pointed out that. root locus. then the equivalent openloop transfer function G o ( j w ) within a unity feedback loop should first be obtained.2 The Nyquist criterion Before discussing the Nyquist criterion in detail. whichever method is preferred. Throughout the remainder of this section.4. 5. It is sufficient though. If the actual closedloop system contains a feedback term N ( j w )which is other than unity. reference t o the transfer function G ( j o ) implies either the openloop function Go(j w ) o r the function G ( j w ) H ( j w ) . the complete Nyquist plot can be sketched as in Fig.1. for the purposes of the Nyquist criterion. the Nyquist criterion provides a test for closedloop stability when that openloop system is connected within a unity feedback system.18 Response plot including negative frequencies . such that a continuous response diagram is obtained. I I ' \ ' .17.4. By plotting the frequency response for an openloop system transfer function. to plot the frequency response of G ( j w ) H ( j w ) . 5. because similar stability results hold. see Appendix 4.g. in common with other techniques (e.340 _ 9 Fig. 5. The Nyquist plot considered thus far (for 0 < w < 00) must be generalized by also considering negative frequencies. and hence the plot does not cross the imaginary axis. a basic assumption is made that a unity feedback loop is to be applied. Armed with the four points of interest.
The closing contour at infinity must always be drawn in a clockwise direction. then the corresponding unity feedback closedloop system is stable. if for a certain frequency a'. So. 5.1 point is encircled in a clockwise direction then the corresponding unity feedback closedloop system is unstable.1 point) in a clockwise direction.17.18 is shown in Fig. 5.164 Frequency response analysis II I \ .18. The closing contour at infinity for the response plot previously considered in Fig. it is also necessary to completely close the contour by connecting up any remaining unconnected contour parts at infinity. As an example consider the response plot shown in Fig. Nyquist stability criterion for stable openloop systems If an openloop system. This results in a negative frequency response plot which is a mirror image.we have that G ( .19 Response plot including closing contour at infinity For all real functions G ( j a ) .3. G ( j o ' ) = a . G ( j w ) . of the positive frequency response plot. about the real axis. As a final extension to the plots considered thus far. When the plot passes through the .1 point in a clockwise direction .(the complex conjugate). see Problem 5.to 0 ) ' \ \ w increasing ( 0 +t o m + ) I I / I c I I I I / / / Y I Fig. \\(m *\ Closing contour at infinity \ \ (w increasing 0.19.1 + j 0 point (known as the . 5.j o ) = G * ( j o ) . such that the following Nyquist stability criterion can be applied. the complete plot forms a continuous closed contour. 5. Once the closing contour at infinity has been drawn.jo')= a + j p . and must be such that the direction of increasing frequency is retained. 5. which becomes that shown in Fig. is stable and its complete frequency response plot does not encircle the .j p then G ( . Conversely if the .\ \ ! ) \ I I \ \ Imag. \ \ I I I 1w increasing too) '.*.
1) and it is required. Example 5. Further. (5. then the corresponding unity feedback closedloop system is stable if N = Np. Hence the Nyquist stability criterion can be generalized to read as follows.N. If the openloop system is unstable then G ( s )will contain N p poles which lie in the right half of the splane.4. then the corresponding unity feedback closedloop system is unstable. if the number of encirclements is equal to the number of righthalf plane openloop poles..i. if there are any righthalf plane closedloop poles. N times. = Number of righthalf plane poles of 1 + G Nz = Number of righthalf plane zeros of 1 + G (these are also righthalf plane closedloop poles) If N is the number of clockwise encirclements of the .1 point in a clockwise direction. such that: N. then: N = Np.4. set s = ju. Using the Nyquist stability criterion find the range of values for K over which the unity feedback closedloop system will be stable. for the closedloop system to be stable.1 point is encircled N times in a clockwise direction and Np> N. such that: G (j w ) = then (jw + l ) ( j w + 2 ) ( j w + 3) K . Let Nzbe the number of righthalf plane closedloop poles. To investigate the system’s frequency response.2 Consider the system openloop transfer function: G ( s )= K (s t l)(s + 2)(s + 3) in which the scalar gain K > 0 can be varied. Conversely if the . then these will also be zeros of the function 1 + G.e. The criterion given only deals with openloop systems which are stable. These poles are in fact also poles of the function 1 + G. that Nz = 0. Nyquist stability criterion If the complete frequency response plot of an openloop system G ( j u )encircles the .The Nyquist stability criterion 165 and does not encircle it then the corresponding unity feedback closedloop system is critically stable.1 point.
It follows that if [K/60] < 1 . the complete frequency response plot for the openloop system of Example 5.20.1. by setting the real part of G ( j w ) to zero it is revealed that the plot crosses the imaginary axis when w = 1 radls. where it can be noted that no closing contour at infinity is required.166 Frequency response analysis and By employing the four important points of interest.1 + j 0 . as far as the Nyquist stability criterion is concerned. then the complete frequency response plot will not encircle the point at . Hence for the real axis crossing. G ( j o ) can be reconsidered in terms of real and imaginary parts. and this results in a magnitude at that point of 1 G ( j w ) 1 = K/60.a’) + 1 ) ( J + 4)(W’ + 9) KW(O’. which can be found by setting the imaginary part of G ( j o ) to zero.2 is found to be that shown in Fig. w = +ill= 3.4.4. as shown in Example 5. which produces a magnitude at that point of I G ( j w ) 1 = K/10.e. t w increasing Fig.317 rad/s. i.4. is the real axis crossing point. and thus the closedloop system will be stable. Of much greater significance. In order t o obtain real and imaginary axis crossing points. real [ G ( j w ) ] = and imaginary [ G ( j w ) l= (W2 (w2 + l)(w’ + 4)(0’ + 9) 6K(1 . 5. 5. [mag.11) Although the imaginary axis crossing point is not immediately of significance.20 Frequency response plot for Example 5.2 .
3. Because the gain margin is uniformly measured in dB. are equally applicable as far as the Nyquist plot is concerned. [ G ( j w ) = k 180" or . the closedloop system will be stable as long as: O < Kc60.4. Thus. In terms of a frequency response plot. If a frequency response plot is also required for negative openloop gain values K. 180" point corresponds to I G ( j w )1 = 1. Imag.3 Relative stability The concepts of gain and phase margin are introduced in Section 5. any gain/magnitude values of G ( j w )obtained from the Nyquist plot must therefore be converted to dB. when / G ( j w ) = +. and are considered here in that context. in terms of /+ Imag. So. and the closedloop system will be unstable. then we wish to know by how much the openloop system gain can be increased before the closedloop system becomes unstable.180". for a positive gain K .4. The 0 dB.1 +jO.21 Frequency response plots for stable and unstable closedloop systems (a) Stable (b) Unstable . 5. with their implementation described in terms of Bode analysis. Both concepts. then there will be a clockwise encirclement of the . 5. however the phase/G(jo) is shifted through 180" for each frequency w : see Problem 5. which relate to the frequency response characteristics of a system. if a unity feedback closedloop system is shown to be stable.1. Icow)= 2 180" I w (a) Fig.The Nyquist stability criterion 167 However if [K/60]> 1. No modification is in fact found as far as the openloop magnitude 1 G ( j w ) 1 is concerned. The gain margin is defined as the amount (in dB) by which the magnitude of G ( j w ) must be increased in order to be equal to 0 dB. by means of finding 20 loglo I G ( j w )I.then this can be obtained in exactly the same way as the method described for positive gain values.1 point.
At this frequency. then 80 (w2 + 1)1/2("2 + 4)1/2(0* + 9)1/2 = = .e. and from this. Also.54".4.21. this corresponds to a decrease or attenuation.838 rad/s. the gain margin is used as a measure of how far the frequency plot is from the . /G(jw2) = .21. /G(jw2) such that the phase margin is . in order to reach the . in order to reach the .4. w2 = 1.62". when the plot crosses the negative real axis. such that the phase margin is 44.188.542 dB. So the gain must be 'increased' by a factor of 3/4. Example 5.4. I G ( j w ) I = K/60 = 1/3. So the gain can be increased by a factor of 3.62". I G ( j w ) I = K/60 = 4/3.2: G(j w ) = K (jw + l)(jw + 2)(jO + 3) 1. The gain margin is therefore 20 lOgl0(3/4) = .135. 5.5 dB. Also. At this frequency.168 Frequency response analysis the Nyquist criterion. .e. when I G ( j w ) I = 1. i. i.1 point. see Fig. when I G(jo)l = 1 then (w2 + 1)'/2(w2 + 4)'/2(w2 + 9)'/2 = 20 and from this. w2 = 3. The gain margin is therefore 20 log10(3)= 20 loglo(l/l G ( j w ) 1) = 9.8. and a > 0.3 Consider the openloop transfer function of Example 5. see Fig. 111 G ( j w ) 1. ~~ Example 5. 111 G ( j w ) 1. The phase margin is defined as the additional phase lag required to make ( G ( j w ) = 2 180" at the frequency for which the magnitude of G ( j w ) is equal to 0 dB. 2. The phase margin is therefore taken as the phase difference between the negative real axis and the phase of the frequency response plot at the frequency for which I G ( j o )I = 1 . When K = 80 the closedloop system is unstable and for / G ( j w ) = k 180".766 rad/s. Note: The phase margin is negative for unstable systems.1 point. Note: The gain must be 'increased negatively' for a n unstable system in order to make it stable. When K = 20 the closedloop system is stable and for / G ( j w ) = 2 180'.2. as indicated by the negative gain margin sign.46'. Note: The gain margin is negative for unstable systems.1 point.4 Consider the system openloop transfer function: K G ( s )= ___ s(s + ff) in which the scalar gain K > 0 can be varied. 5.
This can be considered to imply that the gain margin is.e.22. the gain K can be increased by an infinite amount and the closedloop system will not become unstable. such that G (jo)= then K jw(j0 +a) and / ~ ( j w = . the phase margin can be obtained by finding a solution to I G ( j w )I = 1. To investigate the system's frequency response. w .e.90" . and that a negative real axis crossing. / G ( j w ) = . i.22 Frequency response plot for Example 5. . when K = w ( w 2 + (Y2)1/2 (5. i. i. probably very close to the origin. and so the gain margin is undefined for this system. A sketch of the frequency response plot for the system in Example 5.tan' ) W [a1 + It can be observed directly that the frequency response plot for this system will not cross the negative real axis d G < j w ). 5.180" as w w). that in practice a plant may well not be an 'exact' secondorder system. it is required that the phase margin q5 = 30". 5.4. though.2) But. to all intents and purposes.4.The Nyquist stability criterion 169 Using the Nyquist plot. set s = j w .4. and it can be seen that although the gain margin is undefined for this system. does in fact exist. I I w=oD w increasing f Fig. find the value for K which results in a phase margin of 30".150"' and hag.4 is shown in Fig.4 .e. It is worth mentioning.
The value of phase margin selected here is up at the higher end of the values usually chosen. the Nyquist design can be treated either by finding the equivalent sole cascade controller o r simply by considering the feedback compensator to be in series with the original plant.3 a' will achieve the desired phase margin. and it is this type of requirement which is considered here. In practice. When feedback compensation is used as well as or instead of cascade compensation. the gain margin. Indeed this could result in several designs and redesigns being carried out before a final compromise scheme is obtained.4. O n substituting for w in (5.5 Compensation using the Nyquist plot Applying compensation to a system by means of the Nyquist plot involves a modification of the original openloop frequency response characteristics.2). within a unity feedback loop. 5. however.1 Gain compensation For straightforward and simple performance specifications it is often easiest t o meet the requirements by merely adjusting the system openloop gain K . 5. it is found that an openloop gain of K = 2.150" = . is usually in the range 3 to 12 dB. Typical values for reasonable design are 30" t o 60" for the phase margin.tan I:' [ such that w = J3 a. In practical terms this will most likely be implemented by means of an amplifier in cascade with the openloop plant. . This makes it possible to plot firstly the original openloop system frequency response and secondly the compensated system frequency response as a new openloop system.5. in order to satisfy some previously specified performance objectives.90" . The objectives can be stated simply in terms of gain and phase margins. however. such that a Nyquist plot is made of G ( j w ) H ( j w ) : this is discussed further in Appendix 4.170 Frequency response analysis this occurs when . it may well be that further desirable performance measurements are given such as bandwidth and pole locations and the Nyquist compensation procedure may well need to be combined with such as Bode analysis and root locus techniques. within a unity feedback loop. It is considered that the compensation network is included in cascade with the original openloop system. The closedloop stability and relative stability properties of each system can then be witnessed.
23.3374.4.340 9 Fig.23 Response plot for Example 5. however the gain of the plot at that frequency and its phase margin will be.1 . 5. the frequency at which the phase of G ( j w ) is f 180" will not be affected by a variation in K. without affecting the phase. This value range can be achieved by selecting a gain K such that 0 < K Q 0. and in order to obtain a gain margin of at least 5 dB. as shown in the following example.5)K + 3)(s + 1) whose frequency response plot for K = 1 is considered in detail in Example 5.1 point is encircled in a clockwise direction and hence the unity feedback closedloop system obtained when K = 1 is unstable. In particular. Hence stability of a closedloop system can be ensured by the selection of an appropriate gain value and within limits. if the gain K is increased then this will increase the magnitude of the frequency response plot. the frequency response plot is equal to unity at a hag. find a value which results in a closedloop system which has a gain margin of at least 5 dB and a phase margin which lies between 30" < 4~< 50". and is sketched here again in Fig.437 dB. 5. For a gain K = 0.5623 when the phase of/G(jw) is k 180". I . 5. Using solely the gain K > 0.1 Consider the system openloop transfer function: G ( s )= s(s 20(s + 0.3374. The original frequency response plot (only positive frequencies are shown in Fig.5. Example 5. it is required that the openloop system gain is I G ( j w ) I < 0.23) reveals that the .. uniformly over all frequencies.Compensation using the Nyquist plot 171 Where the plant is of the form G ( j w ) = K E ( j w ) / A( j w ) .5.1. both phase and gain margin specifications can be met. While K = 1 the gain margin is 20 log10(3/5)= 4.
11. So a phase margin which lies within the desired range has been found by the selection of a suitable gain value.175. the gain and phase margin stipulations have been satisfied.584 dB.05. 5.145.172 Frequency response analysis frequency found from: which is satisfied when w = 2.2 Lagllead compensation The application of a cascade lagllead compensator is considered in terms of Bode analysis i Section 5. the effect of the cascade . by means of simple gain compensation.5. such as the lagllead compensation considered in the following section. Thus.62". such that the stability of the unity feedback closedloop system can be tested. and selection of the positions of these singularities. It can subsequently be found (by trial and error for instance) that by employing a gain K = 0. 5.01'. which is certainly more than the required minimum value of 5 dB. At this frequency the phase /G( w ) is: j / G ( j w ) = tan'[:] .968 rad/s.90' .5. by only making use of gain compensation. n the controller adds one pole and one zero to the original openloop system characteristics.05 the gain margin is 20 loglo (12) = 21. with a relevant block diagram given in Fig.086 rad/s and this results in a phase of / G ( j w ) = . coupled with a choice of openloop gain. this means that at each frequency w .tan'[o] = .99'. Essentially. Further.3.3. which gives a phase margin of 34. K.1 if a phase margin within the range 40" < I < 50" was also required.tan'[:] . the frequency response plot is equal to unity at w = 1.1) then the frequency response of the cascaded functions D ( j w ) G ( j w )must be plotted. In terms of the Nyquist plot. Unfortunately it is not always possible to satisfy both the gain and phase margin requirements. If the compensator transfer function is given by: D ( j w )= jw jw +b +a (5.38' Hence the phase margin is only 4. with K = 0. allows for a great deal more flexibility than is provided by means of gain compensation alone.5. Hence for a more general type of compensation $ network it is necessary to employ a slightly more complicated structure. For instance the desire for a gain margin of at least 5 dB but no more than 10 dB could not be achieved simply by gain compensation on the system of Example 5.
which lies between 0 and 00. the effect of the lag compensator is 1. the phase angle of the compensator itself is negative in the midfrequency range.5. 5.2) and when a < b (for a lag compensator). When the system openloop gain K > 0 can also be varied. it follows that the phase will never be positive definite. whereas the effect of a lag compensator is to make the transfer function phase more negative at each frequency. but is 0" at the extremes when w = 0 and w = 00.5. but will multiply the gain at all frequencies by a factor K. To multiply the gain of the original plot by a value which is greater than unity but less than b / a . by a certain amount. For a lag compensator. So. A w = o  w=O Real IDuw) I = 1 w increasing Fig. If a phase shift of larger negative characteristics is required. This can be seen by noting that: / D ( j w ) = tan'[:] tan'[:] (5. There are no axis crossing points and plot end conditions are found to be I D ( j w ) I = b / a when w = 0 and [ D ( j w ) I = 1 when w = 00. The effect of a lead compensator is to make the transfer function phase more positive at each frequency. then several lag Imag.90" (for finite a and b). the magnitude and phase of the transfer function. To shift the phase by an angle which is less than 0" but greater than . then this will not affect the phase angle in any way.24 Nyquist plot for a lag compensator .24. as shown in Fig. 5. with the phase equal to 0" in both cases. for any particular frequency value w .Compensation using the Nyquist plot 173 compensator is to modify. In fact as (5.3) the frequency response plot for a lag compensator is a semicircle. and 2.
5. for the uncompensated system. when the system is compensated. this can be realized by reducing the gain at w = 1.414 rad/s by approximately 20 dB. 5. The gain margin is found to be . reveals a negative phase margin of 30.10. with no compensation applied as shown in Fig. Example 5.1. If it is then ensured that b / a = 10.46 dB when w = 1. this will mean Real * Fig. Note that gain and phase margin results can be obtained either by exact calculations or as approximate values measured from the sketched Nyquist plot. with no compensation.15' (when w = 2..5. the gain margin will be at least 8 dB. Such a decrease in gain is obtained by selecting K = 0. would be unstable.46 dB (when w = 1. on condition that the phase is not also affected to any great extent at that frequency. 5. I f it is ensured that.114 Frequency response analysis compensators can be used in cascade.without affecting the steadystate properties of the system when K = 1. A phase lag compensator can be used to obtain a gain margin of at least 8 dB and a phase margin of at least 30°.cn K = 1.25(a). Such phase and gain results show that a unity feedback closedloop system.2 .414 rad/s).414 rad/s.532 rad/s) and a negative gain margin of . an example of its employment is given by the following.2 Consider the openloop transfer function: G (j w ) = jw(jw + 2 ) ( j w + 1) 20K A frequency response plot for the system. This can be achieved by reducing the overall gain at higher frequencies whilst retaining the same value of gain at low frequencies.25 Uncompensated and (b) compensated frequency response plots for the system of Example 5.10. Restricting ourselves to a single compensator however. w.
this will become the frequency at which the overall gain is 0 dB.e. no such counteraction takes place.749. is counteracted at low frequencies by the compensator gain. Employment of a lag compensator has resulted in a lower frequency at which the gain is 0 dB.532 rad/s for the uncompensated system).Compensation using the Nyquist plot that the overall gain at low frequencies will not be affected.147. as long as little or no phase change takes place at the frequency at which the overall gain was 20 dB. i. with K = 1. Further.1. A choice of 10a = b c 0. the bandwidth has been reduced.0749 would ensure that almost no phase change.e. whereas at higher frequencies.90" . thus meeting the performance requirements. by holding the gain and phase margins steady. along with the lag compensator: D(j w ) = j o 0. then both the gain and phase margins will normally be increased.749 rad/s and at this frequency the phase is: /G(jw) = .tan $1 .387" such that a phase margin of 32. a = 0. The frequency at which the gain is 0 dB (unity magnitude) is now approximately w = 0.05 j w + 0. K.5 dB and a phase margin of approximately 32" are obtained. in terms of slowing down the system's response. 175 0=0 Thus the gain increase.75 rad/s (as opposed to w = 2. This means that. Solving [K= 11 gives w = 0. So. Reduction of bandwidth is a property usually obtained by the use of a lag compensator. and hence a choice of b = O . The resultant effect. error constant values can be increased by means of a lag compensator. due to the compensator. 5. I . occurs at w=O.25(b). and can be looked at in another way.005 + is shown in Fig. obtained by employing a gain K = 0. If the steadystate error constant is held at a steady value.tan'[w] = . at higher frequencies a gain decrease of 20 dB is apparent. with K = 0. . as was done in the example. which is above the minimum required. O 5 . i. Conversely.005 will suffice.61" would be obtained. a gain margin of + 9.
To shift the phase by an angle which is greater than 0" but less than a and b). There are no axis crossing points and plot end conditions are found to be identical to those for the lag compensator. such that they lie in the midfrequency range and therefore d o not affect the specifications already met by the compensator.e.3) also holds.26 Nyquist plot for a lead compensator For a lead compensator.2) still holds. So for any particular frequency value w . This can be seen by noting that (5.26. in order to meet gain margin requirements (a decrease in gain margin will be required). see Problem 5. such as specified error coefficients to relevant input signals. this does not affect the phase angle in any way. At higher frequencies the overall gain is thereby increased. with the phase equal to 0" in both cases. To multiply the gain of the original plot by a value which is less than unity but greater than b / a . which lies between 0 and 00. . The ratio b : a can then be found in order to produce necessary steadystate conditions. The phase shifting properties of the compensator can then be selected. I D ( j w ) I = b / a when w = 0 and I D ( j w ) I = 1 when w = 00. as shown in Fig. the phase angle of the compensator itself is positive in the midfrequency range.5. It follows that the phase will never be negative definite. the effect of a lead compensator is: 1. and 2. + 90" (for a finite When the system openloop gain K > 0 can also be varied.5. then several lead compensators can be used in cascade. is required. K. if a more complicated compensator structure. 5. i. For a single lead compensator. but will multiply the gain at all frequencies by a factor K. but is 0" at the extremes when w = 0 and w = 00. If a phase shift of large positive characteristics (more than 90") is required. the frequency response plot for a lead compensator is a semicircle.5. the tendency being towards a destabilization of the system. this can often be achieved by cascading a number of lag and lead compensators.176 Frequency response analysis w=o Fig. In fact as (5. Conversely. the standard design procedure involves firstly selecting the system openloop gain value. which is more selective about the frequency ranges amplified. 5. however now b < a for a lead compensator.
G (j w ) = 20 j u ( j w + 1) produces a gain (dB)vphase plot as depicted in Fig. while results such as relative stability found from the method. Bode and Nyquist plots both deal with the same data type.tanl[wI Although for the transfer function shown above it is quite straightforward to . is increased by a phase lead compensator. or the function G ( j w ) H ( j w )can be plotted. which are in many ways easier to understand and use.90" . The Nichols chart. deals once more with the same data type but presents it in yet another different way. as shown in the following example.27.6.1 The openloop system transfer function. acceptable but faster response. This is discussed in Appendix 4. The gain in dB of this function can be plotted against the corresponding phase angle values over the range of frequencies o = 0 to w = 00. then either the equivalent unity feedback system must be found and the effective openloop transfer function extracted. in theory over all frequencies. another frequency response technique. Where the closedloop system in question actually employs a feedback term H ( j w ) which is other than unity. 5. thus the bandwidth is increased. but present it in a distinctly different way. Example 5. decreasing the phase and gain margins from those corresponding to a sluggish response to those apparent for a specified. Much of the procedure employed with a Nichols chart is similar to that used for Bode and Nyquist design methods. G ( j o ) . 5. Initially. pertain to the closedloop system when connected in unity feedback mode. The gain for this system is: gain (dB) = 20 loglo whereas the phase is: phase = 20 w(w2 +1 ) I ' Z . In the majority of cases the frequency at which the overall gain is 0 dB. The Nichols chart is therefore not considered here in the same depth as the other two methods.6 The Nichols chart Frequency response analysis of a system involves the use of system gain and phase information obtained over a wide range of frequencies.The Nichols chart 177 Phase lead compensators are usually employed to improve the transient system response. A common feature with all three methods is that the openloop system transfer function response is plotted. consider solely the system openloop transfer function.
178 Frequency response analysis I I I I I I I I I .9oo 78.13O  26.69' 177.43O .01 10 loglo(w' + 1 ) [dB] 0.168.6.2 Table of gain and phase values for Example 5.71' . terms can be calculated separately and merely added.02 13.90" .02 26.9oo .02 6.01 ~ .153.6.03 _ .1 w 0.40 d B 90" I 210" 180" I I .20 loglo(w) . .135' _ 900 63.1 Relative stability Gain and phase margins. at each frequency.69O 87.71° 45O ___Phase(') 95.01 I 2 5 26.40dB I I I I I .02 20 loglo 20 [dB] 20 0 (W 20 I O ~ I O(dB] ) 3. described earlier in this chapter for unity feedback systems. 20dB I I I I I I I I I I I I OdB Gain (dB) I I I ~ I 20dB I I I I I I I I . can also be obtained by reference to the gainvphase plot.15 ~ 2. I ~26.43' . for both gain and phase.27 Gainvphase plot for Example 5. because the gain is plotted in dB.11 20 26.120" 150" Phase Fig.98 23.6.13' 5.03 _ _ 26.02 26.99 _ 13.02 26.9oo 900 tan'(w) 5.04 ~ _ _ Gain [dB] 45.10 loglo(w2 + 1) Table 5.02 6. For the example in question we therefore have: gain (dB) = 20 log10(20).1 obtain the gainvphase plot by simply finding the required values for a number of frequencies.98 14. in terms of Bode and Nyquist plots. 5.
particularly for more complex systems. Example 5. The phase margin is the amount by which the openloop phase can be made more negative until it is equal to k 180'. 5.270" . Nole: Although stability can be claimed for the majority of systems if they exhibit positive gain and phase margins.6. if 0 2 is the frequency at which the gain of the openloop transfer function is 0 dB.2 .180".40 . So. So. the gain margin can be found from: gain margin = 20 logto or I I G h d dB gain margin = 20 loglo I G ( j o l )I dB where I G(j u l ) I can be either calculated exactly or measured on the gainvphase plot.28.20 .180"  140" . for the majority of systems.180".90" 0" Fig. stability of the unity feedback closedloop system is.2 Consider the gainvphase plot shown in Fig. a stricter check must be made using such as the Nyquist criterion to ensure absolute stability. ensured as long as the phase is more positive than ? 180". At this frequency. The gain margin. the phase margin can be found from: phase margin = 180" + / C ( j w z ) where /G(j m ) can be either calculated exactly or measured on the gainvphase plot.28 Gainvphase plot for Example 5. At this frequency. shown as 40 20 0 Gain (dB) .The Nichols chart 179 The gain margin is found when the plot of G ( j o )has a phase angle of . 5. stability of the unity feedback closedloop system is. for the majority of systems. if W I is the frequency at which the phase of the openloop transfer function is . ensured as long as the gain is less than 0 dB (i.e. The gain margin is the amount by which the openloop gain can be increased until it is equal to 0 dB.6. The phase margin is found when the plot of G ( j w ) has a gain of OdB. I G ( j w ) 1 < 1).
showing 2 dB closedloop gain locus (constant I G M ( ~ w ) 1) . Also. then for any particular G ( j w ' ) there will be one. in Fig. placed on phase / G M ( ~ wthen this can be achieved by a variety of G ( j o ) values. M. no closedloop phase values are shown. 5 . if a particular magnitude I GM(jw)I is selected.G ( j w ) ] . 5. Note that the vertical gain and horizontal phase axes apply to the openloop system only. This is because the phase is 0 ' equal to . unity feedback closedloop magnitude and phase value. As an example. This is because the gain is equal to . with no constraint . 2 8 . 5.29 Gainvphase plot. 5. and the relative stability of the unity feedback closedIoop system G ( j u ) / [ 1i. Defining GM(jw) as (5. in ). The plot in Fig.28. shown as 4 in Fig. given by I GM(jw')I and / G M ( j w ' ) However._ _ _ 13. is found to be 10 dB. terms of magnitude and phase contributions.6.29 shows that the openloop gain and phase combinations which produce a closedloop system with a gain of 2 dB.10 dB when the phase is . consider the requirement that 20 loglo I G ( j o )1 = 2 dB. 5. form a locus. is found to be 4.73 1_ 0 Fig.ISO". and only one.6. Phase . at the frequency w I .240"  180" i AT o 2 .1) it follows that for w = w ' .180 Frequency response analysis G. the phase margin.e.2 Closedloop frequency response In a similar fashion to the Bode and Nyquist methods.is found from the plot.140" when the gain is 0 dB. the frequency response of the openloop system G ( j w ) is plotted. i.
30 shows that the openloop gain and phase combinations. each different combination producing a different magnitude I G M ( ~ u1 )but a constant phase/GM(jw).The Nichols chart 181 Fig.861 = 13. let G ( j w ) = 0 dB /60°.180" So. in terms of magnitude and phase.180" corresponds to the gain and phase of the openloop system G ( j w ' ) for some frequency w ' .771 dB/30" Jj .734 dB and / G ( j w ) = .The plot in Fig. GM(jo)= 1/60" ='/3o" 1+1@ = 4. this means that the solution to which is I G ( j w )I = 4. A similar analysis can be carried out by selecting a particular phase/GM(jw). then the unity feedback closedloop gain and phase will be 2 dB at the frequency w' . 5. Note that the vertical gain and horizontal phase axes apply to the openloop system only. 5.734 dB/. say 2 dB /o". then /o".30" closedloop phase locus (constant /GM Taking one point on the locus. variety of different G ( j o ) values. Taking one point on the locus. with no constraint placed on the magnitude I G ~ ( j a ) This phase can then be achieved by a I. which produce a closedloop system with a phase of . no closedloop gain values are shown.30 Gainvphase plot showing .30". form a locus.30°. if the point 13. consider the requirement that / G M ( j w ) = . As an example.
771 dB A Nichols chart is formed by superimposing unity feedback closedloop (a) constant magnitude loci and (b) constant phase loci on the openloop frequency response chart. Note: All closedloop phase contours shown are those of negative phase angles. whereas the curvilinear contours correspond to the closedloop system. e. at the frequency w ' .4. be .60" corresponds to the gain and phase of the openloop system G ( j w ' ) for some frequency a'. on the curvilinear contour.31. 40" signifies a closedloop phase angle of . as shown in Fig. On the Nichols chart.31 Nichols chart So.182 Frequency response analysis Phase Gain (dB: Fig. 5. if the point 0 dB/. as described in 1 . 5. the frequency response of an openloop system can be plotted in the normal way by employing the horizontal phase and vertical gain axes. The horizontal phase and vertical gain axes correspond to openloop system values.then the unity feedback closedloop gain and phase will 30".g.40".
~ ~ ~~~ Example 5.3 and plotted on the Nichols chart in Fig. Fig.32 Nichols chart for Example 5.6.6. from the curvilinear contours.32. 5. the closedloop gain and phase values that the system will produce when connected within a unity feedback loop.3 .2.3 Consider the openloop transfer function: G(j w ) = (jw + l ) ( j w + 2)(jw + 3) 20 The gain (dB) and phase of this openloop system. 5.6.The Nichols chart 183 Example 5. It is then possible to read. over a range of frequencies w rad/s is tabulated in Table 5.
or a plot of G ( j w ) H ( j w )can be made (see Appendix 4). 5.10. that is selecting a compensator in order to reshape the Nichols plot.s ~ 2 1.e.6 dB. it is often possible to provide a satisfactory unity feedback control scheme by merely tailoring the openloop gain K . whichever approach is employed.14 . stability results are the same. then either the closedloop system must be reposed in a unity feedback form. when 1 G M ( j w ) I = 1/2. It may well take several attempts before a particularly desirable shape is obtained.74' . However.87' 4 13. Fig.48O 90' 2. for any chosen value of openloop gain and phase. 5. I I 1 . . As with the sections on Bode and Nyquist compensation. In common with Bode analysis.7 Compensation using the Nichols chart Employing the Nichols chart in order to compensate a system response results in a similar process to Bode and Nyquist compensation. a unity feedback closedloop system then being connected around the cascade combination. 5.13' 3 7.67 . When the feedback N ( j w ) is nonunity.3 Gain and phase values for the system in Example 5.53' The closedloop gain and phase can then be read directly off the chart. Hence any alteration to the system gain can be regarded merely in terms of an addition or subtraction to the original.192. This is because. and they can very often be met with little difficulty. and/or is employed as the compensator. it is considered here that the compensation is applied in cascade with the openloop system.02 . in many cases it is only steadystate and relative stability conditions which are most important.43 .6.184 Frequency response analysis Table 5. Nichols chart compensation has the advantage that logarithmic gain values are plotted. In particular the plot can be seen to intersect with the closedloop .172. i. a variation in gain K has no effect whatsoever on the corresponding phase angle.1 Gain compensation For relatively simple performance specifications.an increase in gain K results in the gainvphase plot being shifted up by an appropriate amount over all frequencies.27 _ 5 18.119.92" I G(jw)l /G(jw) 10. for any frequency. when the closedloop gain locus is equal to .07 205.142.3 w ~ 0.4 6. If the openloop system is given by G ( j w )= K B ( j w ) / A ( j w ) .7.180" phase locus. into a more desirable form.11 refers. whereas a decrease in gain K results in the gainvphase plot being shifted down by an appropriate amount over all frequencies. of an openloop system.
it can be seen from Fig. is shown as the continuous plot in Fig.33.1 185 Consider the openloop system transfer function: G (jo)= (jo+ l)(jo 200K + 2)(jo+ 3) The gainvphase plot.e. with K = 1.5 dB.Compensation using the Nichols chart Example 5.6" whereas the gain margin is approximately . 5.33 that the phase margin is approximately . It is required that K is chosen in order to provide a phase margin of at least 40" and a gain margin of 8 to 12dB.10. i.7.30. K = 0.50" 120" 90" 60" 30" 30 20 / Fig.33 Gainvphase plot for Example 5. 5.7. for K = 1. Because both the phase and gain margins are negative values. Without gain compensation.20 (a) (b) (c) (d) Gain margin Phase margin Gain margin Phase margin J Uncompensated system.1 .1 1 0 Gain (dB) 3 .10 . K = 1 Compensated system. the unity feedback system is initially unstable. 5. Phase 210" 180" 1.
7. in order to obtain the compensated plot. 5. where the compensator transfer function is defined by: jw +b D(j w ) =  jut a when the gain of the openloop system transfer function G ( j w ) is plotted in dB against phase. thus achieving a gain margin of 9.135. as well as a selection of terms a and b. In fact a choice of K = 0.5 dB. The value K is therefore assumed to be the resultant openloop gain value. by an appropriate choice of gain K . a more complicated compensator is necessary.34. In general. For a lag compensator. The gain (dB) versus phase plot for a typical lag compensator is shown in Fig. A check can then be made to find that with K = 0. it is not in general possible to achieve even fairly minimal performance objectives by such an elementary procedure.5O when the system gain is OdB. the larger is the ratio b : a so the more negative is the peak phase value and the larger in dB magnitude is the steadystate gain value (when w = 0). in relation to the gain at higher frequencies.5 dB. If this is reduced by 20 dB. after taking all gains into account. The actual gain value of the compensator cascaded with the plant at any frequency is something which can be selected by K . 5. From Fig.1. for this example.5 dB at a phase angle of . reduces the gain by 20 dB and therefore provides the desired gain margin. a possible solution to many cases being described in the following section. because of the lag compensator. or more likely the controller itself may have a gain associated with it.1.11. Where gain and phase margin values are required to a much greater tolerance and/or where further performance objectives are specified. well within the required range. the gain is initially 10. once a ratio b : a has been chosen. This may be carried out in terms of a modification of the actual system gain.186 Frequency response analysis With the phase at . well above the minimum required. Although. so the effect of the cascade compensator D ( j w ) can be merely added in terms of both dB gain and phase angle. 5.5'.180". and a < b. The lag compensator . For either a lag or lead compensator design. at each frequency. it was possible to provide adequate compensation simply by selecting the gain value K . 5.34 it can be seen that the gain at lower frequencies will be increased. as on the Nichols chart. the phase angle is negative in the midfrequency range. the phase angle is . The phase margin is therefore approximately 44. then the overall system gain will be 9.2 Lagllead compensation The connection of a lagllead compensator in cascade with the original openloop system within a unity feedback closedloop scheme is shown in Fig. it is normally the case that the openloop gain K can be specified.180".
50" Phase . 5. a = 6 .25" Fig.10 .5o' Fig. a = 1 0 5 w increasing Gain (dB) .35 Gainvphase plot for a lead compensator with b = 1.I5 .34 Gainvphase plot for a lag compensator with b = 6 .Compensation using the Nichols chart 187 15 LO w increasing Gain (dB) 5 w=OD / 0 4 .25" Phase . 5.
In general.188 Frequency response analysis therefore emphasizes lowfrequency modes at the expense of highfrequency modes. with no compensation applied. it can be seen from Fig. 5. without steadystate conditions being affected.7. because of the lead compensator. This design can be achieved by means of a phase lead compensator. Note that gain and phase margin results can be obtained either by exact calculations or as approximate values measured from the Nichols chart.The gain (dB) versus phase plot for a typical lead compensator is shown in Fig. on condition that the phase is not also affected to any great extent at that frequency. By taking into account possible use of the gain K. 5.472 radls). thus speeding up the system response and thereby increasing the bandwidth. with no compensation.04 dB is obtained by selecting K = 4. the larger is the ratio a : b. so the more positive is the peak phase value and the larger in dB magnitude is the steadystate gain value (when w = 0). The lead compensator therefore emphasizes highfrequency modes at the expense of lowfrequency modes.35 that the gain at higher frequencies will be increased. if it is ensured that when the system is compensated the gain margin will lie in the required range. the phase angle is positive in the midfrequency range. Such phase and gain margin results reveal that a unity feedback closedloop system.907 rad/s) and a gain margin of 21. Therefore. Suppose that a requirement is made for the system to be compensated such that the gain margin lies in the range 8 to 12 dB and the phase margin lies in the range 25" to 35".472 rad/s by approximately 12 dB. In fact an increase in gain of 12. along with a suitable choice of gain K. 5. A Nichols chart plot for this system.35. For a lead compensator. With the compensator defined as jw+ b D( j w ) = jw+ a . This reveals a phase margin of 60. in relation to the gain at lower frequencies.4" (when w = 0. thus slowing down the system response and thereby decreasing the bandwidth. In common with Bode and Nyquist designs.2 Consider the openloop transfer function: G(jw)= jw(jo + 2 ) ( j w + 10) 20K with K = 1.6 dB (when w = 4. should such a more complex form prove necessary or useful. it is possible with a Nichols chart design to apply a compensator which contains both lag and lead networks. is shown in Fig. would be rather sluggish. and a > 6. this can be realized by increasing the gain at w = 4.36(a).  ~~~ Example 5.
e. at higher frequencies a gain increase of 12.20 (b) (a) Fig. frequencies by the compensator gain.36 Nichols chart plot for Example 5.7. As long as little or no phase change takes place at the frequency for which the . So.1 a 4 K (J=O Thus the gain increase. D ( j ( J = .2 (a) Uncompensated system (b) Compensated system the overall gain at low frequencies will not be affected as long as a / b = 4.10 .=1 = ) b . whereas at higher frequencies no such counteraction takes place. is counteracted at low.04 dB is apparent. 5. i.Compensation using the Nichols chart Phase 189 1" dB 30 20 lo Gain (dB) 0 . K.
due to the compensator.36(b) in terms of a phase angle of much less than 90°.4" would be obtained.04 dB] 20 rad/s and at this frequency the phase is: / G ( j o ) = 90" tan'[:] tan'[. A choice of 4b = a < 0. The design has therefore been accomplished in terms of the high and low frequency properties of the compensator.04 dB.6.36(b). the midfrequency phase shift being placed where it is of no consequence. been arranged to occur at frequencies much lower than those at which the gain and phase margins are determined.1 is shown in the Nichols chart plot of Fig. occurs at w = 2.e. this will become the frequency at which the overall gain is 0 dB. and hence a choice of b = 0.12. I will suffice. a gain margin of 9. with K = 4. The resultant effect obtained by employing a gain of K = 4 along with the lead compensator: D(j w ) = j w + 0. The phase shifting property of the lead compensator has. i. Further. 5. This can be seen on Fig. 5.] =  154. 5.454 rad/s.025. Solving IWw)l [K= 11 gives w = 2. error constant values can be reduced by means of a lead compensator. the bandwidth has been increased.454 = w(w2 + 4)'/2(w2 + 100)'/2 = 0. then both the gain and phase margins will normally be reduced. which is within the stated range of 25" to 35".6" such that a phase margin of 25.36(a) the phase angle is never less than 90". by holding the gain and phase margins steady. and can be seen as speeding up the system's response. as was done in the example. a = 0. with K = 1.25 [ = .190 Frequency response analysis overall gain was . The standard design procedure for a single lag compensator. thus meeting the performance requirements. If the steadystate error constant is held at a steady value. whereas for the uncompensated system in Fig. Employment of a lead compensator has resulted in a higher frequency at which the gain is 0 dB.453 rad/s (as opposed to w = 0. see Problem 5. An increase in bandwidth is a property usually obtained by the use of a lead compensator.907 rad/s for the uncompensated system). Conversely.12. in the example.2" are obtained. The frequency at which the gain is OdB (unity magnitude) is now approximately w =2.025 jw + 0.9 dB and phase margin of approximately 27.245 would ensure that almost no phase change. involves .
the tendency being towards a greater extent of relative stability. Hence if a gain margin of 10dB is found for a system by employing Bode analysis. such that they lie in the midfrequency range and therefore do not affect the specifications already met by the compensator. in particular the stability properties. Stability and relative stability closedloop results can though be taken from the graphical plot. At higher frequencies the overall gain is thereby reduced. exact calculations can be made from which both the stability properties and graphical plots are obtained. or from ideal transfer function calculations. For each of the methods.Summary 191 firstly selecting the system openloop gain value. this is the same data required for all three of the techniques considered. Whether the frequency response data has been obtained from realworld experimentation.e. For each of the three techniques. in the examples used. whichever method is used. of a unity feedback closedloop system. 58 Summary The techniques described in this chapter have all been concerned with analyzing a system in terms of its frequency response. that is one of assessing the properties. it is likely that a number of lag and/or lead compensators must be cascaded in order to provide the degrees of freedom necessary to solve the problem. the gain margin will still be 10 dB if either Nyquist analysis or the Nichols chart are used instead. The different techniques are merely different ways in which the data can be presented. and in most practical cases the requirements will indeed be this fundamental. it is unlikely that an ‘exact’ transfer function will be known. in order to meet gain margin requirements (an increase in gain margin will be required). For many practical realworld systems however. such as specified error coefficients to relevant input signals. .and the a and b compensator coefficients. Concepts of stability and relative stability are also common to the methods. The ratio b : a can then be found in order to produce necessary steadystate conditions. The phase shifting properties of the compensator can then be selected. It may be however that either tighter margin limits must be met. by employing the characteristics belonging to the openloop system frequency response. gain and phase information being found over a wide range of frequencies. i. the gain margin and phase margin values are not altered because a different method is used. If the system transfer function is well known. The performance specifications were given as fairly straightforward phase and gain margin requirements. the basic procedure carried out is essentially the same. t o an extent which allows for a good graphical plot to be made. Each of the methods discussed is a different way in which a graphical representation can be given to system frequency response data. or the compensated frequency response plot is required to pass through further defined points. In these cases the frequency response data is collected by experimentation. K. the design of a suitable lag or lead compensator in order to meet specified performance objectives was considered. Although it might be possible to meet these demands by simply fiddling around with the gain K.
4 Employ a Bode plot to find a phase lead cascade compensator.3 5. which produces a unity feedback closedloop system with a minimum gain margin of 10 dB and a minimum phase margin of 30".5 Obtain a frequency response plot for this system when 00 < K < 0. then the steadystate output signal is also a sinusoid. For the uncompensated openloop system in Problem 5.1 Find the maximum phase shift due to a phase lead compensator. with K = 1: By means of a Nichols chart design. Consider the openloop transfer function: K G ( s )= (s + l)(s + 2)(s + 3) 5.2 1OK s(s+ l ) ( s + 2 ) ' K= 1 5. and the frequency at which it occurs ( a < b for a lag compensator).7 By means of a Nyquist design.5.41 produce a phase margin in the range 30" t o 40" and a gain margin in the range 8 dB to 12 dB. and by using the Nyquist criterion. find the phase lag compensator and gain K which will produce a phase margin in the range 30" to 40" and a gain margin in the range 8 dB to 12 dB. For the uncompensated openloop system: G ( s )= 20K s(s + 2)(s + 10)' K=l 5.6 5. find the minimum phase shift due to a phase lag compensator. such that y ( t ) = Y cos(wt + 4). The compensator is given by D(j w ) = jw +a where a > b. . where Y = I G ( j w ) I U and / C W ) = 4. find the range of values for K over which the unity feedback closedloop system will be stable. Consider the openloop transfer function: G ( s )= jw +b 5.192 Frequency response analysis Problems 5. Show that G ( . Similarly. and the frequency at which it occurs. Show that when a sinusoidal input signal u ( t ) = U cos w t is applied t o a stable linear system. of magnitude Y . where G * ( j w ) denotes the complex conjugate of W w ) . find the phase lead compensator and gain K which v.j w ) = G * ( j w ) .
over which the closedloop system will be unstable.6 exp( . and ascertain the damping ratio. When a step input of magnitude 0. and the bandwidth. find the range of values for K.Problems 5. What value of 6 is required for such a design? . the output response is equal to [6 . when the system is connected within a unity feedback loop. Design a lag compensator in order to achieve a phase margin of approximately 45" and a minimum gain crossover frequency of w = 1 rad/s.2) 2K 5.13 For a unity feedback system. find the gain and phase margins by means of a Bode plot.9 A unity feedback system has an openloop transfer function G ( s )= s(s + 2)(s + 5) 200K 5.8 193 Sketch Bode plots for the systems whose openloop transfer functions are: (a) G ( s ) = s(s + 1) (s + I( + 6 ) )s (b)G ( s )= (s + 2)(s + 3) G ( s )= s2(s + l)(s + loo) s + 10 5.12 By means of a Bode plot. find the phase and gain margins. find the value of a which results in a gain crossover frequency of w = 100 rad/s. A phase lead compensator of the form is employed such that the phase margin is approximately 45".4)(s + 0. Obtain the Bode plot for this system and find the phase and gain margins. The openloop transfer function of a unity feedback system is G ( s )= (s + 2)(s + 0.1 1 By means of a Bode plot .10 Obtain a Bode plot for this system when K = 1 and find the phase and gain margins. Also.t ) ] units. Given the openloop transfer function G ( s )= ~ s(s 3(s+ 1) + a) 5.when K = 1 .04 units is applied to a certain openloop system.5) s2(s + 5)(s + 3) 5. The openloop transfer function for a particular system is G ( s )= 30(s + 0.
For the openloop system transfer function G ( s )= s(s + 1O)(s + 25)(s + 40) 0.37 5.15 Also.194 5. the Bode gain is approximately 6 dB when w = 60 rad/s. and the bandwidth. Obtain a Bode plot for this system. By means of a frequency response plot.37 and obtain a Bode plot in each case. J. R I = 100 kQ = R2. . find the values of K for which the unity feedback closedloop system. and find the phase and gain margins. Fig.17 For a system with openloop transfer function G ( s )= s + b ~ s(s+ 1)Z O<b<m 5.5 find a cascade compensator.16 in order to obtain a gain margin of approximately 12 dB and a phase margin of approximately 40". The input signal V ( s ) and output signal Y ( s ) both correspond to voltage transforms. C.5. = 2 pF and CZ= 3 pF.14 Frequency response analysis A unity feedback closedloop system has an openloop transfer function of: C ( s )= s(s 3 2 K ( s + 10) + 4)(s + 80) 5. Find the input to output openloop transfer functions for the electrical networks shown in Fig. of the form 5.18 find the three different forms which a Nyquist plot can take. whose openloop transfer function is G ( s )= s(s + l)(s + 2) K (a) has a gain margin of 6 dB and (b) has a phase margin of 4' 0.
McGrawHill (Schaum).. C. 1976. pp. . find the phase margin for a unity feedback system whose openloop transfer function is 41 0 C ( S ) ___ = (s + 5. J. R. 133.. find a suitable phase lag cascade compensator which will produce a gain margin of approximately 10 dB and a phase margin of approximately 45”. Brockett.Principles o automatic control. ‘Frequency domain stability criteria . Proc. IEEE Trans. Stubberud. L . . E. Brogan. Measurement and Control. PrenticeHall..5) 1 5.D. J . (c) the magnitude and phase of any asymptotes. pp. L. IEEE Trans.21 Find a phase lead cascade compensator which will produce a phase margin of approximately 45” and a gain margin of approximately 10 dB. 18990. 1986.22 by means of: (a) Nyquist design. 1972. . Brockett. Douce. (b) Nichols chart design.20 and find (a) the frequency of real axis crossing. 25561. W. Eveleigh. pp.. J. Hodder and Stoughton. IEE. J. L.19 Sketch the frequency response plot for the openloop system G ( s )= s(s + l)(s + 0. ‘Closedloop transient response from the openloop frequency response’. Healey. M. R. No. L. W.. AC10. 1978. pp. DiStefano. AC10. Bell System Tech.Part II’. 40713. Feedback and control systems. I. W. 4. J. By means of a Nichols chart design. H. f Hind. 1940. where the original openloop system is: G ( s )= s(s + 5)(s + LO) K Further reading Bode. 1991. J. Tata McGrawHill. ‘Relations between attenuation and phase in feedback amplifier design’.Part I’. pp. Pt... J. 1965. where the original openloop system is: G ( s )= s(s + 2)(s + 8) ’ 40K K=l 5.. and Williams.. Modern control theory. ‘Frequency domain stability criteria . (d) the gain margin and (e) the phase margin. (b) the frequency at which I G ( j o )I = 1.Further reading 195 5. 1981. on Automatic Control. 42154. and Willems. 3028. on Automatic Control.1965. W. A. By means of a frequency response plot. V. W. and Willems. ‘A note on frequencyresponse measurement’. Control systems design. R.
J . systems and transforms. Control engineering. J. IEEE Press. 1991. Oxford University Press. McGrawHill. and Harbor R. Frequency response methods in control systems. K. Tech. 1984. Schwarzenbach. 1991. H.. C. G.. J . ‘The frequency response method: its relationship to transient behaviour in control system design’. . Multivariable feedback design. J. Hunt. L. McGrawHill. and Gill. 1932.. B. F. Raven.. Synthesis of feedback systems. Chapman & Hall. Westcott. pp. 1991. R. of Soc. Nyquist. 12647. MacFarlane. 0. . L. Trans. Edward Arnold. 4th ed. D. of Instr. Mechatronics. V... 1974. 1989. 1990. 1952. Morris. N. 1963.196 Frequency response analysis Horowitz. M. 4. Phillips. I. 1979. Maciejowski.. No. Signals. AddisonWesley.. H. A. C.... Introduction to control theory. Automatic control engineering. Feedback control systems. Jackson. 11324. Bell System Tech. M. pp. 3. Jacobs. Academic Press. 2nd ed. H.. ‘Regeneration theory’.. J. F. PrenticeHall.. AddisonWesley. System modelling and control.. 1990.
Once in the statespace it is possible to transform one representation into another by means of operations on the state variables in question. rugby or any other) in which the system input is effort on the part of the teams. and to show how the system descriptions obtained when using the state are arrived at. often scalar numbers. The output is the score in terms of points or goals at the end of the game. All design has been based on the differential equations describing the relationship between input and output. depending on the state variables chosen. . the state is not always a measurable variable. It is shown how the roots of the transfer function denominator can be important in these transformations and the possibilities for moving to a state representation with no obvious link to the original transfer function are suggested. which if known along with the system input will result in complete knowledge of firstly any future states and secondly the system output. The approach taken in this chapter is to start from the original sdomain transfer function in order to obtain the time domain statespace model for a defined state vector. in the football match example it could be the team’s morale during the match for instance and. The state of any particular system is not necessarily unique. In general the state is a set of values. as in this case.6 State equations 6. The purpose of this chapter is to introduce a further variable.1 Introduction In the previous chapters emphasis has been placed strongly on an input/output sdomain transfer function description of the system. then. consider a football match (American. When considering one system. whereas the system state could be the score at any point during the game. often these equations have been built up from individual system components. and hence simpler representations can be found that exhibit certain system properties. a large number of statespace representations are possible. subject to the dynamics of the system also being known. namely the system state. As an example of the state of a system. unlike the system input and output.
to be equal to the output y ( f ) .4) can be written as 1 0 a2 0 a1 (6. let D 3 y ( t ) + azDZy(t)+ aiDy(t) + U O . defining the state.2.2.3) This is a rather simple case to serve as an example of a statespace system description. x l ( t ) .it follows that (6.2.2 State equations State equations from transfer functions Consider a system with one input and one output. Y ( ~= b o ~ ( t ) ) (6.2. 6.1 Matrix representations Assuming an extended version of equation (6.u ( f ) U U (6.2) Then. That is: 1 b y ( t ) = .2.1) can. let XI ( 0= A t ) Xz(t)= D~l(t)=Dy(t) ~ 3 ( f = DXz(f) = D2y(t) ) Then.J q t ) .4) Now however. Also i ( t ) = (d/dt)y(r) = Dy(t). the description will now be extended to cover more generalized and complicated expressions.2.2. a singleinputsingleoutput (SISO) system. but also on the rate of change of the output.2. Equation (6. however. whose output depends not only on the system input.5) .2).198 6. y ( t ) the system output at time t and u ( t ) the system input at time t.1) where a and b are scalar values.2. (6. also be written in the form: y ( t )= a y ( t ) b u ( t ) + (6.
2. It can be seen from (6. the initial state. is then introduced such that: 1. &) = Ax(r) Bu(r) + and y ( t )= Q ( f ) (6. as can be seen when D 3 y ( t )+ a2D2y(t)+ a l D y ( t ) + a o y ( f )= bzD2u(t)+ b l D u ( t ) + bou(f) (6. is often equal to the null matrix.*+ alD + a o ) y ( t ) u + = ( b m D m + r n .I D " .9) The states can now be defined as: XI ( t )= z ( t ) ~ 2 ( f= D z ( f ) ) x3(0 = D2z(t) which gives a representation equal to: 1 (6. in more general terms. all its elements are zero. z ( t ) .8) The variable. which gives a direct relationship between system input and output. is known. u ( t )= (b2D2 blD + bo)Z(f) + (6.1 1) .2. In the equations considered. u ( f )= ( D 3 a2D2 + + U I D+ a o ) z ( t ) 2. It will be seen that the matrix D.2.8) that a general form of the transfer equation is: [:.2. i. This is not necessarily the case in general. one of the states is made equal to the output.2.e.7) or ( D 3+ a2D2 + alD + a o ) y ( t )= (b2D2+ blD + b o ) u ( f ) (6.State equations from transfer functions 199 or.2.6) + Du(t) where A is the coefficient matrix and x(2) is the state vector at time 1.10) a0 a1 a2 and y ( t ) = [bo 61 brl_x(t) in which x ( t ) = x2(t)  and _x(to).~ D m+'* * b  + btD + bo)u(t) (6.' .2.I (D"+ ~ .
The differences between representations are dependent on the method of state definition.2 .b n . 6 1 . though. .12) can result in state equations of the form: u(t) (6.2.10).. This is. the two descriptions. whereas Mesh 2 has a current i 2 ( t ) . this simple transformation from one description to another does not exist and a return to the Doperator description by means of an inverse Laplace transform is necessary. however. the function (6. when considering the sdomain transfer function. y ( t ) . If it is assumed. and system output.1. assuming zero initial conditions. that initial conditions are zero.. If nonzero initial conditions are apparent. are identical when sJ is replaced by DJ.2. 6. Hence. is merely another way of writing the transfer function from input to output. bn11 .2. u ( t ) .2 Transfer functions The differential equation (Doperator) description of the relationship between system input. Mesh 1 has a current i l ( t ) flowing through it.2.2.2. just one of an infinite number of possible statespace representations of the transfer function (6.12). 6.13) and Y ( t ) = [bo. and follows as a generalization to (6.3 Worked example Consider the RLC circuit shown in Fig.200 State equations 6. Doperator and sdomain.
Ri/Lz Ri/Li 0 Ri/L2 z ( t ) + 0 Ri/Li 1 :. and the output signal y ( t ) = uo(t). However.3 Solution to the state equations Consider the state equation g t ) = &(t) + Bu(t) (6. 6. all states. for all t 2 to. xs(t0) = il(to) and u ( t ) = ui(t) are known for all t 2 lo. x l ( t ) . The state equations can then be written as: & ( t ) = .1) . uo(t). without being measured.1/Lz and y(t)= [: l/C .Solution to the state equations 201 ebb6 h(0 v. [ 1 u(t) 0 OIz(t) [ X l ( t ) . can be calculated.x2(t).xz(t0) = i2(t0).(t) ___) iz(0 c == Assuming that the input signal u ( t ) = ui(t).T ( t )= X In this example the states have physical meanings and can be measured if desired.x3(t)l where .as the state. including the output signal. the currents i l ( t ) and i2(t)can be classified as the states xj(f) and x z ( t ) with the output.3. as long as xl(t0) = uo(to).
the initial time.5) An alternative approach in the calculation of the state transition matrix is to take the Laplace transform of (6.3.3. This equation must therefore be solved such that under these conditions a unique state vector is found. equation (6.to)l_x(to) x (6.1 The state transition matrix The n x n state transition matrix.4) The state transition matrix is then expressive of the natural system response when the input is zero.t o ) = expM ( t . is the solution to equation (6.3.3. if inverse Laplace transforms.A ] is nonsingular and therefore invertible: ( s ) = X (6.3.1) with the input u ( f ) set to zero. and relates the state vector at time t to the state vector at time to. # { t ) . .3.202 State equations The righthand side of this equation consists of two parts. Where to is considered to be zero: dJ(t) = exp[A (t)l (6.3.'_x(O) (6. B u ( t ) . becomes: 6 ( t )= LT*( [SI A ] . it satisfies the homogeneous equation $(t) = &(t) (6.A ] .2) With initial conditions at time to. In its present condition.3.A]')_x(O) x (6.z(0)= A&(s) (6.3. 6.2).3.3) in which 4 0 .6) By rearranging this expression: [sl.8) Now. to give: s&(s) .3.3. & ( t ) .' ]= exp[A(t)] (6. is known as the forcing function.t0>1 (6.3. for any A .9) for all t 2 0 such that the state transition matrix. the solution to this equation (6.' " s l .10) . d ( t ) .7) [sl. That is.1) will not give an explicit state vector value due to h ( t ) on one side of the equation and _x(t)on the other.3.( t ) = g .A]_X(s) _x(O) = On the assumption that [ s l . is known as the homogeneous part whilst the second. denoted by 9'[ are taken 1.2) is ( t ) = exp[A ( t . the first of these. B and u ( t ) .
A t = ( e A ) .' ( t ) =e . i.tl) d(t1 .' ( f ) = qq.Solution to the state equations and the zeroinput response with initial time set to zero is therefore: x ( t ) = W)_x(O) 203 (6.. via the state transition matrix. d(0) = exp[0] = I (b) 9(0 * 4 ( t ) . .3. has been found. it is now shown that this matrix has several important properties.1 = 4 ( .to) Further properties follow from those already given. .3.e.12) .. . In this section it is considered that the forcing function is nonzero.4 ( t )= eAteAt eA1 . the state transition matrix as a solution to the state equations. of the homogeneous part of the equation.to) = eA ( t i . but will not be considered here.t ) * 9 .. a matrix in which every coefficient is zero. assuming zero input. 4 ( t ) = d r n ( t ) = 9 ( t m ) This holds because: 4 ( t ) 4 ( t ).t l ) Ii Wl . 6.z(0)= A&(s) + BU(S) (6. eAli.eA(tztu) = 902 .. ) 631: S&(S) ..3. * 4(f2 . then by taking the Lapiace transform of ( .t o ) @ti (d) .3. in the previous section.t o ) = 9(t2 ..+ At e +A1 and emAt mAt = (eAt)" = 6 " ( t ) or emAl = (eA)tm = 9 (tm1 (c) 4 .l ~ ) ~ A (hi) . these are as follows: (a) If A is the null matrix = O(n x n). eAtu = eAt>eAti~ .3 General solution of the state equation Thus far the solution.t ) 4(t2 . .11) 6.2 Properties of the state transition matrix Having found.eAti.81 + .
14) can be written as: (6.3.3.A ] is then: [sl.3.such that: but.' ] .A ] .A ] . the following is obtained: ( t ) = g .1 1 3 s+4 The inverse of this matrix [sl.3.13) is identical to (6.3.A ] .7)Bu(7) d7 (6.13) where the first part of equation (6.4 Worked example A unit step function is applied to the system whose state equation is Find the state transition matrix and hence solve the state transition equation. By taking the inverse Laplace transform of (6.A ] .A]')_x(O) 9'( x + [sl. :]=[s . The solution is as follows: slA=[O o s B][.' B U ( s ) (6.3.3. (6. d J ( t )= 9'[ s l .3.13). however.A ] . a further equation arises in that: [ ~ lA]'Eu(s)=  and taking the inverse Laplace transform of this results in: .' = The state transition matrix can now be obtained by taking the inverse Laplace [ transform.16) 0 6.' B U ( s ) ) (6.'_X(O) + [sl.204 State equations Rearranging this equation gives: &(s) = [sl. If all the elements in the initial state vector are also zero.8). then the zerostate response is given by: ( t ) = x i' + ( t .3.' ([sl.IS) where the initial time is assumed to be zero.14) and using the representation for the state transition matrix.
The inputloutput relationship is therefore given by: Y ( s )= [ C[Sl.A ] in its inverse form appears as the basis for the transfer functions found. In fact. the value of these elements at any subsequent time t can be found. the denominator of the expression (6. the term [sl.4. also be written in the form of (6. Taking the Laplace transform of the initial output state equation (6.A I D IslAl U ( s ) (6. 6.The characteristic equation 205 The overall solution to the state transition equation is therefore: Hence.6): Y ( s )= C & ( S ) + DU(s) (6. this follows from the transfer function definition.1).6) .A ) must be nonsingular.A I = sn+ an. initial conditions ) must be set to zero.2. if the initial values of the elements in the state vector at time to = 0 are known.A ] ) B + (sl. where this is not so: I sl A I =0 (6.13) to be dependent on initial state values and the system input.3.4.5) and this is called the characteristic polynomial of the matrix A . _X(s).3) C ( a d j [ s l .1s".12) as: .4)is: I SI .4.and this leads to: Y ( s ) = C [ S ~A]'5(0)  + [ C[d A]'E + D)U(S) (6.4 The characteristic equation In the previous section it was shown how the state of a system at any time t can be found in terms of the input function and the initial state values.A ] . ~ ( 0= 0.4) Hence.4.e.4.4.2) To obtain the final transfer function relating input to output.1) The state vector. This matrix is also apparent in the overall transfer function relating system input to output as can be seen from the following.s ) Y( (6.4.' B + D ) U(S) This can. though.' + + als + a0 (6.4. For the existence of a finite transfer function the matrix ( s l . however.2. i. was determined in (6. this equation can therefore be substituted into (6.
substituting for X I = 5 . i.8) Example 6.206 State equations and this (6.4. .7) (6. X I = 5 .6) is termed the characteristic equation of the matrix A . using equation (6. A2 = 2. the denominator of the input to output transfer function is set t o zero in order to obtain the characteristic equation.4. [XI3 1 XI4 2 ] [.4.7s + 10 = 0 The eigenvalues are thus.4. Eigenvectors are associated with the previously obtained eigenvalues such that if Xi denotes the ith eigenvalue of matrix A . The roots of the numerator of the input to output transfer function are. Firstly the eigenvector corresponding to X I .7). 6. its associated n x 1 dimensional eigenvector. Uir is found to be such that (XilA)vi=O (6.4. These can now be used to obtain the eigenvector.1 Eigenvalues and eigenvectors The roots of the characteristic equation are referred to as the eigenvalues of the matrix A o r the poles of the input t o output transfer function.1 Let the n = 2 dimensional matrix A be defined as: The characteristic equation is therefore obtained from: or sz .I =O or.4. on the other hand.e. called the zeros of the transfer function.
. + BnIsnl It follows though.I A ".A ) (sIA~ (6. a(s)=IslAl =S"+anIs"L+.4..4.4.A ) * adj(s1 A) (6. Hence the eigenvalues are X i A2 = 2.2 CayleyHamilton theorem The CayleyHamilton theorem states that every square matrix A satisfies its own characteristic equation.10) (s"+ anls"' + + U ~ S +& ) I = (d. that by rearranging (6. ( n = 1) or less.9) where the adjoint of (sl.+als+ao then a () 6 A ~ + an .+ + A + a01 = O Proof: The inverse of (sl. In other words. with associated linearly independent eigenvectors = 5.4. Secondly. such that the matrices Bi can be defined as: adj(s1 . if A is an n x n dimensional square matrix resulting in a characteristic polynomial.1.A ) = Bo + Bls + B~s' + .A ) is a matrix whose elements are all polynomials in s of degree less than n.I = adj(sl . It will be seen in the following section how the eigenvectors are of great importance in matrix diagonalization. i.4.A ) can be obtained as (sl..A ) .4.12) . 6. the eigenvector corresponding to must be found from: and this is satisfied when U Z I= 2.The characteristic equation 207 A2 This is satisfied if ut 1 = U I Z = 1.9) (6. u22 = .e.4..11) or by using (6.) * (Bo + B ~ s + A + Bnls"') (6. Firstly however a result fundamental to the study of linear systems must be introduced.10) Isl A I * I = (sl.
4.7 s + 10 = 0 Substituting A for s leads to a ( A ) = A 2 .ABI = a l l .ABo = aoz The first equation can now be multiplied throughout by A " .7 A + 101=0 Therefore. and to find A . and its characteristic equation is a(s) = s2 .4. the following relationships result: En[ = I En2 AEn1 = an11 En3 . the final (bottom) equation being multiplied by I.' gives: A71+10A'=O or A .1 satisfies its own characteristic equation.z ) + Z + A (Bo . by equating like powers of s in (6. = A"Enl + A " . This results in A"+ anlAnl + u~ZA"' + + U ~ + a01 A a * .~ .2 1 Bo .208 State equations Hence. multiplying both sides of this equation by A .l using the CayleyHamilton theorem.A E n .7A + 101 = 0 or Also A 2 .A E I ). Example 6.AEn2 = ~ ~ . the second equation bl ~ n I and so on.j .' ( B ~ .l = &(71.13: The CayleyHamilton theorem is thus proved.4.2 To show that the matrix A in Section 6.12).4.A ) + a ( ~ ) = [ 11 14 21 18]  [ 14 [ 10 0 =O .A E ~ AI") + ( B n .A& =0 (6.
5. Further operations on the system matrix will now be considered in the following section.5.2) By substitution of (6.5. especially with regard to the A matrix. It can be the case.1 State variable transformation Let the ndimensional system be characterized by the state equations: g ( t ) = &(t) + Bu(t) (6. these being termed canonical forms.3). therefore.2 it was commented that for any particular transfer function there are an infinite number of possible statespace representations. Either Q&'(t)= A Q g ( t ) + Bu(t) + Q'Bu(t) (6. In order to simplify any subsequent analysis and design it is preferable to transform the state vector into a condition such that the A matrix corresponds with one of a number of simplified forms which exhibit specific properties.3) or  = Q'AQx'(t) .3 I These properties show that this method can be used to obtain A k where k is a large number and successive multiplications of the A matrix would prove time consuming.5 Canonical forms In Section 6.1) and y ( t ) = C x ( t )+ Du(t) where A is an n x n coefficient matrix. 6. that the state representation obtained results in a complicated expression. these being dependent on the definition of the state vector itself.5.5. This section describes firstly the method of transforming a state vector and secondly a number of possible simplified A matrix forms.2 0.Canonical forms 209 Hence 0.2) into (6.5.1) a new statespace representation of the same system is obtained (6. 6.5. The state vector x ( t ) can then be transformed to the state vector ~ ' ( tby means of the relationship ) x ( t )= Q x ' ( t ) (6.1 0.4 0.
a . linear system is said to be in phase variable canonical form if the state equations are written such that: TX Looo 1 0 0 1 .2. as described in 6. 6. Summarizing (6.’ ( f ) x (6.5.5.2 Phase variable canonical form A single input. al. It can be seen that the phase variable form is in fact the same as that described in Section 6.210 and State equations y ( t ) = C Q z ’ ( t )+ D u ( t ) These equations hold on condition that Q is an n x n nonsingular matrix. This phase variable form. Also C’=CQ and D’=D Finally it must be noted that as y ‘ ( t ) = y ( t ) .4) +mu’([) B’ = Q. timeinvariant. the parameters U O . . important for this definition. and Y ( l )= C x ( t > The form of the C matrix and the existence of the D matrix are not. the equations can also be written as: g ’ ( t )= A ’ & ’ ( t )+ E ’ u ’ ( t ) and y ’ ( t ) = C ’.l being the coefficients of the characteristic polynomial. 0 (6.*B where A =Q~AQ.2 for the representation in the statespace of an sdomain system transfer function.2. therefore. ...5.2. in which A ‘ and A are termed similar matrices.the system output and input are unaffected by any state variable transformation. . and u ’ ( t )= u ( t ) ..5) lo Oa 1 Oa2 * . can thus be obtained directly from the system transfer function with numerator coefficients appearing in the C matrix.5. simple because of its sparsity. .an.3).
Canonical forms 6.. A. resulting in a diagonal matrix.7) to hold. Then (6. n .. q 2 . It follows that: Q A = AQ Let the n columns of (6...7). which consists of the corresponding eigenvalues.5. . XZ. Only eigenvectors corresponding to distinct eigenvalues are independent. Qn1 in which qi = Ui i = 1.5. the n associated eigenvalues X I . this is a consequence of (6....8) 0 be denoted by q. a matrix Q must be obtained which satisfies: A = A ' = QIAQ (6. A. . Hence Q = [q1. .. thus where A has multiple order eigenvalues it may not be possible to calculate a diagonal similar matrix. where i = 1. such that: exp X l t exp X 2 t exp A t = exp [. .8) must satisfy: which is true when X i is an eigenvalue of the matrix A . which has to be true for (6. The state transition matrix will.4. =[ 0 (6...5.6) To find a diagonal matrix which is similar to the original matrix A . therefore.5.5. are also the elements of the main diagonal. The matrix Q can therefore be found by means of the eigenvectors of A .5.5.1. with an associated eigenvector qi( = ui). From (6.9) it is apparent that a nonsingular matrix Q can be obtained only when matrix A has n linearly independent eigenvectors. also be diagonal.3 Diagonalization 211 For any ndimensional diagonal matrix... n .4.2.7) where A is a diagonal matrix and Q is that value of Q which when used as an operator on A causes A ' to be diagonal. this follows from the definition of eigenvalues given in Section 6. but when A has n distinct eigenvalues a diagonal transform can always be found.
. consider the matrix A=[ F I I 0 L 2 3 2. the total number of which is equal to the number of linearly independent eigenvectors. each eigenvector being associated with one block only. . the (fl x r i ) dimensional Jordan block associated with this eigenvalue is given by: [I 0 i : : . and hence eigenvalues XI = 3 (a double eigenvalue) and XZ = 1.5 1 1 2 :] 0 The characteristic equation is found from I X I . however.9 = 0 which has roots equal to ( X . an offdiagonal unity term resulting in the block corresponding to the multiple eigenvalue. The Jordan canonical form matrix is therefore There are thus two Jordan blocks.A I as: I XI A I = X3 . XZ. if Xi is an rith order eigenvalue. as described in Section 6.3)2(X .7Xz + 15X. X3.1). X I .5.212 State equations 6. In general. The Jordan canonical form matrix is found to be As an example.4 Jordan canonical form A matrix..J Where ri = 1. This form copes with the possibility of multiple order eigenvalues by means of a set of Jordan blocks.3 with eigenvalues along the main diagonal is a special case of the Jordan canonical form.5. assume matrix A has three eigenvalues. where XI and X3 are single ordered but A2 is triple ordered. As an example. The overall matrix in Jordan canonical form is then made up of Jordan blocks aIong its ‘main’ diagonal. the Jordan block reduces to the eigenvalue Xi.
also be written as a linear combination: (6. XI.6.6.2) Anmtl where only one multiple eigenvalue.Decomposition 6. in terms of multiple order eigenvalues.6.3) such that the state vector can be defined as .6..6 Decomposition 2 13 Having described the Jordan canonical form in the previous section.' + Us) * a * + + b l + bo ~ + ~ I + a0 S '1.1 Partial fraction decomposition This particular method. its relationship with the original transfer function.2sn' sn+ an. (6.2 during the introduction of statespace techniques.2).4) (6. Here. This can. is present and is of order m. however..Isn' + b. .xn(t)l x in which the individual states are found as follows: (6.. will be made use of in describing the various methods available with which a statespace description can be obtained from the original transfer function.6..6.6) .. 6.s ) b n . they are looked at more formally such that the advantages and disadvantages of each method can be considered.1) in which signal initial conditions are assumed to be zero.6. a special case being when all eigenvalues are distinct. however. when multiplied throughout by U ( S ) .I S ~ .5) but also (6. also be represented in partial fraction form as: . can The sum obtained in (6. + (sxl) Cml Cm C + + +m tx12 sXI s + S Cn (6.f ( t )= [ x l ( t ) . whereby a diagonal A matrix will result. One or two of these methods have already been considered in Section 6.6. also known as parallel decomposition. is designed to result in a Jordan canonical form representation.x 2 ( t ) . Note: When m is unity all the eigenvalues are distinct. Let the transfer function to be decomposed be defined as: Y( .) Us V(S) c1 (Sl)m + .
5.6.6.6.214 Stale equations and substituting for this in (6. which becomes: im(t) = hlXm(t) + u(t) I I (6.5).8) or SXI ) = XIXI( s ) + XZ(S) (s In the time domain. XI (s)= X z ( S ) S . this can also be expressed as sXI(S)AiXi(s)= XZ(S) (6.6.6.6.11) .7) as an example. As an example consider the mth state.7) and finally Xn(S) = 1 S Xnm+I U(S) Taking the first equation of (6.XI similarly (6.1) in which the matrices are defined by: I I I I I I I I I B= 1 I 0 I I I I I 1 (6.6. this becomes.9) The same transform can be taken for the remaining n . by taking the inverse Laplace transform: s (1) = X I X I ( t )+ xz(t) i (6.10) Writing the states in vector form results in the standard ndimensional state characterization of a system (6.1 states.
6. .cnl and D= 0 p=nm+l The A matrix is hence of the Jordan canonical form.13) Then by letting Xl(s)= X ( s ) and X i ( s )= sXil(s).6.I s n .13) i becomes: U ( S )= SXn(s) anlXn(s) + or by rearranging + + al&(s) + ~ o X I ( S ) . This may not always be the case and so a more direct approach.6.6. and will reduce to a diagonal form only when no multiple eigenvalues are present.' + + U ~ + u ~ ) x ( s ) S + * a Multiplying out the second of these equations leads to: U ( S )= S " X ( S ) + a n . to obtain the eigenvalues (poles) of the transfer function.an.~+ and + IS + bo)X(s) (6. this results in Y ( s )= (bnlsnl + b n . n equation (6..2. as was carried out in Section 6.= 2 .2 Direct decomposition The previous description relies on the fact that it is possible.6.' X ( S ) + + U ~ S X ( S )+ U ~ X ( S ) * a * (6.a1X2(s) (6.. without too much difficulty.2 ~ " . .6.IXn(S)  aoX1(s)+ U ( s ) When a time domain description is required.. must be taken.. Let the top and bottom of equation (6. then if numerators and denominators are equated. 6..1) be multiplied by X ( s ) .* S " ..14) sXn(S) = . the statespace representation can be built up such that .12) U ( S )= (s" + an.Decomposition Also 215 cT= where [CIS ~ 2 .
have been formulated by means of state transformations such that where a particular state vector gives rise to a problematic description. Problems 6. 6. rather than the more standard input/output transfer function description. Fig. There are various other methods of decomposition dependent on the conditions applicable to both system numerator and denominator.12). has been considered. Find a statespace model for this system by taking the two mesh currents and the output voltage as states.7 Summary The fundamental concept of the statespace as a basis for controller design.216 State equations in which the matrix C is formulated directly from the first of the equations defined in (6. which much simplify further analysis.1 Consider the electrical circuit shown in Fig. The direct method of decomposition therefore results in a phase variable canonical form realization. eigenvalues. Canonical matrix forms. 6.2 with an input signal V i ( t ) .6. a statespace model can be obtained directly. diagonal forms of state matrix. although where multiple roots are found certain modifications must be made using Jordan forms. by starting with the transfer function parameters. and secondly a method in which n o assumptions are made concerning factorization. By locating the roots. A . can be obtained. Emphasis is placed on specific important aspects such as closedloop stability via assignment of the transfer function denominator roots. The statespace framework built up in this chapter will now be employed in the control of systems. this can be converted t o a canonical form in order to clarify the situation. In this section two methods have been considered. 6. It has been shown how. of the transfer function denominator.2 . through complicated parameter structures. firstly a method which relies upon all the factors of the system denominator being available.
which is the rate of water flow but of tank 2.2 217 The transfer function of a position control WardLeonard unit is given by:  m) V ( s ). has a head of water equal to height r. is given by uo = 3p.3 Given the statespace description 6.3 Water flows into tank 1 at a rate u (the input) and this tank. 6.4 find the input u ( t ) to output y ( t ) transfer function. Describe the tank system by means of a statespace model. vo $ Fig. . Water flows out of tank 1 and into tank 2 at a rate u l . which has unit crosssectional area. the two states being the tank heads t and p . Consider the two tanks shown in Fig. where u1 = 2r. 6. Tank 2 has unit crosssectional area and a head of water equal to height p.Problems 6.s(R + Ls)(B + Js) KIK 2 Find a statespace model for this system if the states are given by: 6. The output.3.
6.10 and solve the state transition equation. . Can a phase variable canonical form be obtained if a = O? Consider the statespace model gt)= [ 0 1 3 12 0 7 .11 Reconfigure the state equations into phase variable canonical form if Q! = 2 and the same input u ( t ) is applied.9 and solve the state transition equation.8 Find the input u ( t ) to output y ( t ) transfer function for this system.]N)+u ( t ) 6 [.7 6.5 State equations Consider the system transfer function: (D3 + 6D2 + 11D + 6 ) y ( t )= (D3 + 9D2 + 14D + lO)u(t) relating system input u ( t ) to output y ( t ) .I Find a matrix Q such that ( r ) = Qz'(t) x transforms the system to phase variable canonical form. Find the state transition matrix for 6. For the statespace model of Problem 6. Find the state transition matrix for 6.218 6. obtain a Jordan canonical form state description by carrying out an appropriate state transformation.6 Consider the statespace system model: go= 1 + E] 0 u(t) 1 y ( f ) =11 0 ll_x(t) 6. Consider the statespace model 6. Reconsider this transfer function in terms of an sdomain representation and find a n equivalent statespace model in which: i ( t )=  A x ( t )+ B_u(t) y ( t ) = cx(r) + ou(t) 6.
0 .15.1.' .4 . x(0) = 0.12 Consider the statespace model 219 6. What is the value of ~ ( t ) t = O? when What conclusions can be drawn from your results? By means of the CayleyHamilton theorem. Also. A statespace model of a particular system is given by: Y ( 0 = [1 lIx(t) 6. . x(0) = 1 in statespace form.13 ) with u ( t ) = 0.I Fig.14 Find y ( t ) if u ( t ) = 1 for t 2 0. 631) For the system in Problem 6. find y ( t ) if u ( t ) = 1 for t 2 0 and xl(0) = xz(0) = 0. 6. for all t. but x1(0) = .3t).15 Calculate exp(At) and write (Dz + 3D + 2)x(t) = exp( .find y ( t ) if u ( t ) = 1 for t 2 0. x2(0) = 0. find y ( t ) if u ( r ) = 1 for t 2 0. x 1 ( 0 ) 1 and = x2(0) = 0. After 1 second it is found that x ~ ( t = 2 and x2(f) = 1.show that this is equal to exp[A(t)].Problems 6. ?+IIF.16 6.A ] .17 By noting that the state transition matrix is equal to the inverse Laplace transform of [ s l . x1(0)= 1 and x2(0) = 0. Also. see ( .4. Find the eigenvalues and eigenvectors for the matrix A = [ 2 3 '1 6. 6. For the system shown in Fig. find a statespace description by employing two states and calculate the state transition matrix. calculate the state transition matrix for Y ( t ) = [1 OI_X(t) 6.
. Fig.. Roy. D. M. 1965.. J .. DeRusso.5. Macmillan. F. Pergamon Press. 1970. J. R. C. Macmillan. and Close. 6. Feedback control theory.19 1 Il_x(t) For the system shown in Fig. P.20 Consider the system transfer function: 6 7 G ( s )= ___ + ( s + 3 ) 3 (s+3)2 ~ 2 ++. and Tannenbaum. SpringerVerlag. Francis. Finite dimensional linear systems. find a statespace description by employing two states and calculate the state transition matrix. and Piasco. Brockett. R .P. A. John Wiley. K. W. 1986.3 s+ 5 s+4 Show that the Jordan form statespace description is: Further reading Blackman..).. 6. P.5 6. SpringerVerlag. J. . 1992. 1977. Introduction to statevariable analysis. 1981. Signal processing for control. 1988. Delchamps. and Jones.18 State equations Obtain the input u ( t ) to output y ( f ) transfer function for the system described by: 1 and u(t)= [1 6. Statespace and inputoutput linear systems. B. J. M . Classical and modern control.220 6. Stare variablesfor engineers.M. P.. Elloy. . Wiley.. (eds. F. Doyle. Godfrey.
M. 1980.. Analysis and design o discrete linear control systems. 1967. McGrawHill. 1994. Control.Further reading 221 Healey. R. Olsson. W.. 1970. p p . G. No. PrenticeHall. Nelson. Wiley Eastern. Moscow. f Ogata. K. (ed. McGrawHill. 90512. 1972. Wolovich. and Piani. New York. K. 1973. G. Prime. PrenticeHall. E. Linear multivariable systems. ‘A simple method for finding the Jordan form of a matrix’. H. PrenticeHall. on Automatic Control. State space analysis of control systems. ‘On the general theory of control systems’. W. B. Wiberg. 120. H. Congress Aut. Kalman. W.. 1969. Theory andproblem of state space and linear systems. 1987. O’Reilly. Linear systems.. AC17. IEEE Trans.. Modern concepts in control theory.A. Rubin.. D. SpringerVerlag. 1960. Sinha. 48193. A.. McGrawHill. R. Proc.. T.. 8. 1456. 1970. PrenticeHall. State space and multivariable theory.). IEE.. . 1992. Kailath. Rohrer. ‘Study of methods of computing transition matrices’.Circuit analysis: an introduction to the state variable approach.. N. V. J. Multivariable controlfor industrial applications. H.. Proc. Control systems. pp.. Peter Peregrinus Ltd. Computer systems for automation and control. Kucera. Rosenbrock. IFAC 1st Int. New York. 1974. 1971. 1991. pp.
both the system output and the state variables are resultant values. i. Relationships between inputoutput transfer functions were shown and various manipulations possible when in a statespace format were discussed. by considering a . In general the only control we have over a given system is provided by the control input u ( t ) which we can choose. The theory behind this is that if the standing of the balance is known (probably this means how much it is in the red) then an appropriate deposit can be made. In fact we would like to be able to do this for all the state variables. However. the system inputs and outputs. Consider for example a bank balance for which no bank account statement. What we would really like is to achieve a control over the system.7 State variable feedback 7.e. One further important factor is whether or not we can actually see all the states of the system. such that we can alter a state variable by taking it from its present value to a value we desire. usually this means by some direct or indirect measurement. simply by an appropriate choice of control input. output and state. i.e.1 introduction In Chapter 6 we reconsidered a system in terms of three variable types. namely input. no direct state information. Then the system is said to be completely state controllable. although often it is simply called completely controllable. as yet. we would have complete control of the system by being able to choose an input u ( t ) which would move any state vector from one value to some other arbitrary value after a certain time. However. is given. In fact it is only this chapter in which the ideas of providing a controller scheme and in particular a feedback control system based on a statespace framework are introduced. not really been uncovered. If we cannot measure directly certain of the states. it must be possible to reconstruct each of them simply in terms of the available information. a certain time later. Purchasers of this book will be able to feel pleased in that they have helped to prove the practicality of this theory as far as my own state variable is concerned. then the state of the account in terms of the present balance can be calculated by noting the withdrawals (output) and deposits (input) made.e. i. in the analysis carried out thus far the advantages and flexibility provided by a statespace basis have.
Controllability and observability 223 general state vector. the result of the control applied. and finally methods are described in which the state vector can be estimated or observed by employing the input and output variable data. the complete vector can be obtained from the system input and output information available over that same time period. obtaining a reconstruction of the state vector. 7. In this chapter the concepts of state controllability and observability are presented in terms of the statespace system descriptions previously introduced.2) . by means of a statespace description.2. where the input u ( t ) and output y ( t ) are connected by a middle man. in that it may be required to find out what has happened even if we cannot do much about it.1) A schematic diagram of this basic system is shown in Fig. when.6).1.1 Concepts of controllability and observability General expressions for a statespace system description were given in the previous chapter.2. The overall design for a controller is. In some cases the importance of one of these two measures may well dominate. dependent on the control action required and (b) the observer. (a) the control.2. 7.2 Controllability and observability The problems of controllability and observability are very often treated together and indeed this seems sensible in that both are tests applied to a feedback control system in order to ascertain whether or not we can actually dictate exactly system performance. Controllability is concerned with whether or not we can control a system. these are repeated here as: $ ( t ) = & ( t ) + Bu(t) y ( t ) = C g t ) + Du(t) (7. following this. in (6. the system is said to be completely state observable. or vice versa. split up into two parts. details are given of some of the different types of controller design possible with emphasis on the approaches taken for pole assignment and optimal control. tests for controllability and observability which can be carried out on a particular system are described by making use of the state matrices available. the state _x(t). The loop can be closed by deriving the input in terms of system response information.2. 7. or work out. whereas observability is concerned with whether or not we can see. over any finite time period. In the later sections of this chapter. in this case in terms of the state variable vector _x(t).Let the control input be defined as: (7.
4 results in nonzero steadystate state variables.224 State variable feedback Fig.2.1) this results in the equation g t ) = [ A .2. The addition of a nonzero external input as shown in Section 7. this is the case when the input is derived solely by state feedback.2) into the state equation of (7.2. In this case it is desired that all elements in the state vectcr tend t o zero.BFJ&(f) (7.1.2.4) which shows that for a given system A. we would like t o be able t o select a vector F such that the state vector _ x ( f ) settles down t o a steady value.2. 7.to)lx(to) (7.1) is said to be state controllable . and a n initial state condition ? ( t o ) . and this results in ( t ) = exp[(A x  BFNt . for a single input system with a n ndimensional state vector. 7. But.2.3) can be found in a similar way to the procedure carried out in the previous chapter.B. A solution to the equation (7.3.3) A schematic of this is given in Fig. can we choose an F such that _ x ( f ) settles down? This is in fact the same as asking the question: ‘is the system controllable?’ Definition of controllability: a system defined by (7. I Block schematic of a statespace system description where F i s a 1 x n feedback vector.2. On substitution of the control input from (7. see Section 6.
controllability is concerned with whether the state variables will settle down or not.to)l_x(to) X (7.2. no external input is present as shown in Fig. however. whereas by applying an external input.to)l_x(to) (7. However. subject to certain conditions on the A and C ) matrices. Observability of a system is dependent on whether or not information from all of the state variables can be obtained at the system output. just what the input is makes no difference to whether or not it is actually possible to find the vector. from (7.1) is said to be state observable if the system state vector ~ ( tcan be obtained over any finite time period. .e. Definition of observability: a system defined by (7.not necessarily zero. possible then the system is observable. remembering that u ( t ) = 0. By means of state feedback only.2. as considered thus far. as we would hope.5) given that ?(to) is dependent in some way on y ( t ) .1).Controllability and observability 225 if a control u ( t ) exists which will change the state &(to) to any desired state &(t)over the finite time period to to t: where t > to. 7. this will dictate the steady values themselves: zero external input. It is important therefore that we can obtain the state vector by means of system input and output information only. e. and thence ~ ( tfrom the same source. resulting in a state variable solution from (7. it is possible for _x(to) to be found from y ( t ) .2. Further. The concept of observability is extremely important in that for many statespace system descriptions the state vector is chosen to be some mathematical tool required to meet specific needs.2.2.6) by exp[AT(t . When a system has more than one output then every state must have an effect on at least one of the outputs.6) What is then desired is to be able to obtain _x(t)over any time period. The inclusion of system input knowledge in a definition of observability is in a sense a red herring. If one or more states of a particular system cannot be observed at the output then that system is said to be unobservable. if indeed this is. ) from complete knowledge of the system input and output over the same time period. In order to simplify things the input u ( t ) can be set to zero. to to t l . then the final state value ~ ( tis)required to be zero. This definition generalizes the idea of controllability such that from any initial state value x(t0) we wish to be able to take the state to a particular value _x(t).2. starting at time to. we have y ( t ) = C exp[A(t . hence leading to a quite complex relationship for each state variable in terms of any physically real measurements.1) of . where only state feedback is employed.to)]CTand then integrating.2. It is fairly straightforward to show (see Problem 7.5) On substitution of this into the output equation of (7. in that the system input does not in any way determine itself whether a system is observable or unobservable. as the actual input value does not affect the observability of a system. A system for which the definition of controllability is shown not to apply is said to be uncontrollable. In order to calculate the state vector we do need to know the value of the system input.U )= exp[A ( t .2.g. one not generated by feedback.1) that by premultiplying both sides of (7. i. means zero steadystate requirements for a controllable state vector.
2. Examples are given of how to apply the tests. 7.e.2 Tests for controllability and observability Although several tests are possible in order to determine controllability and observability conditions of a system. i. however no proofs are shown as these are considered to be a little too detailed for this text and indeed can be found in several of the referenced works. It can be noted that the concept of controllability concerns the input u ( t ) and state x(?) along with matrices A and B. C=[l 11 such that so v.2.1 Consider the n = 2 state variable system A= [ O 2 3 '1. = [ B . :I which is nonsingular. These associations will now be ) further developed in terms of tests that can be carried out in order to ascertain whether or not a system is controllable and/or observable. is defined as: v.1) is completely state controllable if the n x n controllability matrix has rank n. . B= [.2.226 State variable feedback it will still be possible to calculate x ( t ) for any known input as long as it is possible to do so for the zeroinput conditionsdescribed.2. if it is nonsingular. this is of no direct concern. AB. The controllabiiity matrix V .7) Nore: The nonsingularity condition does not apply to systems with more than one input. it has rank 2. AB] = [.. whereas the concept of observability involves the output y ( t ) and state ~ ( talong with matrices A and C. A " . Test for controllability: a system defined by (7..I. those most commonly encountered and perhaps therefore simplest are described here.e. .' B ] (7. i.. As we are only considering single input systems here. Hence the system given in this example is completely controllable. Example 7. = [ B .
I.AB] = [: I : ] which is singular. Hence the system given in this example is uncontrollable.Controllability and observability Example 7.=[. this is of no direct concern.2.=[&I=[ '1 which is singular.e. The observability matrix VO is defined: (7.I: C=[O I] such that so v. 2 . i. 3 Consider the n = 2 state variable system A = [ 2 O 3 '1. 2 2 c=[l 11 such that CA=[2 21 so .1) is completely state observable if the n x n observability matrix has rank n. it has rank 1 . Test for observability: a system defined by (7. B = 2 [. . .2. As only single output systems are considered here.8) Note: The nonsingularity condition does not apply to systems with more than one output. it has rank 1 .2.2 227 Consider the n = 2 state variable system 0 A = [1 J. Example 7 . = [ B . it is nonsingular.
possibly to make use of advantageous properties.3 Invariant theorems It was described in the previous chapter how the state variable vector could be transformed to an alternative form.e. J. held by the new form.2.4 Consider the n = 2 state variable system A= 0 [. as well as the conditions shown in the examples it is quite possible for a system to be both controllable and observable or to be neither controllable nor observable. it has rank 2.228 State variable feedback Hence the system given in this example is unobservable.2. Example 7. Consider the original state representation to be (7.1.2. Note: This is the same system as that in Example 7.2.2.2. 7. Note: This is the same system as that in Example 7. such as simplicity. B = [. The results of the controllability and observability tests are completely separate from each other. where it was shown to be controllable. Hence the system given in this example is completely observable.I.9) resulting in the new state description i ’ ( t )= A‘x’(t)  + B’u(t) y ( t )= C ’ z ’ ( t )+ Du(t)  (7.1) with a new state vector _x’(t) related to the original _x(t)by means of _x(t)= Qz’(t) (7.2. So this system is controllable but unobservable.2. where it was shown to be uncontrollable.10) . So this system is uncontrollable but observable. 2 CA= [l 31 c = [O 11 such that so vo= [g][: :] = which is nonsingular. i.
is of full rank then not only is that particular system controllable but also the controllability matrix Vc'for the system described by (7.' A Q . C ' ]is also completely state observable.B] is uncontrollable then so is [ A ' $ ' ] . Invariant theorem of controllability: If the pair [ A .Controllability and observability in transfer functions 229 where A' = Q . So.' B ] or V.'= [ B ' .' A " . . A consideration of this last point is the purpose of this section.2. Q . It is also true that if [ A .' = Q'Vc Therefore.2. Proofi Similar to the proof given for the invariant theorem of controllability. B ' ]is also completely state controllable.10).' B and C' = CQ with Q nonsingular. 7.5. But any one particular system is the same system irrespective of whether we prefer to think of it in terms of the statespace rather than as a transfer function..2.'.' A ' Q it follows that V. 2. A ' B ' . then so too will be V.. the controllability matrix for a system described by (7. B' = Q .2. on condition that Q'is nonsingular.. In this chapter controllability and observability have been introduced as system characteristics with specific dependence on a statespace framework.3 Controllability and observability in transfer functions In the previous chapter. hence the system described by (7.1) is controllable then so is (7.2..10). It is also true that if [A. ( A ' ) ' B .. C ' ] . Proof: If Vc.'= [ Q . Section 6. Q .10) is: V..(A')"'B'l and as ( A = Q . of full rank..' A 2 B .e. if the system (7. having uncovered the tests for controllability and observability in the statespace. q is completely state observable then the pair [ A ' . see Problem 7. . i.e. i. q is unobservable then so is [ A ' .1) is observable then so is (7.e. i.10) is also controllable.2 in particular. The following two statements can then be made: 1. the idea of statespace system descriptions was introduced by means of a transfer function approach.2. QIAB. if the system (7. these must necessarily also have a meaning in terms of a transfer function description. B ]is completely state controllable then the pair [ A ' .' B . Invariant theorem of observability: If the pair [ A . l).2.
is given as Y ( s )= G ( s ) * U ( S ) (7. Minimal realization: A statespace system description 1A .3.2) On substitution of the state vector into the transformed output equation of (7.3. is the minimum possible order. 2.i. C .1 Before discussing controllability and observability in terms of a system transfer function it will be shown how a transfer function.3. In fact for any G ( s ) the order n of the state vector chosen is not unique.2. 7.3.'BLl(s) (7. )is a realization of the transfer function system description G ( s ) if G ( s )= C [ s l .230 State variable feedback Transfer function realization 7.3) and thus if the openloop transfer function relating input to output.e.A ] . considered openloop in this section. On taking Laplace transforms of (7.12).5) However.l B + D ] U ( s ) (7. D . we have Y ( s ) =( C [ s / . assuming zero initial conditions on the state vector.3. the order of the state vector. which leads to the second statement. It will be seen in Section 7. can be simply obtained in terms of the state matrices.3.2 how the idea of a minimal realization is linked with the controllability and observability of that system. l ) . Realization theorem: A statespace system description [ A .1).3.4) two basic statements can be made. simpler transfer function. it follows that s X ( s )= A X ( s ) + B U ( s ) (7. for any one G ( s ) many possible selections of ( A . D ) exist.3.2 Polezero cancellation The controllability and observability of a statespace system description are dependent on there being no common factors between the numerator and denominator of G(s). B . D ) is a minimal realization of the transfer function system description G ( s )if (7.3. B .A ] . This can best be seen by an example. poles and zeros of a transfer function which are identical and can therefore be cancelled to produce a lower order. . 1. C . B .1) or X ( s ) = [ s l .2. depending on the state vector specified.' B + D (7.5) holds and n.2.A ] . C . defined previously in (6.
. some are neither: each particular case being dependent on the state vector. which is a minimal realization and which defines the transfer function with no polezero cancellations.5'1. which is in fact also a minimal realization of G(s). C'=[l]=C. and hence statespace description. results from the realizations (a) to (e) are as follows: (a) (b) (c) (d) (e) uncontrollable uncontrollable uncontrollable controllable controllable observable observable unobservable unobservable observable n=3 n=2 n =2 n =2 n=l This list shows that for any transfer function. B=K]. some are both controllable and observable.Controllability and observability in transfer functions 23 1 so consider the transfer function: G ( s )= s+2 (s + 2) =. 0 0 2 (d) A = [ . by making use of (7. some are observable. provided that the 'extra' poles and zeros produced cancel in the final transfer function.3. It can be seen from the list that although (c) is an uncontrollable. selected.2. C'=K]. in the case of (a) this is observable.2.6 (e) A = [ .s2 + 5s + 6 It can be seen that this transfer function has a common polezero at s = .1 (s + 2)(s + 3) (s + 3) .2. By applying the test for controllability and observability introduced in Section 7. In fact this leads to the general rule for system transfer functions. of the many possible statespace realizations some are controllable. the dimension n and nature of the state vector being different in each case. Note D is the null matrix in each case. The only realization in which no polezero cancellations occur is (e). . some are not. some are not. The only system description which is both controllable and observable is though (e). Further. unobservable description a higher order state vector can be found which is either controllable or observable. After toying around for a while with the above set of realizations it is apparent that it is relatively easy to obtain realizations for higher values of n .3 1 . n=2 n=l B=[l].5) it can easily be verified that the following are all statespace realizations of G(s).
2.3 is a modification of that in Fig. arising from controllability and observability concepts. Complete controllability and observability in a transfer function. controllability: can we actually control the system as desired by manipulating the state vector. assuming the state vector to be available. point to the fact that very often it is not possible to witness everything that is taking place in a system simply from its inputoutput transfer function. Specifically when polezero cancellations occur. this effectively corresponds to information on the system lost somewhere between input and output. as depicted in Fig. and is independent of the type of control action selected. hopefully improve. The diagram shown in Fig. G ( s ) . A transfer func tion. and includes a state U Fig. if it is possible to obtain a controllable and observable realization for that system. observability: can we obtain enough knowledge about the state vector? This topic is considered more fully in Section 7. Firstly. 7.4 State variable feedback It has been shown thus far in this chapter that the problem of controlling a system in terms of a statespace framework can actually be split up into two separate subproblems. and this type of feedback control. The two definitions made. G ( s ) . if the cancelling modes are suitably acted upon in the statespace realization. 7. 7. Note: A minimal realization of a system is a controllable and observable realization. we consider how to provide a state feedback control scheme in order to alter. as desired. 7.232 State variable feedback 1.represents a system which is uncontrollable or unobservable or both if at least one polezero cancellation takes place.3.5. the performance of the system under control.3 State variable feedback . Several methods which can be employed on a statespace system description in order to achieve a suitable control action will now be discussed. A transfer function. By means of a statespace description this information can be retrieved in terms of the state vector and hence a system which cannot be controlled simply by consideration of its transfer function can be controlled. represents a system which is both controllable and observable if no polezero cancellations occur. Secondly. A second rule follows directly from this: 2. 7. irrespective of how that vector has been obtained? That is. is the subject of this section.
this results in &(t)= [ A .1). In terms of state variable feedback design the pole placement scheme can be described as one where we are given the desired closedloop pole positions and must select the state feedback parameters in order to achieve those pole positions.4. u ( t ) is a scalar reference signal and H i s a gain which we can select. but is though not the main subject of this discussion.4.4.2) in which F is the (1 x n ) dimensional vector of state feedback parameters.2) into this equation.4. 7.1 Pole placement Pole placement control is concerned with the specification of the poles of the closedloop system. Further.2 effectively it was assumed that the reference input u ( t ) was zero.State variable feedback 233 feedback parameter vector. i. and the single control input u ( t ) is defined as u(t)= . this is commonly referred to as the regulator problem. that is the roots of the characteristic equation.3) (7. this latter choice will be considered briefly a little later in this section.4.4. However. In Fig.4.2) into the state equation (7. 7. The closed loop in this sense refers to the relationship between the reference value u ( t ) and the actual output value y ( t ) . also included is a reference input vector u ( t ) . something which we as the controller designer can choose.4. the output was desired to be zero. we find that y ( t ) = [ C .4. then we can tailor that performance as required if we can arbitrarily select the pole positions. In general it is usually required that the output y ( t ) follows the desired value o ( t ) . the criterion placed on the way this reference following is carried out then defines the type of control action that is needed. On substitution of the control input equation (7.Fx(t) + H u ( t ) (7.e. It is in fact these n feedback parameters which we must find in order to achieve the desired pole positions. The basic philosophy is that if we know how the closedloop poles affect the system dynamics and hence performance. or in other words the closedloop denominator coefficients. introduced via a feedforward gain H.4) then on substitution of the control input (7.1) where _x(t)is the (n x 1) dimensional state vector.5) . F.DFl_x(t) D H v ( t ) + (7.BF]_x(t) BHu(t) + It also follows that if the output equation is defined by y ( t ) = C x ( t )+ Du(t) (7. The general statespace description has a state equation g t ) = A g t ) + Bu(t) (7.
BF.B + D ) H V ( S ) ' and it can be seen from this that the closedloop poles are in fact the roots of the equation: (sZFJ= I d .4.4.8) Equation (7. In order to show the operation of pole placement by example.4.this closedloop system is completely defined by the two equations (7.F ] .DF ] [ sl .F ] . so X ( s )= [ s l . controllable system with A= [ 2O 3 '1: B= [.5).4. On substitution of (7.8) shows the possibility that all of the closedloop poles can be arbitrarily placed by selection of F and that this is dependent on the state matrices A and B only.6) into the output equation transformed from (7. and the state vector is assumed to be initially zero. which leads to the following. assume that the desired closedloop response 01 a secondorder system is given by the denominator s2 + 5 s + 4 = (s + 4)(s+ 1) Example 7.7) Y ( s )= 1 [ C . Pole placement by state variable feedback: The n state feedback parameters contained within the vector F can be selected in order to obtain an arbitrary set of n closedloop poles if the pair ( A .4.I B H V ( s ) (7.3)and (7.1 Consider the n = 2 dimensional.5).3)in terms of its Laplace transform.4.234 State variable feedback Fig.4.] .4.6) where F = A .4 Closedloop system with state feedback A diagram showing the statespace system description with feedback connected is shown in Fig. Viewing (7. B j is controllable.4. it follows that (7. 7.4.4. 7.A +BFI = O (7.
i. for this uncontrollable system it is not possible to obtain a state feedback parameter vector F which will achieve the desired positions.2 Consider the n = 2 dimensional.BF] = 2. But [ A . and fl+ 2 = 4 . uncontrollable system with and F= so I1 f f21 [sl. Hence.Fl f2 = sz + ( f z + 3)s + ( f . + 2 ) fz= 2 In order to obtain the desired closedloop pole positions it is required that +3 =5. Example 7. .e. ft =2 which results in F = [2 21.e.F ] = S+fi f2+2 S + f2+3 1 f2=2 f2=2 In order to achieve the desired closedloop pole positions it is required that f2+3=5 ft+f2+2=4 fl+ or or fi+ f i + which result in one equation but two unknowns.4.f. i.3 .f2 and such that I sl. .State variable feedback Then we have also that the state feedback parameter vector is given by F=[fi 235 A1 where f 1 and f2 are to be found in order to achieve pole placements.
236
State variable feedback
Note: The selection of the gain H i s usually carried out in order to ensure that the actual output y ( t ) is exactly equal to the reference u(t) once steadystate conditions are reached, i.e. H i s chosen to counteract the inherent steadystate gain of the closedloop system, So it is therefore straightforward, with reference to (7.4.7), and assuming a step input, to choose N as:
H = { [ c  OF] sl  F ]  B + D ) 1 1s = 0 [
'
(7.4.9)
This calculation of H therefore must come after the selection of the vector F and hence F also.
7.4.2 Optimal control
In Section 7.4.1 a controller was designed with the intention of ensuring that the closedloop system response is of a form that we want, irrespective of the openloop characteristics of the system, and this is done by forcing the closedloop denominator to assume a specified polynomial. But this is just one particular criterion on which a controller design can be based. In fact many requirements can be made for the closedloop system, including fairly straightforward and perhaps obvious limitations such as maximum and minimum possible input values, e.g. due to the type of amplifier used, or the allowable settling time, i.e. the time taken for a response to settle to approximately its final value. Perhaps one of the most commonly encountered and widely researched overall control performance objectives is that of optimal control, and in this section a glimpse is given of the possibilities open to the designer by making use of the state variable feedback ideas previously introduced. An optimal controller is one which is best or most favorable, and this automatically implies that it is best in terms of a particular sense, i.e. in terms of a defined performance requirement or specification. So the statement that a certain controller is optimal is, on its own, fairly meaningless unless it is also stated in which sense the control is optimal. Very often the performance specification is made in terms of a mathematical function, the optimal control then being the one which minimizes (or maximizes if appropriate) this function. For a single input system a cost function defining performance could be simply:
CF=
[
J I,
u * ( t )d t
(7.4.10)
where the function is considered over the time inverval ti = initial time to tf = final time, and the initial and final time state vectors are assumed to be available. A minimization of this particular function by a selection of the control input to be zero over the time period specified, results in a cost function which is also zero. A straightforward requirement that the controller be designed in order to minimize the magnitude of the control input therefore results in a rather obvious solution, i.e. no control input at all. The input will be a signal, e.g. voltage or velocity, and hence the cost function (7.4.10) is then
State variable feedback
237
proportional to the energy expended in controlling the system; we can therefore lo), minimize the energy expended, as required from (7.4. by not expending any at all. An alternative cost function could be one in which the state vector is seen as the important variable and in particular the error between the state at the initial time ti and the final time tf; note that the ndimensional vector at time I f , _x(tf)is termed the origin of the statespace. This cost function can then be defined as
CF=
s" s
xT(t)_x(t) t d
(7.4.11)
I,
such that if this is minimized by an appropriate choice of control input, the vector z(ti) will approach the vector _x(tr)with very little overshoot. However it will most likely be the case that certain states are considered to be more important than others, e.g. a particular velocity or acceleration might be critical whereas another might be of relatively little importance. Weightings can therefore be placed on certain states and/or certain combinations of states, which means that the cost function (7.4.11) becomes
(f
CF=
xT(t)Qz(t) dt
(7.4.12)
ti
where Q is an n x n weighting matrix and is usually symmetric with all positive, real coefficients. An interesting and commonly encountered special case of (7.4.12)arises when Q = CTCand D = 0, see (7.4.4).In which case the cost function described (7.4.12) refers to the selection of an input signal which minimizes the error with respect to time, between the actual output, with initial conditions at time ti, and the final desired output at time tf. However, whether in terms of the output or the states of (7.4.12),choosing a control input u(t) which minimizes this cost function may well result in a set of high magnitude values for u(t); ti < t < tf, with large and rapid variations occurring during the time interval, and in practice this may well result in actuator saturation and hence nonlinear characteristics. It appears appropriate therefore to combine the cost functions derived in order to put emphasis on both the control input and the states, such that an overail function can be described as
CF=
s
(I
&'(t)QS(t)
I,
+ u T ( t ) R u ( t ) ]d t
(7.4.13)
in which (a) a multiinput system is allowed for, and (b) the real and usually symmeiric and positive definite weighting matrix R is used to put emphasis on different inputs and on the input(s) with respect to the states. This latter point refers to the relative costing placed on the inputs as opposed to the states, i.e. the importance of (7.4.10) opposed as to (7.4.12),and points to the fact that this latest cost function, when minimized by appropriate choice of control input, results in an action which is a compromise between the two extremes considered. The cost function (7.4.13) termed a quadratic performance index and when it is is wished to find the control input u(t) from a linear system (7.4.1)as a solution to tbe cost
238
State variable feedback
Fig. 7.5
LQ feedback control
function minimization for ti = 0, tf = 00, then this is referred to as the linear quadratic control problem ( L Q ) . Note R is a scalar for a single input. In order to consider an overall closedloop transfer function, an external reference input signal, u ( t ) , is required, and this is generally introduced to the control input in the form of (7.4.2). The external input does however not affect the optimality of the control employed; this is obtained by the choice of the state feedback parameters within the vector F. In particular, the optimal control which minimizes the cost function (7.4.13) is given by (7.4.2) in which
F = R'BTK
(7.4.14)
where K is a positive definite matrix and the unique solution to the algebraic matrix Riccati equation
Q = KBR'BTK A T K  KA
(7.4.15)
A block diagram of the overall system is shown in Fig. 7.5. Unfortunately it is not a simple matter to obtain the solution (7.4.14) and (7.4.15) from the cost function (7.4.13), this problem is however dealt with in several texts (see D'AZZO and Houpis (1983), for example) and is not considered further here. Finally it must be noted that the closedloop poles of the optimal system are the roots of the equation:
1 SI
A
+ BF(= 1 d  A + BR'B'K~ = o
(7.4.16)
and these lie in the left half of the splane. Further, the minimum value of the cost function is then xT(ti)Kx(ti).
7.5
State estimation
The state vector has thus far been assumed to be either known or directly measurable, the control input being formed as some multiple of this known vector. The state vector is though very often selected merely as a mathematical entity with very little obvious physical relevance, and is therefore generally not measurable. Hence, an alternative
State estimation
239
scheme must be employed in order to calculate or more likely estimate the value of each state vector element. The control input can then be formed as a multiple of the estimated state vector which is used in place of its true but unavailable value, hopefully the estimate being a reasonably good one. The mathematical tool employed in order to estimate the state vector is called an observer and, as will be described in this section, the measurable system inputs and outputs are used in the estimation procedure such that the estimated state vector is linearly related to the actual, but unknown, vector. This means that an error will exist between the actual and the estimated state values, one property desired of an observer then being that this error tends towards zero. It will though be shown that it is not simply a case of trying to reduce the error to zero as quickly as possible because this can lead to problems due to poor noise rejection. The selection of observer characteristics is therefore an involved task and it must be remembered that if the observer is part of a feedback path, which will very likely be the case, then the observer characteristics will directly affect the closedloop system characteristics.
7.5.1
Asymptotic estimation
The linear system under consideration is defined by
(7.5.1)
which is identical to (7.2.1) except that D = 0. The matrices A , B and C are assumed to be known, as are the system input and output, u ( t ) and y ( t ) respectively. However the state vector _x(t)is unknown, all that we do have is an estimate of this vector denoted by $ ( t ) , which means that by considering the output equation in (7.5.1) we can in fact write that as
9(t)= cgt)
then
Q t ) = V ( t ) E(t)
(7.5.2)
in which P ( t ) is an approximation to the output signal but, as it can be calculated from
R(t), is a known value. The error between the two outputs, actual and approximate, is
(7.5.3)
and by referring to the state equation of (7.5.1) it seems reasonable that R(t) can be derived from
i ( t ) = Ag( t ) + B u ( t ) + G l ( t )
(7.5.4)
where the estimated state vector is subject to the same system characteristics as the actual state vector, the term G l ( t ) accounting for any error. By an appropriate choice of the parameters in the vector G so the state estimate can be made to vary significantly even for only small errors I(t), when G is large, or to vary only slightly even for large errors,
240
State variable feedback
when G is small. So how should G be chosen? A further insight into the problem is given by considering the error between the actual and the estimated state, defined here by
E ( t ) = _x(t) g 
t)
(7.5.5)
If (7.5.4) is subtracted from the state equation of (7.5.1) this results in
B(t) = A  ( [ )IB u ( t ) X i( f )  
A g ( t ) B u ( t ) G / ( t )
(7.5.6)
and by including (7.5.5) and the output equation of (7.5.1) this becomes
i ( I ) = ( A  GC)E(f)
(7.5.7)
which has a solution
€ ( I ) = exp( [ A  G C ] ( t tO))e(fO)
(7.5.8)
where to is the time at which initial conditions are taken. It is therefore only the error present at the initial time to, & ( t o ) , which causes the remaining error & ( t )for all following time. Further it can be seen from (7.5.8) that the response of the error r ( f ) is dictated by the eigenvalues of the matrix (A  GC), these being called the observer poles, such that if there are any roots which lie in the right half of the splane, the error equation will be unstable resulting in the error heading off t o infinity. For a n observable system though, the observer poles can be arbitrarily chosen and hence the observer can not only be selected as stable but can also have a specified rate of convergence to zero, i.e. the exponential decay can be chosen. The error will therefore asymptotically tend t o zero, which means that the estimated state vector will become equal t o the actual state vector, and this will occur whatever the system input and output values are.
7.5.2
Observer and system stability
Irrespective of the type and nature of the control action in operation, if a n observer is introduced such that an estimated state vector is used in place of its unknown actual value then the closedloop system n o longer depends on the actual value but rather is affected by the characteristics associated with the estimated version. It must be noted though that for a n asymptotic observer, which reduces the estimation error t o zero as t P 00, then under steadystate conditions there will in fact be n o difference whatsoever in the closedloop system. The most important question that must be asked is this: Assuming a particular control system design t o be otherwise stable, will the closedloop system remain stable when a n observer, which is itself stable, is incorporated? In order t o obtain a n answer t o this problem, consider again the system equation of (7.5.1) along with the state feedback equation
U(l)= 
F&(t)
(7.5.9)
State estimation
241
with vector F containing the state feedback parameters. If it is initially assumed that the state vector x ( t ) is known, then on substitution of the control input of (7.5.9) into the state equation of (7.5.1) it follows that
& ( t ) = ( A  BF)A(t)
(7.5.10)
However, if the estimated state vector g ( f )is used in place of ~ ( t )its actual value, in , order to obtain the control input from (7.5.9), then on substituting u ( t ) defined as
U(t)=
 FZ(t)
(7.5.11)
into the state equation of (7.5.1), we have that
Z ( f ) = A g t )  BFZ(I)
i ( t ) = GCx(t)+ ( A  B F  G C ) g ( t )
and in matrix form this is
(7.5.12)
Also, if the observer equation (7.5.4) is considered, when u ( t ) is substituted it follows that
(7.5.13)
(7.4.14)
This means that the characteristic equation for the overall closedloop system can be calculated from characteristic eqn = det
[ I; s l  ( A  BFSi BF
GC)
1 1
(7.5.15)
Note: See Section 6.3.1 for details. By elementary matrix operations (see Problem 7.7) the determinant is more simply reconfigured as: characteristic eqn = det
sz ( BF G C ) A
(7.5.16)
or
characteristic eqn = det(sl  A
+ B F )  det(s1  A + GC)
and this means that the characteristic polynomial of the overall system is found to be the, characteristic polynomial of the controller, see (7.2.4), multiplied by the characteristic polynomial of the observer. By appropriate selection of the vector gains F and G therefore, a stable closedloop denominator, characteristic polynomial, can be obtained. Further, if the system is completely controllable and completely observable then all the roots of the characteristic polynomial can be arbitrarily chosen. The fact that the characteristic equation of the closedloop system is obtained simply as a combination of the controller and observer polynomials is important in that the controller design can be carried out regardless of the nature of the state vector, e.g. it can be assumed that we know its actual value, and also the observer can be designed
242
State variable feedback
irrespective of the type of control action to be applied. This is known as the separation principle. As a final note to this section it must be pointed out that no reference signal was included in the control input defined in (7.5.9) and (7.5. I I). By a simple addition to the input u ( t ) , as was done in (7.4.2), such a reference can be introduced and this results in only a trivial extension to the equations (7.5.12) and (7.5.13): see Problem 7.8.
Example 7.5.1
To show how the design of an observer is carried out consider the system described by the equations (7.5.1) in which:
A = [  30
2
'I; [;I;
B=
CT=
[A]
An observer, as defined by (7.5.4) can then be employed in order to obtain an estimate of the unknown and unmeasurable state. For this system the gain vector is therefore
where the selection of the values gl and gZ is the object of this example. The poles of the observer are obtained from the characteristic equation given as the determinant of
and the determinant of this matrix is
S'
+ (5 + gi)s + (6 + 2g1 + gz)
The gain values gl and gz can now be selected in order to obtain a required form for this polynomial, i.e. in order to place the observer poles in their desired positions. Let us assume that the requirement is for the poles to lie at s =  6 , and s =  1. This is satisfied by choosing g l = 2 and gz =  4 which results in a characteristic equation
S'
+ 7s+ 6 = 0
Finally, the observer can be then written as
remembering that because
i(f) A g ( t ) + E u ( t ) + G l ( t ) =
the actual design procedures employed for controllers and observers providing the subject matter for Sections 7.3.5 it was however pointed out that the state vector is very often not available.e. Tests for controllability and observability merely point to whether or not a system is controllable and/or observable.to)l_x(to) (b) ~ ( t= )C Z ( ~ ) 7.4 controller analysis was carried out by means of state variable feedback. and that it does not show signs of instability. u ( t ) = 0.3) &t) =(A 243 G C ) Z ( t )+ h ( t ) + G y ( t ) 7. The methods described for controller design were firstly pole placement and secondly LQ optimal.4 ( t . i. i.6 Summary In this chapter some of the basic facilities provided when a statespace system description is employed have been described. two very commonly encountered techniques. It is required though that the state estimate be as near as possible to its actual value. In the first instance the ideas of controllability and observability were introduced and it was shown how simple tests can be made by using the system matrices. an estimation of the vector is therefore made and this estimate is employed.t~)c'] then integrating both and sides of the equation. that it is a good estimate. show that ? ( t o ) can be found in terms of y ( t ) . by prernultiplying (b) by exp[A'(t . The terms controllability and observability therefore have a meaning in so far as system transfer functions are concerned and this particular topic was discussed in Section 7. in order to find out whether the system under consideration is controllable and/or observable. making the assumption that the state vector was actually available to be fed back. for state variable feedback.5. Problems 7.Problems this is also.e.5. . for all t. these points were also covered in Section 7.4 and 7. instead of the actual but unknown value.5.1 for a general input signal u ( t ) and show that the same conditions must hold for the calculation of _x(to) to be possible. Any statespace description of a system can also be reconfigured in terms of a transfer function polynomial relationship between input and output. What conditions must be assumed to hold for this calculation to be possible? Repeat Problem 7.1 If the input to a particular system is zero.2 where to is the initial time considered. In Section 7. by means of (7. In Section 7. and the system is described by the equations: = (a) _x(t) expI.
where Q= [i :] and R=l The optimal LQ problem is solved by using the matrix Riccati equation: Q = .C'] also completely state observable.6 Find the value of a that will cause the system to be unobservable. If the pair [ A . then prove that the transformed pair [A '.A ~ K KA where + KBR~B'K u ( t ) = .@ ( t ) and the observer vector G defined in . find the minimal order canonical statespace representation. What is conditions must be assumed? The statespace equations for a particular system are: and it is required that the quadratic cost function m CF = 0 [_x'(f)Qz(t)+ u ( t ) R u ( t ) l d t must be minimized. For a system described by g ( t ) = A x ( t )+ B u ( t ) At) = Cx(t) with a control input u ( t ) = .5 7.R'BTKx(t) 7. C ] for a given system is completely state observable.7 Find the optimal state feedback and obtain the openloop and closedloop eigenvalues.3 State variable feedback Consider the statespace problem model: 7. Consider the statespace model: W)=2 : y(t)=[.4 Show that the system is completely state controllable and completely state observable. With this value of a.1 [ 1 0 2 0 Y]A30+ 1 [L] u(t) 1 O]x(t) 7.244 7.
s l . that this can be written more simply as: char. A system is modeled in the statespace by: 4 4 2 . eqn = det(s1. on the characteristic equation (7.A + GC) 7. Choose F such that both closedloop eigenvalues (poles) lie at s = . .GC) Show. the system can be made closedloop stable. Find the values of CY which mean that an observer cannot be constructed for the system A t ) = t(.1 .5. What effect does the inclusion of a reference signal have on observer equation (7. 3. Test the system described by 7. eqn = det [“i. i. such that the four closedloop eigenvalues lie at s = 4. defined in (7.8 For a system described by g ( f )= A x ( t ) + B u ( t ) Y ( t ) = Cx(t) the control input is u ( f ) = .( A .1.5. 1 1 N 7.2 and .A + EF) * det(s1. where u ( t ) = . show that by a suitable state feedback design u ( t ) = . it is an openloop unstable system.B F .e.4).2 and . such that the estimation error decays with eigenvalues at s = .Problems (7.F’(t). Also.@ ( t ) + H u ( t ) where u ( t ) is a scalar reference signal and H i s a gain.5.9 7.4).15)? Show that the statespace system described by: 7. by elementary matrix operations. the overall closedloop system characteristic equation is: 245 char.11 for controllability and observability properties.5.4. hence.12 When a=O find an observer vector G.F x ( t ) .4) and.10 has at least one eigenvalue which lies in the right half of the splane. Find the state feedback vector F.
2 and . . such that the input is driven to xT(log 2) = [O 01 from ~ ‘ ( 0 = [ . find the state variable feedback u(r)= . u(f) = .~ (=1(0 01 subject to the cost function ~ ) CF=7. For the system: design an asymptotic observer with eigenvalues at s = . By a suitable state transformation. but has at least one eigenvalue which lies in the right half of the splane. where a is a constant. Find a control input u ( t ) = a.14 where B’ = Q . ) A system described by = [“ 0 0 .t 2 0.(t)+ x ‘1 [. it is openloop unstable. whilst not affecting the other two eigenvalues.’ B and = A ’ x ’ ( t )+ B ’ u ( f ) . such that if _ x ( f ) = Qz’(t) B‘ = [. Also.F . Find the state variable feedback vector F. which moves the right half plane eigenvalue to s = .17 Find the minimizing control and minimum cost. For the statespace model of Problem 7.e. find a vector F which will cause the closedloop system to be unobservable.F x ( t ) such that the closedloop system has eigenvalues at s = .16 1 2 1 ‘ 0 u 2 ( r ) dr 7.3.13 then Find Q .I u(t) is subject to the initial conditions ~‘(0) [ 2 21. assuming that a state feedback of the form u ( t ) = .15 is completely state controllable. The state must be driven = to . i.] x’(f) 7. Show that the system described by 7.1 41.& ( t ) + v ( t ) is applied.3 . A statespace system description is given as: 7. repose the system description in phase variable canonical form.5 and  6. ( t ) .3.246 State variable feedback Show that this system is completely state controllable.
Problems If. . 7.' B K when bo = 0.m 7. 7. Consider the system with openloop transfer function: G ( s )= (s + l)(s + 2)(s + 5) 2 The system is connected in the closedloop configuration in Fig.18 247 find the values of For the system: 01 for which an asymptotic observer cannot be obtained.19 test for controllability and openloop stability (eigenvalue positions must all be in the left half of the splane). Lp F Fig.' B & ( t ) .0 1= 5a2 and Finally. q= . a I .8~2. Find u ( t ) in terms of a2. Find a statespace representation for the system.1. calculate the system closedloop eigenvalues. The LQ problem is to find the optimal control which minimizes the cost function. 7. instead.6 where H ( s ) = K s + 1. Find the state variable feedback which produces closedloop eigenvalues which both lie at s = .6 7. bo and q.1.3 and . and use this to obtain the closedloop transfer function. we have that y ( t ) = 11 1 . Consider the system equation: (a2D2 + aiD)y(t) = bou(t) relating system input u ( t ) to output y ( t ) . Find a matrix Q such that a phase variable canonical form can be obtained for the system by making a state transformation. CF = Jm O ( y 2 ( t + qbou2(t)) t ) d for some value q. and obtain R . Select a value for K such that the closedloop eigenvalues lie at s = 4. by means of the control u ( t ) = .20 Find an equivalent statespace representation.R .
Statefunctions and linear control systems. Signal processing for control. PrenticeHall. 1983. Kwakernaak. J. 1971. Introduction to linear system theory. D. Kailath. and Narendra. Macmillan. and Tannenbaum. World Scientific Publishing Co.. 135. 1992. Linear conrrol system analysis and design... 1962. G. Kalman. Schultz. 1970. A. zeros and feedback: state space interpretation’. Ho. on Automatic Control. on Automatic Control. 6. SpringerVerlag. Linear multivariable systems.. K.. ‘Poles. 1. ‘An introduction to observers’. 12935. C. Wolovich. AC10. J. 1965.. SpringerVerlag. 1991. H. Luenberger. F. W. observability. (eds. 1988. Muftivariable and optimal systems. K. Computer systems for automation and control. 1980. Interscience Publ.. D. D. in LaSalle et al. Wileylnterscience. AC16. 1983.). D.. and Houpis. Brockett. on Military Electronics. 1964. McGrawHill.. R. G. pp. An observability criterion for a class of linear systems’... Linear algebra. 1981. H. Delchamps. E. W.. McGrawHill. ‘Controllability. IEEE Trans. J. L. Y. J . MIL8. 1986. 58295. Approximate Kalman filtering. R. 1974. . (eds. G.. IEEE Trans. 7480. Academic Press. P.. Godfrey. A. T. A. Tropper. D‘Azzo. K. Nelson. ‘Controllability of linear dynamical systems’. G. Doyle. pp. AC16. Kucera. and Melsa. G. Holt. D. . 1993. 1979. V ... Chen. O’Reilly. on Automatic Control. 1971. on Automatic Control. C.. J. .Linear optimal control systems. SpringerVerlag. 596602. PrenticeHall. and Piani. T. pp. pp. Feedback control theory.. Francis. 1966.. IEEE Trans. p. No. ‘Observing the state of a linear system’. Chen. S . and Mitter. and Jones. Willems.).. Analysis and design of discrete linear control systems. Luenberger. M. Contributions to direrential equations.. pole allocation and state reconstruction’. R. Rinehart and Winston. R. Statespace and inputoutput linear systems. IEEE Trans. Observersfor linear systems. B. H. C. J.. AC11. and Sivan. C . Linear systems.248 State variable feedback Further reading Bonnell. D. IEEE Trans. 1992. 1972. Academic Press. S. 1967. Owens. Olsson. G. PrenticeHall.
It also allows for more adaptive controller schemes in which control parameter tuning can take place. or samples. To change a particular setting only one line of the program needs to be altered or acted upon. the controller therefore being formed with this as the desired objective.6. Stability of systems in the zdomain and the effects of sampling are considered in the following section. . 8. the change in mode of controller operation results in another theoretical base which is necessary to deal with understanding the principles of controller design procedures. meanwhile. The later sections. Essentially. 8.2 the computer’s eye view of the system to be controlled is introduced leading to zdomain representations of discrete time systems. deal with forms of controller design in the discrete time case. In general this merely means discrete views of system input and output signals. In Section 8. the controller action is obtained partly by means of software operations. This can even result in a certain amount of rewiring. Section 8. however. By employing a computer in the control loop.5 being based on state variable methods in particular.1 Introduction Once an analog controller has been designed and built it is generally a permanent structure. Section 8.1.2 Sampled data systems Most digital computer systems are designed as that of Fig. concentrates on choosing closedloop denominator coefficients or closedloop pole locations. be accomplished by adjustment of a potentiometer or by arrangement of switch settings to arrive at a desired gain value.Digital control 8. This results in a great deal more flexibility and speed. However. although in certain cases the system state may also be available. Change in operation must. the computer can only take brief glimpses. of the realworld system under control.
resulting in a digital representation of its analog value.at time t . U(I) I ly. 8. Its input is defined here as u ( t ) . 8. it is typed into the computer. at time t . the system output becomes y ’ ( t ) when eventually sampled by the computer. this indeed is the case for virtually all control applications. 8.1 itself consists of several smaller functional parts. The analog value y ’ ( t )must therefore be converted into a digital (possibly binary representation) form o n which the computer can operate. i. and the measured output value. The system (plant) to which the computer is connected is a realworld analog system. it will initially be in digital form. 8. every T seconds. Once the digital value has been obtained. and this is immediately fed into the D/A converter resulting in a new u(t) Fig. a finite time being taken before a particular sampled analog value can be represented in digital form. the program is used to calculate a new plant input value u ( t ) . The program itself must then obtain the error. v ( t ) . Conversely. A reference output value. 8. whereas the program carrying out the control operations requires signal values to be in a digital form such that it can work with discrete events. This conversion is carried out in hardware using a n analogtodigital (AID) converter. I Block diagram showing a computer applied to a continuous time plant The plant itself is a realworld system operating in continuous time. because of the sensors applied. ~ ( t ) . However. The computer shown in Fig.(l~l * Computer Fig. due to measurement errors. v ( t ) . Fig. ~ ’ ( t The way in which e ( t ) is ). once the correct plant input has been calculated from the programmed algorithm.250 Digital control 1 Reference output value. see Fig. and its output as y ( t ) .2 Sampling operation of the computer . e ( t ) . used in order t o obtain the plant input.e. between the reference output. is fed into the computer by software means. being the control algorithm held within the rest of the software.2 therefore contains a digitaltoanalog (D/A) converter on the computer output side The standard procedure is t o use a realtime clock such that y ’ ( t ) is sampled periodically.2.
only a relatively small and therefore finite number of u ( . whereupon the process is repeated.2.1).. let the error signal at that time be e ( k ) and the new control (plant) input be u ( k ) . is sampled. ) and e ( . f(l). however. . The program can not only make use of these values. u ( k . therefore.Sampled data systems 25 1 analog value being passed to the plant..2...2)..1) + .kT). e ( k ) .2) in which fi: i = 1.2) + +f"U(k . The program then waits until T seconds have expired from the time of the previous sampling. in order to obtain the error between actual and desired outputs. at a specific time kT the value of this sampled signal is therefore f ( k T ) 6 ( t . the new control input u ( k ) can be found as a function of past input values and past and present error values. u ( k . + gme(k. it follows that the sampler output consists of a series of impulses of varying magnitude.1 Delay operator representation It is assumed here that once a sample of y ' ( t ) has been taken its digital value is subtracted from the reference input.. are scalar coefficients. v ( t ) . but also previous values it has calculated for the plant input. equation (8. e(k . 1. . 8.. The A/D time along with the program time plus the D/A time must in total then be less than T seconds. At one particular time instant.kT)be the unit impulse function at time t = kT. This method of representation is termed a difference equation and can be used.1) can be reduced to: u ( k ) = f .2 Ztransform analysis Let 6 ( t ..l ) .2. if it is supposed that the initial time t = to = 0.1) + f2u(k . to represent many physical systems.. be e ( k . This series can continue until e(0) and u(O). n and gi: i = 0. e(0)) (8. 8. Also. ) terms to be m + 1. ) terms to be n and the number of e ( . and the same will be true T seconds later still. Assuming the number of u ( .1) In practice. m. .. let the error signal obtained one sample period ( T seconds) prior to this. The exact means by which difference equations can be employed in a transfer function description will now be investigated.2. that is: u ( k ) = f ( u ( k .. At time r = k . where T is the sampling period..1) with a control input at the same time of u ( k ... When a particular continuous time signal.n ) + goe(k) + gte(k .l ) .rn) (8. u(0). ) terms are actually required. If it is considered that the  . when t = k. In T seconds time a further error value will be obtained.2. often with only a few parameters (four to five)..
1) term from the final value operation and the ( z . . firstly the final value theorem which is useful later on for steadystate response calculations. secondly the initial value theorem and finally the effect of a time shift on the original signal. by taking the Laplace transform of this: g { ~ ( t )2 f ( k ~ )esTk = ) k=O  (8. Final value theorem lim (f(t)) 1m = lim ?+I l(z.1)..3 Properties of the ztransform Three properties associated with the ztransform are considered.. of f(t) can hence be written as: f’(t) = k=O c f ( k T ) q t .2. Tables of certain ztransforms are to be found in the Appendices. to = 0. can be found in terms of its steadystate (final) value simply by summing the coefficients of F ( z ) . as many works cover the underlying theory.252 Digital control time origin. 8.l ) F ( z ) ) on condition that F ( z ) converges when I t \> I and all the denominator roots of (1 . denoted by the operation 2 .’ + f(ZT)z2 + . the ztransform of a unit step function being z / ( z . is defined by 11 (8. This is due to the fact that the ( z . the overall sampled version. leaving F(1) as the steadystate .2.4) The term z is now defined as z = e=T such that the ztransform of f(t).5) or F ( z ) = f(0)+ f ( T ) z .z ) F ( z ) have magnitude less than unity.1) term in the unit step function cancel each other.2. f ’ ( t ) .kT) m m (8. obtained as the multiple of F ( z ) and the unit step. This means that the output signal.3) or. It is not intended to investigate the underlying mathematics of this operation further here however.2. A special case of the final value theorem occurs when a unit step input is applied to the transfer function F ( z ) . such as Shannon’s sampling theorem.
This section will initially introduce the stable region for system poles. as an equivalent to the stability region in the splane. in which the z is replaced by unity. Subsequently transforming systems represented in the sdomain to their zdomain equivalent will be . Just as with continuous time systems. Initial value theorem lim If(t)l f+O = lim zm IF(z)I This condition applies only when the limit exists.nT)Yk in which k ' = k . Time shifting theorem This follows from: m Z ( f ( t .Stability analysis 253 value. . where z = 1. 8. F ( 1 ) . further controller design being dependent on the results obtained.considered when the input to the system is sampled. This mode of expressing the steadystate value. i.n T ) ]= k=O f(kT.e.n. analysis to determine whether or not a discrete time system is stable is carried out at the initial design stage.3 Stability analysis Perhaps the most important characteristic of a system is its stability or lack of stability. is often encountered in digital control literature.
Equivalent stability regions between s.2 Zero order hold (ZOH) When a computer is used in the control of a system.' then determines the magnitude of 2.3 Stability regions (a) splane (b) zplane 8.3. whereas elwT determines the phase. showing stability occurring within the unit circle of the zplane.before being acceptable in digital form to the . i. this area being equivalent to the left half of the splane. 8. critical stability occurs when roots are such that 1 z 1 = eo = 1. this means that critical stability occurs in the zplane when a root is of magnitude 1 . a negative real part of s. In the zdomain.3. 3 . Stable region Fig.254 Digital control 8. see Fig.1) eU T e JUT Stability in the splane is given in the region represented by u < 0.1 The unit circle A linear continuous time system. that is when u = 0. the system output has to be fed into an analogtodigital converter. 8 . represented in the zplane. To summarize. characterized by its sdomain representation. Therefore the corresponding stability region in the zdomain is such that 1 z 1 = eoTwhere CJ < 0.e. is stable if all its poles lie within the unit circle of the zplane. Critical stability in the splane is found when roots lie on the imaginary axis.4.and zplanes are drawn in Fig.e. i. with any phase angle. therefore. In the previous section (8. the magnitude of z must always be less than unity.2) it was defined that the sdomain is related to the zdomain by means of the equality z = e" which is also = e(u +J~)T = (8. 8. The value en. is stable if all its poles (denominator roots) lie in the left half of the splane. This is true whether the system is in openloop or closedloop mode. n o restriction being placed on phase. a discrete time system.3.
3 Worked example for ZOH system A system is considered to be sampled every 0.5 seconds  4 ZOH u(') * G'(s) YU) . 8. While these calculations are being carried out. see Fig. and this value is held.5. If the sampled period is Tseconds.4 Staircase effect of a ZOH o n a continuous time signal computer. 8.Stability analysis A 255 Sampled signal Analog signal Representation I T 5T 10T 15T Time * Fig.e . by means of a zero order hold. as shown in Fig.3.3. it is the sampled and held value that is still being processed. given by G ' ( s ) and the zero order hold.2) where l/s is the unit step and eTs/sis the unit step delayed by T seconds.s T ) s2(s + 2) e(r) /\ T = 0. then Tseconds later the new analog signal multiplied by unity minus the old value becomes the new ZOH output. The ZOH can therefore be regarded as a unit step function with Laplace transform: (8. by unity.4. The overall output signal value is thus obtained by multiplying the ZOH by the system transfer function and the input signal. that is G(s)= (1 . 8.5 seconds. is denoted by G ( t ) . The analog signal is sampled. although the actual analog signal value can vary. The ZOH has the effect of multiplying the analog signal at a particular sampling instant. prior to being passed through to the A/D converter and used in program calculations. 8. in which the system transfer function is given by: G'(s) = 1 ~ s(s + 2) The ztransform of the overall transfer function.
3679) (9. . especially under steadystate conditions.4 Digital compensation A computer can be thought of as perhaps the most flexible. a digital compensator can be regarded as any device which samples its own input signal and provides a sampled version of its output.06606 (Z.z Is%  1 +2J where Z ( 1 .0.I + zz. In general. that is the response to a change in plant input signal.197~ 6.5 seconds. this particular system is critically stable.1. Although. In this section the time response. A table of further common ztransforms can be found in the Appendices. Then and T = 0.256 Digiral conlrol such that G(z) = (1  zl).l)(z . I t is then shown how this transfer function can be used as the starting point in the design of a compensator. Hence resistorcapacitor networks. however. having a pole actually on the unit circle (disc).606)lOO. for example.36792+ 0. hence 1 or which is G(z) = or C(Z) = 0. closedloop stability remains of great importance. of a group of digital compensators. digital compensation is generaily used for obtaining desirable transient performance.eFTJ 1 = z'.091972+ 0. 8. of a discrete time transfer function is considered.3679 From the analysis given in the previous section. can also fit the digital compensator definition. due to programming possibilities. as in the continuous time case.
a l y ( k .2) in which the time of the most recent output value is considered to be k rather than k + 2 as was the previous case.Digital compensation 8.4.2) +hu(k  1) + b2u(k .4. This is the main means by which a system response t o a specified input can be calculated.2) can be regarded is such that: y ( k ) = . when in the form of equation (8. If the transfer function relating . FirstIy rewrite the equation as: (1 + alz*+ a z ~ .1 ) .1).2) is obtained directly from equation (8.~ ) y (= (biz'+ 6 2 z . also be written as: (z2 + alz + a2) * y ( k ) = (612 + 62) u ( k ) or y(k  + 2) + a l y ( k + 1) + a 2 y ( k ) = h u ( k + 1) + b2u(k) + bzu(k .' ) u ( k ) k) or This form of equation (8.4. however.1 = k + 1 and j .4.4. therefore be written in either of the forms shown. (8. but inclusion of a new variable is thought to complicate matters.4. dependent on exactly where in time the specific time instant k is chosen. This procedure could be carried out by defining a new variable j as j = k + 2.3) By this method.1) (8. simply by multiplying through the transfer function.1 Transfer function time response 251 Consider the secondorder discrete time transfer function given in equation (8.2) (8.a2. at time instant k . as is shown by the worked example of Section 8.4. It suffices in this section merely to generalize the statements made regarding the secondorder transfer function specifically detailed.2 = k .2) However.2) will now be considered.4.4. this is the same as: y ( k )+ a l y ( k . Two forms of equation (8. hence j .62 are scalar coefficient gain values.1) in which 0 1 . the value of the system output can be calculated by multiplying past output and input values (regressors) by their associated scalar coefficients and summing the results.4.2. 61. T h e second mode in which equation (8. by Z".2) = b l u ( k .4.1) simply by multiplying top and bottom of the equation by Y 2 A n nth order transfer function can .1) + azy(k .azy(k . This equation can.1).
such that u(0)= unity. i.4. n.6y(k2)+2u(k l)+~(k2) Now. if bl. a delay of j + I sample periods is required before a change of input value has any effect on the output. equal to the magnitude of step input. as will be seen in the next section.4y(k.6~(0) 2u( 1) + ~ ( 0 ) + such that this reduces to: 1~ y(2)= . . is termed G ( z ) then this can be defined as: (8.4y( I ) . 2 By this means a table can be produced showing the new output value at each sampling instant. ~~ ~ ~~ 8.. i = I .5) As described in the previous section. . however. y ( k ) . to system output. . Significantly.4.1) ) ) where k = 1 . we have: y ( 2 ) = . see Table 8. u ( k ) . It follows then that: ~ ( 1= .2 Time response . The above equation therefore reduces to: y(1) = 2u(O) = 2 when k = 2.worked example Consider the secondorder system defined by: (8.1)0.1.6y( . As the input was zero up until k = 0. In this particular example it is apparent that once transient conditions have dissipated.1 . digital compensation cannot really cancel out this inherent delay in a system.4..258 Digital control system input. that the system output becomes unity under steadystate conditions. let a step occur at the input from an initial value zero up to unity.. the output value one sampling period after the step input has occurred is considered. terms can be zero to match the definition given.0. y(0) is also zero. it is assumed that all outputs up until k = 0 are also zero. As will be seen shortly. this can also be written as: y ( k ) = 1.1. bj terms are all zero. However.1) + 2u(O) + U( ..4) i where certain of the b and ai. 4 ~ 2 + 2 + 1 ~ 0 . It is not generally the case.1. as y ( 0 ) relies only on values up to and including k = .0..1 . all of which are zero.e. the steadystate output value is unity.. 4 ~ ( 0.
6 Compensator D ( z ) within a sampled data control system .04 1. its continuous time form.I ) k _ _ 0 0 2 0.87 1.Digilal compensation Table 8.04 1. 8.16 0.2 1. 8.84 1.4. It can be seen from Fig. ~ ( .01 y(k) ~ _ 0 2 0.96 1.6. 8.97 8.52 0.10 0.2 k u(k 2) ___ 0 I 2 3 4 259 u(k  1) 0 0 1 1 1 0 I I I 5 6 7   1 ~ 0 0 0 2 0.09 0.52 0.10 0. the computer program can be such that it provides a rational transfer function in the sdomain.1: assuming that y ( t )= y ’ ( t ) .04 1. D ( z ) as a multiple of the error between actual and reference outputs.1 Table of output values for the worked example of Section 8.4.6 that the closedloop transfer function is: (8.96 I .09 0.16 0.87 1.84 1.2 y(k  1.87 I . where G ( s ) is the multiple of the zeroorder hold and the ztransform of G ( s ) .01 0.52 0.01 I .09 0.01 1.10 8  9 10 II 12 1 13 I 14 I I I I 0.75 1. This is shown in Fig.01 2) .84 1. 8.75 1.2 I .16 0.3 Compensator design Consider the computercontrolled system shown in Fig.4.75 I .96 1.6) such that if a certain secondorder system is defined by: G ( s )= 3 ~ s(s+ 1) input  f Fig.
1 Let the compensator be: mZ)= ( z + A ) D(z)G(z) = K ( z .8467) h + such that the closedloop denominator is then: z2 + z(0.3196K + X .6065) As an example of compensator design.X) and this must be equated with the desired denominator z 2 + 0.1 ~ or Hence G(z) = (Z  0.1) + (0.1) . resulting in the matrices [0*3196 0.32 + 0. T = 0.A ) + 0.1.2706 .:I = which give values of K = 2.z')z [s2(s3+ 1.3721.3196(z + 0.0.l)(z + X) and part (a) of the design problem has been accomplished.8467) I)(z .32 + 0.3196K(z + 0.6) to be: Y(t)= (Z2 +~ 0. The closedloop expression relating u ( t ) to y ( t ) is found from (8. and (b) to choose a gain K such that along with a correct choice of A. the closedloop denominator becomes z2 + 0.6065) (z + 0.260 Digital control when the sampling period.4.5 seconds. The final compensator is thus: D ( z )= 2.[ [. X = 0.1 '1 ].5419) . its zdomain representation becomes: G(z) = (1 .0.6065) The openloop gain therefore is: 0.3196K(z + 0. assume this particular problem to be (a) to provide a compensator. satisfying part (b) of the design problem.6065 is replaced by an openloop pole at z = X.3196K(~ 0.8467) ( z .5419.2706K .8467) u(t) ( .3721(~ 0. D(z). such that the openloop pole at z = 0.
Consider the discretetime transfer function (8.State variable forms 8.4) Xnl(k+ I ) = x n ( k ) x. reasons for selecting the one described will hopefully become apparent in Section 8.' y ( k ) = y ( k .5. 8.1) + + a. The state variables are firstly defined such that subsequently the characteristic equation and a return to the polynomial description can be obtained. ~ .3) The state variables can now be introduced by a suitable choice. This section gives details of a statespace framework for use with discrete time systems.a l x n ( k ).5 State variable forms 261 In the two previous chapters a statespace framework was firstly introduced and then employed in controller design work.5.4). therefore one representation only will be considered.5." ) y ( k = (b1zI + + bnZ")u(k) y ( k ) + a . This is no less true about discrete time statespace descriptions. In this chapter. discrete time systems have been discussed.y(k . .a2xn. specifically concerning itself with a transform into the statespace from the original polynomial format.4.1) + b2u(k . dependent on the state vector chosen. ( k + 1) = X2(k) X2(k + 1) = x3(k) (8.~ ( k) 1. This was done entirely with continuous time systems in mind. however.n ) (1 a * * * in which z .5. in which: (8.6. in particular with reference to an input/output discrete time polynomial transfer function.1) or This can also be written as: or (8.2) + + bnU(k .n ) = b l u ( k .1 State variables for discrete time systems It has previously been mentioned that for any one system there are an infinite number of possible statespace representations. Let the state variables be defined as: x .i) (8..xl(k) .(k + 1) = u ( k ) . y ( k .a.5.2) ) + a1zl + + a . other statespace forms being obtainable from this by means of state variable transforms.
' B u ( k ) ~ (8. B and C matrices.. x i ( k )+ b.(k) (8. and b.6) to be:  ~ ( +k1) = Ax(k)+ Bu(k) Then q ( k ).4).5) In matrix form. ...5.262 Digital control with the output equation: ~ ( k= )b . i.5. Remembering the first of these state equations from (8. This is also true for the state vector itself (8.5. such that given any A . parameters.1) also occur. here they are for convenience a.5.R ) . B= (8.e..: i = 1.5. and b. in the A and C matrices.1 .. the definition of which was based on the initial transfer function parameters.7) L and  a.9) .A g k ) = B u ( k ) (8. the only difference in the A and C matrices being the method of defining the a.5.ixz(k) + + bix. the discrete time statespace representation can be written directly from the discrete time transfer function.( k )( z [ ..6) ...2 Solution of state equations The discrete time state equation can be solved. One of the important properties of this particular state representation is that the coefficients occurring in the transfer function (8..5.5.  a. the values of scalar state coefficients in the state vector can be found.8) or .: i = n .0. and b. 8. I .. n whereas in Chapter 6 they are defined as a.. . .  a2  al These equations therefore take on a similar form to the phase variable continuous time state equations described in Chapter 6. i. with n o modification.5.e. the state equations can thus be written as: (8.
35y(k .11) (8. denoted by x(O).2 .5.13) Consider the system defined by: This can also be written in the form: y ( k ) = 1.zl AI= 0.A I = 2" + a l z n .5.5.2Xz(k)0.5.3 Worked example + an= 0 (8.5.Su(k .12) where the characteristic equation is given by the denominator of G(z) as: I z l .1) .6).2) But by using the state vector definitions introduced in (8.A)'Bu(k) (8.l + 8. 2 z + 0 .5. the elements in this vector are: Xl(k + 1) = X 2 ( k ) x 2 ( k + I ) = u ( k ) + 1. namely: y(k)= W k ) the following expression is obtained: y ( k )= C(2l.12). 3 5 1. Firstly: .A ) .5.' B or (8.1) it can be seen that G ( . it will be transformed back to G(z) by means of (8. z ) =C ( d .2) + ~ ( k1) + O.9) into the second equation of (8.2y(k . contains only zero elements.35 Z ' 1 = z 2 .5.4).1 .10) But from (8.0.State variable forms 263 in which it is assumed that the initial state vector.5.5. Substituting for (8.35Xt(k) or with an output equation: To check that this discrete time statespace representation is a true one.
6 Pole placement design The poles of a system transfer function are the factors of its denominator polynomial or roots of its characteristic equation.6.3) . Pole placement for discrete time systems will be considered firstly by means of state feedback and secondly by straightforward polynomial controller design. ftj19 .A ) ) B = 0.6.6.1) and ~ ( k= )W k ) Let the control input at time instant k be: u ( k )= F&(k)+ u ( k ) (8. in this section it is therefore assumed that this is the case.. G(z)..2) where u ( k ) is the reference input signal at time instant k and F is the 1 x n dimensional feedback matrix of the form: F= [ftl.5 + z which is certainly equal to the numerator of the transfer function.1 State feedback pole placement Assume the linear system is described by the state equations:  x(k+ I) = A(k) x + Bu(k) (8. The control criterion based o n pole placement is such that given an openloop system transfer function the feedback provided will cause the closedloop poles t o lie in previously specified positions. f t l (8. It is only possible t o d o this when the nth order system in question is completely controllable. 8.264 Digital control which therefore equates the denominator and hence the characteristic polynomial of the t w o representations. 8. The numerator of G(z) should be equal to: Hence: C ( a d j ( z 1 ..6.
BFx(k) + B u ( ~ ) ~ or (8.' B v ( k ) (8.( k ) [ ~ l] . The closedloop denominator from (8.. it can be seen that for an ndimensional system: (8. l).6) In pole placement design it is only the poles. f2 .6. If a choice or selection of the n closedloop poles has previously been made.5 and defined in (8.i .8) are equal to those defined in the pole polynomial (8.( +k1) = Ax(/?). i = 1.6..1). .6.BF. 0 1 fi 1 .5..6. the desired closedloop pole polynomial can be expressed as : T ( z )= zn+ tlz"' + t2Zn2 + + t.an. (8.6.6.4) .9) and hence: 0 F=ABF= I .ai . and hence the denominator of the closedloop equation (8.. i.e.. the following is obtained: .6. f i ..6. *** (8.(12.6).7) in which ti.FI = I zZ ( A . on which our interest is centered.8) It follows then that f i .6.2.7).Pole placement design 265 By substituting for (8. B and C matrices employed in Section 8. by multiplying out these factors. Thus: (8.F ] .6.7). . the final closedloop form relating reference input to system output is: y ( k ) = C[ZZ.6.' B u ( k ) and when this is substituted into the second state equation of (8..F ] .6) is given by: I tl.an1.10) 0 f . the output equation. f n must be selected such that the coefficients obtained from (8. .B u ( ~ ) ~ F = in which F = A .2) in the first equation of (8. Using the A .n are previously defined and hence known values.6.5) x(k)= [zl...BF) 1 (8.6. . f n .6..
l.7. did not appear in (8. > + . where the statespace description is not in this particular phase variable form the bi coefficients will indeed be present in the state feedback parameter.4 and z = 0.5 must be acted upon.6 and z = 0.* + (ani + fil. The problem can therefore be regarded as one in which the openloop poles situated at z = 0.9~ 0.it was shown that the coefficients of the transfer function numerator polynomial. b.. IzlF(= Thus [ 0 0 : z. These polynomials can + also be written in factored form such that: A (Z)= (Z .5)  and T ( z )= (2 .7). xz(k) u(k) .6.2 0.6)(~ 0..13) 8.42. from which it follows that for complete pole placement: (8. that the state equations are in the form of (8.5.12) can now be equated with those of the specified pole polynomial (8. hence let A ( z ) = z 2 .9 ] [ ] [*q] +[ ...7).6.1.12) The coefficients of (8.6.~)(Z 0. It will be considered here.0. < alIl+ & I . can be the written as: [ x i ( k + 1) = O I ~2(k+ 1) 0..2 Worked example using state feedback When using the phase variable statespace system description as was done in Section 8.6.2.13). such that in the closedloop transfer function the poles are placed at z = 0..In general. an+f..i )Z+ ( 0 .7) state representation.6. however. and T(z)= z 2 .6.7) in which A ( z ) is the system denominator whereas T(z) the desired closedis loop denominator.266 Digital control Therefore Z 1 0 0 . fi.5. by means of the controller. excluding output equation. + j n ) (8. a2 0 1 +ft z+ at :I (8.1. %h.6.F 1 = t l l + (a1 +j i )z” I + (a2 + j i ) z ” + .32 + 0.6.0.11) I ZI . From (8. calculations thus leading to a more difficult problem.O.
13) to be: f l = a l . 3 ~ + 0 .Pole placement design The state feedback parameters can be found from (8.e.I I F ( z ) = z"' + f1znZ + .6.3 Polynomial pole placement The system is defined as in (8.4 = 0.6.0.15) These polynomials are applied in the form of: or F ( z ) u ( k )= D ( z ) y ( k )+ o ( k + n .. y ( k ) = G ( z ) u ( k ) = u ( k ) B(z) A (1) (8.6.9 f2 = a2 12 261 + 1. u ( k ) ..5.. However. The control input.2 .7) as T ( z ) .I ) . i.3 = 0.0.1 . 4 2 = T(t) 8.22 The control input at time instant k is therefore u ( k )= 0. Also Thus Hence IzZFI = z 2 .14) with the pole polynomial specified by the operator and defined in (8.t l = 0.42 = . general feedback polynomials D ( z ) and F ( z ) are defined here to be: D ( z ) = doZ"' + diZ"' and + *.6.*+fn(8.+ d.22X1(k) + 0.4X2(k) + U ( k ) where u ( k ) is the reference input signal.can take on one of several forms to provide feedback of past and present output values and feedback of past input values.1 ) where the coefficients of D ( z ) and F ( z ) are calculated to provide a closedloop pole .6..
.6.17) can be written in terms of the available coefficients as: [ ( z 2 .6. the polynomials A ( z ) and T ( z ) are those defined in the worked example of Section 8..has now been modified to the desired polynomial.] 0.2.6.2 1 0.42)~ di From this. 8. T ( z ) .0.18) It is noticeable that although the openloop denominator.o":]= [i:. the openloop numerator B ( z ) remains intact as the closedloop numerator.6. with T ( z ) as its denominator by: (8.92 + 0.6. A ( z ) .This can be seen if (8. The same example as that used in the previous section with a B ( z ) polynomial included this time. in the closed loop.268 Digital control polynomial equal to T ( z ) .2] = (Z + O. these will be of the form: D ( z ) = doz + dl and F ( z )= z + fl  Hence equation (8.17) in which the z"I term is required t o cancel with that on the top line of (8. the coefficients are calculated as: =  0.1667 . Then.6.6.15) is substituted into (8.acting on v ( k ) .2333. do = 0.1. from the definitions of the controller polynomials. It follows that the closedloop equation must be represented. this is: [ fi 1 0.6.32 +0.S)(doZ + di)]= (2' . will now be employed to show the mode of operation of the polynomial pole placement controller.5 F] [a.4 Worked example showing polynomial pole placement Consider the system in which B ( z ) = z + 0.15).'  (8.2)(2 + f i ) In matrix form.5 0 .5.14) t o give: or Then the coefficients of F ( z ) and D ( z ) are calculated such that: [ A ( z ) F ( z ) B(z)D(z)l= T ( z ) z " .9 0.
0933 i Hence the control input is provided by utilizing equation (8.0. Problems 8.0.16672 . Microprocessors have more recently continued this trend and it is envisaged that special purpose devices or machines based on VLSI chips will retain its impetus.l).5zI 1. this can be implemented by obtaining the control input at instant k from: ~ ( k=)0.2333) On multiplying both sides of the above equation by ( z .0.2333u(k . This chapter has merely acted as an introduction to the enthralling world of digital control which includes controllers whose parameters can vary in line with any system alterations while the controller is still in operation. i.0. that is: u ( k )= (0. however.0933y(k .7 Summary Digital control. following a unit step input: (a) G ( z )=0.19. y ( k .5 . content itself with looking at one or two more special forms of digital control.52I 1 .1667y(k).1 z .0.5 Z .1) + ~ ( k ) where u ( k ) is the reference input at instant k and u(k . 8.0933)y(k) ( Z .0.0.Problems 269 and d = .5 (c) G ( z ) =  0. especially in the form of computer control.0.9 . y ( k ) . This book will. as described in the remaining chapters. a larger proportion of controllers are now based on digital control design philosophies.1) + 0. constitutes a continually larger slice of the control systems cake.0.2333) + ~ ( k ) ( Z .1) are known values. The controller can therefore adapt or tune itself to variations in the plant under control and can lead to more robust controllers with far less necessary operator supervision.e.1 Sketch the output response of the following openloop systems.0. .2333)z'.5 (b) G(z)=z + 0.6.
7)/(z . Use (a) G ( s ) =  1 s+ 1 1 (b) G ( s ) = .1 8.0. within unity feedback loop.4.l ) ( z .2 Note: It is assumed that the function G(z) already includes the effect of a zerc order hold.(s) 8.5) (€9 G ( z ) = ( z + o * l ) 8.5(z + 0. following a unit step input.0.6 = 3 s+2 Consider the system shown in Fig. find the output time response whe T = 0 . when the ztransform ( + zeri order hold) is taken.5 For the closedloop system of Problem 8. z3 . 2 seconds and G.252 + 0.1 1O(z + 0. 8.s.82’ . Find the digital compensator of the form D ( z ) = K(z + b ) / ( z+ a ) ..0.2 Investigate the effect on the stability of a system. 1 10 (e) G ( z ) =  z+o.7. Find the output time response. 8. such tha when cascaded with a system G ( z ) = 0. 8 .4 as starting points.0 .9 z.3 8.9.8 where G(z). 7 8.5) ( f ) G ( z ) = ( z + o*l) ~O(Z 0. the closedloop denominator is given b.(s) = ___ s(s + 6) * 5 I I Fig.270 Digital control (d) G ( z ) = 0. which includes the effect of . for the closedloo1 system shown in Fig. where G ( s )consists of a zero order hold in cascade wit1 the openloop system I G.
when K ( z + 0.6) G ( z )= (1.9)(~ K ( z .l)(z + 0.3) G ( t ) = (1.0.10 Find the openloop system transfer function G ( z ) .2) 8.Problems 211 I Fig.0. Plot the root loci for this system for variations in K when K' = 0.8 For the closedloop system of Problem 8. Find the system transfer function and hence obtain a statespace representation which is observable.0. i. 8. plot the zplane root loci for changes in gain K when K ( z + 0.noting that a zero order hold must be included.6.5 seconds and 8. A = [ 12 B=[8] and CT=[I] 19 8 Show that this system is unobservable. is given by G ( z )= (Z . For the closedloop system of Problem 8.8 zero order hold.e. .3) 8. plot the root loci for this system. within the unity feedback loop.5) + 0. with T = 0.5.l)] is cascaded with G ( z ) .K ' ) / ( z . over which the closedloop system is stable.7 Show how the closedloop root positions move around in the zplane as the gain K changes.9 Consider the closedloop system of Problem 8. Consider the discretetime system in statespace form ( k x + 1) = A x @ ) + Bu(k) Y ( k )= Cx(k) where A PI. find the range of values for K .l)(z .4.1) 8.6. A digital controller of the form D ( z )= K [ ( z .
8 ? j O . by plotting the root loci for varying 00 > K > 0. 8. find a vector F such that the closedloop pole z 3 + 0.024 . 8.13 For the system of Problem 8.0.9 where G I( z ) is the multiple of a zero order hold and the ztransform of s'.192 . If N i s a unit step disturbance input.e. By employing state feedback u ( k )= . find the maximum value of K for which the closedloop system is stable. For the discretetime state equation 8.1 I D ( z )= K ~ 1 .0.' ) and K is half its value at the limit of stability. with D ( z ) = K .z . 5 seconds. 8.11. the sampling period T = 0 .14 find a transform matrix t o place the state equation in canonical form.9 8. and G2(z) is the ttransform of (s+ 2)'. calculate the output signal response t o the same disturbance. find the effect of the disturbance on the system output. by calculating several output signal values with respect to time. set K to equal half its value at the limit of stability.& ( k ) polynomial is: + u(t).12 8.zl N Fig. Also. obtain a vector F which produces a characteristic polynomial with roots at z = 0. What happens if an integral action controller is employed instead? i. If D ( z ) = K . For the discretetime statespace system description If u ( k ) = . 1 .212 Digital control Consider the system shown in Fig.F x ( k ) . If D ( z ) = K / ( 1 .62' .
.).. PrenticeHall. R. 1991. McGrawHill.. J. 1984. 1965. H. PrenticeHall. Kuo. PrenticeHall. 1993. Marcel Dekker. D. and Powell. Discretetime systems.. V. S.. J. Ragazzini. Applied digital control. 1958. 59. Stoten.pp. PrenticeHall.. pp. Smith.Digital control systems. Part D. P. Strejc. Malik. Applied control... 141. J. ‘Generation of discrete and continuous time transfer function coefficients’. 1972. Phillips. L. and Nagle. G.F. Holt.. AddisonWesley. Academic Press. A. D. Sampleddata control systems. K. AddisonWesley. 1994. Scranton. R. and Franklin. K. 1984. C. G2(2)is the ztransform of s . Soeterboek. AddisonWesley. A. and Hope. SpringerVerlag. J.~ the sampling period T = 2 seconds. G. f Franklin. R. R.. 1973. f Tou. B. G.. Tzafestas. H. C. Bishop. F.. J. 1980.11591172. and What is the system output response to a unit ramp input signal u ( t ) . and Martens. Dorf. Leigh.PrenticeHall. C. 1975. Warwick. 0. Discrete time and computer control systems. Digital control system analysis and design. Int. P. J. E. and Harrison.). B. Intext. R. Jury. 1981.. Rinehart and Winston. G. (ed. A. PrenticeHall. Isermann. McGrawHill. P.. Digital control o dynamic systems. D. and Rees. Sampled data control systems.Digital computer process control. Cadzow. Time domain analysis and design of control systems. A.. R..Predictive conrrol. P. I.. Digital control systems. 1992. T. GI 2 ) is merely a zero order ( hold. ‘Generalised discrete control system design method with control limit considerations’. Peter Peregrinus Ltd. L. 1959. PrenticeHall.Digital and sampleddata control systems. G. 1958. Franklin. Industrial digital control systems. and Wittenmark.. Cadzow..15 213 For the system diagram shown in Problem 8. John Wiley and Sons Inc. G. T..Further Reading 8. Chen.. 1991. 1984. 1980.11. Introduction to discrete linear controls. J. and what steadystate error is caused when N is a unit step disturbance input? Further reading Astrom. C. IEE.. Katz. Digital control systems. Digital control using microprocessors. Proc. . New York. (eds. 1970. B. New York. J. Computer controlled systems. S. J. John Wiley and Sons. 3947. Control. 2nd ed.1994.. 1981. L. 1986. State space theory o discrete linear control.
In Section 9. Hence system performance can be investigated. In its simplest form it can be employed merely as a collection point for plant input/output information. are considered. it is only when the computer is connected within a feedback loop that it has an effect on system performance. when compared with conventional analog controllers. both within and outside the expected operation range. rather than merely being a tool for the purpose of digital controller implementation. thus any problem areas or improvement possibilities can subsequently be taken care of as far as the actual system is concerned. and often also in size. while still remaining within the field of control systems analysis. computers can be used for far more than simply data logging or system simulation. In this chapter various areas for the employment of computers are outlined.2. the use of a computer as a central element in a closedloop control system is discussed. This can mean however that the computer is employed simply as a digital control device.1 introduction In Chapter 8. and although excellent data display capabilities are usually proffered by the computer. Included are facilities for signal processing. by means of a discussion as to the various roles that a computer can take. At a higher level though. When not within a feedback loop. in terms of the simulation. however. In fact complete packages are available which can take as input plant input/output data. and advantages. when compared with the system itself. .Computing methods 9. Further. the computer can take on a more intelligent role and can be used as an expert system to make decisions concerning the type of controller to be employed or action to be taken when fault conditions occur. Computers have a much wider applicability in the field of control systems. with the simulation. Simulating a realworld system with a computerbased model almost always results in a reduction in cost. tests can be carried out not only under all likely operating conditions but also under some alternative special conditions.3. Several different aspects of computerbased control systems are considered in Section 9. multiplying inputloutput signals with controller coefficients and sending the resultant control signal to the plant. system simulation by means of computers is covered.
9. leading to greater expense. An alternative approach is to model the required system in the form of a computer simulation. Such computeraided control system design packages invariably contain selftesting procedures in order to ensure the validity of the system models obtained and to check on the final overall controller performance by means of a simulation trial run. e. modifications can be carried out on the simulation and when an idea for a new design is found to work successfully. Electronic circuits are used as basic building blocks. construction and analysis of a complete control system. controller design and system simulation. rather than the actual physical components in the case of the system itself. in that if the simulation does not represent the original system with a certain amount of accuracy.2. can be both timeconsuming and expensive. Indeed. in the analysis stage.System simulation 275 system modelling. with a simulation it is possible to apply different signals and disturbances without worrying unduly over the system blowing up or causing physical harm. then results obtained from the simulation may well not hold for the original system. Also. forming links between signals. Nevertheless. The simulation of systems by means of a computer is very useful when cost is restricted. 9. the simulation can still be employed to get a feel for the characteristics of the actual system. it can then also be tried out on the actual system with a much greater degree of confidence. with the latter two elements being based on results obtained from the first two. where the term analog is used to signify the fact that the voltage and current signals employed are analogous to the signals found in the original system. The simulated system can then be modified with very little effort and no further expense as far as rebuilding or restructuring components is concerned.g. This is particularly true if it is found. system simulations are now almost always realised .2 System simulation The design. we cannot all afford a space shuttle. or even an element of a more complex process. Accuracy is an important factor for the simulated model. Although analog models and analog computers are still encountered. that the system is not quite right and some modifications are required. in that the purchase of the actual system under investigation may well not be a possibility.1 Analog models Analog simulations are obtained by means of an analog computer. either in terms of an analog hardware design or as a digital softwarebased scheme.
then the continuoustime signal remains constant during that T seconds. Therefore. a computer is an extremely important tool by which means the system can be simulated not only to check results already available. when simulation of a continuoustime system is to be carried out.276 Computing methods using digital computers. this latter assumption is not in general true.1 Consider the RC system shown in Fig. i. The main reasons for this are perhaps improved accuracy and reliability. is described by the equation: . although standardization within establishments and research groups is commonplace. allowing system models to be varied arbitrarily and for a large selection of input signals to be applied. leading to the conclusion that a digital computer simulation of a continuoustime system will be subject to modeling errors due to the discretization process. In fact further errors will also be present due to the use of a computer. a basic assumption being made that with a sampling period of T seconds. 8. When a system is represented in terms of a mathematical model.2. The vast majority of modernday simulation studies are carried out by means of software development on digital computers. but also to develop improved controller performance. The concept of a delay operator system description is discussed in Section 8. shown in Fig.1. Hence.2.e.2. allowing for multiuser access to specific system simulation packages. by means of appropriate approximations. A wide variety of computers are available for programming and an everincreasing choice of programming languages is apparent. in response to a unit step in input voltage.1 and the method of sampling a continuoustime signal by means of a zero order hold. As discussed in the previous chapter. approximations must be made in order to obtain discretetime signals. such as a differential equation. 9.4. although digital computer simulations are usually far more accurate than those obtained by analog computer. Example 9. a digital computer operates on discrete signal values. without the need for amplitude scaling because of saturation effects. along with increased flexibility. caused by truncation of real numbers to fit the computer’s finite storage length. where that model is in the form of a continuoustime equation.4. Hence. as is shown in Fig. no further discussion is therefore given here on analog models and readers are directed to Gordon (1978) for further details. Unfortunately. where the output voltage. it must be remembered that errors do exist and that in some cases these errors can be critical. 8.2 Software models In the design and analysis of a complicated control system. only a finite number of bits are used to represent each real number. then this must be converted into a difference equation. 9.
i.uo(kT .(t) voltages are constant in value over each inter sample interval lasting T seconds.1) is therefore given as: .such that v.2.1) is a differential equation which represents the output voltage with respect to time. with the assumption being made that the ui(t) and u.2.T ) T (9.(f) = i(t)R + u.System simulation u. by defining the time index in terms of sample periods rather than as an absolute value: (9.4) A discretetime approximation of the differential equation (9. is assumed to be equal to its value when sampled at kT .(t) : uo(to) = 0 277 and as this means that (9.2) The derivatives of the output voltage can then be approximated in terms of the gradient between the output voltage value at t = (kT . This is certainly true for ui(t). A difference equation approximation of (9.3) or.2..T.2. uo(f) = o.(kT .2.e.(&T) .T.T ) : (kT . as it is a unit step input applied at t = to. The value of voltage u&) up to a particular time instant kT where k is an integer number of sample periods T.2.1) < t < kT (9.1) is required.T) and the new value at t = kT. from the previous time instant kT .
T ) . then equation (9. as expected from the original system in Fig. A simulation of the simple RC system can therefore be carried out quite quickly by means of a digital computer once the initial program. 9.2.2. C and R. R and T .1 0 1 2 3 4 0 0. without the need for time consuming rewiring.271 0. based around the recursive equation (9. which might be necessary with the actual system. .(k .(kT .4.I ) ] (9. etc.1) = ui(k).(kT . Also. when CR = 10 and T = 1 . R or T can then easily be made and the program rerun.1. The output voltage signal u.T ) ] T = u. the magnitude of step input.(k) = 0.19 0.u. then uo(k) = uo(k .9u0(k  1) + 0. at any discretetime instant t = kT. ( k T ) . [ l  uo(k .1 0.(t.2 and therefore (9. The first four iterations are however shown. and given the initial value of u.5) remembering that u. any arbitrary values can be selected for C.5) can be recursively evaluated in order to find the time response of the RC circuit for a particular choice of values T. Any modifications to C.5) is not laboriously recursively calculated here.218 Computing methods S [ u .5).1 ) + 0.we are not restricted into using the handful of capacitors which are physically in stock. A similar form of time response is considered in the worked example of Section 8.T ) or Uo(k) = uo(k .(t) is now described.1 .). in terms of its previous value. has been written.1) + T [1 CR .2.T ) = 1 = ui(k .l ) ] or u. with the output voltage heading asymptotically towards a final value equal to unity.uo(kT .uo(kT .3439 The exponential rise is thus already apparent.2.
however. 9. 9.1 RC circuit for Example 9. Using a computer for control purposes provides a great deal more flexibility than is otherwise obtainable. merely monitoring the behaviour of a system over a period of time. i. can be restructured. but also the role of other associated controllers.2. the text by Banks (1986) details many such schemes. controller which assesses and coordinates data from several sources and which takes a “commonsense”. possibly themselves computer based.2.e. be used merely as a “dumb” controller by evaluating a set of equations and giving out an appropriate actuator signal. 9.I In the example given.3. approach to its calculations. various ways are considered in which a computer can be used either directly or indirectly in the control of a system.1 Data logging Probably the simplest means by which a computer can be made use of in a control system is for data logging purposes as shown in Fig. to use a computer as the basis for a more intelligent. For more complex control systems and/or where improved accuracy is required. in the simplest case. The trend is. This can mean. merely an alternative means for system reproduction. other approaches to the digital approximation problem can be considered.3 Computer control In Section 9. active. 9. By means of a computer therefore.Computer control 279 Fig. The computer can. not only can the type of control action and objective be modified due to certain occurrences. for instance.2 the discussion is centered on the use of computers as a basis for control system simulations. In this section. although it can also mean that a computer calculates the required control action by fulfilling the role of feedback element. expert. a fairly straightforward firstorder differential equation was approximated by a simple gradient technique. .
it can use the data to estimate the coefficients in a structured model of the system. Although remaining within the general classification of a data logger. 9. possibly interconnected. in response to a logged input signal.2.2. systems and will provide detailed graphical output which is representative of the data collected. Remaining within the framework of Fig. Alternatively. is the use of a computer as a fuuZt detection device. 9. if the computer is used as an aid to carrying out corrective action on the system input or the system itself. effectively this means that it is being employed within a feedback loop and is therefore no longer a simple data logger. as shown in Fig. Examples of simple fault conditions are. (a) increasingly large output oscillations and (b) output signal stuck for a long period of time at a high or low value. despite input signal fluctuations. However.2 it is very often the case that the computer will actually log data from several. information obtained over a period of time may well be used as the basis of a control system reformulation. 9. can be detected and signalled by means of an alarm. When it is predicted that fault conditions will occur if the present course of action is continued. an output pattern. then an alarm can be given. it is possible for certain actions to be taken within the computer. By monitoring the output signal over a period of time. . which is of extreme importance. One such case. or for a new model to be calculated when it is eventually decided that the old model is no longer accurate enough. the data collected from a system can be used to predict performance or output position. more powerful. which has previously been denoted as that caused by a fault condition. although this may well mean that a mathematical model of the system is also formulated and it is from this model that the prediction is made. However. i. With such a system as shown in Fig. passing values periodically to a central.e. 9.2 A computer connected for data logging When in this mode the computer takes no direct active part in the control of the plant. Where the plant itself is a fairly simple system it is likely that a lowcost microcomputer will be used as a local data logger.280 Computing methods Computer Input 1 4 Plant I 1 output Fig. machine. without a direct feedback link being provided to the system. it is also possible for the computer to use the data collected in order to calculate a mathematical model of the system. Any further data collected can then be used to assess the goodness of the model calculated and to either cause the model coefficients to be modified in order to improve the model.
It was only in the 1960s that it became viable to replace hardwired analog controllers with digital computer counterparts. If an analog controller has been working successfully on a plant. Flexibility in control action: computer program modification rather than analog rewiring. several control loops can be coupled in any desired mode. Although DDC schemes can readily be programmed in common highlevel languages such as Pascal. their role was merely that of a supervisor. g . Improved user interface: information can be displayed graphically on a VDU when required. without limitations relating to physically realizable schemes. a clock which periodically gives an output . Reduced cost when further control loops are added: extra loops merely need to be connected up to the same computer.1 and 8. Coupling of control loops: simply by linking different programs. 6 . ambient temperature. when computers were first employed for control purposes. e. rather than employing a large panel for the display of limited analog information. By these means the operator is able to refer directly to input/output variables and important plant/controller features.2.Computer control 9. 3. i. Analog controller performance can often deteriorate considerably with time. without the need to know intricate programming details. Basic. This is particularly true as the initial cost of implementation of the computer will most likely be very high. a further assumption must be made however that the computer is required to contain an onboard realtime clock. reasons for replacing the controller with a computerbased scheme must be convincing.e. 2. 8. several specialpurpose languages have been derived. Although carrying out a similar function. when compared with the analog format. more analog controllers would otherwise be required. sounding an alarm under fault conditions. for some time. Adaptive control: the controller can be modified in realtime when operating conditions. some of which are as follows: 1.2 Realtime control 28 1 Initially. Increased controller complexity: versatility in programming allows for an infinite number of possible controllers. Schematic diagrams of a computerbased control system are shown in Figs. It is then also possible for the operator to select an appropriate control algorithm from those available. Fortran. as in the analog case. the computer in a DDC scheme forms part of the control loop and has a direct controlling action on the plant. But the computerbased controller has several advantages over analog schemes. vary. 4. often referred to as direct digital controllers (DDC). along with a suitable sampling period for the plant under control. 5 .3.
where online control is desired. taken at the sampling instants. The computer must therefore operate in realtime. then there is little use in the computer providing a control action which is updated once every three days. 9. Controller forms employed for realtime control are based on the zoperator description. and are not considered further here. to the system to be controlled. if a system responds with a time constant of 3 seconds to an input signal change. the controller can then be switched online and hence will take over system control.3.g. This does not in general mean though that the operator can immediately take a long vacation. As an aid to proof of controller viability it is most sensible to first carry out a simulated plant application (see Fig. because a supervisory role will normally be required in the first instance. via appropriate interfacing. The sampling period must therefore be directly dependent on the response speed of the plant. that is the time frame of the plant. the computer can be connected.3). samples are required much less frequently. changes in control action must be carried out at a frequency which is found to be most suitable for the system. e. and this presents a major problem for realtime control. As soon as the operator is content that everything is in order. thus complex updating controller calculations can only be carried out if they can be completed within each sampling period.3. as discussed in Chapter 8. If the plant under control has a relatively fast response pattern then samples must be taken frequently. based essentially on the time required by the computer for calculation purposes. whereas the exact time at which a sample is taken is referred to as a sampling instant. . The main difference between a computerbased (DDC) control scheme and an analog controller is that the former employs a control action which is derived from discretetime plant input and output signals.282 Computing methods pulse. Before the controller is actually switched online however. when a computer is connected online to a system in order to control that system.3 Simulated application Once a suitable computerbased control (DDC) algorithm has been selected. See Astrom and Wittenmark (1984) for a more indepth view. The time between two consecutive samples being taken is referred to as the sampling period. it is most likely the case that the plant manager will need to be convinced that the programmed controller will definitely work on the plant in question. However. The analogtodigital converter can then be synchronized. whereas when the plant’s response pattern is slow. In Section 9. such that the system output measurement is sampled with a period equal to an integer number of clock pulses. 9.3 the concept of a simulated application is discussed and in this case it is usual for the computer to employ its own time scaling.
whereas its computerbased model will be a discretetime approximation. With a simulated application it is possible to try many different control schemes on a particular plant model. a simulated application can take only a fraction of a second. nonlinear effects. if a controller causes a simulated plant to “blow up”. So although the control of a realworld plant may take several days to analyze. then this will only be a numerical “blowup” rather than a completely destroyed real plant. when compared to the relevant realworld system is dependent largely on the accuracy with which the mathematical plant model mimics the realworld plant. a mathematical model of the plant to be controlled is stored in the computer. a lesson will have been learnt without adversely affecting the plant itself. Note: This is also the plant input. If a certain controller is found to work well. Conversely if a controller is found to work very badly within the confines of the simulation. The plant itself will however. and is not limited by any realtime restrictions placed by the online control of a realworld system.e. due to the slow response time of that plant. or to tailor one specific control scheme in order to improve performance. being retained within the computer and merely fed to the plant model rather than to interface circuitry. Similarly the evaluated model output is immediately fed back into the controller program. with the computer being used to not only calculate the control signals but also to calculate the plant model response. almost certainly. So. It is often important therefore to include. The programmed control algorithm can then be run with the controller output. A . All of the signals connected with the simulation therefore are discretetime signals due to their remaining within the computer spectrum. then it can later be applied to the actual system. The simulation speed is then only dependent on the speed of the computer.Computer control 203 r I I I I I I I I L I Computer 1 spectrum controller I Fig 9. The complete simulation process therefore takes place within the spectrum of the computer.3 Simulated application With a simulated application. be a continuoustime system. such as saturation and hysteresis. The accuracy and usefulness of a simulated application. where necessary. for the plant output to continually increase in value to unacceptable levels. and can be remedied by simply restarting the program with a different set of parameters. within the plant model. i.
284 Computing methods simulated application is therefore a good means by which controller program bugs can be ironed out. It is generally much cheaper to simulate a plant with a computerbased plant model.3.. 9. or is not. a great deal is lost in terms of flexibility and ease of modifications.. it is worth noting that although the simulated applications considered here are completely computer based. one can deduce the general application areas of the controller. 9. To conclude this section. via a simulation. Also. detailed displays can be obtained. while information from a simulation is retained within the computer.4. This technique adds a certain amount of realism to the simulation. and to apply the controller to the model. one final form of simulated application is that encountered when Computer 2 in Fig. discover any likely problems and most importantly prove that the controller is. 9. 9. as shown in Fig. a viable proposition.. It is also possible for two (or more) computers to be used. when compared with a realworld implementation. As a followon from this.4 Multiuser simulated application There are many advantages to carrying out a simulated application. not the least of which is cost. hence providing an aid to further analysis. In order to carry out a simulated example. The effect of operational problems such as . one type of controller on many different plant models. rather than to buy in the actual plant. r1 r . either in graphical or numerical forms. however. It is also possible to try out. in that the interfacing circuitry must bridge the digital and analog worlds.4 is replaced by an analog model representation of the realworld plant. no matter how complex it might be. This is especially true when one is not completely certain of the operating effectiveness and reliability of the controller.4 Expert systems and artificial intelligence In practice it is not usually possible to straightforwardly implement a DDC algorithm. this is not necessarily always the case.. By these means. one containing the controller program and one containing a program representing the plant model. the two computers must then be connected together via appropriate interfacing.i Ll Computer 2 Fig.
In the particular framework of this discussion however. for the time being at least. which some people have. relating to a welldefined subject area. Such a decision making mechanism is called an expert system. 3. it would be wrong to forget about the human expert. particularly because of the decisional logic which can be introduced via software. The knowledge engineer’sjob is to interrogate experts and to make use of case studies in order to obtain detailed answers and solutions. Most cases will require occasional updating. 2. particularly in the event of * “Intelligence is the ability. retrieval and operation. This information must then be translated into a data set which is suitable for storage. indeed there may well be cases in which it is preferable. procedure and logic information must be extracted from one or more experts and stored efficiently in a data base. 9. Such a requirement makes the implementation of computerbased digital controllers amenable to artijicial intelligence techniques. stored in a data base. Once an expert system has been constructed.5. improving and gap filling by human experts. as well as relevant mathematical inequality tests. MIT). . with a fallibility equal to that of an expert. This decision making mechanism can.Computer control 285 nonlinearities.’ Relevant application areas are to be found in the fields of robotics and intelligent sensors. artificial intelligence is referred to in the sense of a bank of knowledge. to make other people regard them as being intelligent” (Kevin Warwick).g. A heuristic fact storage/retrieval procedure. and more directly in software development for automatic programming and the expression of natural language. With a desire to operate a digital controller successfully within an overall plant control scheme. if the plant output goes above a certain level then an alternative backup controller must be switched in. suitable strategy. Relationships between the facts. depicted in Fig. The person who constructs an expert system in such a way is referred to as a knowledge engineer. This means that in order to construct an expert system. to retain a role for the human expert. in its simplest form. An expert system is made up of three main parts: 1. The performance of an expert system is dependent on the size and quality of the data base. The knowledge of relevant facts. In its widest sense artificial intelligence can be defined as “the science of making machines do things that require intelligence if done by people” (Marvin Minsky. be regarded merely as software jacketing around the digital control algorithm. logical decisions must be made as to the action taken when a particular operational problem occurs. The main purpose of this stored knowledge is to produce a response to certain stimuli which exactly mimics that which would be exhibited by a human expert. But to be truly effective the logic should be based on heuristics resulting from both experience and common sense. e. spurious noise signals and manual/automatic switching must somehow also be taken into account. transients.
286 Computing methods changes in operating function or ambient conditions. (b) can apply common sense. artificial experts also have distinct advantages in that they (a) are consistent. Experts Expert Fig. 9. (c) can take a broad view.5 Construction of an expert system User 4 + Inference engine Fig. 9.6 Expert system organization . (d) can adapt to different circumstances and (e) can use sensorial observations. Human experts retain several advantages over artificial (computerbased) experts in that they (a) are creative. (d) are permanent and (e) are easily documentable. However. (c) are affordable. (b) can be easily transferred.
The different forms are expanded in Example 9. A reverse path also exists however. In the forward path one commences with a trigger. The inference engine. 9.6. By these means one can locate. data can be compared with “known” values in order for a justifiable conclusion to be drawn. Deduction: i f A then B. The most commonly encountered expert system rule is the “if. so expect B if A.7 relates. then” structure. 9. Subsequently a cause is pointed to which means that a certain fault is apparent. a hypothesis is then suggested. < fault > confirmed by <test > . and the knowledgeldata base. The knowledge base/data base level. and this commences with a test which points to a number of faults.1 Consider the following example of an expert system in operation.3. 9. 2 .g. In terms of the DDC algorithm.3. The fault can then be confirmed by a test and/or repaired by an appropriate action. If < fault > then do <action > . 2. as shown in Fig.Computer control 287 An expert system is usually organized in terms of two levels. 2. In the forward path the four rules applied are: 1. The purpose of this is such that once a hypothesis has been selected. e. The inference engine provides an interface between the user (by user is meant the program which “calls” the expert system). and be prepared for. where A then B. a reason for the alarm is given. 3. This conclusion can then be acted upon. <hypothesis > caused by < fault > . again with reference to the knowledge base (see Fig. an expert system serves more in a supervisoqdrnonitoring role and also modifies controller action.6). e.1 Example 9.. there is support for A.g. 1. in what could be a drastic way. and the action which needs to be taken . although this can take on one of several forms: 1. Induction: if for a large number of cases. Fig. 3. 4. . when certain triggering conditions apply. a particular trigger. Abduction: ifB. one of which is confirmed. an alarm condition. The usefulness of an expert system is dependent on the depth with which its knowledge base details the triggers which can possibly occur. If < trigger > then suggest <hypothesis > . when we have if A then B.
.g. The basis for fuzzy logic is the idea of a fuzzy set which contains all the possible values of a particular variable.p. So if F is the set and f is an individual value in F.p. say 50 m. F = {f}. (Hypothesis1 Suggestion Trigger Confirm Test Fig. 55 m. e.p. All of these factors are in fact very important in general fuzzy controller design. although this is dependent on (a) the car. and (c) the problem domain. which might have a value of 0.h could then be assigned a value.h is either a fast speed or it is not! However a fuzzy set has a grade of membership associated with it such that each value in a fuzzy set.p. which means that it is not such a fast speed as 100 m. on or off. say 0.As an example consider the speed of a car with F being the set of all fast speeds for that car.288 Computing methods when each trigger occurs.p.1].h.h.3. or a trigger occurs and the expert system suggests an incorrect action in response.e. 55 m.83. F. Normal logic would simply state whether a variable was or was not a member of a set. i.12. so F is the set of all possible values off.1 Fault Repair Action 9. i.3. is associated with a real number in the range [0.5 Fuzzy logic controllers Digital computers are based on standard logic.h to 120 m. The variablefcould then range from.e. If a trigger occurs for which the expert system has no information. whereas it is often the case when controlling a realworld system that we would like the temperature to be “fairly warm” or the height of liquid to be in the middle. then the operative quality of that system should be questioned.7 Flow diagram for Example 9. (b) the person involved. 1 or 0. These concepts led to the development of fuzzy logic as an alternative computer based approach to control. 9.
9.10 Quantised membership function . 9.8 Triangular membership function Grade of Membership I Values Fig.Computer control Grade of Membership 289 Values Fig.9 Continuous membership function Grade of Membership Values Fig. 9.
A fuzzy control procedure. is perhaps the most commonly encountered version. 9. In order to achieve fuzzification.8. This method has the advantage that for online implementation it is very fast. all of the shaded area being considered as within the computer. with the resultant fuzzy controller output being converted back to a realworld value via a process termed “Defuzzification” .8 to 9. representing the grades of membership in a fuzzy set.12 shows an example range of functions used to convert a measured speed value into its fuzzy equivalent. Using this figure.g. e. 15 cm.h. can then be applied. The quantised membership function of Fig. In order to apply a fuzzy controller therefore. 9. Variables associated with realworld control situations are measurable values.10.10 is relatively easy to define by splitting the overall range of values considered into a number of subranges. gF = 0) .2. are fuzzy values. 9. It does have the advantage though of including a wide range of values with a low grade of membership.p. The triangular membership function. shown in Fig. whereas x . gM = 0. 63 m.p.9. as will shortly be described. medium temperature. a number of functions describing the grade of membership for a particular value are defined. and x. this process is called “Fuzzification”.1 1 are thus realworld. System Fuzzy Controller  I I Fig. however it can be extremely difficult to decide on the exact shape of the function and the online calculation of a values’ grade of membership can be quite time consuming.290 Computing methods Dependent on the problem encountered. 9. The error and control input in Fig. a realworld speed of 40 m. shown in Fig. where it must be remembered that in terms of a computer implementation the fuzzy controller input will be a sampled signal via an analogtodigital converter. this being completely specified by only 3 values. 9. Figure 9. etc. so different types of functions. Three typical functions are shown in Figs.67.1 I .h would become (gL = 0. often the function being realised simply by means of a lookup table. 9. 9. The continuous membership function. whereas the controller output will be fed through a digitaltoanalog converter. such that the measured value in question is merely converted into its fuzzy equivalent. are required.11 Fuzzy logic controller The overall implementation of a fuzzy logic controller is shown in Fig. gives a more direct mathematical relationship. whereas fuzzy systems are based on ideas such as fast speed. it is necessary to firstly convert the realworld value of interest into its fuzzy equivalent. measurable values.
(e). Example 9.2 Consider a rulebased controller described by five rules (a) . O(max) Speei (m. gM = 0. (b) IF e is Medium THEN u is Medium. gF = 0) whereas a realworld speed of 100 m. and gF = grade of membership offast speeds.4. The hypothesis part then assigns a qualitative value to the controller output in the form of an output reference set.h. gM = 0. by example. if <trigger> then suggest <hypothesis > .3. Once a signal has been fuzzified it can be operated on by the fuzzy controller.e. would become (gL = 0.12 Fuzzification It is worth pointing out that normally a number of signals are actually passed through the fuzzification exercise. 9. where e = error.) Fig.p. The rulebased system takes on the form described in Section 9. Clearly the extent of overlap between functions and the range of values selected is very much problem dependent. gM = grade of membership o medium f speeds. (a) IF e is Large AND 6e is Large THEN 6u is Large. gF = 0.3.h. The trigger part of this mechanism is based on fuzzy logic operators.2 is used to indicate. .92. with different membership functions apparent for each signal. how a rulebased controller is applied. i.p.6).3. i.2. and (b) Relational systems. So. as well as speed we could have acceleration and distance provided as signals. Also a realworld speed of 55 m.Computer control 29 1 where gL = grade of membership of low speeds.p. the two most commonly encountered forms of which are (a) Rulebased control. derivative and integral forms. and when grouped together these can realise quite a complicated expression. u = control input.h would become (gL = 0.e. Example 9.
14 Fuzzy control output reference set (6u) .2 g6eS = 0.1) = 0. Assume.4 g6eL = 0. which results in the output reference sets shown in Figs.6 geL = 0.5) = 0. (e) IF e is Small AND 6e is Small THEN u is Small.6 that u is Medium.292 Computing methods (c) IF e is Medium THEN 6u is Medium. that fuzzification gives: geS = 0.13 Fuzzy control output reference set (u) Grade of Membership 0 5 10 15 20 input derivative (bu) Fig. 0.3 geM = 0.4 that 6u is Medium.2. 9. ye = min (0.2 that Y~ = 0. 6u is Large.6 that 6u is Medium. 0.13 and 9. for a particular example. 9.1 g6eM = 0.14 A / / \ \ / / 4 / I I I 0 5 " 10 \ 15 20 control input (u) Fig.5 The overall degrees of truth obtained from the five rules are then: (a) (b) (c) (d) (e) ya = min (0.1 that u is Small. 9. yd = 0. (d) IF 6e is Medium THEN 6u is Medium.3. yc = 0.
such that g. and these can be summarized as follows: If &(x) and &(x) indicate the grade of membership of a variable x in functions A and B. gA’(x) is the complement of gA(x). containing degrees of truth associations.4.3. Rulebased fuzzy controllers suffer in the same way as expert systems in general. An alternative method for fuzzy controller implementation is the use of relational systems.’(x) = 1 . gB(x). as hypothesis support. the relational array. particularly in terms of grades of membership and degrees of truth.Computer control 293 Two key points arise from Example 9.3. Here qualitative inputs are related to qualitative outputs by means of an array. &(x) is a subset of &(x) if &(x) 5 gB(x). in that obtaining the rules can cause distinct problems. the higher of these values should be employed.e. The latter point. gB(x)) this is the OR function. this is the AND function. gc(x) is the intersection of gA(x)and gB(x)when gdx) = min(gA(x). This is the case in Example 9. It can also be a very expensive procedure if system down time is involved.2. for all x .2 for the hypothesis that 6u is Medium which receives yc = 0. a very similar measure to the grade of membership. The method is also not tied down by the particular “experts” involved. not simply the AND statement found in Example 9.gA(x). gc(x) is the union of &(X) and &(x) when g () C’ = max(gA(x). A big advantage of such an approach is that the relational array can be obtained through experimentation. as can be the case for rulebased fuzzy controllers.3. is merely giving an indication of how true a fact is dependent on how true the rule’s trigger is. then A is empty if &(x) = 0. i. in which case the results can be quite inaccurate and certainly dependent on which expert has been asked.2. for all x . the degree of truth. One further point on rulebased fuzzy controllers is that if. A number of fuzzy logic operators exist.6 and yd = 0. firstly the AND operation carried out within the trigger part of the rules and secondly the resultant degree of truth of the hypothesis part. two degrees of truth are realised for the same fact. . particularly if “experts” must be monitored or interviewed. It is. therefore. finding input/output signal pairs. such that yc overrides yd in this case. Recent research in this area has therefore concentrated on selforganizing fuzzy controllers in which the rules modify themselves.
[max(min(O. 0. e = array of grades of membership of fuzzy controller input.3. Each reference set must however be .7.. Once an input is provided.(j) = grade of membership ofjth input. 0. gR(i.6))l = [0.294 Computing methods The basic definition of a relational equation is: u = R o e (9. 9. the relational array is given by: 0. g.3 0.6 [yYz)] .min(0.7 [0. 9. 0.13. For a particular fuzzy controller.min(0. ( i .2).3] The final element of a fuzzy logic controller is that of defuzzification. the relational array is employed to obtain: Y"(i) = max [ m W .3. g.6)) 0. j ) .2) The concept of the relational array is perhaps best shown by means of an example. this being Example 9. The output from either a rulebased fuzzy controller or a relational array. is in terms of a number of reference sets.l.1 0 3 . (9. as shown in Figs.2).3. R = fuzzy relational array.3. 0.j) = relational array value relatingjth input to ith output set.14.3 Example 9 3 3 .1 = Given that the input to the fuzzy controller is Find the degree of truth of the output set: Yu(l) =  max(min(0.1) in which u = array of degrees of truth of fuzzy controller output.(j))l i where yu(i) = degree of truth in ith output set.3.
as shown in Example 9.3) Cl yi i= in which yi is the degree of truth of set i and ui is the characteristic value of set i. (b) Centre of Area Method.3. If several maximum points exist then the mean value of these points is found. Although this solves the problem of large defuzzified signal swings.3.15.4 indicates a specific case. For the fuzzy mean method the controller output signal value is thus found from: n (9.75 2 = 12. in the reference set. A weighted mean of the characteristic values is then taken.25 + 8. (c) Fuzzy Mean Method. This method involves a bisection of the area under the overall reference set function. Three methods of defuzzification are most commonly encountered: (a) Mean of Maxima Method.25 Mean of Maxima = 16. the defuzzified value for this set is found to be: (a) Mean of Maxima: Maximum = 0.3. it does require a relatively high computational effort. subject to such a maximum existing.3.5.5 . from value = 8. in the reafworld control input. is employed as the defuzzified signal.4.4 Consider the overall reference set shown in Fig. and is achieved by assigning each individual reference set a characteristic value. 9. as a crisp value. This is very much of a compromise between the previous two methods. often this coincides with the largest reference set value. Example 9.75 to value = 16. This is a fairly simple method in which a single maximum value. One negative point with this method is that small changes in fuzzy controller input can cause large changes in the defuzzified signal. Example 9.Computer control 295 converted into a single figure to be used.
5 these both being midrange points.5 Characteristic value of function (ii) = 12. The output value is then found from: value = 0.5 + 0.5 0. However.5 * 12.296 Computing methods (b) Centre of Area: Centre of area = 11 (see Problem 9.3 + 0.15 Reference set for defuzzification To conclude this section on fuzzy logic controllers. Choice of reference sets is left up to the user/operator and yet such a choice is extremely important in that a poor choice will result in poor control. 7.3 . 9.1). (c) Fuzzy Mean Method: Choose: Characteristic value of function (i) = 7. whilst there is no defined way of making a good choice. .5 = Grade of Membership A A 0. it is worth pointing out that although such controllers are now quite popular they are not so “exact” in their calculation or in the control action produced. with most problems it is usually possible to select reference sets which will provide reasonable control.3 / \ 0 5 1 0 15 20 value Fig.
These can range from slight numerical inaccuracies in calculations to loss of control due to saturation effects. A natural solution to this problem is to employ parallel processing techniques both in the form of specialpurpose hardware and concurrent programming. polynomial coefficients. numerous operational problems will most likely appear. Finally. Numerical problems are caused principally by the finite word length. A certain amount can.3. this information should be fed back to the controller. then the controller will continually integrate the error between actual and desired output signals. All of these factors combined could provide an impossible schedule between sampling instants. in terms of operational difficulties. not only of the computer data bus and memory. If such a signal occurs. Actuators are usually fairly limited in both their range and slew speed and sometimes can contain a dead zone. A different aspect. when . Their cause can be seen as either due to rounding and/or truncating the analog representation in order to convert it to its nearest digital form.g. When a controller contains integral action. usually be done within the control algorithm itself. in order to aid the parallel implementation. detailed in the previous section. when the computer is controlling a realworld plant. hence leading to further errors when calculations take place. is the problem caused by the time taken for calculations to take place. by suitable modifications. is highlighted where appropriate. although very often these problems are due primarily to the signal actuators rather than the plant itself. possibly as an expert system trigger. Assume for a moment that the output signal actuator has saturated. then the resultant error will be continually integrated producing very large. the PID controllers discussed in Chapter 10 are an example. a small control signal change will not produce any corresponding actuator change. will themselves be merely digital representations stored with finite precision. e. as soon as a calculation takes place in which those numbers are concerned so further errors will be introduced because of the finite precision of the calculations.6 Operational problems 297 When a computerbased digital controller is implemented on a plant. Quantization errors are caused by the representation of an analog signal by means of a digital base in terms of a finite number of digits. unless sequencing and priority conditions are applied. the controller parameters. i. this is called integrator saturation (or reset windup). however. It may well be necessary for one computer to service several control loops as well as operating a logical expert system testing sequence. but also of the analogtodigital and digitaltoanalog conversion mechanisms.e. this usually presents several operational problems in its own right. With regard to the plant under control. increasing. controller values. on top of which the display of information is required and also communication with the operator. One possible solution to this is to employ only the integrating effect of the controller. Subsequently when two analog numbers are stored digitally.Computer control 9. A range of the most common problems are discussed here and their relationship with an expert system structure solution.
9. or alternatively the presentation of a further menu may occur from which another selection must be made. Indeed such spikes can also be witnessed when controller coefficients are manipulated online. The final operational problem considered here is the effect of manuallautomatic transfer. Unfortunately a solution to this problem is not simple.298 . because of the many apparent advantages. if a command driven package is employed. in order to give a concise but comprehensive outline of the relevant stage such that users’ problems can be dealt with. confusing detail. and involves the use of a statespace system description in order to match not only the manual and automatic control signals. Programming was based around a block diagram formulation combined with a selection of fixed operations. A good set of explanation facilities at all levels of operation is therefore of great importance within the package itself. whether they are run on a standalone personal computer or a multiuser computer with timesharing facilities. Such a choice of integrator on/off can be made within an expert system hierarchy. More flexibility can be obtained with a computerbased system simulation. on condition that the menu is not cluttered with too much irrelevant. The user is then able to interact directly online with the package such that any modification made is immediately reported on by the computer. whereby the user can take appropriate necessary action.4 Computeraided design As the availability of computers became more widespread in the 1960s. When switching over from manual to automatic control of a plant. In this case the user selects from a list of available commands. The most successful simulation schemes are in fact those implemented as an interactive package. For all types of system simulation work. at once.Computing methods the error is small.the user can then choose one of the options listed in the menu. and this is certainly possible with a computerbased system. The interactive capabilities can be presented to the user in one of a number of ways. it is important that model parameters can be altered easily. it does necessitate a higher level of experience on the user’s part. Once a selection has been made it may then be the case that the computer will take immediate action.e. so digital system simulations replaced analog simulations. Often learning to work with a particular package can be a harrowing and timeconsuming period. although modeling in terms of mathematical equations was soon introduced. i. One possibility is the presentation of a menu at each particular stage of simulation . and although this leads to much greater efficiency in operation. The menu approach is extremely useful for inexperienced or nontechnical end users. but also the state vectors at the instant of transfer. The help command is invariably . transient signal spikes can occur. within defined minimum and maximum bounds.
model form. it is not merely simulation work for which a computer is useful. 9. in which various input signals are applied to the plant and the corresponding output signals recorded. ranging from specific numerical values dredged from the depths of an algorithm.Computeraided design 299 employed for this purpose. Although for continuoustime designs. can be carried out by means of an appropriate package.1 Modeling and identification Of primary importance in the study of system performance and controller design is the construction of a mathematical model which adequately describes the behaviour of that system. It is in fact common practice to use the multiexperiment approach. Making use of the collected input/output data. both continuoustime and discretetime. In general. Presentation of simulation package results is usually available in several modes. However. it is usually the case that a time saving is obtained. data analysis and model testing. Although a plant model can by found be investigating the internal plant structure. to the help facility. in order firstly to choose a suitable model structure and secondly to select . something which was sadly neglected in earlier packages and which can still be variable in quality when a new package is introduced (until initial problems with the package have been found and removed or avoided). or graph. and/or introduction. this can mean that certain discrete approximations are necessary. with due consideration being given to problems such as noisy signals. it is important that these input/output data pairs are taken within the normal operating range of the system. to vary with time as different stages of the simulation are reached.4. backup documentation is also an important feature. System identification is the procedure of modeling systems mathematically. statistical tests are applied to a set of plant inputoutput data. and is generally also available for menudriven packages. in order to estimate coefficients within a plant model is termed system identijication. However. because of the complexity and variability in model structure which can be represented. particularly when interactive capabilities are available. almost all of the control system studies and controller design procedures described in this book.g. it is usually the case that a complete picture cannot be obtained in this way. e. due to unknown or immeasurable factors. it is unfortunately the case that the data obtained from a realworld system will invariably be corrupted by noise emanating from random fluctuations both in the plant itself and in the measuring devices employed. Due to the numerical and graphical properties associated with a computerbased facility. Interactive graphical facilities allow for a diagram. to colorful graphic displays of timevarying signals. a Nyquist diagram. Computerbased procedures are more of a necessity nowadays where system modeling is concerned. To be representative of the system under consideration. despite any associated noise. As an extension. thus presenting an uptodate picture of the study being carried out.
If the structure of the model is wrong. hence the square of the prediction error is in general used. N. Essentially. Many procedures exist for the minimization of the sum of squares function. the most commonly encountered method being that of linear least squares. of the prediction error signal squared are summed. The vast majority of identification procedures are based around the idea of making the magnitude of the prediction error as near to zero as possible. because of the disturbance effects. see Fig. This process has the advantage that noise effects on the model are reduced considerably. Plant input 4) * error Model Predicted Prediction L_) Fig. and this sum of squares is employed as a function to be minimized by the selection of model coefficients. Conversely. then no matter how much time is spent finding values for the coefficients. a number of samples. Such techniques are . however. the error produced is dependent on how close the model is to representing the actual plant. a large prediction error will occur. If the outputs from the plant and model are then compared. are not necessarily of use because variations in some of the coefficients may well have less of an effect on the prediction error than will the noise signal. i. As many computer packages are available for system identification.16 Prediction error generation Prediction error models are perhaps the most useful for identification purposes. with many coefficients. not enough coefficients are employed. complex models.300 Computing methods appropriate coefficients in a model with the chosen structure. the main problem for the user is often achieving sufficient realworld data for the package. 9.16. 9. whereby the same input signal is applied to the plant itself as is applied to a mathematical model of the plant.e.
Computeraided design 30 1 however.2 Controller design System evaluation and design can be split up into roughly four areas: (a) signal processing. a controller design package is required to be not only operative in its own right. So. which is merely required to operate on a given data set in order to produce a representative model.17 Requirements for a CAD package Controller design t package then includes facilities for all four of these areas. ideally suited to the computer.4. as shown in Fig. (c) controller design and (d) system simulation. 9. 9. although many schemes merely concentrate on one or two of the topics. but also to tit in with identification and simulation routines.17. It is therefore necessary for . 9. A complete computeraided control system design Real world plant data Signal A  Modeling and identification to Simulation Presentation of results Fig. (b) modeling and identification.
302
Computing methods
the model structures, employed as a basis for the controller design stage, to be of a compatible type with those employed for identification procedures. Hence, as part of the input to the controller design process, a plant model must be entered, whether this model is from a known, reliable source, or whether it is directly obtained as the output from an identification procedure. Note: As shown in the diagram, it might be the case that data output from the signal processing stage is to be combined with a known plant model, for controller design purposes. Due to the large number of possible controller actions, the controller design stage can attempt to offer an extremely wide variety of choice, although very often only a few different techniques are available (indeed in many cases only one scheme is on offer). It is usually the case that for any particular controller technique selected, e.g. assume that a PID controller (described in Chapter 10) has been chosen, then the coefficients of that controller can be found in one of a number of possible ways, e.g. pole placement or ZieglerNichols. Also, once the controller coefficients have been finalized, jacketing and compensation terms must be sorted out. Hence, the effect of time delays, nonlinear behaviour, disturbances and loop interactions can all be dealt with in this stage of the design procedure. Selectable actuator gains and interfacing specifications can also be decided upon in the control design stage, some by the user and some within the package program, in order that the overall control system response is manipulated to satisfy realworld engineering constraints. Once the controller has been finalized, it can either be employed within a simulation study, from which useful operational details, such as performance and problem areas, can be ascertained, or coefficient information can be output from the CAD package such that the controller can be used directly in an online, most likely computerbased, control system.
9.5
Summary
In this chapter it has been shown how computers can be used as efficient tools in many aspects of control systems study and design. Three important features have been looked at in detail, the first of these being the use of computers, whether analog or digital, to simulate the behaviour of systems. This is an extremely important issue in that largescale plant, when represented by a computer model, can be investigated and modified by means of a flexible benchtop exercise. It also allows for the study of certain plants, which for cost and/or security reasons would normally be unavailable for analysis. The second aspect of computer usage considered in this chapter was the employment of computerbased control systems, in which the computer itself makes up part of the overall system. Present trends are for more intelligent, faster controllers with the ability to adapt to changes in the environment. Hence, as well as discussing operational details
Computeraided design
303
of realtime online computer control systems, intelligent systems based on a knowledge store employed within an expert system framework were also considered, Finally, various aspects of computeraided control system design were outlined. Initially the use of CAD for system simulation work was highlighted, although system modeling and controller design packages were also covered. The diversity of CAD tools available to the control engineer makes it impossible to do other than give a flavour of the techniques on offer; the prospective user should however beware of the number of userunfriendly packages which require specialist knowledge before they can be operated.
Problems
9.1 9.2 9.3
9.4
Verify that in Example 9.3.4 the centre of area method realises a value of 1 1 . Find defuzzified values for the control input, u, in Fig. 9.13, using the three methods described. Find defuzzified values for the change in control input, 6u, in Fig. 9.14, using the three methods described. A system has been modelled using a relational equation with the output and input described by two fuzzy reference sets. The resultant relational array produced by identification is:
9.5
9.6
Given that the input to the system, described in terms of its reference sets, is e = [0.8, 0.21, what is the output? A fuzzy logic controller is to be constructed for the control of an electric motor. Outline an initial design for such a controller and give examples of rules to be applied, if a rulebased system is employed. For a fuzzy controller based on a relational model, the controller output, in terms of degrees of truth of output sets, is given for two different inputs as:
If the array of degrees of truth of input sets is:
(a) e = [0.4, 0.21, and (b) e = [0.3, 0.41
suggest an appropriate fuzzy relational array.
304 9.7
Computing methods
9.8
0.1 0.2 0.3
R
=
0.6 0.5 0.4 0.7 0.2 0.4
9.9
What advantages are apparent if a computer contrc, scheme is employed in place of a straightforward digital (noncomputer) control implementation? What are the disadvantages? 9.10 Consider the statespace system matrix
2.1 6.1 A =(Lo 2.J
which is nonsingular, and therefore invertible. However, if the matrix elements are measured values which are stored within computer memory, each element will have errors, e.g. quantization, rounding, associated with it. For the example given, show how these errors can result in a singular, noninvertible matrix. 9.11 In the construction of an expert system for a home heating system, what important rules should be included? 9.12 In the construction of an expert system for an automated vehicle driver, with regard merely to vehicle guidance along a well defined roadway, what factors must be accounted for? Having constructed such an expert system, are there any occurrences that have not been taken into account but which a human expert could easily cope with, e.g. adverse weather conditions? 9.13 An important aspect of many complex control systems is to ensure that even if one system element fails, the overall system remains in operation. Apart from alarm sounding, how can a computer control scheme provide fault tolerant control?
Further reading Further reading
305
Aleksander, I. and Morton, H., ‘Artificial intelligence: an engineering perspective’, Proc. IEE, 134, Pt.D, 1987, pp. 21823. Artwick, B., Microcomputer interfacing, PrenticeHall, 1980. Astrom, K. and Wittenmark, B., Computer controlled systems, PrenticeHall, 1984. Banks, S. P., Control systems engineering, PrenticeHall Int., 1986. Becker, R. G., Heunis, A. J. and Mayne, D. Q.,‘Computer aided design of control systems via optimisation’, Proc IEE, 126. 1979, pp. 57384. Bennett, S. and Linkens, D. A. (eds.), Real time computer control, Peter Peregrinus Ltd, 1984. Boullart, L., ‘An introduction to artificial intelligence and expert systems’, Proc. Int. Seminar on Expert Systems and Artijicial Intelligence in Industry, Antwerp, 1986. Bratley, P., Fox, B. L. and Schrage, L. E., A guide to simulation, SpringerVerlag. 1983. Brdys, M. A. and Malinowski, K., Computer aided control systemdesign, World Scientific Publishing Co., 1993. Clayton, G. B., Operational amplifiers, Butterworths, 1979. Driankov, D., Hellendoom, H. and Reinfrank, M., An introduction tofuzzy control, SpringerVerlag, 1993. Efstathiou, J., ‘The chocolate biscuit factory’, Journal A , 27, No. 2, 1986. Fisher, D. G. and Seborg, D. E., Multivariable computer control, NorthHolland, 1976. Gordon, G., System simulation, PrenticeHall, 1978. Hartley, M. G. (ed.), Digital simulation methods, Peter Peregrinus Ltd, 1975. Jacoby, S. L. S. and Kowalik, J. S . , Mathematical modelling with computers, PrenticeHall, 1980. Millman, J. and Halkias, C. C., Integrated electronics, McGrawHill, Kogakusha, 1972. Minsky, M., The psychology of computer vision, McGrawHill, New York, 1975, pp. 21 177. Ng, T. S., ‘An expert system for shape diagnosis in a plate mill’, IEEE Trans. on Industry Applications, pp. 10571064, 1990. Norton, J. P., An introduction to identification, Academic Press, 1986. Pedrycz, W., Fuzzy control andfuzzy systems, 2nd ed., Research Studies Press, 1992. Peel, D., ‘Adaptive control and artificial intelligence’, Journal A , 28, No. 3 , 1987, pp. 1437. Pressman, R. S. and Williams, J. E., Numerical control and computeraided manufacturing, John Wiley, 1977. Rosenbrock, H. H., Computer aided control system design, Academic Press, 1974. Sandoz, D. J., ‘CAD for the design and evaluation of industrial control systems’, Proc. IEE, 131, No. 4, 1984. Shoureshi, R. and De Silva, C.W. (eds.), Intelligent control systems, American Society o f Mechanical Engineers, 1992. Stout, T. M., ‘Economic justification of computer control systems’, Automatica, 9, 1973, pp. 919. Tzafestas, S. G. (ed.), Applied control, Marcel Dekker, 1993. Warwick, K. (ed.), Implementation o selftuning controllers, Peter Peregrinus Ltd, 1988. f Warwick, K. (ed.), Applied artificial intelligence, Peter Peregrinus Ltd, 1991. Warwick, K., Karny, M. and Halouskova, A. (eds.), Advanced methods in adaptive control for industrial applications, SpringerVerlag, 1991. Williams, T. J., ‘Computer control technology  past, present and probable future’, Trans. Inst. MC., 5, No. 1, 1983. Winston, P. H., Artificial intelligence, AddisonWesley, 1984. Zanker, P. M. and La Pensee, S., ‘Design of a computer controlled electrohydraulic servosystem’, Proc. IEE Conference on Control and its Applications, 1981, pp. 96101.
I0
Threeterm controllers
10.1
Introduction
When first attempting to control a particular plant, in all but the simplest cases, it is very soon appreciated that a controller which consists solely of a gain, o r amplification, value is not versatile enough to cope with even the most basic requirements made for overall system performance. O n the other hand there are a large number of elaborate controllers which satisfy a particular mathematical function, require high accuracy and may well take several days to calculate on the most expensive computers. Clearly there is room for a fairly straightforward controller type which although it is relatively simple and cheap t o operate, can also deal successfully and reliably with the essential control requirements. To be acceptable this controller must be able to cope adequately with the vast majority of plants to which it is applied and prove its worth through many years of service, often in a n environment for which it is not particularly designed and to d i c h it is not previously accustomed. Such a controller is the threeterm or P I D controller which has been for many years by far the best known and most widely used control structure for industrial applications. I t is applicable in either continuous o r discretetime modes and is the most likely form of control system that a practising engineer will encounter. In a sense, threeterm control is just one form of compensation which can be compared directly with the leadlag compensation described in Chapters 4 and 5 , o r with more specific schemes such as optimal, deadbeat or pole placement control. However, once the basic structure has been decided upon as consisting of threeterms, i.e. Proportional + Integral + Derivative (PID), the selection of gains associated with each of these terms is still to be found, and although in many cases methods such as root locus, Bode o r Nyquist could be used to meet the desired control objective, it may well be that the controller parameters must be chosen to satisfy a n optimal o r pole placement criterion. It must be pointed out that in many practical situations an exact mathematical description of the plant is not available and this may result in the P + I + D gains being modified online until satisfactory performance is achieved. In this chapter the advantages and disadvantages of the threeterm control structure are investigated in a bottomup approach. Firstly, a n attempt is made to apply simply a proportional controller, and it is shown how this can be seen as positive as far as
Proportional control
307
closedloop stability is concerned, but very little help in terms of steadystate output following or transient response tailoring. It is then shown how the inclusion of both integral and derivative action can deal with these requirements, without the need for more elaborate schemes. Indeed, if an elaborate controller is to be used in practice, it should first be shown that the controller performance is superior to that of a threeterm controller in at least one important aspect.
10.2
Proportional control
There are many ways in .which a controller can be connected to a system in order to provide the required compensation. One common element between the different approaches is that the error between the actual system output and the desired reference value for the output needs to be available for manipulation. Perhaps the simplest and most commonly encountered method is that of cascade or series compensation, and this is shown in Fig. 10.1. It is apparent in the cascade controller that the error e ( t ) ,between the actual output, y ( t ) , and the desired reference, u ( t ) , is obtained directly and used as the input to the comgensator/controller, D, whose output, u ( t ) , is then the input to the system/plant to be controlled. This chapter is concerned primarily with the nature of D, some of the forms it can take, and its effect on the system, G. As a slight aside, an alternative but nevertheless popular form of compensation connection to that proposed in Fig. 10.1 is the minorloop feedback controller shown in Fig. 10.2, in which the unity negative feedback loop is preserved. The controller,
I
Fig. 10. I
Controller
Plant
I
Cascade compensation
Fig. 10.2 Minorloop feedback compensation
308
Threeterm controllers
however, loses a lot of the intuitive appeal provided by the cascade connection, and will only be of indirect relevance here, as the controller is no longer a simple operator on the difference between what we want and what we have. With reference to the cascade controller of Fig. 10.1, a fairly obvious and primitive attempt at control of the system would be to take into account the error at time t, e ( t ) , then to provide a large corrective action when e ( t ) is large and a small corrective action when e ( t ) is small. This is a nice and simple procedure because it only involves a controller D which consists entirely of a multiplying constant.
10.2.1 Continuoustime proportional controller
By connecting up the proportional controller in its cascade form, the controller input is the error, e ( t ) . Its output, u ( t ) , is made up of a single term, which is simply the signal e ( t ) multiplied by the controller gain, K,, as in equation (10.2.1)
u ( t )= K , e ( t )
(10.2.1)
The closedloop equation relating u ( t ) to y ( t ) is, when the control input of (10.2.1) is applied, given by
Y ( t )= ___ u ( t ) GKp 1 + GKp
(10.2.2)
or in terms of its sdomain representation
Y ( s )=
1
+ G(s)K,
GWKP
v(s)
(10.2.3)
So a signal has been applied in an attempt to control the system, such that the larger the output error, the larger the control signal  but does this simple form of control provide a good result at the end of the day, i.e. does the output signal, y ( t ) , follow the input signal, u(t)? Consider firstly just what happens in the steadystate when all transients have settled down, i.e. when f + 00. It is obvious that for the simplest of systems, when G is just a straightforward gain value, e.g. G = K , then
Y ( 0= KKp 1 + KK,
~
U(f)
( 10.2.4)
Therefore as long as KK, % 1, whatever type of input signal is applied it follows that the output will be just about equal to the input. In fact it is fairly irrelevant just how large K is, we can ‘in theory’ make K, as large as we like, i.e. use a very ‘high gain’ indeed, in order to achieve the desired results. So as well as being a simple control action this scheme has other hidden benefits in that if the system is not quite how we had thought, i.e. if the actual system gain K is a little bit different from the designed value, possibly due to ageing or a modification, then it does not matter much as long as K p is a very high value.
i. as an example consider a system of the form: (10. consider (10. such that in order to obtain KK.5).2. then if the time delay 7 = 0 for simplicity. If a step input of magnitude 3 units is applied to such a system (10. a period of time elapses before any event or change at the output will be witnessed in response to an event or change at the input.2.7) or in terms of the characteristic equation: .2.5). if it is not possible to use a high gain K p then an offset between desired and actual output values could occur in the steadystate.e.Proportional control 309 The ‘high gain’ results obtained for a simple system in fact extend to much more complicated cases. So. In practical terms therefore.2 and Kp = 10.001. + (10. then the steadystate output signal. As an example. but what are the drawbacks? A first point to take into account is of a practical nature in that the gain K p will most likely be a simple amplifier . is found from: lim y ( t ) = lim s Y ( s ) = s /m s0 1 + C(s)K. W)K. which we wish to be equal to 3 units. the denominator is given by: (71s + ~)(TZS1) + KK.s7) is a pure time delay. The results have thus far pointed to the benefits of proportional feedback.5) in which the term exp( . K p is not a suitable control selection to be used for steadystate following of a reference signal.2. For many systems we might have that K = 0.2.3) and the example system (10.2.2. s lso but lim G ( s )= K s0 so 10 lim y(t)= 1 3 KK. + KK. The system G(s) is typical of the transfer function obtained for a heating system with a measured temperature output value.4) when v ( t ) = 3 units. A second point to note is the effect that the choice of K p has on the transient response and closedloop stability of the system.it may therefore not be possible to obtain the high gain required before saturation occurs. s 1 it would be required that K p > lo6 which is quite a tall order for a simple amplifier.6) with K = 0.6) which gives a similar result to (10.2. it then follows that the output would be 2 units in the steadystate and not the desired 3 units. (10. With reference to the closedloop equation (10.
> . Hence on condition that 71. as tends to zero the response to a step input becomes more oscillatory and conversely as { increases the response becomes more heavily damped producing less overshoot. it is required that: KKp> . In summary. we appear to fall between two stools in that both steadystate and transient conditions need to be catered for. required for steadystate following. D ( t ) .9) and for a given system T~ and 72 will be constant.8). the frequency of oscillations in response to a step input will increase.8) For any given natural frequency. only. In particular a high gain K.72 > 0 the closedloop system will be stable as long as KK. the oscillations would go on for a long time.7) with (10. In Section 10.2. The increase in anwill however result in a decrease in { which means that less damping of the oscillations will occur. The digital version of a proportional controller will however be considered briefly before going on to Section 10. This means that an increase in controller gain K. connected up in the mode depicted in Fig. by employing only a proportional control scheme.1.2 Discretetime proportional controller With the controller.2. this results in very poor transient control and steadystate conditions will only actually be reached in the distant future.3. and these requirements cannot both be satisfied with K. with an extra design feature employed to cope with the steadystate requirements. would result in a very low value for j and therefore the closedloop system would perform with oscillations which occurred more frequently and were only slightly damped. we have that: (10.3.1 which would certainly be satisfied for a high gain system with KK. will increase a n r i.e.3 10 Threeterm controllers For system stability it is easy to see that as all coefficients of the denominator polynomial must be positive. Section 3.3.3 a slightly more complicated controller will be investigated. Comparing (10. But what of the transient performance of the system? This is best seen by making reference to the transient response characteristics shown in Chapter 3. i.2.. 8. in which the secondorder denominator is considered to be of the form: s2 + 2 C W " S + u: (10.6 in which G ( z ) is the discretetime version of the system to be controlled assuming a zeroorder hold is .2.2. 10.e. If Kp is chosen large to provide approximate steadystate following. wn. > 1.
So we have once again that u ( t )= K.)2 + 0.3) =. can be found by setting z to unity in G ( z ) such that by means of (10.5(1 .10): noting that D(1) = K p = D ( z ) .6) G(l) = and hence (1 .~(Z 0.Proportional control present.11) which is identical to u ( k )= D ( z ) e ( k ) (1 0. However this will only be true if Kp is large or. Firstly let us consider steadystate conditions which. It is left as an exercise. in that now the error signal e ( t ) will contain a measured output signal y ( t ) which retains a steady value between samples.4. the proportional controller must be ‘high gain’ in order to remove.0. . and also how a ‘high gain’ K p will result in very lightly damped oscillations.2.2 For steadystate following this value is required to be 3 units.0..2. consider as an example the discretetime system transfer function O. any steadystate error.3) (1 0. then the closedloop equation is given by: 311 (10.6) ‘‘ ( ) = ( z.e(t) where now u(t)=D(z)e(t) (10.2. as for the continuoustime case.0 2 ( . Now 0.1 5.2)(1 + 0.13) which is assumed to include the effect of a zeroorder hold.1). The discrete version of a proportional controller is in fact just a discretized version of the continuoustime equation (10.2.12) In order to try and find out the effect of a particular gain selection K.2. see Problem 10. as much as possible. in terms of the unit circle.10) as shown in (8. in order to equal the stated reference input.1.2. to show how the selection of Kp will affect closedloop system stability. for a threeunit step input u ( f ) .6).
saturation occurring if the input signals are large enough. if the error is zero then so is the input signal.e. namely stability. not too much of a desirable feature when for instance the practical requirements of an input signal actuator are considered.e. Let us consider the steadystate following properties of a closedloop system with proportional feedback alone. 10.2. It was stated that in order to obtain no error between the desired and actual output values in terms of steadystate conditions. steadystate following and a desired transient response. then on a typical British day the outside temperature will be a lot colder than this. i. What is required therefore is a different type of control action which provides a constant control input when the error is zero and which changes its value when an error does exist. i. It was shown however that by using proportional control alone these requirements could not all be satisfied. whereas if the error is small and negative it will be multiplied by the large Kp to produce a large and negative control input signal. This means that at any instant. the actual room temperature will not be 20"C. further it will only have linear properties over a certain range. So something more than a simple proportional controller is necessary in order to deal with steadystate offset. It will most likely be the case that the actuator employed is not able to change instantaneously from a large negative value to a large positive one. So the requirement of our room heating controller is a constant heat input signal in order to obtain zero output error. So if a straightforward proportional controller is employed. using a proportional controller alone will not provide sufficient flexibility such that the operator may specify closedloop system performance. for the actual room temperature to be equal to the desired value of 20 " C . requiring heat to be input to the room. a steadystate error will exist. Because of heat losses from the room it will be necessary to continually input heat in order to keep the temperature at 20 "C. the operator can ensure.3 Proportional + integral control It was highlighted in the previous section that several fairly rudimentary requirements can be asked of a closedloop system. Obviously the result of this is very wild control input variations caused by only slight fluctuations in output error. for certain systems. if a constant input signal is necessary there must be an output error exhibited. But with a proportional controller alone.1) or. if the error is small and positive it will be multiplied by a large Kp to produce a large and positive control input signal. in order to provide a constant heat input. closedloop stability by an appropriate choice. then the proportional feedback gain must be a very high value. looked at in a different way. By this means. if the desired temperature value is altered to 25 "C it would be expected that the .3 12 Threeterm controllers The results obtained from a discretetime proportional controller correspond directly with those obtained from a continuoustime implementation and it is obvious that even for very simple systems. If we desire the temperature of the room to be 20 O C . see (10. and in order to see just what properties are required let us consider the problem of heating up a room. However.
3) The benefits of employing an integral action controller can be seen if steadystate conditions are investigated. the overall closedloop equation is given by (10.Proportional + integral control A 3 13 b Time Fig. It is required that the output signal will also be equal to 3 units.3. holds a memory of all the error signals which occur up to and including that at time t = i ~ .1) U(i)=Ki 1 f l e ( t )d t (10.2) such that when this forms the controller D in Fig. the application of a step input having magnitude 3 units. Consider. i.3.In terms of the Laplace transform the integrating controller can be restated as: U ( s )= Ki E ( s ) s ( 10. Ki.3.3 for an error e ( t ) .3. selected in Equation (10.1) 0 So the integrating action.3. once started at time t = 0 . 10.3.1. 10.3).1 Continuoustime proportional + integral controller The slope of the control input signal during an error period is dependent on the integrating gain.e.3 Integrating action of a controller necessary constant heat input would be greater than that necessary to keep the room at 20 "C. as was done previously. 10. V ( s )= 3/s in (10. In fact an integrating action has just the desired property and achieves the type of control input u ( t ) as shown in Fig. 10. in the steadystate and this can be seen . with no error.
However. which consists of a proportional + integral control action. But there were other basic requirements made on the closedloop system.6) and now there are problems. it could take several years for a room to be heated up to 20°C).3. Clearly. the closedloop characteristic equation becomes s271 + s + KKi = 0 (10. namely stability and transient performance. So using integral feedback alone does not necessarily mean that stability is lost.e. if Ki is selected to be positive and large in order to obtain low damping.3.3 14 Threeterm controllers from lim y ( t ) = lim s Y ( s )= s 103 s0  and thus the output will be equal to 3. this is not generally the case and hence the usual procedure is to use a twoterm controller. and secondly KKi > 0. room temperature will swing wildly. regardless of the value of Ki or the complexity of G(s).4) When this system is cascaded with an integrating controller (no proportional action).e. although for the simple system (10.4). reaching possibly 30 "C and 10 "C). 10.3. using integral control alone results in nice steadystate following properties but throws caution to the wind as far as transient conditions are concerned. . this will result in a high frequency of oscillations (i. Selecting Ki to be positive and small in order to obtain a low frequency of oscillations will result in a high damping factor (i. the transient performance of the closedloop system that will result is defined by ( 10.5) where Ki is the integral gain. which is true for a stable openloop system (10. So lim y ( t )= 3 = the desired output value I'm Hence steadystate following of the desired output value is obtained by the use of an integrating controller. However.4) closedloop stability could be easily obtained. Consider the firstorder system represented by (1 0.3. Further. as shown in Fig.3. In order to obtain a stable closedloop system it is directly apparent that all the coefficients of the closedloop system must be positive and that this is satisfied as long as firstly 71 > 0.4. or Ki > 0 for K > 0.
the time constant Ti then specifies how quickly it is attempted to achieve zero error .3. In terms of the Laplace transform this can also be written as V ( s )= ]I[ 1 + S T .Ti large means a long time elapses before the error reduces to zero whereas Ti small means that the reduction to zero error occurs swiftly. however. i. It is fairly straightforward to implement integral action by means of an active network with reasonable accuracy. i. are in practice P + I controllers. The overall philosophy of P + I control can be viewed in the following way. By applying this controller to the firstorder system in equation (10. is not this generally the case. Although transient performance in the closedloop can also be specified for many simple systems.9) and on condition that KKi > 0 and KK. the transfer function differs.e. see (10. by the inclusion of a zero in the numerator of (10. E ( s ) Kp T i s (1 0. is still present. the . The proportional gain K p is then selected in order to achieve a stable closedloop system with properties of relative stability. disclosed by the common s factor in the denominator.9).1 for stability. especially those for which transient conditioning is not important.e.Proportional + integral control 3 15 U Control input Fig. an extra controller term is now present for transient conditioning.3. however. 10.3. as was shown in the example.3. an operational amplifier integrating circuit. extremely stable or marginally stable.4 Proportional + integral controller The equation for this twoterm controller is then u ( t ) = Kpe(t) + Ki [ 1:' e ( t )dt] (1 0.7) or u ( t )= K p [ e ( t ) + 1:' e ( t ) dt] in which Ti = Kp/Ki is the integral time constant. however it must be remarked that the vast majority of process controllers.3. > .3.8) where it can be seen that the integral action.8). closedloop the characteristic equation is found to be: s271 + (KKp + 1)s + KKi = 0 ( 10. The integral term is included to remove any steadystate error. when compared to the solely integrating controller.4).
10.3. whereas lowfrequency modes are given a high weighting. Fig. the lowfrequency end of the spectrum.2. The effects of sampling a continuoustime signal are in fact shown in Fig.8).3. is shown in Fig. a predefined sampling period T later. The discretetime representation of the same continuoustime error curve is.e. which was to deal with the steadystate system response. however.2. that such an integrating implementation has a limited operating range. when it is replaced by the new value. The best way to describe such a response is perhaps as a highfrequency attenuator and this can be understood directly to mean that highfrequency modes of the closedloop system are given a low weighting by this controller. 10. 8.20 dB/decade \ 20 log K . for the equation (10.3 16 Threeterm controllers Magnitude dB \ slope = . adding up the results. It must be remembered.4. a certain amount of signal conditioning may also be required. curve from time I = 0 to t = t l .5. and as quite large magnitude inputs may be required in order to obtain relatively small output fluctuations. one in which the error value at a particular sampling instant is held until the next sampling instant.5 Bode plot for a P + I controller same cannot be said for the derivative action considered in the next section. e ( r ) . whereas the integral part is a lot more complicated in that in terms of its continuoustime realization the process of integration is one of finding the total area under the error. 10. This is consistent with the main aim of employing the integral action. though. i. The proportional part is simply a repeat of that considered in Section 10.2 Discretetime proportional + integral controller The implementation of a discretetime proportional + integral controller is by no means an exact procedure and it is perhaps best to take the proportional and integral elements separately. The best we can hope to d o therefore is to approximate the integral term by finding an approximation to the total . dictated by the magnitude of the power supply. One final point worth taking into account as far as the P + 1 continuoustime controller is concerned is its Bode plot which.
and. greater accuracy being obtained generally at the cost of greater complexity.11 ) Before proceeding it must be remarked that the method of approximation carried out in order to find 4 ( n ) in (10.3. in which the value of continuoustime signal when sampled at time instant k T is held and therefore considered to be the true value until time ( k + 1 T The two items should in no way be seen to be contradictory. 10. then any error prior to that instant can be ignored and hence the initial summation of error values will start at zero.10) it follows directly that: 9 ( n )= 9 ( n . the method ). The integral is then 4 ( n )= 1 tl e ( t )d t = 0 1' 0 e(t)dt and by approximating this with the error terms available. If it is assumed that we are only interested in the integral from the initial time t = 0 onwards.1) + T e ( n ) (10. where n is equal to an integer multiple of the sampling period T. 10. This is in direct contrast to the sample and hold techniques described in Chapter 8.3. just described for integration purposes merely being a way of approximating a particular signal over a given time.3. From (10.Proportional + integral control 3 11 area under the error curve from manipulations on the discrete error values that are at our disposal.6 and integration is carried out over the period 1 = 0 to t = t~ = n. The sampled error signal is shown in Fig. many other approximations based on different philosophies are possible.6 Sampled values of the continuoustime error . There are quite a number of possible methods for carrying out this approximation.3. we have $(n)= T kT= T 5 e(kT) (1 0. Although the trapezoidal method is often used. the simpler technique of rectangular integration is perhaps the most popular and hence this is described directly.1)Tto kT.10) where k is an integer. Error ! Fig.10) assumes that the error at time instant k T has in fact been the error over the period (k . as has been stated earlier.
2.Kp ~ 2 I and if the system G ( z )of (10.2.12) it follows that u(n) .3.2) + Te(n . where the control input is given by (10.13) where Ti is the integral time constant.I)] and by combining this latter equation with (10.3. which will become zero.3.2 is now combined with the integral part given by (10.12) is introduced. It is sufficient to note that as the . This means that in the steadystate lim y ( t ) = lim v ( t ) t'm 1 . as was hoped for with the inclusion of an integrating term.3.1) term in the denominator.11) and subtracting the result from (10.2.3 I8 Threeterm controllers Returning to the P + I controller.3.2. i We have also that the previous control input was ~ ( n1) = K p e ( n.z'1 = K .0 and thus steadystate following is achieved directly. and z'u(k) = u(k . Consider now the example previously investigated for the discretetime controller of Section 10.1 2) where u(n) is the control input to be applied at time instant n and K is the integral gain.1) = K p [ e ( n ) e ( n .u(n . if the proportional element obtained in Section 10. it is only really necessary to consider the (t.1) + Te(n)l ( 10.11). however.2. (10.11) as u(t)= D(z)e(t) or u ( k )= D(z)e(k) in which the P + I controller of (10. can also be written as u(k)[l .1) = control input at the previous sampling instant.3.1) + K i [ d ( n.l)] + KiTe(n) which in terms of a general time instant of k sample periods. the final closedloop expression is then Steadystate conditions can simply be investigated by setting z to unity in the above transfer function. The denominator has however become of third order and hence a simple analysis in terms of natural frequency and damping is not possible. this results in ~ ( n=)K p e ( n )+ Ki[d(n .13) means that D ( z )= where K z .
4. for a straightforward step reference input. then for steadystate conditions so lim y ( t ) = 2 = lim u ( t ) fm 1m I Fig. it will have three roots. and we have only two controller f parameters. the actual output signal level will automatically follow in the steadystate. 10. 10. i and K p .7. It may be the case however that the system to be controlled has itself an inherent integral action and thus. because of the common s' factor in the denominator. However. and hence an integral type of controller has been of importance. it was found to be straightforward to achieve steadystate following of the desired output value without placing restrictions on the type'of system response. in order to obtain greater flexibility in the type of transient response obtainable. a further term is required in the controller and this is the subject of the next section. This property is shown for a simple proportional controller on the position servomechanism of Fig. to link with the proportional feedback. which can be selected in order to place those roots in arbitrary positions in order to achieve a desired transient response. 10.4 PID control So far in this chapter the discussion has concentrated primarily on the steadystate following capabilities of control systems. The closedloop transfer function of Fig. IO. The openloop transfer function for the servomechanism is which means that this is a type 1 system.1) and it can be seen directly that if V ( s ) is a 2unit step input.PID control 319 system is third order. The conclusion then is that by the introduction of integral feedback.7 Position servomechanism with proportional control 1 .7 is however (10. and this includes closedloop system stability.
the transient response of the system (10. If the system to which derivative action is applied is type 0 however. firstly it has introduced a zero into the closedloop system at a point s = . Having said that.K p / K d . To summarize then. and this only serves to confuse someone learning about a topic. it is not possible with only the one controller parameter. incorporated in the mode described in the next section.. no common denominator s factor. The output position will in this case be dependent not on the actual desired position value but rather on its rate of change . The closedloop transfer function of this system with proportional and derivative (P + D) feedback is then (10.4. the effect is merely a cancellation of the inherent integrating action of the system. i. With solely derivative action applied to the type 1 system of Fig. 10. as shown in Fig. that a further important feature arises from the closedloop equation (10.4. i. then no cancellation of sterms will occur and hence an s factor due to the derivative action will appear in the closedloop numerator. a point selected by the two controller . Such is the case with the terms derivative control and velocity feedback. 10.e. they mean exactly the same thing.2) or The effect of the derivative term is twofold. a system which has its own inbuilt integrator. I ) in that although by a correct choice of K p it can be ensured that the closedloop system is stable. K.not generally a good feature.e. with Kd as the derivative gain. in place of the proportional term. without the need for an integrating type of controller. If the same attempt is made with derivative action alone then the result is very much dependent on the system type. thus producing problems of steadystate offset.8.4.320 Threeterm controllers Hence steadystate following of the desired output value is achieved for a step input on a type 1 system.4.1) can be characterized exactly as desired. use of derivative feedback alone is not generally a sensible arrangement and hence in the first instance it will be considered that the derivative term is coupled with a proportional element. It is worth pointing out here though. in the previous section an attempt was made to employ integral control alone without incorporating any proportional element. It is in fact the case that with one further controller parameter.1 Proportional + derivative control There are instances in many subjects when different terminologies are used to describe things which are exactly the same. i. to define the type of transient response which we might like.7.e. 10. because as far as control systems analysis of position control systems is concerned.
9. while lowfrequency components are attenuated. In this example: an= 1.O and damping factor t = 0. consider that we have a position servomechanism in which K = 5 and 7 1 = 0. i. highfrequency components are considered to be of importance and are given greater emphasis by the controller.e.5.06 Hence the controller becomes Kp+KdS=O.5. then it can be seen that for this type of controller.PID control 321 I I Fig.7 = ( 1 + 5&). and then to select Kd in order to satisfy the damping factor requirements.4.l[l 0.l and t = 0. Kd = 0. It must be noticed also that inclusion of the derivative action. 10. Secondly. 10. as is shown in Fig. as this gave a further insight into performance .8 Position servomechanism with proportional + derivative control parameters. The P + D controller.3) .. and this is really of much greater importance. it has given an extra degree of freedom as far as choice of the closedloop denominator coefficients is concerned. Now the closedloop characteristic equation can be written as: s 2 + s (1 KKd) 71 + 5= 0 71 where So it is a straightforward process to select K p in order to achieve the required a n . the integrating property of a type 1 system has not been openloop cancelled.o=jR. The design requirements are then to select K p and Kd such that the closedloop system has a response with natural frequency con = 1 .=O.JT. and hence allows for selection of the closedloop transient response as desired.6~1 Previously with the P + I controller it was found advantageous to consider the controller's Bode plot. If this is done with the P + D controller. As an example of how the design of a P + D controller could be very simply carried out. has not altered the steadystate following characteristics of the closedloop system to a step input.7. which is usually written in the form u(s)= Kp[ 1 + TdS]E(S) ( 10. along with the proportional term. SO so K.see Fig. 10.
steadystate position following properties may not be present in the closed loop. 10. the closedloop transient response of that system. and in this scheme the benefits of all three types of control action can be obtained. no matter how complicated the system might be. unless the system is of a certain type. because of its lowpass filtering action. Each element has a purpose for which it is particularly useful. without the drawbacks exhibited by the individual terms. By combining the derivative term of the P + D controller (10. Td = Kd/K. 10.4. then ensures that the higher the frequency of a signal. However. exactly as desired.322 Threeterm controllers Magnitude dB A 0 where the derivative time constant. It has been shown that when P + D control is applied to a system it is possible to characterize.2 Proportional + integral + derivative control In this chapter so far the benefits of the individual proportional. The inclusion of a derivative term in the controller therefore will serve to highlight any noise present while at the same time attenuating the actual system signals. The derivative term is based on a multiple of the rate of change of error and therefore can have little effect on steadystate values.. the general form of Proportional + Integral + Derivative (P + I + D) control will now be considered. and these cover the basic requirements for almost all linear control systems. tends t o remove the effects of noise. This in fact is the converse of the P + I controller which. the more it is amplified. Two notes must be made from this: firstly the gain of the P + D controller tends to infinity as the frequency tends to a high value.3) with the P + I . whereas the integral term is based on the value of error over a relatively long period of time and therefore can have little effect on transient values. For this reason. A block schematic of the P + I + D controller is shown in Fig. whereas noise signals present will almost certainly have fairly substantial highfrequency components.4. and realistically it is not possible to achieve this with standard active devices.10. integral and derivative control elements have been shown and major drawbacks of certain combinations have been pointed out. secondly the system input/output signals will generally be of relatively low frequency.
4. as the accuracy of the controller can be no better than the accuracy of our plant model. Here however. the overall threeterm P + I + D controller has the error to control input equation U ( S )= K p 1 + + T S d .3.5).7) The overall controller now has three coefficients. a set of coefficient values can be obtained. (10. it is wished to make a further point in that it is often the case that an exact representation of the plant to be controlled is not available. for example in terms of its timedomain representation e ( t ) dt +T d dt (10. the three separate coefficient gains can be seen. two . 10.8) it can be seen that.4. Kp.4.4) or (10. Ti and Td.s i [ ] E(S) (10.6) or with a common 'integrating' denominator s factor (10.5) where in the latter form of the threeterm equation.4. carrying out an elaborate exact controller calculation by using inaccurate plant parameters seems rather rash. 10. This might be due to one of a number of reasons.PID control 323 Fig.10 P + I + D controller controller (10. Whatever the reason. for example actually obtaining a good plant description could prove to be too expensive or the plant may well have deteriorated due to ageing. as given in Fig. indeed this was the approach taken for both the P + I and P + D controllers previously described.10. It can be noted that the threeterm controller equation is to be found in various forms. integral and derivative coefficients are to be found. An example could be given to show how for one specific plant.4. which must be selected by the operator to meet some desired criterion or objective. In fact many approximate methods for the selection of proportional.
324 Threeterm controllers Table 10.11 . A further consideration which must be made is that for certain plants it can be extremely dangerous to cause the output to oscillate. much experimental evidence. (0. it must be noted however that these are not necessarily the best of all possible coefficient values. The ZieglerNichols procedure then requires that the gain K . 0. The critical value of gain K . In this section the procedure used in order to implement the ZieglerNichols method is described.1. 10. I ZieglerNichols threeterm controller ..1. and has been proved to work effectively by.coefficient values 1 Proportional P + 1 P+I +I ) controller controller controller K.1).  0. such as mechanical breakdown. What is of importance is (a) the lag time and (b) the . 10. Perhaps the most widely employed method for the setting of P. The final steadystate value of the step response is of no importance as far as the controller parameters are concerned. in that this may well result in physical plant problems. certainly for process control.2. it is merely found as the system output in response to a step input of sensible magnitude. The method is based largely on. with the controller consisting simply of a proportional gain term.Also the frequency of the oscillations which occur is measured as wc with a period Tc = 2a/w.5Tc < I) Ti Td  0. Unfortunately the method described for controller coefficient calculation has one or two drawbacks when applied to some plants. O.. In terms of the values Kc and T. P + I or P + I + D controllers. is that detailed by Ziegler and Nichols..6 < 01 0.125Tc essential properties being that they are both relatively simple and fairly inexpensive to implement. Note that no controller parameters are varied or modified in order to achieve the step response. see Fig. the ‘best’ values for proportionalintegralderivative terms are shown in Table 10. Fig. and none is given here. which requires solely that a plant step response (also called a reaction curve) is obtained. this is particularly true when there is an amount of noise affecting the system and/or measurements. and the information necessary to calculate each of the controller coefficients is tabulated. be increased until the closedloop system just oscillates. where necessary it is possible to employ an openloop alternative. as in equation (10.45Kc aK. i.5K. Rather than use a closedloop technique for ZieglerNichols tuning. The system is set up in the form of a simple cascade compensator. the system output will just begin to oscillate. In practice it can be extremely difficult to measure exactly the point at which the onset of oscillations occurs. hence very little background theoretical detail is available.e.83T. at which oscillations just startlstop is then denoted as K.
shown in Fig. from the measured P and I L values. 10. although this is rather more straightforward than the integral approximation .8) where y is the step input magnitude. the latter of these being dependent on the magnitude of the step input signal. The two values Kc and Tc required in order to calculate the controller parameters can be found by the ZieglerNichols formulae. also called the reaction rate.e.PID control 325 tI Output magnitude I I I / / / / Reaction curve P = gradient / tL Time / I I I / / I ' / / / / / I I IV Fig. i. is simply the gradient between the nearest measurement e ( k T ) . as (10. t ~ and gradient.4.T).4. In this section it therefore suffices to consider merely the discretization of the derivative term.2. 10. the derivative of the error at time instant kT = n. and to subsequently append this to the P + I controller.3. e ( k T . 10.12.especially so if the idea of finding a differential or difference is associated directly with the derivative. An approximation is required for the derivative term.showing lag time. gradient. where k is an integer number of sample periods T .3 Discretetime P + I + D controller The discretetime version of a P + I controller was considered previously in Section 10. and the previous measurements. the derivative is . By referring to Fig. where it was shown how approximations were necessarily made to the integration procedure in order to incorporate such a technique in a discretetime format. P . 10.11.11 System response to a step input .
it remains only to append the derivative term when multiplied by the derivative gain Kd.e(kT .10).1)T.1) = K.u ( n Kd + e(n .12).1) is the integral summation of errors at time instant ( n . The control input therefore becomes ~ ( n=) K .9) or alternatively de(n) dt e ( n ) . O ( n ) = * ( n .11) follows that ) it u ( n ).T ) T (10.e ( n . to the control input equation at time instant n .2 ) T remembering that (10.11) where + ( n . found previously for the P + I controller as equation (10. between them: e ( k T ) .12 Backward differencing operation taken t o be the difference between the two values divided by the time. i.3.2) + Te(n .1 from (10.3.4. e ( n ) + K i [ 4 ( n  (10. We have also that u ( n .1) T (10.2)] and on subtracting this equation for u ( n . the general P + I + D equation at time .1)+ T e ( n ) .l). 10.4.326 Threeterm controllers Time ( k .e ( n . By rearrangement of this difference equation.e(n .3.4.4. T .e.11)holds.l)T kT=n Fig.10) In order to obtain the complete threeterm controller.l)] + Kd[e(n. defined in ( 1 0.1) + Ki[+(n.
and thus the total computation per iteration only amounts to three multiplications and three additionslsubtractions as can be seen in Fig.:I [ 3 (10.~ u ( = u(k . Firstly the problem of noise is highlighted by the fact that the derivative effect acts directly on the error term which Fig. in Fig. can be written as 321 u(k)[l .PID control instant t . for discretetime control.13) which means that the threeterm controller equation can be written in the form. The implementation of the discretetime P + I + D equation of (10..13. etc.+ .1) = the value of control input one ~) ) sampling instant prior to time t.4. 10.Kze(k .1) 1 + 2 [(. it must be remarked that the threeterm controller discussed here does exhibit one or two drawbacks.12)is fairly simple to carry out in that it is only necessary to store an error signal value for two sample periods before discarding. 10.1) + &e(k . similarly z .2[] )$] in which the integral time constant Ti = Kp/Ki and the derivative time constant T d = Kd/Kp. u ( k ) = u ( k .4.13 Schematic of the discretetime threeterm controller ( t = kT) .2. 10.13 are given as: K2 = K p [1 + 2 +] (10.4.z . Also the terms multiplying the error values would usually remain fixed once an initial selection has been made. Remembering that z'u(k) = u(k .12) + e(k .I) + K l e ( k ) . The multiplying factors K I .2) Although fairly straightforward to develop.e ( k .' ] = K p e ( k ) 1 + . KZ and K .
A second problem is the production of a large error derivative value every time a change in desired output. action is known as ‘derivative kick’ and although it is common to both continuous and discretetime controllers. This temporary.2 seconds is a pure transport delay.328 Threeterm controllers consists of the difference between a noisefree reference signal and a noisecorrupted measured value. it will be assumed that G ( s )= (s + 50 5)(s + 10) (10. The ‘derivative kick’ effect can be reduced by filtering the setpoint input o r by employing an alternative derivative approximation scheme.1 . The engine transfer function is given by loe .1 + jl0. in order to place the roots of the closedloop system denominator at s = .100.5). However. set point. by taking a n approximation to the transport delay. but potentially dangerous. as in Fig.14.2s G(s)= ~ s + 10 where 0. 0 0 By cascading the controller of (10.15) be employed.14) and applying a unity feedback loop.4.4.4.4. thereby attenuating the highfrequency noise components which would be accentuated directly by the derivative action. Ki D ( s ) = Kp + .14.4 Worked example A n example is given for the design of a continuoustime threeterm controller and is based on a problem given in Kuo for the control of the airlfuel ratio in a n engine.jl0. it is particularly problematic in the latter. 10. i.e..4.15) with the engine equation (10.14) It is required that a threeterm controller of the form shown in (10. 10.14 Engine airlfuel ratio controller . closedloop equation the Fig. however in many cases the derivative action is applied solely to the fedback measured value rather than the error and indeed threeterm controllers can also be found where the same is true of the proportional action.4.0. ~~~ ~ ~ ~ 10. . value occurs. 10. As shown in Fig. A sensible way round this problem would be to filter the measured value before feeding the signal back.+ K ~ s S (10. the threeterm controller acts o n the error between the desired and actual airlfuel ratios and is required to achieve a predefined response pattern. by means of a truncated power series expansion.
this is a direct result of the use of an integral action controller. 10. Various problems such as ‘derivative kick’ and noise were mentioned as far as PID controller implementation is concerned. These conditions are both caused by the derivative action. and thus the three controller parameters must be found in order to satisfy the equation: S(S~ + 15s + 50) + 5O(Kds2+ K ~ + Ki) = (S + 100)(~ 20s + 200) s + + K ~ )+ 50Ki = s3+ 120s’ + 2. that if a step input is applied to the openloop system (10. It is fairly easy to check. a derivative term is also introduced. a steadystate error will result. this is often found to be the case in practice.1 It is useful to note that the derivative coefficient is much smaller in magnitude. by employing the final value theorem. and yet if the same step input is applied to the closedloop system of (10. however other problems can occur due to the other terms. resulting in P + I controllers being sufficient in many cases.16) no steadystate error will appear. and simple examples have been given to show how a typical design can be carried out.200s + 20. However. For many systems this is complex enough. Finally. and the inclusion of an integrator solves this problem. by far the most widely encountered procedure is that of threeterm or PID control. hence initially an attempt was made to use solely a proportional controller.Summary relating desired to actual air/fuel ratio is found to be Y ( S )= S(S 329 50(KdS2+ KpS + Ki) + 5)(S + 10) + 50(KdS2 + KpS + Ki) U S1 (10. A complete chapter has been designated here specifically for the introduction and description of both continuoustime and discretetime threeterm controllers.4. for this example.5 Summary Despite the availability of many different control design techniques. particularly in process control systems. The buildup has been such in this chapter as to attempt to get away with as simple a controller as possible.14).16) But it is only the denominator of this equation which specifically concerns us for design purposes.000 s or s3+ (15 + 50Kd)~’+ 50( 1 Therefore our design requirements are satisfied if. the three controller coefficients are: Ki=400.4.4. in order to achieve greater flexibility in the transient design procedure. Kp=43 and Kd=2. it can be the case that a steadystate error exists between the actual and desired output values. one example .
Hence. where u ( t ) = K .2.2 with D ( z ) and G ( z ) cascaded within a unity feedback loop. For the system: '(') = ( z . and much of the reason behind its employment must lie in the fact that satisfactory performance is invariably achieved along with reliability and robustness.5(~ 0.3 within a unity feedback loop. It is required that the closedloop poles be placed at s = . will affect closedloop stability. Fig.2)(2 0. Problems 10. 10.15). very large will result in lightly damped output oscillations.3 30 Threeterm controllers being integral 'windup' which is apparent when an offset exists persistently between the actual and desired output values. and how a choice of K.0.7 and s = . K and K d such that this requirement is satisfied. it must be remembered that the extensive use of a PID structure is due only partly to its simplicity. While an offset is present the integral part of the controller output will increase and therefore lengthen considerably the time taken for the closedloop system to settle down.3 5 j . l Find Kp.15 . show how the choice of K.2. Consider a continuoustime P + I + D controller cascaded with the system G ( s )= ~ 5 s(s + 3) 10. when compared with alternative controller frameworks. find the resultant closedloop denominator. Finally. i Proportional + derivative action can be applied either to the error between the desired and actual output signals or directly to the actual output signal alone. e ( t ) .1 Consider the discretetime proportional controller discussed in Section 10. as shown in the diagrams (Fig. 10.6) + 0.3) 10.
A threeterm controller is of the form: Show that if Ti = 4Td the frequency response of the controller has the form D(j w ) = Kp[l + jwT]’ j2w T where T=2Td 10. 10. find the time solution of the i transient ramp error for the case in which Kp = 14 and K = 8.16. Consider a P + I controller D ( s ) = Kp(l + l/s) which is cascaded with the openloop system: G ( s )= (s + 2)(s + 5) 25 10. in terms of steadystate errors. Calculate the value Kp which will ensure a steadystate velocity error of 10%.17.7 For the closedloop system shown in Fig. firstly when a unit step input is applied.4 Consider the effect of the different controller connections.6 For the closedloop system shown in Fig. 10. 10. find values for K p and T to d provide ideal transient performance and a 1% ramp error. and secondly when a unit ramp input is applied.5 within a unity feedback closedloop system.Problems 331 10. 10.17 . Fig.
Find the range of values for the proportional gain K.18.18 10.11 within a unity feedback loop..5. of the form [l .y ( l ) .12 with e ( t )= v ( t ) . Disturbance. show that all three controller coefficients are required. and the cascaded functions are connected within a unity feedback loop.10 Find the integral gain K necessary to ensure that the steadystate ramp error for i the closedloop unity feedback system is 5 % at most.4 k jS.8 Threeterm controllers Consider the use of a threeterm controller D ( s ) in the system shown in Fig. K and K d such that the closedloop poles are placed at s = .50 and i s = .' ] u ( k )= [ K I.13). N(s) I Fig.5 and 0. see (10.' + Kjz*Ie(k) is cascaded with the openloop input u ( t ) to output y ( t ) transfer function 10. 10. Find Kp.K 2 2 . 10. Find K p . Is there any steadystate offset in the output signal following a unit step input? Consider the openloop system described by the discretetime equation: . +0.. also find the range of values for the proportional gain K p over which the closedloop is stable. A discretetime P + I + D controller. A continuoustime P + I + D controller is cascaded with the system G ( s )= (s + 10 l)(s+ 2) 10.9 A P + I controller is employed in cascade with the openloop system which ias a transfer function: 3 G ( s )= s2 + 4s + 8 10.z . Note: D ( s ) is identical to that in Problem 10.5.9. v ( t ) is a reference input.4. K and K d such that the i closedloop poles are given by z = 0. such that the closedloop system is stable.332 10. If it is required to track disturbance steps with zero steadystate error and also to exhibit overall closedloop system stability.
0.1. a P + I controller is applied. i.14 i.5.e. 10.within a unity feedback loop. Consider the system shown in Fig.19 with D ( s ) = K p + K i / s .20 with a = 2 and D ( s ) = K P + Ki/s. any further numerator or denominator terms will mean that closedloop pole positions cannot be arbitrarily specified.Problems 333 10.7071 is desired.4. this can only be accomplished by a digital P + I + D controller when cascaded with a system G ( z ) . Ki = 1.12).5.20 .01. (b) K p = 10. Ki = 0. I f it is required that the steadystate value of Y ( s ) is equal to 100 with 2% error. + 0. if the system is no more complicated than: + 10. If V ( s ) is a ramp input of slope = 0.5 and .13 By using the general P + I + D control equation (10. Ki = 0. N ( s ) I W I I I Fig.19 10. KZ and K3? If a P + D controller is cascaded with G ( z ) .0. 10.15 Consider the system shown in Fig. i. Disturbance.9? Show that if it is desired to place the closedloop poles of a system in specified locations. Fig. can the closedloop poles be placed at z = . only a P + D control is employed.e. then K Z = K I + K3. 10. the disturbance N ( s )can in this case be regarded as a ramp of slope = 0. (c) K p = 1. find values for the step input magnitude and the proportional and integral gains when a closedloop damping ratio = 0.e. find the steadystate ramp error between desired V ( s ) and actual Y ( s ) outputs when: (a) Kp = 20. show that if Ki = 0. 10. What problems does the employment of a P D controller therefore raise if an attempt is made to place the closedloop poles of a unity feedback system by the selection of K I . Also.
pp. 1975. Proc. Proc. A. ‘Simple tuning of discrete PI and PID controllers’. J. Proc. 2603. 20. R. and Hagglund. 4th IEE Workshop on SelfTaning and Adaptive Control. D. 3328.. 1987/33. pp. AC31. Digital control systems.Jones. on Industrial Electronics.334 Threeterm controllers Further reading Astrom. Franklin. 1979.). and Nomikos. 1985. ‘PID controller design for a TIT0 system’. ‘Automatic tuning of simple reactors with specifications on phase and amplitude margins’. 1986. and McKeown. 4. Inst. P. . 64551. E. . T. A. . 1984. D. B. No. Shinskey. 1983.. Instruments and control systems. 1987.. pp. Y.. Smith. PrenticeHall. T. A. C. IE31. Research Triangle Park. pp. ZEE. pp. 89. 22734.. P. G. IEEE Trans. J. Warwick. K.. 14. and Kelly. D. 4. 1982. 134. Gawthrop.. pp. J.. No. B. Japan. D. 31 1314...D. Industrial digital control systems. pp. Thomas. J. 3. ‘Tuning PI and PID digital controllers’. 1987. C. O. No. K . IEE. Kuo. Lopez. 18492. Porter. W. P. (eds. and Murrill. 5 . AddisonWesley. Pt. K.. 75968. pp. G. 1969. pp. Automatic control systems. Tuning and control loop performance. ASME Paper 52A58. M. Smith C. C . L. 9.. 1972. on Automatic Control. IEEE Trans. 1976. pp. Part D. B. Ortega. Sandoz. Control system principles and design. J. Warwick. Fensome. N. W. Control Engineering. Doebelin. 647. ISA Trans... USA. Kyoto. 42.. ‘Tuning controllers for noisy processes’. Digest No. IEE Conference ‘Control 85’ 1985.. Pemberton. Roberts. ASME.. ‘Controller tuning from simple process models’.1942. and Rees. P. pp. 130. 3944. McGrawHill. NC. No... Gawthrop. IFAC Congress (1981). K. and Atherton. 1983. H. J. pp. Pt. pp. L. IEEE Control Systems Magazine. F. 141. A. P. 1952. No. 11 1120. ‘PID methods in selftuning control’. Electronics and Power.. Nishikawa. ‘Understanding 3term controllers’. N. 292304. 1994. 1990. 647 (July). and Nichols.D. W.. 2019. Fertik.. P. Ziegler. ‘Automatic tuning of commercial PID controllers’. M. Proc. 1983. 4. G... D. ‘Selftuning PID controllers: algorithms and implementation’. 3441. ‘New desaturation strategy for digital PID controllers’.. 1991. 1975. ‘PID: the logical control algorithm’. Tech. ‘Applications of selftuning control in the chemical industry’. ‘Optimum threemode controller settings’. and Thompson. M. ‘A method for autotuning of PID control parameters’. 1986. McAlpine. John Wiley and Sons. H. S .. Instr. 4. Automatica. Proc. 647 (May) and pp. J.and Sannomiya. Pessen. M. R. ‘Optimum settings for automatic controllers’. Proc IEE. Trans. Corripio. 3826. 1984. H. A.. Process control systems. G. 2nd ed. and Martin Jr. MC. McMillan. ‘Realtime expert tuners for PI controllers’. B. Trans. ‘PID selftuners: some theoretical and practical aspects’. B. 2. p.. Peter Peregrinus Ltd. Zhuang. D. No. ISA.
it can be written as: (A1.2) where The array A is called a matrix and in more general terms. 1) for which a solution is required for the three unknowns X I . form as: AX= y (Al. as an aid to problem solving.Appendices Appendix 1 Matrix algebra In this appendix. by considering m simultaneous equations with n unknowns. e. x3. X Z . These equations can also be written in an alternative.g. I Basic notations Consider the three simultaneous equations allxl + 012x2 + 013x3 = y1 + 022x2 + 023x3 = y2 431x1 + 0 2 2 + 0 3 3 = y 3x 3x 3 021x1 (Al.3) . matrix. A1 . some of the basic rules of matrix algebra are covered briefly.
. . I= [::y. Al.. and is denoted by the symbol 0. the matrix is termed an identity matrix.6).50 1 1 (Al.e.e. in which that element lies. 0 . as shown in (A1. such that when m = n .2 Matrix operations The transpose of a matrix A . when all the elements of a matrix are zero. examples being both x and y in (Al. aij = 0. whereas when m = 1. for all i and j .. K' 0 azz' . Ol 0 (A I .5) 17 This is taken a stage further if the matrix coefficients aij = aji = 0 for all i Z j . Such a matrix is called a diagonal matrix.2). an example being given in (Al.7) *\O 0 1 0 Finally. . is found by interchanging the rows and . the m x 1 matrix is called a column vecfor (or column matrix).. i.6) O'U L O o .5). ] . ' (Al. .. with the indices i and j denoting the row and column respectively. the 1 x n matrix is called a row vector (or row matrix) as shown in (A1.3).. * . When the index n = 1. as was the case in (A1. 2 4 A = [ :4 5 .4) A special case of a square matrix exists when aij = aji.4): w= [Wl. denoted by A T . as shown in (A1.2). i. A . Wlll (Al.. and is usually denoted by the letter I . w2... leaving only those elements of A on the principal diagonal. it is called a square matrix. in which case the matrix is called a symmetrical matrix.336 Appendices in which the coefficients aij are termed the matrix elements. .J When all the principal diagonal elements of a diagonal matrix are equal to unity. is called a n m x n matrix. then the matrix is called a null matrix. aii = 1..
+ ann (Al. the transpose of a row vector is a column vector and the transpose of a column vector is a row vector. is denoted by Aij. where A U . As an example consider the matrix 5 then AT=[: 4 1 3 :] 5 Also. aij 337 A=[: 3 . its determinant is found by means of cofactors and minors.8) The truce of a square matrix is the sum of the elements on the principal diagonal. whereas when the determinant of a matrix is not equal to zero.g.e. for the n x n matrix trace A = all + a22 + . e.e.5 = 2 The determinant of a matrix is a scalar and is only applicable to square matrices.g. if the vector then xT = [ X I X Z X ~ ] . as a result. that matrix is said to be nonsingular.( . Mij is the 1 ) i + j ~ ~ ~ minor of aij and is the determinant of the square array which remains when the ith row and jth column of the original array are removed. The cofactor of an element aij. or vector. When the determinant of a matrix is equal to zero. (AT)T=A (Al.24 0 i] + Uji. i. The determinant of matrix A is written as 1 A 1.e. that matrix is said to be singular. e.9) For the matrix A= [: 3 2 4 0 5 381 (Al. 10) trace A = 3 + 4 .. if 011 012 013 . For an n x n matrix. i.Appendix I columns of the matrix. in the matrix A . Note: Taking the transpose of a transpose produces the original matrix.. i.
18) + 8( . So [a : 3 2 1 = a l l A l l + a 1 2 A ~+ a13A13 2 by n] selecting the first row = 3(20) + 2(5 . I The determinant of matrix A= can be found from 1 A ( i = 1). which confirms the previous answer. i. if A = [ QII 012 013 a21 a 2 2 031 1132 .158. So = 2(5  18) + 4(15 .ll) such that the determinant of a n n x n matrix A is given by either n I A 1 = i= I or n UijAij.A23 ZI by selecting the Therefore I A I = . The adjoinr of a square matrix A is found by firstly replacing each element by its cofactor and secondly by transposing.158.24) I A 1 = U Z ~ A+ a22A22 + &. or second column ( j = 2).e. for any one selected column j I A 1 =j = I OijAij.48) Therefore 1 A I = . for any one selected row i Example A 1 .338 then Appendices T h e determinant of a general 2 x 2 matrix is (Al.
it follows that in many cases A B # BA This is easily verified by a simple example. So C=A+ B=B+ A (A1. the element Cij is found by multiplying the elements of the ith row of A and the jth column of B and summing the results. are carried out by simply adding (or subtracting) the corresponding elements. . consider the case when B = I for example. e. consider 2 2 = + Cij = a i j .] [:g . consider c= 3 [ 6 Note: By the definitions of multiplication given.I 4 10 2 6 7 Multiplication of two matrices requires that the number of columns of the first matrix is equal to the number of rows of the second. 3 6 For subtraction however. This means that if A is of dimension m x n and B is of dimension n x p .g. +. then C will be of dimension m x p . which are of the same dimensions. try (Al. . (Al. As an example.Appendix I 339 Al.13) where the ehIentS of c a r e Cij = = 1 a i k b k j . So C= AB (A1 .6f 3 !I[: .] [.e. 4 + b i j . As an example.3 Matrix manipulation Addition and subtraction of matrices. the two matrices d o not therefore necessarily have to be of the same dimension.bij. 14) However. such that t: !I=[. c = [ . i. 14) is certainly not a general rule.12) cij = aij where the elements of C are c= [: . .
g. if C = XA where X is a scalar. Firstly. i.]=[I2 01 Multiplication of vectors can take on two forms. when a column vector premultiplies a row vector the number of elements in each vector need not be identical.16) This can easily be verified with a simple example. e. The product of two transposed matrices is equal to the transpose of the product of the matrices when commuted.340 Appendices Multiplication of a matrix by a scalar results in all the elements of the matrix being multiplied by the scalar.e. 2 Find the inverse of the matrix A= [a 3 2 31 .e.g. so. The inverse of a nonsingular square matrix A is defined as: (A1 .18) Example A l . then xy = i. an n x m matrix results. then However. then Cij (A1 . consider the n x 1 column vector x and the 1 x m row vector y.15) = Xaij. e. consider 6 3 c=3[: . consider the n x 1 column vector x. A ~ = BB ( A)~ ~ (A 1. when a row vector premultiplies a column vector. As an example.17) and has the property that: AA~ A . in which case the number of elements in each vector must be identical.~ AI = = (A1.
4 Simultaneous equations solution The concept of a matrix description was introduced in terms of writing a set of simultaneous equations ( A l .' = 1 1 5 ' [ 20 13 24 10 33 12 . l to be I A adjoint of A is however adj A = I = . and Y= [ 1q 2 so . X = [ i i ] . A will be a square matrix of dimension n x n.:] Al. 3 Find a solution t o the set of simultaneous equations given by: where A= [:: 6 3 i]. The [ 20 13 24 10 33 12 38 . X= (A 1 .:] 38 The inverse of A is therefore A . and A .158. Example A l .e. It also follows that A'(Ax)= A'y i. 1) in an alternative form (Al.' exists. As long as A is nonsingular.Appendix I 341 The determinant of A was found in Example A l .1 9) Aly which gives a unique solution for x .2). a solution will then exist t o the set of simultaneous equations. namely AX=Y On the assumption that the number of equations n is equal t o the number of unknowns in the vector x .
2). Nelson. then its rank is equal to the row (or column) dimension of the largest (in terms of dimension) minor within that matrix. Finally in this section. It can also be said that a nonsingular matrix is of full rank. such a matrix is singular.e. Tropper. M . T. i. x2=3 and x3= .23 x ..21) Further reading Kailath. By this means. Linear sysferns.(3X 1 + 2 X 10 . rank = n for a n n x n nonsingular matrix.1 4 ~ 2)=::  ~ 3 = .20) and A / = /A = A (A. .2 ) = .1 I t is assumed. So.:. Linear algebra. that the square matrix A is nonsingular. its determinant is zero.342 Appendices and adj A = [: 6 6 3 I f I:.. 23 which means that 1 2 23 14 and hence XI = . the matrix A=[: ! :] has rank = 2. two useful matrix relationships are: I A l IBl=IABI (A 1.\ (6 X 1 I x 10 . PrenticeHall. \ ( . 1. whilst every minor of higher order is zero.~ ~ The solution is therefore: XI =2. . Inc. 1980. to ensure that a solution exists to the equation (A1. if the matrix A possesses a nonzero minor of order r .3 ~1  + 3 ~ 1 0+ 2 4 . i. A.e. however. If.2) = 'i: 15 X2=. 1978. then A has rank r .
a/)cos W / exp( . unit impulse at / = 0 I .a t ) s(s + a) m sin w f COS(Wf S w 7z + q4) sin(w/ + 6 ) cosh w t sinh wt exp( .a f ) ((11 (s+ a) ‘ I) ( n .Appendix 2 Appendix 2 343 Tables of common transforms Table A2.’ ) / ( n .I)! exp(af) (s+ a)’’ a I .I)! Laplace transform.at)sin w t exp( ar)sin(wr W (s+ a P + W * + cp) where y = a + w/tan 6 . unit ramp f ( ” . I Laplace transforms Function. unit step f . F(s) i.exp( . Y ’ I f ( f ) ] I or s” SI S‘2 S” exp( . f(/) [for I 2 01 6 ( f ) .e.
exp( (w.a / ) s i n ( w / +6 ) 6 = [(A  a)’ w ’ ] + p = (a’ w’)”’ + ”’ 1 s[(s+ a)’+ w’l (s + Y) 6 (s + p ) [ ( s+ a)’+ w’l I” Ian 6= W 0.1 For an example of transform with zero order hold present./) a a’+ s2 + 2rw..344 Appendices Table A2.e.. F ( s ) i../)sin(au./)sin(au.z lo? w I. subject to a sample period of T seconds..1 (contd. f(/) [for I 2 01 Laplace transform./ a 1 a’+ I +6) <’= r = cos 6 Example A2. V’[J(/)l I s[(s W = + a)’+ W ’ I atan@and p = (012 + w 2 ) ” 2 exp( .exp( . The overall transfer function..exp( .a)’+ w’] exp( .) F u n c t i o n . consider the original system transfer function G. .S T ) ] s(s + 3) . with zero order hold included is G(s) = 1 [ I .. p =I  1 .(w. 6 = [ (0 .(s) = __ s+3 which is to be cascaded with a zero order hold.a / ) e x p j(w/ + 6) I ( s + a)’+ w * 2 [(s + a)cos 6 ..w sin 6 + j ( s + a ) s i n 6 + w c o s $1 WI.s + w.
22 exp(.22 cos w T + I z(z .I)lzexp(aT)] t sin wT sin W I cos w l W sz+w2 S 72 w WI exp( .~ a ~ ) z2 .1 is a further example.a T ) I Z .z exp( .Appendix 2 Table A2.a T)cos w T 2’ . f r o m the above table.aT)cos W T + exp( 2 a ~ ) z2  such that The ztransform for the function l/s(s + 3). Therefore if it is the case that a zero order hold is present.~ x D ( . f ( t ) [for I 2 0 1 Laplace transform F(s) 1 345 ztransform. it is assumed that no zero order hold is included in the required transform.af)sin (s+CY)*+wZ exp( at)cos wl  S+ a (S+ff)2+Wi .a t ) I + .3.2 ztransforms Note: In the following table of ztransforms.~ T N .a TI1 1 I] s2(s + a ) z[l .3.exol .21 cos W r + I zexp(aT)sin W T r’2zexp(a~)cos w ~ + e x p ( . F(z) ( T = sample period) I Zk c b(1.a T ) ] 4 2 .kT) exp( .exp( .exp(. is found.at) I (s + (Y I . Example A2.ar)i2 t l I . with particular reference to Example 8. Function. to be: zri .[exp(at)ff s(s + 01) a  zexp(aT) Tz exp( . the overall transform should be obtained as described in Chapter 8.exp( .cos W T ) Z~ .kTs) S’ I I z 1 S2 rn T2z(z + 1) 2(t I)’ Tz IZ 2 exp( of) f s3  (s+a) 1 Z  exp( .
though this might mean that the range of operation is limited. definitions of linearity being given in Chapter 2.3 T ) ] which can be evaluated for any specified sample period T. For any one nonlinear system however.2.exp( . Despite this. A3. nonlinear effects only being witnessed when I e ( t ) I > I el I. a point is reached with some systems where nonlinearity effects are too large to ignore.2. i. Because there are many different ways in which a system can exhibit nonlinear behavior. to a good approximation. and unfortunately it is not always possible to know exactly the range of values to which the system will be subjected. nonlinear.range . they break at least one of the rules of linearity. Section 2.346 Appendices giving an overall transform of: G ( 2 )= 1 exp(3T) 3 [t. Appendix 3 Nonlinear systems The study of control systems carried out in this book is concerned with those systems which can be regarded as linear.1. In practice the vast majority of systems to be controlled are. The amplifier characteristics will very likely result in amplifier output saturation caused by the application of too large an input signal. is given. J Cascade nonlinear amplifier While the signal e ( t ) remains within the . . see Fig. other than by classifying particular types of nonlinearity. A3. it is difficult to build up a general theoretical background. However.e.3. v :? Ed1 amplifier U .1 will operate as a linear system.y system Fig. to an extent. and they must be taken into account. A3. Example A3. p q . it is apparent that the overall system in Fig. and the methods employed for their control.1 Consider the case of an amplifier cascaded with a linear system as shown in Fig. A3. it is often sufficient to consider a system to be linear. the response of that system will depend on the type and magnitude of the input signal applied.e l < e ( t ) < + e l . In this appendix a brief overview of nonlinear systems.
3 Examples of nonlinearities (d) Relay (hysteresh + dead zone) . A3. A3.Appendix 3 W. amplifier input Fig. transducer errors and gear backlash are examples of inherent nonlinearities. motor stiction.2 Nonlinear amplifier characteristics (saturation) Amplifier saturation. amplifier output 347 t I I + el t e(0. These are nonlinear effects which are part of a system we wish Output T (a) Dead zone I (b) Diode Output I i (c) Friction I Fig.
348 Appendices to control. as an example. ) ) Let u l ( t ) = 3 and u l ( t ) = 4 .2 As an example of an operational nonlinearity. Similarly. especially when high accuracy is not a requirement. A speed control system. A3.the cost of such an ‘improvement’ being a significant reduction in system gain. However. this relationship can be made to look like that in Fig. Another type of nonlinearity is encountered simply by multiplying together a number of linear system signals. then u l ( t ) + u z ( t ) must produce y ~ ( t+ y ~ ( t ) .u (a) Openloop Fig.4 Speed control system characteristics (b) Closedloop .3. in practice. may well have an input voltage to output speed relationship such as that in Fig.4(b) simply by selecting a suitable proportional feedback gain. we must also have that if u l ( t ) produces y l ( t ) and v ~ ( t produces y z ( t ) . result in the linearization of an openloop nonlinear system. if an input u l ( t ) produces an output y l ( t ) . Also. in which both the input signals are variable. in this case. Indeed. However. output speed output speed Control input. In general. the application of a feedback control loop can. and an input u z ( t )produces an output yz(t). From the definition of a linear system. an example being that shown in Fig. A3. So. consider the linear system y~( t ) = u t ( t ) u t ( t ) . this is termed an operational nonlinearity and is explained by means of the following example. effects which are neglected. The result of employing feedback. let u z ( f ) = 2 and u z ( t ) = 5.2 with further examples given in Fig. Example A3. such that y l ( t ) = 12. is to drastically reduce the effect of nonlinearities .4(a) when in openloop. they are effects which can be represented by means of an input to output nonlinear system function. then an input U I( t ) + u z ( t ) must produce an output y l ( t )+ yz(t). in many cases. u Desired input. therefore for two different input signals yz(t) = uz(f) u z ( t ) . A3. A3.  Inherent nonlinearities are very often. and ) y l ( t )+ y z ( t ) = 22 does not result from the multiplication of the sums. u t ( t ) + U Z ( ~ = 5 and u l ( t ) + u z ( t ) = 9. such that y 2 ( t ) = 10. A3.
the aim being either to simplify system design or merely to meet specific performance requirements.2. A3. Example A3.1 Analysis at the extremes The system is investigated by looking only at the extreme values (highest and lowest) of each of the variable parameters. only positive values need be considered for the amplifier Fig. Consider the following example. it is often possible t o linearize the description in a way which produces roughly equivalent results. overall specifications however still remain in terms of stability requirements. the simple onoff type of control applied with a domestic thermostat. in which the variable parameter is the nonlinear amplifier stage. Where a description of a particular nonlinearity is quite complicated.Appendix 3 349 When designing a controller for certain systems. it can be useful t o introduce an intentional nonlinearity. with el = 1. possibly in terms of several differential equations. Because u ( t ) is a step input. e. A3. A3.5. and u ( t ) = 1 when el = 1 = e ( f ) .3 Consider the system shown in Fig. then it is probably best to employ a computer in order to find the equation solutions necessary. The characteristics of the nonlinear amplifier are shown in Fig. steadystate values and transient response tailoring. Analysis and design of nonlinear control systems can take on one of several forms dependent o n the extent and type of nonlinearities present. Several approaches to equivalent linearization are now discussed. An example of such a nonlinearity is a closedloop system in which the control action is only applied when the output exceeds either an upper or lower limit. following the application of a 10 unit step input. I f the nonlinear effects are more straightforward though.g. from v ( t ) = 0 t o u ( t ) = 10.5 Cascade nonlinear amplifier for Example A3. A3. in which a n amplifier is cascaded with the linear system G ( s )= K ~ s+ 1 where the gain K > 0 must be selected t o ensure that the steadystate output value is of magnitude 10 units 2 5 % . whether they be linear o r nonlinear themselves.3 .
A3. Immediately following the application of the 10 unit step input. Two extremes are then of interest: (i) O < e ( / l ) < 1. the error e ( r ) will be less than I and the output signal will increase from 9 to 10 units following the equation found in (i). The steadystate output value will therefore be and if it is required that 9. is shown in Fig. u ( t ) = 1 from f = 0 and Y = GU. a value of K = 25 will suffice.6 Time response of the system in Example A3. Once y ( t ) has reached a value of 9 units.350 Appendices characteristics.5 < IOK/(I + K ) .5. A3. the output signal y ( / )will start to increase from zero as in (ii). A3. when the error e ( t ) will be greater than unity.3 . Thus (ii) e ( r ) > 1. the overall response of the system shown in Fig.6.e. With this value of K . with 1=0 I 1 =o Fig. u ( t l ) = e ( t l ) f r o m 1 1 = O ( w h e n y ( t 1 ) = 9 : f 1 = 0 ) . i. following a 10 unit step input.
Typical amplifier characteristics are shown in Fig. the describing function of a nonlinear element can be found by obtaining the first harmonic component of the Fourier series which represents the nonlinear . is approximated by assuming that only the first harmonic (fundamental) component of its output response is significant (employing Fourier series analysis).) so where Usti and Vs0 Y + 7 = uv + ti v + ui. The describing function of a nonlinear system element is the complex ratio of the first harmonic components. Consider the effect of operational nonlinearities.3 Describing functions In the describing function method the isolated nonlinear element. in the system under consideration. U t j + But the product ti0 can readily be neglected. A3. A3. such that Y+y=(U+ts)(V+i. leaving Y = uv with an error y=liv+uo and this errorg is the error incurred by assuming that the operational nonlinearity is approximately linear. for example see Fig. the describing function is N = 'U h E I where 41 is the phase shift exhibited by the first harmonic component of u ( t ) .2. A3.7.Appendix 3 A3. a basic assumption being made that the system under consideration is more or less linear at the operating point and any nonlinear effects are negligible.7. Let both U and V vary by small amounts a and tj respectively from the operating point. Consider the effect of applying a sinusoidal input to a nonlinear amplifier with saturation. V. such as those encountered about the fixed operating point Y = U. In general.2 Analysis about a fixed operating point 35 1 Linearization of a nonlinear function about a fixed operating point is a commonly encountered procedure in practice. A3. In the amplifier considered in Fig.
A3. I ) .9. the describing function itself then is the ratio of the output (first harmonic) t o input sine wave. As long as harmonics of second order and above are sufficiently attenuated. the closedloop system transfer function is given by (A3. Period = T seconds Fig. A3. the describing function of the ideal relay shown in Fig.7 Input and output signals for a saturating amplifier output I Input lM Fig. A unity feedback control system which contains one o r more nonlinear elements can then be represented as shown in Fig.352 Appendices I (a) Sinusoidal input e ( / )= Esin w i Period = T seconds I (b) Nonsinusoidal output u ( t ) = U I sin(of + 41)sin 2wt + U2 sin(2wt + 62) . A3. the describing function can be calculated by accounting for the first harmonic due to each. etc. As a further example.8 Ideal relay characteristics elements’ output response. Where several nonlinear elements are present in a system... A3.8 is where E is the peak magnitude of the input applied t o the relay.
3. The characteristic equation of the closedloop system (A3.4) or [G(jo)]'= A similar stability analysis to that carried out for Nyquist design (see Sections 5.' as an inverse Nyquist plot. then the unity feedback closedloop system is stable. 2. I).1/N or .3) (A3.1/N locus.1 + j 0 being of importance. rather than the point . then the unity feedback closedloop system output may well exhibit a limit cycle. Note: A limit cycle is a sustained oscillation at a fixed frequency and with constant amplitude. given an openloop transfer function G ( j w ) and a describing function N. as is the case with the straightforward Nyquist plot. If the complete G ( j w ) locus does enclose the . or by plotting [ G ( j w ) ] . the locus of . If the G ( j w ) and Nloci intersect.1) is then 1+NG=O which is 1 G =  N N (A3. 1. However. Stability criteria If the complete G ( j w ) locus does not enclose the .2) and K > 0 is a scalar gain.1/N locus. Stability properties can then be considered either in terms of a normal Nyquist plot of G ( j w )over the range of frequencies w . 5 ) can be employed to investigate the stability properties of a system defined by (A3.Appendix 3 353 Fig. A3.It must be remembered that in general G ( j w ) and N are characterized by both magnitude and phase components.4 and 5 . the unity feedback closedloop system output will increase to some finite limiting value (possibly caused by saturation). .9 Unity feedback system with nonlinear element where (A3.N respectively is of importance.
and E will continue to increase along the .1/N loci intersect. 5 . If the signal E is slightly t o the left of that at A then this is within the stable region (ii). the G ( j w ) and .10.354 Appendices hag. At the point y. if E is slightly to the right of that at X then this is within the unstable region (i). where E is the peak magnitude of the signal input to the nonlinear element with At the point X. However. and E will continue to decrease. However. t Real Fig. any variation in E from its value at that point being sufficient to cause further movement along the . A3.l/Nloci intersect. If the signal E is slightly to the left of that at y then this is within the stable region (ii).10 Nyquist plot with  1 / N locus Consider the plot in Fig. system performance can be modified to meet specifications made by reshaping the G ( j w ) loci with a suitable choice of linear compensator. if E is slightly to the right of that at y then this is within the unstable region (i). A3. The limit cycle at the point y is called a stable limit cycle. and E will tend to decrease back towards its value at y.1/N loci towards the point y. any variation in E from its value at that point resulting in a tendency for E to return to its value at y.4 Consider the openloop transfer function: . and E will tend to increase towards its value at y.1/Nlocus away from X. The limit cycle at the point X is called an unstable limit cycle. ~~ Example A3. the G ( j o )and . In a similar fashion to the studies carried out for Nyquist design procedure in Section 5 .
the methods then give stability results for a unity feedback closedloop system. (b) Bode.1. and that simply by decreasing the gain K by a suitable amount. This is true of the continuoustime tests described in Chapters 4 and 5 .= ' j w (j w + 2) 10 for K= 1 The frequency response of the inverse function is shown plotted in Fig.' locus n o longer intersects with the . (c) Nyquist and (d) Nichols.N locus. A4.' is increased over all frequencies. resulting in a stable unity feedback closedloop system when the [ G ( j w ) ] . A3. A4.N IOCUS Then [ G ( j w )] . For each of these it is necessary t o consider initially the openloop transfer function. the magnitude of [ G ( j w ) ] .a function of s o r j w as appropriate.Appendix 4 hag. one stable and one unstable. Appendix 4 Unity feedback loops A large number of tests employed t o study the stability of a closedloop system require that the system contains a unity feedback loop. such than when connected within a unity feedback loop. the Fig. It can be seen that two limit cycles exist.1 Unity feedback loop * . I 1 along with the frequency response of a nonlinear element N. namely (a) root loci. as shown in Fig. 355 . Consider the openloop transfer function t o be GO.
Go =I + Go (A4.2. So I+GH= BB + K K A A =O BB Stability of the general closedloop system can then be investigated by considering the roots of BB + KKAA = o (A4.1) how the general form of Fig. A 4 2 Appendices General closedloop system closedloop transfer function is given as G.3 56 c Fig. A4. this means that the equivalent GOcan be found from (A4.3) as: .3) As far as stability of the closedloop system is concerned.4) Let A G=KB where A . The relationship is found to be: 1 + G [ H .4. A and B are polynomials. it is the roots of the characteristic equation which are of importance.1) In practice however. In the general case the roots are obtained from: 1 + GH=O and .A H=KB  (A4. A4.2 can be reconsidered in terms of a unity feedback loop. the closedloop system may well take o n the more general feedback form shown in Fig. The closedloop transfer function in the more general case is then: (A4.5) With G and H a s defined. where G. B . by equating GOwith G and H.2) It was shown in Chapter 3 (Section 3.11 G (A4. K and K are constants.
For the stability tests discussed in Chapters 4 and 5 it is therefore sufficient to plot the characteristics of GH. as though it were in fact the openloop transfer function itself. subsequently to plot the characteristics of this new function.6) The resultant characteristic equation is therefore the same. (A4.6) is identical to (A4.Appendix 4 357 So the stability of the equivalent unity feedback system can be investigated by considering the roots of which are found from BB+MAA=Q (A4. on the appropriate diagram.e. Hence stability results will be the same if G H i s employed. and . i. whether the general system characteristic equation 1 + GH = 0 or the equivalent unity feedback characteristic equation 1 + GO= 0 is employed.5). rather than to firstly find the equivalent GO.
.
140 decomposition. 284 asymptotes Bode. 190 biological control. 211. 336 compensation Bode. W. 99 bandwidth. 279 control systems. 307 CayleyHamilton. 44. H. 209 capacitor. 41 w e ofarrival. 336 diagonalization.206 polynomial. 226 convolution integral. 298 continuoustime. 107 break frequency. 275 analogous systems. 143 critical damping. 93 branches. 331 determinant. 33542 analog models. 184 Nyquist. 21I . 147 DC motor. 73 acluator. 337 deterministic. 107 attenuation. 109 canonical form. 108 of departure. 95. 251 derivative. 170 root locus. 256 firstorder. 28 1 adjoint of a matrix.. 250 adaptive control. 159. 146 computer control. 46. 27. 108 armature motor control. 21 comer frequency. 125 complex roots. 149 digital. 11 control system. 205 closedloop control. 250 damping factor. 126 Nichols. 104. 338 algebra of matrices. 70. 84 artificial intelligence. 223 test. 143 breakaway point. 140 automatic control history. 295 characteristic equation. 15 controllability. 9 block diagram.Index acceptable damping. 5 Bode diagram. 67 stability. 43 deadzone. 207 centre of area. 13959 bounded input/output. 75 column matrix or vector. 96 DIA converter. 3 A/D converter. 35 cascade compensation. 25 Bode. 6. 309 operator. 81 diagonal matrix. 68 response. 35 auxiliary polynomial. 274 computeraided design. 143 mot loci. 213 defuzzification. 347 decibel. 320 describing function. 290 delay. 95. 94.
16. 163 Nyquist plot. 167. 172. J. 125 Maxwell. 313 integrator saturation. 4 mean of maxima. 151 mass. 203. 300 linearity. 73 signal. then. 35 Hunvitz. 19 index (of performance). 228 inverse Laplace. 34655 nonsingular. 42 errors general. 71. 7. C. 104 margin (gaidphase). 225 value theorem. 158 margin. 2. 347 ideal damping. 336 Nyquist criterion. 184 crossover point. 6 frequencies. 24973 direct decomposition. 11. 13796 friction. 11 field control. 336 if . 33 impulse response. 22 least squares. 152. 72. 353 linear approximation.126 Fourier. 127 criterion. 160 difference operator. 45. 287 impedance. 435 multivariable system. 15. 290 models. 290 fuzzy logic control. 43. 46 history of control. 81 nonlinear. 203 . 46 fmal value theorem. 198 maximum overshoot. 206.360 Index state. 281 discretetime.. 34 inertia. 163 Nichols chart. 261 disturbance. 250 mechanical systems. 32 motor. 347 fuzzification. 4 frequency response. 151. 252 firstorder system. 2 fault detection. 186 Laplace operator. 17 limit cycle. 81 logarithmic plot. 285 lagllead compensation. 206 electromechanical. 154. 153. 11. 45. 80 eigenvalues. 6 integral control. 253 input signal. 7. 200 digital control. 284 exponential response. 211 eigenvector. 139 magnitude compensation. 33542 state representations. 295 gain compensation. 24. 17. 49 natural frequency. 297 invariant theorem. 12. 170.. 95102 hysteresis. 2 inductor. 18 knowledge engineer. 288 mean. 37 matrix algebra. 179 gears. 337 null matrix. DC. 202 Nyquist. 72 identity matrix. 40 inference engine.. 25 1 differential operator. 295 measurement. 340 Jordan form. 287 initial conditions. 280 feedback. 215 digital controllers (DDC). 355 inversion (matrix). 147 negative feedback. 35 membership function. 10. 22 load disturbance. 11 expert systems. 17791 noise. 212 Kirchhoff. 84 generator. 64.
275 singular. 60 roots. 202 variables. 86 . 197 steadystate error. 12. 5862 threetern controller. 336 system identification. 47. 121 sensors. 151. 337 sinusoidal input. 121 rotational motion. 264 poles. 250 servomechanism. 4. 74 output. 57. 210 plant. 6 regulator. 16. 117 response. 137. 249 sampling period. 282 simulation of control systems. 240 openloop control. 10231 sensitivity. 82. 93. 37 square matrix. 101. 10 polar plot see Nyquist plot pole placement. 209 transient error. 63 random signal. 178 relational systems. 759 response. 159. 26 terminology. 24. 26 symmetric matrix. 223 test. 336 361 rulebased controller. 71 step error. 230 polynomial function. 281 reference input. 233. 253. 73 relative stability. 93. 79 response. 58. 31930 parabolic input. 179 variable form. 66 overshoot. 95102 row matrix (vector). 127 criterion. 65 delay (transport lag). 60. 236 order of a system. 240. 139 software models. 10 overdamped response. 62. 261 statespace.63 partial fraction decomposition. 57. 293 relay. 350 varying control. 347 secondorder system. 347 resistance. 291 sampleddata system. 31219 P + I + D. 81 realtime control. 104 polezero cancellation. 116 proportional control. 337 transfer functions. 21 trace. 239 feedback. 33 response time. 70 root loci. 5. 30634 time constant. 76. 348 8plane. 65. 36 summing (point). 282 saturation. 151 lead. 10 test input signals. 306. 94. 68 Routh array. 336 stability. 198. 65 sensitivity. 22243 transition matrix. 137 compensation. 165 observability. 93. 276 speed control. 70 simulated application. 95. 79 input. 93 phase angle. 307 quantization errors. 299 takeoff point. 197 spring. 353 state P + D. 151 margin. 65 rise time.Index Nyquist stability. 59. 227 observer. 5 optimal control. 63 stiffness. 200 transformation of state vector. 213 performance. 77. 297 ramp error.3202 P + I. 847 set point see reference input settling time. 167. 62. 168. 80 input. 39 estimation. 61. 104 lag.
237 zero order hold. 359 transpose. 202 translational motion. James. 58 response. 254 ztransforms.362 Index unobservable. 231 unstable. 68 unit circle. 58 . 20 step input. 93 velocity feedback. 73 uncontrollable. 324 zplane. 3 weighting matrix. 254 impulse. 255 zeros. 336 type. 25 1 transition matrix. 320 Watt. 93. 104 ZieglerNichols. 23 1 undamped response.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.