Vibration Control of Active Structures: An Introduction (Solid Mechanics and Its Applications)

Vibration Control of Active Structures Edition SOLID MECHANICS AND ITS APPLICATIONS Volume 96 Series Editor: G.M.L. ...

Author: A. Preumont

97 downloads 1026 Views 13MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

Vibration Control of Active Structures Edition

SOLID MECHANICS AND ITS APPLICATIONS Volume 96 Series Editor:

G.M.L. GLADWELL Department of Civil Engineering University of Waterloo Waterloo, Ontario, Canada N2L 3GI

Aims and Scope of the Series

The fundamental questions arising in mechanics are: Why?, How?, and How much? The aim of this series is to provide lucid accounts written by authoritative researchers giving vision and insight in answering these questions on the subject of mechanics as it relates to solids. The scope of the series covers the entire spectrum of solid mechanics. Thus it includes the foundation of mechanics; variational formulations; computational mechanics; statics, kinematics and dynamics of rigid and elastic bodies: vibrations of solids and structures; dynamical systems and chaos; the theories of elasticity, plasticity and viscoelasticity; composite materials; rods, beams, shells and membranes; structural control and stability; soils, rocks and geomechanics; fracture; tribology; experimental mechanics; biomechanics and machine design. The median level of presentation is the first year graduate student. Some texts are monographs defining the current state of the field; others are accessible to final year undergraduates; but essentially the emphasis is on readability and clarity.

For a list of related mechanics titles, see final pages.

Vibration Control of Active Structures An Introduction Edition

by

ANDRÉ PREUMONT Université Libre de Bruxelles, Active Structures Laboratory, Brussels, Belgium

KLUWER ACADEMIC PUBLISHERS NEW YORK, BOSTON, DORDRECHT, LONDON, MOSCOW

eBook ISBN: Print ISBN:

0-306-48422-6 1-4020-0496-6

©2004 Kluwer Academic Publishers New York, Boston, Dordrecht, London, Moscow Print ©2002 Kluwer Academic Publishers Dordrecht All rights reserved

No part of this eBook may be reproduced or transmitted in any form or by any means, electronic, mechanical, recording, or otherwise, without written consent from the Publisher

Created in the United States of America

Visit Kluwer Online at: and Kluwer's eBookstore at:

http://kluweronline.com http://ebooks.kluweronline.com

” . . . le travail éloigne de nous trois grands maux: l'ennui, le vice et le besoin. ” Voltaire, Candide (XXX)

This page intentionally left blank

Contents xv

Preface to the second edition Preface to the first edition

xvii

1

Introduction 1.1 Active versus passive 1.2 Smart materials and structures 1.3 Control strategies 1.3.1 Feedback 1.3.2 Feedforward 1.4 The various steps of the design 1.5 Organization of the book 1.6 References 1.7 Problems

1 1 5 7 7 9 10 12 13 15

2

Some concepts of structural dynamics 2.1 Equation of motion of a discrete system 2.2 Vibration modes 2.3 Modal decomposition 2.3.1 Structure without rigid body modes 2.3.2 Structure with rigid body modes 2.3.3 Example 2.4 Transfer function of collocated systems 2.5 Continuous structures 2.6 Guyan reduction 2.7 References 2.8 Problems

17 17 18 20 20 22 25 26 30 31 33 34

3 Actuators, piezoelectric materials, and active structures 3.1 Introduction 3.2 Proof-mass actuator 3.3 Reaction wheels and gyrostabilizers vii

37 37 38 40

Vibration control of active structures

viii 3.4

3.5 3.6 3.7 3.8

3.9 3.10 3.11

Piezoelectric actuators 3.4.1 Constitutive equations 3.4.2 Linear actuator 3.4.3 Laminar actuator 3.4.4 Laminar sensor 3.4.5 Example: Tip displacement of a cantilever beam 3.4.6 Spatial modal filters Passive damping with piezoceramics Active cantilever beam Active truss Piezoelectric shell 3.8.1 Two-dimensional constitutive equations 3.8.2 Kirchhoff shell Finite element formulation References Problems

4 Collocated versus non-collocated control 4.1 Introduction 4.2 Pole-zero flipping 4.3 Collocated control 4.4 Non-collocated control 4.5 Notch filter 4.6 Pole-zero flipping in the structure 4.7 Effect on the Bode plots 4.8 Relation to the mode shapes 4.9 The role of damping 4.10 References 4.11 Problems 5 Active damping with collocated pairs 5.1 Introduction 5.2 Direct Velocity Feedback 5.2.1 Lead compensator 5.3 Acceleration feedback 5.3.1 Direct Velocity Feedback 5.3.2 Second order filter 5.3.3 SISO system with many modes 5.3.4 Multidimensional case 5.4 Positive Position Feedback 5.4.1 SISO system 5.4.2 Multidimensional case 5.5 Integral Force Feedback 5.5.1 Modal damping

40 41 46 48 51 52 53 54 55 58 63 64 65 68 70 73 75 75 76 78 80 80 82 84 84 87 88 88 91 91 93 94 96 97 97 98 100 101 101 102 103 104

CONTENTS

6

ix

5.6

Remarks 5.6.1 Controllability, observability 5.6.2 Actuator and sensor dynamics 5.7 References 5.8 Problems

107 107 107 109 110

Active vibration isolation 6.1 Introduction 6.2 Passive isolator 6.3 The “sky-hook” damper 6.4 Force feedback 6.5 Flexible clean body 6.5.1 Free-free beam with isolator 6.6 6 d.o.f. isolator 6.7 Decentralized control of the 6 d.o.f. isolator 6.7.1 Remarks Pointing control 6.8 Vehicle suspension 6.9 6.10 References 6.11 Problems

113 113 114 116 118 119 121 124 127 128 129 131 133 134

7 State space approach 7.1 Introduction 7.2 State space description 7.2.1 Single degree of freedom oscillator 7.2.2 Flexible structure 7.2.3 Inverted pendulum 7.3 System transfer function 7.3.1 Poles and zeros 7.4 Pole placement by state feedback 7.4.1 Example: oscillator 7.5 Linear Quadratic Regulator 7.5.1 Symmetric root locus 7.5.2 Inverted pendulum 7.6 Observer design 7.7 Kalman Filter 7.7.1 Inverted pendulum 7.8 Reduced order observer 7.8.1 Oscillator 7.8.2 Inverted pendulum 7.9 Separation principle 7.10 Transfer function of the compensator 7.10.1 The two-mass problem 7.11 References

137 137 139 139 140 141 142 144 145 146 148 149 149 151 153 154 155 156 156 157 158 159 162


x

8

7.12 Problems

163

Analysis and synthesis in the frequency domain 8.1 Gain and phase margins 8.2 Nyquist criterion 8.2.1 Cauchy’s principle 8.2.2 Nyquist stability criterion 8.3 Nichols chart 8.4 Feedback specification for SISO systems 8.4.1 Sensitivity 8.4.2 Tracking error 8.4.3 Performance specification 8.4.4 Unstructured uncertainty 8.4.5 Robust performance and robust stability 8.5 Bode gain-phase relationships 8.6 The Bode Ideal Cutoff 8.7 Non-minimum phase systems 8.8 Usual compensators 8.8.1 System type 8.8.2 Lead compensator 8.8.3 PI compensator 8.8.4 Lag compensator 8.8.5 PID compensator 8.9 References 8.10 Problems

165 165 166 166 167 170 171 171 172 173 174 175 178 181 183 185 185 187 187 189 189 189 190

9 Optimal control 9.1 Introduction 9.2 Quadratic integral 9.3 Deterministic LQR 9.4 Stochastic response to a white noise 9.4.1 Remark 9.5 Stochastic LQR 9.6 Asymptotic behaviour of the closed-loop 9.7 Prescribed degree of stability 9.8 Gain and phase margins of the LQR 9.9 Full state observer 9.9.1 Covariance of the reconstruction error 9.10 Kalman-Bucy Filter (KBF) 9.11 Linear Quadratic Gaussian (LQG) 9.12 Duality 9.13 Spillover 9.13.1 Spillover reduction 9.14 Loop Transfer Recovery (LTR)

193 193 193 194 196 197 197 198 200 201 202 204 204 205 205 206 209 210

CONTENTS

xi

9.15 Integral control with state feedback 9.16 Frequency shaping 9.16.1 Frequency-shaped cost functionals 9.16.2 Noise model 9.17 References 9.18 Problems

211 212 212 215 216 217

10 Controllability and Observability 10.1 Introduction 10.1.1 Definitions 10.2 Controllability and observability matrices 10.3 Examples 10.3.1 A cart with two inverted pendulums 10.3.2 Double inverted pendulum 10.3.3 Two d.o.f. oscillator 10.4 State transformation 10.4.1 Control canonical form 10.4.2 Left and right eigenvectors 10.4.3 Diagonal form 10.5 PBH test 10.6 Residues 10.7 Example 10.8 Sensitivity 10.9 Controllability and observability Gramians 10.10 Relative controllability and observability 10.10.1 Internally balanced coordinates Model reduction 10.11 10.11.1 Transfer equivalent realization 10.11.2 Internally balanced realization 10.11.3 Example 10.12 References 10.13 Problems

221 221 222 222 224 224 226 226 227 228 229 230 230 231 232 233 234 235 235 237 237 237 238 240 241

11 Stability 11.1 Introduction 11.1.1 Phase portrait 11.2 Linear systems 11.2.1 Routh-Hurwitz criterion 11.3 Liapunov’s direct method 11.3.1 Introductory example 11.3.2 Stability theorem 11.3.3 Asymptotic stability theorem 11.3.4 Lasalle’s theorem 11.3.5 Geometric interpretation

245 245 246 247 248 249 249 249 251 251 252


xii

11.4 11.5 11.6 11.7 11.8 11.9

11.3.6 Instability theorem Liapunov functions for linear systems Liapunov’s indirect method An application to controller design Energy absorbing controls References Problems

253 254 255 256 257 259 259

12 Semi-active control 12.1 Introduction 12.2 Magneto-rheological (MR) fluids 12.3 MR devices 12.4 Semi-active control 12.5 Open-loop control 12.6 Feedback control 12.6.1 Continuous control 12.6.2 On-off control 12.6.3 Force feedback 12.7 References 12.8 Problems

263 263 264 266 267 268 268 268 273 274 275 275

13 Applications 13.1 Digital implementation 13.1.1 Sampling, aliasing and prefiltering 13.1.2 Zero-order hold, computational delay 13.1.3 Quantization 13.1.4 Discretization of a continuous controller 13.2 Active damping of a truss structure 13.2.1 Actuator placement 13.2.2 Implementation, experimental results 13.3 Active damping generic interface 13.3.1 Active damping 13.3.2 Experiment 13.3.3 Pointing and position control 13.4 Active damping of a plate 13.4.1 Control design 13.5 Active damping of a stiff beam 13.5.1 System design 13.6 The HAC/LAC strategy 13.6.1 Wide-band position control 13.6.2 Compensator design 13.6.3 Results 13.7 Volume displacement sensors 13.7.1 QWSIS sensor

277 277 278 279 280 281 282 282 284 286 287 288 290 290 291 293 294 295 297 299 299 302 303

CONTENTS 13.7.2 Discrete array sensor 13.7.3 Spatial aliasing 13.7.4 Distributed sensor 13.8 References 13.9 Problems 14 Tendon Control of Cable Structures 14.1 Introduction 14.2 Tendon control of strings and cables 14.3 Active damping strategy 14.4 Basic Experiment 14.5 Approximate linear theory 14.6 Application to space structures 14.6.1 Guyed truss experiment 14.6.2 JPL-MPI testbed 14.6.3 Free floating truss experiment 14.6.4 Microvibrations 14.7 Application to cable-stayed bridges 14.7.1 Laboratory experiment 14.7.2 Control of parametric resonance 14.7.3 Large scale experiment 14.8 References

xiii 305 308 309 316 319

321 321 323 324 325 327 329 329 331 331 334 335 335 336 338 344

Bibliography

347

Index

361


Preface to the second edition

My objective in writing this book was to cross the bridge between the structural dynamics and control communities, while providing an overview of the potential of SMART materials for sensing and actuating purposes in active vibration control. I wanted to keep it relatively simple and focused on systems which worked. This resulted in the following: (i) I restricted the text to fundamental concepts and left aside most advanced ones (i.e. robust control) whose usefulness had not yet clearly been established for the application at hand. (ii) I promoted the use of collocated actuator/sensor pairs whose potential, I thought, was strongly underestimated by the control community. (iii) I emphasized control laws with guaranteed stability for active damping (the wide-ranging applications of the IFF are particularly impressive). (iv) I tried to explain why an accurate prediction of the transmission zeros (usually called anti-resonances by the structural dynamicists) is so important in evaluating the performance of a control system. (v) I emphasized the fact that the open-loop zeros are more difficult to predict than the poles, and that they could be strongly influenced by the model truncation (high frequency dynamics) or by local effects (such as membrane strains in piezoelectric shells), especially for nearly collocated distributed actuator/sensor pairs; this effect alone explains many disappointments in active control systems. The success of the first edition confirmed that this approach was useful and it is with pleasure that I accepted to prepare this second edition in the same spirit as the first one. The present edition contains three additional chapters: chapter 6 on active isolation where the celebrated “sky-hook” damper is revisited, chapter 12 on semi-active control, including some material on magneto-rheological fluids whose potential seems enormous, and chapter 14 on the control of cable-structures. It is somewhat surprising that this last subject is finding applications for vibration amplitudes which are nine orders of magnitude apart (respectively meters for large cable-stayed bridges and nanometers for precision space structures). Some material has also been added on the modelling of piezoelectric structures (chapter 3) and on the application of distributed sensors in vibroacoustics (chapter 13). I am deeply indebted to my coworkers, particularly Younes Achkire and Frédéric Bossens for the cable-structures, Vincent Piefort for the modelling of piezoelectric structures, Pierre De Man and Arnaud François in vibroacoustics, Ahmed Abu Hanieh and in active isolation and, last but not least, Nicolas Loix and Jean-Philippe Verschueren who run with enthusiasm and competence our spin-off company, Micromega Dynamics. I greatly xv


xvi

enjoyed working with them, exploring not only the concepts and the modelling techniques, but also the technology to make these control systems work. I also express my thanks to David de Salle who did all the editing, and to the Series Editor, Prof. Graham Gladwell who, once again, improved my English. André Preumont Brussels, November 2001.

Preface to the first edition

I was introduced to structural control by Raphaël Haftka and Bill Hallauer during a one year stay at the Aerospace and Ocean Engineering department of Virginia Tech., during the academic year 1985-1986. At that time, there was a tremendous interest in large space structures in the USA, mainly because of the Strategic Defense Initiative and the space station program. Most of the work was theoretical or numerical, but Bill Hallauer was one of the few experimentalists trying to implement control systems which worked on actual structures. When I returned to Belgium, I was appointed at the chair of Mechanical Engineering and Robotics at ULB, and I decided to start some basic vibration control experiments on my own. A little later, SMART materials became widely available and offered completely new possibilities, particularly for precision structures, but also brought new difficulties due to the strong coupling in their constitutive equations, which requires a complete reformulation of the classical modelling techniques such as finite elements. We started in this new field with the support of the national and regional governments, the European Space Agency, and some bilateral collaborations with European aerospace companies. Our Active Structures Laboratory was inaugurated in October 1995. In recent years, with the downsizing of the space programs, active structures seem to have lost some momentum for space applications, but they gave birth to interesting spin-offs in various fields of engineering, including the car industry, machine tools, consumer products, and even civil engineering. I believe that the field of SMART materials is still in its infancy; significant improvements can be expected in the next few years, that will dramatically improve their recoverable strain and their load carrying capability. This book is the outgrowth of research work carried out at ULB and lecture notes for courses given at the Universities of Brussels and Liège. I take this opportunity to thank all my coworkers who took part in this research, particularly Jean-Paul Dufour, Christian Malekian, Nicolas Loix, Younes Achkire, Paul Alexandre and Pierre De Man; I greatly enjoyed working with them along the years, and their enthusiasm and creativity have been a constant stimulus in my work. I particularly thank Pierre who made almost all the figures. Finally, I want to thank the Series Editor, Prof. Graham Gladwell who, as he did for my previous book, read the manuscript and corrected many mistakes in my English. His comments have helped to improve the text. André Preumont Bruxelles, July 1996. xvii


Chapter 1

Introduction 1.1 Active versus passive Consider a precision structure subjected to varying thermal conditions; unless carefully designed‚ it will distort as a result of the thermal gradients. One way to prevent this is to build the structure from a thermally stable composite material; this is the passive approach. An alternative way is to use a set of actuators and sensors connected by a feedback loop; such a structure is active. In this case‚ we exploit the main virtue of feedback‚ which is to reduce the sensitivity of the output to parameter variations and to attenuate the effect of disturbances within the bandwidth of the control system. Depending on the circumstances‚ active structures may be cheaper or lighter than passive structures of comparable performances; or they may offer performances that no passive structure could offer‚ as in the following example. A few years ago‚ the general belief was that atmospheric turbulence would constitute an important limitation to the resolution of earth based telescopes; this was one of the main reasons for developing the Hubble space telescope. Recently‚ it has been demonstrated that it is possible to correct in real time the disturbances produced by atmospheric turbulence on the optical wave front coming from celestial objects; this allows us to improve the ultimate resolution of the telescope by one order of magnitude‚ to the limit imposed by diffraction. The correction is achieved by a deformable mirror coupled to a set of actuators (Fig.1.1). A wave front sensor detects the phase difference in the turbulent wave front and the control computer supplies the shape of the deformable mirror which is required to correct this error; the time slice required to perform the computations is so small that the atmospheric turbulence can be considered as frozen. Prototypes of such control systems have recently been tested successfully for ground-based astronomy. The foregoing example is not the only one where active structures have 1

2


proved beneficial to astronomy; another example is the primary mirror of large telescopes, which can have a diameter of 8 m or more. Large primary mirrors are very difficult to manufacture and assemble. A passive mirror must be thermally stable and very stiff, in order to keep the right shape in spite of the varying gravity loads during the tracking of a star, and the dynamic loads from the wind. There are two alternatives to that, both active. The first one, adopted on the Very Large Telescope (VLT) at ESO in La Silla, Chile, consists of having a relatively flexible primary mirror connected at the back to a set of a hundred or so actuators (e.g. Wilson et al.). As in the previous example, the control system uses an image analyser to evaluate the amplitude of the perturbation of the optical modes; next, the correction is computed to minimize the effect of the perturbation and is applied to the actuators. The influence matrix J between the actuator forces and the optical mode amplitudes of the wave front changes can be determined experimentally with the image analyser:

J is a rectangular matrix‚ because the number of actuators is larger than the number of optical modes of interest. Once the modal errors have been

Introduction

3

evaluated‚ the correcting forces can be calculated from

where is the pseudo-inverse of the rectangular matrix J. This is the minimum norm solution to Equ.(l.l) (Problem P. 1.1). The second alternative‚ adopted on the Ten-Meter Telescope (TMT) of the Keck observatory at Mauna Kea‚ Hawaii‚ consists of using a segmented primary mirror. The potential advantages of such a design are lower weight‚ lower cost‚ ease of fabrication and assembly. Each segment has a hexagonal shape and three computer controlled degrees of freedom (2 pointing and 1 piston); the control system is used to achieve the optical quality of a monolithic mirror‚ to compensate for wind disturbances and minimize the impact of the telescope dynamics on the optical performance (Aubrun et al.). As a third example‚ also related to astronomy‚ consider the future interferometric missions planned by NASA or ESA (one such a mission‚ called ”Terrestrial Planet Finder” aims at detecting earth-sized planets outside the solar system; other missions include the mapping of the sky with an accuracy one order better than that achieved by Hypparcos). The aim is to use a number of smaller telescopes as an interferometer to achieve a resolution which could only be achieved with a much larger monolithic telescope. One possible spacecraft architecture for such an interferometric mission is represented in Fig.1.2; it consists of a main truss supporting a set of independently pointing telescopes. The relative positions of the telescopes are monitored by a sophisticated metrology and the optical paths between the individual telescopes and the beam combiner are accurately controlled with optical delay lines‚ based on the information coming from a wave front sensor. Typically‚ the distance between the telescopes could be 50 or more‚ and the order of magnitude of the error allowed on the optical path length is a few nanometers; the pointing error of the individual telescopes is as low as a few nanoradians (i.e. one order of magnitude better than the Hubble space telescope). Clearly‚ such stringent geometrical requirements cannot be achieved with a precision monolithic structure‚ but rather by active means as suggested in Fig.1.2. Let us first consider the supporting truss: given its size and environment‚ the main requirement on the supporting truss is not precision but stability‚ the accuracy of the optical path being taken care of by the wide-band vibration isolation/steering control system of individual telescopes and the optical delay lines (described below). Geometric stability includes thermal stability‚ vibration damping and prestressing the gaps in deployable structures (this is a critical issue for deployable trusses). In addition to these geometric requirements‚ this spacecraft would be sent in deep space (perhaps as far as the orbit of Jupiter) to ensure maximum sensitivity; this makes the weight issue particularly important. Another interesting subsystem necessary to achieve the stringent specifications is the six d.o.f. vibration isolator at the interface between the attitude

4


Introduction

5

control module and the supporting truss; this isolator allows the low frequency attitude control torque to be transmitted‚ while filtering out the high frequency disturbances generated by the unbalanced centrifugal forces in the reaction wheels. The same general purpose vibration isolator may be used at the interface between the truss and the independent telescopes; in this case however‚ its vibration isolation capability is combined with the steering (pointing) of the telescopes. The third component relevant to active control is the optical delay line; it consists of a high precision single degree of freedom translational mechanism supporting a mirror‚ whose function is to control the path length between every telescope and the beam combiner‚ so that these distances are kept identical to a fraction of the wavelength (e.g. ). These examples were concerned mainly with performance. However‚ as technology develops and with the availability of low cost electronic components‚ it is likely that there will be a growing number of applications where active solutions will become cheaper than passive ones‚ for the same level of performance. The reader should not conclude that active will always be better and that a control system can compensate for a bad design. In most cases‚ a bad design will remain bad‚ active or not‚ and an active solution should normally be considered only after all other passive means have been exhausted. One should always bear in mind that feedback control can compensate for external disturbances only in a limited frequency band that is called the bandwidth of the control system. One should never forget that outside the bandwidth‚ the disturbance is actually amplified by the control system. In recent years‚ there has been a growing interest in semi-active control‚ particularly for vehicle suspensions; this has been driven by the reduced cost as compared to active control‚ due mainly to the absence of a large power actuator. A semi-active device can be broadly defined as a passive device in which the properties (stiffness‚ damping‚ ...) can be varied in real time with a low power input. Although they behave in a strongly nonlinear way‚ semi-active devices are inherently passive and‚ unlike active devices‚ cannot destabilize the system; they are also less vulnerable to power failure. Semi-active suspension devices may be based on classical viscous dampers with a variable orifice‚ or on magneto-rheological (MR) fluids.

1.2

Smart materials and structures

An active structure consists of a structure provided with a set of actuators and sensors coupled by a controller; if the bandwidth of the controller includes some vibration modes of the structure‚ its dynamic response must be considered. If the set of actuators and sensors are located at discrete points of the structure‚ they can be treated separately. The distinctive feature of smart structures is that the actuators and sensors are often distributed‚ and have a high degree of integration inside the structure‚ which makes a separate modelling impossible

6


(Fig. 1.3). Moreover‚ in some applications like vibroacoustics‚ the behaviour of the structure itself is highly coupled with the surrounding medium; this also requires a coupled modelling. From a mechanical point of view‚ classical structural materials are entirely described by their elastic constants relating stress and strain‚ and their thermal expansion coefficient relating the strain to the temperature. Smart materials are materials where strain can also be generated by different mechanisms involving temperature‚ electric field or magnetic field‚ etc... as a result of some coupling in their constitutive equations. The most celebrated smart materials are briefly described below: Shape Memory Alloys (SMA) allow one to recover up to 5 % strain from the phase change induced by temperature. Although two-way applications are possible after education, SMAs are best suited to one-way tasks such as deployment. In any case, they can be used only at low frequency and for low precision applications, mainly because of the difficulty of cooling. Fatigue under thermal cycling is also a problem. The best known SMA is the NITINOL; SMAs are little used in vibration control, and will not be discussed in this book. Piezoelectric materials have a recoverable strain of 0.1 % under electric field; they can be used as actuators as well as sensors. There are two broad classes of piezoelectric materials used in vibration control: ceramics and polymers. The piezopolymers are used mostly as sensors, because they require extremely high voltages and they have a limited control authority; the best known is the polyvinylidene fluoride (PV DF or Piezoceramics are used extensively as actuators and sensors, for a wide range of frequency including ultrasonic applications; they are well suited for high precision in the nanometer range The best known piezoceramic is the Lead Zirconate Titanate (PZT).

Introduction

7

Magnetostrictive materials have a recoverable strain of 0.15 % under magnetic field; the maximum response is obtained when the material is subjected to compressive loads. Magnetostrictive actuators can be used as load carrying elements (in compression alone) and they have a long lifetime. They can also be used in high precision applications. The best known is the TERFENOL-D. Magneto-rheological (MR) fluids consists of viscous fluids containing micron-sized particles of magnetic material. When the fluid is subjected to a magnetic field‚ the particles create columnar structures requiring a minimum shear stress to initiate the flow. This effect is reversible and very fast (response time of the order of millisecond). Some fluids exhibit the same behaviour under electrical field; they are called electro-rheological (ER) fluids; however‚ their performances (limited by the electric field breakdown) are significantly inferior to MR fluids. MR and ER fluids are used in semi-active devices. This brief list of commercially available smart materials is just a flavor of what is to come: phase change materials are currently under development and are likely to become available in a few years time; they will offer a recoverable strain of the order of 1 % under an electric field‚ one order of magnitude more than the piezoceramics. The range of available devices to measure position‚ velocity‚ acceleration and strain is extremely wide‚ and there are more to come‚ particularly in optomechanics. Displacements can be measured with inductive‚ capacitive and optical means (laser interferometer); the latter two have a resolution in the nanometer range. Piezoelectric accelerometers are very popular but they cannot measure a d.c. component. Strain can be measured with strain gages‚ piezoceramics‚ piezopolymers and fiber optics. The latter can be embedded in a structure and give a global average measure of the deformation; they offer a great potential for health monitoring as well. We will see that piezopolymers can be shaped to react only to a limited set of vibration modes (modal filters).

1.3

Control strategies

There are two radically different approaches to disturbance rejection: feedback and feedforward. Although this text is entirely devoted to feedback control‚ it is important to point out the salient features of both approaches‚ in order to enable the user to select the most appropriate one for a given application.

1.3.1

Feedback

The principle of feedback is represented in Fig.1.4; the output is compared to the reference input ‚ and the error signal‚

of the system is passed

8


into a compensator and applied to the system The design problem consists of finding the appropriate compensator such that the closed-loop system is stable and behaves in the appropriate manner. In the control of lightly damped structures, feedback control is used for two distinct and somewhat complementary purposes: active damping and model based feedback. The objective of active damping is to reduce the effect of the resonant peaks on the response of the structure. From

(Problem P.1.2)‚ this requires near the resonances. Active damping can generally be achieved with moderate gains; another nice property is that it can be achieved without a model of the structure‚ and with guaranteed stability‚ provided that the actuator and sensor are collocated and have perfect dynamics. Of course actuators and sensors always have finite dynamics and any active damping system has a finite bandwidth. The control objectives can be more ambitious‚ and we may wish to keep a control variable (a position‚ or the pointing of an antenna) to a desired value in spite of external disturbances in some frequency range. From the previous formula and

we readily see that this requires large values of GH in the frequency range where is sought. implies that the closed-loop transfer function is close to 1, which means that the output tracks the input accurately. From Equ.(1.3), this also ensures disturbance rejection within the bandwidth of the control system. In general, to achieve this, we need a more elaborate strategy involving a mathematical model of the system which, at best, can only be a low-dimensional approximation of the actual system There are many techniques available to find the appropriate compensator, and only the simplest and the best established will be reviewed in this text. They all have a number of common features:

Introduction

9

The bandwidth of the control system is limited by the accuracy of the model; there is always some destabilization of the flexible modes outside (residual modes). The phenomenon whereby the net damping of the residual modes actually decreases when the bandwidth increases is known as spillover (Fig.1.5). The disturbance rejection within the bandwidth of the control system is always compensated by an amplification of the disturbances outside the bandwidth. When implemented digitally‚ the sampling frequency must always be two orders of magnitude larger than to preserve reasonably the behaviour of the continuous system. This puts some hardware restrictions on the bandwidth of the control system.

1.3.2

Feedforward

When a signal correlated to the disturbance is available‚ feedforward adaptive filtering constitutes an attractive alternative to feedback for disturbance rejection; it was originally developed for noise control (Nelson & Elliott)‚ but it is very efficient for vibration control too (Fuller et al.). Its principle is explained in Fig.1.6. The method relies on the availability of a reference signal correlated to the primary disturbance; this signal is passed through an adaptive filter‚ the output of which is applied to the system by secondary sources. The filter coefficients are adapted in such a way that the error signal at one or several critical points is minimized. The idea is to produce a secondary disturbance such that it cancels the effect of the primary disturbance at the location of the error sensor. Of course‚ there is no guarantee that the global response is also reduced at

10


other locations and‚ unless the response is dominated by a single mode‚ there are places where the response can be amplified; the method can therefore be considered as a local one‚ in contrast to feedback which is global. Unlike active damping which can only attenuate the disturbances near the resonances‚ feedforward works for any frequency and attempts to cancel the disturbance completely by generating a secondary signal of opposite phase. The method does not need a model of the system‚ but the adaption procedure relies on the measured impulse response. The approach works better for narrowband disturbances‚ but wide-band applications have also been reported. Because it is less sensitive to phase lag than feedback‚ feedforward control can be used at higher frequency (a good rule of thumb is ); this is why it has been so successful in acoustics. The main limitation of feedforward adaptive filtering is the availability of a reference signal correlated to the disturbance. There are many applications where such a signal can be readily available from a sensor located on the propagation path of the perturbation. For disturbances induced by rotating machinery‚ an impulse train generated by the rotation of the main shaft can be used as reference. Table 1.1 summarizes the main features of the two approaches.

1.4

The various steps of the design

The various steps of the design of a controlled structure are shown in Fig.1.7. The starting point is a mechanical system‚ some performance objectives (e.g. position accuracy) and a specification of the disturbances applied to it; the controller cannot be designed without some knowledge of the disturbance applied to the system. If the frequency distribution of the energy of the disturbance (i.e. the power spectral density) is known‚ the open-loop performances can be evaluated and the need for an active control system can be assessed. If an active

Introduction

11

system is required‚ its bandwidth can be roughly specified from Equ.(1.3). The next step consists of selecting the proper type and location for a set of sensors to monitor the behaviour of the system‚ and actuators to control it. The concept of controllability measures the capability of an actuator to interfere with the states of the system. Once the actuators and sensors have been selected‚ a model of the structure is developed‚ usually with finite elements; it can be improved by identification if experimental transfer functions are available. Such models generally involve too many degrees of freedom to be directly useful for design purposes; they must be reduced to produce a control design model involving only a few degrees of freedom‚ usually the vibration modes of the system‚ which carry the most important information about the system behaviour. At this point‚ if the actuators and sensors can be considered as perfect (in the frequency band of interest)‚ they can be ignored in the model; their effect on the control system performance will be tested after the design has been completed. If‚ on the contrary‚ the dynamics of the actuators and sensors may significantly

12


affect the behaviour of the system‚ they must be included in the model before the controller design. Even though most controllers are implemented in a digital manner‚ nowadays‚ there are good reasons to carry out a continuous design and transform the continuous controller into a digital one with an appropriate technique. This approach works well when the sampling frequency is two orders of magnitude faster than the bandwidth of the control system‚ as is generally the case in structural control.

1.5

Organization of the book

Structural control and the emerging area of smart structures belong to the general field of Mechatronics; they consist of a mixture of mechanical engineering‚

Introduction

13

structural mechanics‚ control engineering‚ material science and computer science. This book has been written for structural engineers willing to acquire some background in structural control; it has been assumed that the reader is familiar with structural dynamics and has some basic knowledge of linear system theory‚ including Laplace transforms‚ root locus‚ Bode plots‚ Nyquist plots‚ etc... Readers who are not familiar with these concepts are advised to read a basic text on linear system theory (e.g. Cannon‚ Franklin et al.). Some elementary background is also assumed in signal processing. Chapter 2 recalls briefly some concepts of structural dynamics; chapter 3 considers the modelling of some active structures; chapters 4 to 9 are mostly devoted to the design of single input single output (SISO) compensators with an increasing degree of complexity‚ both in the frequency and in the time domain (state space); this part includes active damping with collocated actuator/sensor pairs (chap.5) and active isolation with decentralized controllers (chap.6). Chapter 10 is devoted to controllability and observability; chapter 11 discusses stability; chapter 12 discusses the semi-active control; chapter 13 describes some applications and discusses briefly some issues in transforming a continuous controller into a digital one. Chapter 14 is entirely devoted to the control of cable structures. Each chapter is supplemented by a set of problems; it is assumed that the reader is familiar with some computer aided control engineering software such as MATLAB or SIMULINK. Chapters 1 to 8‚ part of chapter 9 and some applications of chapter 13 could constitute a one semester graduate course in structural control.

1.6

References

J. N. A UBRUN ‚ K. R. L ORELL ‚ T. W. H AVAS & W. C. H ENNINGER ‚ Performance Analysis of the Segment Alignment Control System for the Ten-Meter Telescope‚ Automatica‚ vol. 24‚ no. 4‚ 437–453‚ 1988. R. H. CANNON‚ Dynamics of Physical Systems‚ McGraw-Hill‚ 1967. G. F. FRANKLIN‚ J. D. POWELL & A. EMANI-NAEMI‚ Feedback Control of Dynamic Systems‚ Addison-Wesley‚ 1986. C. R. FULLER‚ S. J. ELLIOTT & P. A. NELSON‚ Active Control of Vibration‚ Academic Press‚ 1996. M. V. GANDHI & B. S. THOMPSON‚ Smart Materials and Structures‚ Chapman & Hall‚ 1992. P. A. NELSON & S. J. ELLIOTT‚ Active Control of Sound‚ Academic Press‚ 1992. K. UCHINO‚ Ferroelectric Devices‚ Marcel Dekker‚ 2000.

14


R. N. WILSON‚ F. FRANZA & L. NOETHE‚ Active Optics. 1. A System for Optimizing the Optical Quality and Reducing the Costs of Large Telescopes‚ Journal of Modern Optics‚ vol. 34‚ no. 4‚ 485–509‚ 1987. General literature on control of flexible structures R. L. CLARK‚ W. R. SAUNDERS & G. P. GIBBS‚ Adaptive Structures‚ Dynamics and Control‚ Wiley‚ New York‚ NY‚ 1998. W. K. GAWRONSKI‚ Dynamics and Control of Structures - A Modal Approach‚ Springer‚ 1998. C. H. HANSEN & S. D. SNYDER‚ Active Control of Sound and Vibration‚ E&FN Spon‚ London‚ 1996. D. C. HYLAND‚ J. L. JUNKINS & R. W. LONGMAN‚ Active Control Technology for Large Space Structures‚ J. of Guidance‚ vol. 16‚ no. 5‚ 801–821‚ Sep.-Oct. 1993. D. J. INMAN‚ Vibration‚ with Control‚ Measurement‚ and Stability‚ PrenticeHall‚ 1989. H. JANOCHA‚ editor‚ Adaptronics and Smart Structures (Basics‚ Materials‚ Design and Applications)‚ Springer‚ 1999. S. M. JOHSI‚ Control of Large Flexible Space Structures‚ Lecture Notes in Control and Information Sciences‚ Springer-Verlag‚ vol. 131‚ 1989. J. L. JUNKINS‚ editor‚ Mechanics and Control of Large Flexible Structures‚ vol. 129‚ 1990. J. L. JUNKINS & Y. KIM‚ Introduction to Dynamics and Control of Flexible Structures‚ AIAA Education Series‚ 1993. L. MEIROVITCH‚ Dynamics and Control of Structures‚ Wiley‚ 1990. D. K. MIU‚ Mechatronics - Electromechanics and Contromechanics‚ SpringerVerlag‚ 1993. R. E. SKELTON‚ Dynamic System Control - Linear System Analysis and Synthesis‚ Wiley‚ 1988. D. W. SPARKS JR & J. N. JUANG‚ Survey of Experiments and Experimental Facilities for Control of Flexible Structures‚ AIAA J.of Guidance‚ vol. 15‚ no. 4‚ 801–816‚ July–Aug. 1992.

Introduction

1.7

Problems

P.1.1

Consider the underdeterminate system of equations

15

Show that the minimum norm solution‚ i.e. the solution of the minimization problem such that is

is called the pseudo-inverse of J. [hint: Use Lagrange multipliers to remove the equality constraint.] P.1.2 Consider the feedback control system of Fig.1.4. Show that the transfer functions from the input and the disturbance to the output are respectively

P. 1.3 Based on your own experience‚ describe one application in which you feel an active structure may outclass a passive one; suggest a configuration for the actuators and sensors.


Chapter 2

Some concepts of structural dynamics 2.1

Equation of motion of a discrete system

Consider the system with three point masses represented in Fig.2.1. The equations of motion can be established by considering the free body diagrams of the three masses and applying Newton’s law; one easily gets:

or, in matrix form,

The general form of the equation of motion governing the dynamic equilibrium between the external, elastic, inertia and damping forces acting on a non-gyroscopic, discrete, flexible structure with a finite number of degrees of freedom is where and are the vectors of generalized displacements (translations and rotations) and forces (point forces and torques) and M, K and C are respectively 17


18

the mass, stiffness and damping matrices; they are symmetric and semi positive definite. M and K arise from the discretization of the structure, usually with finite elements. A lumped mass system such as that of Fig.2.1 has a diagonal mass matrix. The finite element method usually leads to non-diagonal (consistent) mass matrices, but a diagonal mass matrix often provides an acceptable representation of the inertia in the structure (Problem P.2.2). The damping matrix C represents the various dissipation mechanisms in the structure, which are usually poorly known. To compensate for this lack of knowledge, it is customary to make assumptions on its form. One of the most popular hypotheses is the Rayleigh damping:

The coefficients

2.2

and

are selected to fit the structure under consideration.

Vibration modes

Consider the free response of a undamped (conservative) system of order is governed by If one tries a solution of the form eigenvalue problem

and

It

must satisfy the

Because M and K are symmetric, K is positive semi definite and M is positive definite, the eigenvalue must be real and non negative. is the natural frequency and is the corresponding mode shape; the number of modes is equal to the number of degrees of freedom, Note that Equ.(2.5) defines only the shape, but not the amplitude of the mode which can be scaled arbitrarily. The modes are usually ordered by increasing frequencies

Some concepts of structural dynamics

19

From Equ.(2.5), we see that if the structure is released from initial conditions and it oscillates at the frequency according to always keeping the shape of mode Left multiplying Equ.(2.5) by one gets the scalar equation

and, upon permuting

and

we get similarly,

If we substract these equations, taking into account that a scalar is equal to its transpose and that K and M are symmetric, we get

which shows that the mode shapes corresponding to distinct natural frequencies are orthogonal with respect to the mass matrix.

It follows from the foregoing equations that the mode shapes are also orthogonal with respect to the stiffness matrix. The orthogonality conditions are often written as

where is the Kronecker delta index is the modal mass of mode Since the mode shapes can be scaled arbitrarily, it is usual to normalize them in such a way that If one defines the matrix of the mode shapes the orthogonality relationships read

To demonstrate the orthogonality conditions, we have used the fact that the natural frequencies were distinct. If several modes have the same natural frequency (as often occurs in practice because of symmetry), they form a subspace of dimension equal to the multiplicity of the eigenvalue. Any vector in this subspace is a solution of the eigenvalue problem, and it is always possible to find a set of vectors such that the orthogonality conditions are satisfied. A rigid body mode is such that there is no strain energy associated with it It can be demonstrated that this implies that the rigid body modes can therefore be regarded as solutions of the eigenvalue problem (2.5) with


20

2.3 2.3.1

Modal decomposition Structure without rigid body modes

Let us perform a change of variables from physical coordinates coordinates according to where

to modal

is the vector of modal amplitudes. Substituting into Equ.(2.2), we get

Left multiplying by we obtain

and using the orthogonality relationships (2.8) and (2.9),

If the matrix is diagonal, the damping is said classical or normal. In this case, the modal fraction of critical damping (in short modal damping) is defined by One can readily check that the Rayleigh damping (2.3) complies with this condition and that the corresponding modal damping ratios are

The two free parameters and can be selected in order to match the modal damping of two modes. Note that the Rayleigh damping tends to overestimate the damping of the high frequency modes. Under condition (2.12), the modal equations are decoupled and Equ.(2.11) can be rewritten with the notations

The following values of the modal damping ratio can be regarded as typical: satellites and space structures are generally very lightly damped 0.001 – 0.005), because of the extensive use of fiber reinforced composites, the absence of aerodynamic damping, and the low strain level. Mechanical engineering applications (steel structures, piping,...) are in the range of most dissipation takes place in the joints, and the damping increases with the strain level. For civil engineering applications, is typical and, when


21

radiation damping through the ground is involved, it may reach depending on the local soil conditions. The assumption of classical damping is often justified for light damping, but it is questionable when the damping is large, as in problems involving soil-structure interaction. Lightly damped structures are usually easier to model, but more difficult to control, because their poles are located very near the imaginary axis and they can be destabilized very easily. If one accepts the assumption of classical damping, the only difference between Equ.(2.2) and (2.14) lies in the change of coordinates (2.10). However, in physical coordinates, the number of degrees of freedom of a discretized model of the form (2.2) is usually large, especially if the geometry is complicated, because of the difficulty of accurately representing the stiffness of the structure. This number of degrees of freedom is unnecessarily large to represent the structural response in a limited bandwidth. If a structure is excited by a bandlimited excitation, its response is dominated by the normal modes whose natural frequencies belong to the bandwidth of the excitation, and the integration of Equ.(2.14) can often be restricted to these modes. The number of degrees of freedom contributing effectively to the response is therefore reduced drastically in modal coordinates. Now, let us consider the steady state response to a harmonic excitation The response will also be harmonic with the same frequency, substituting into Equ.(2.2), we find that the complex amplitudes of F and X are related by The matrix is the dynamic generalization of the flexibility matrix and is called the dynamic flexibility matrix. Equ.(2.16) can be obtained alternatively by Fourier transformation of Equ.(2.2). If one considers the same problem in modal coordinates, the modal amplitudes will also be harmonic, and, from Equ.(2.14), we obtain

From Equ.(2.10),

Comparing Equ.(2.16) and (2.18), we obtain the modal expansion of the dynamic flexibility matrix:

where the sum extends to all the modes. expresses the complex amplitude of the structural response of degree of freedom when the structure is

22


exposed to a steady state harmonic excitation at degree of freedom the structure has no rigid body modes, we can write Equ.(2.19) for obtain the modal expansion of the static flexibility matrix:

If we

For a limited frequency band if we select in such a way that the dynamic expansion can be split into the contribution of the low frequency modes which respond dynamically, and that of the high frequency modes which respond statically:

Using Equ.(2.20), we can transform the foregoing result in such a way that the high frequency modes do not appear explicitly in the expansion:

The static contribution of the high frequency modes to the flexibility matrix is often called the residual mode; we shall denote it R. It is independent of the frequency and introduces a feedthrough component in the transfer matrix: part of the output is proportional to the input. In terms of Laplace transform, the numerator and the denominator of some components of have the same power in (such a transfer function is said to be not strictly proper). We shall see later that truncating the modal expansion of a transfer function without introducing a residual mode can lead to substantial errors in the prediction of the open-loop zeros (e.g. see Fig.3.15) and, as a result, of the performance of a control system.

2.3.2

Structure with rigid body modes

The approximation (2.22) applies only at low frequency, If the structure has rigid body modes, the first sum can be split into rigid and flexible modes; however, the residual mode cannot be used any more, because no longer exists. This problem can be solved in the following way. The displacements are partitioned into their rigid and flexible contributions according to where and are the matrices whose columns are the rigid body modes and the flexible modes, respectively. Assuming no damping, to make things formally


23

simpler, and taking into account that the rigid body modes satisfy we obtain the equation of motion

Left multiplying by and using the orthogonality relations (2.6) and (2.7), we see that the rigid body modes are governed by

or Substituting this result into Equ.(2.24), we get

or where we have defined the projection matrix

such that check that

is orthogonal to the rigid body modes. In fact, we can easily

24


P can therefore be regarded as a filter which leaves unchanged the flexible modes and destroys the rigid body modes. If we follow the same procedure as in the foregoing section, we need to evaluate the elastic contribution of the static deflection, which is the solution of

Since the solution may contain an arbitrary contribution from the rigid body modes. On the other hand, is the superposition of the external forces and the inertia forces associated with the motion as a rigid body; it is self-equilibrated, because it is orthogonal to the rigid body modes. Since the system is in equilibrium as a rigid body, a particular solution of Equ.(2.30) can be obtained by adding dummy constraints to remove the rigid body modes (Fig.2.2). The modified system is statically determinate and its stiffness matrix can be inverted. If we denote by the flexibility matrix of the modified system, the general solution of (2.30) is

where is a vector of arbitrary constants. The contribution of the rigid body modes can be eliminated with the projection matrix P, leading to

is the pseudo-static flexibility matrix of the flexible modes. On the other hand, left multiplying Equ.(2.24) by we get

where the diagonal matrix is regular. It follows that the pseudo-static deflection can be written alternatively

Comparing with Equ.(2.31), we get

This equation is identical to Equ.(2.20) when there are no rigid body modes. From this result, we can extend Equ.(2.22) to systems with rigid body modes:


25

where the contribution from the residual mode is

Note that is the flexibility matrix of the system obtained by adding dummy constraints to remove the rigid body modes. Obviously, this can be achieved in many different ways and it may look surprising that they all lead to the same result (2.35). In fact, different boundary conditions lead to different displacements under the self-equilibrated load but they differ only by a contribution of the rigid body modes, which is destroyed by the projection matrix P, leading to the same Let us illustrate the procedure by an example.

2.3.3 Example Consider the system of three identical masses of Fig.2.3. There is one rigid body mode and two flexible ones:

and From Equ.(2.27), the projection matrix is


26

or

We can readily check that

and the self-equilibrated loads associated with a force (Fig.2.3.a)

applied to mass 1 is

If we impose the statically determinate constraint on mass 1 (Fig.2.3.b), the resulting flexibility matrix is

leading to

The reader can easily check that other dummy constraints would lead to the same pseudo-static flexibility matrix (Problem P.2.3).

2.4

Transfer function of collocated systems

Consider a diagonal component of the dynamic flexibility matrix of a undamped system. It is the transfer function between the generalized force (point force or torque) and the response of the corresponding degree of freedom (respectively displacement or angle). If the force is generated by an actuator, and the displacement is measured by a sensor, the actuator and sensor are collocated. We shall see later that control systems using collocated actuator/sensor pairs usually enjoy special properties which are the consequence of alternating poles and zeros along the imaginary axis. To establish this property, consider the diagonal term of Equ.(2.34):


27

is real since the system is undamped, and all the numerators in the modal expansion are positive. As a result, it is readily demonstrated that

The behaviour of is represented in Fig.2.5: The amplitude of the transfer function goes to at the resonance frequencies (corresponding to a pair of purely imaginary poles in the system). In every interval between consecutive resonance frequencies, there is a frequency where the amplitude of the frequency response function (FRF) vanishes; in structural dynamics, these frequencies are called anti-resonance frequencies; they correspond to purely imaginary transmission zeros in the open-loop transfer function It is worth mentioning that A harmonic excitation at an anti-resonance frequency produces no response at the degree of freedom where the excitation is applied; the structure behaves as if an additional restraint had been added. The antiresonance frequencies are in fact identical to the resonance frequencies of the system with the additional restraint (Fig.2.4), and the acting force is identical to the reaction force in the additional restraint during the free vibration of the modified system. In contrast to the resonance frequencies, the anti-resonance frequencies do depend on the actuator location. If the diagram of another diagonal component is examined, the frequencies where the plot goes to (resonances) are unchanged, but the frequencies where do change (as do the natural frequencies of the modified system when we change the location of the additional restraint). In all cases, however, there will be one and only one anti-resonance between two consecutive resonances.

28


Observing Fig.2.5 and Equ.(2.36), we see that neglecting the residual mode is equivalent to moving the downwards by Doing that, all the anti-resonance frequencies tend to be overestimated by an amount which depends on the slope of in the vicinity of the zeros. We shall see later that errors in the spacing between poles and zeros of the open-loop system can lead to substantial errors in the performance of the closed-loop system. For a system with non-collocated actuator and sensor, the numerators of the various terms in the modal expansion of become they can be either positive or negative, and the property (2.37) is lost. Interlacing poles and zeros are no longer guaranteed in this case. In system theory, it is customary to write the system transfer functions in the form

In the present case, the undamped collocated system has alternating imaginary poles and zeros at and (Fig.2.6.a). The transfer function reads

If some damping is added, the poles and zeros are slightly moved into the left half plane as indicated in Fig.2.6.b, without changing the dominant feature of


29

interlacing. A collocated system always exhibits Bode and Nyquist plots similar to those represented in Fig.2.7. Each flexible mode introduces a circle in the Nyquist diagram; it is more or less centered on the imaginary axis which is intersected at and the radius is proportional to the inverse of the modal damping, In the Bode plots, a 180° phase lag occurs at every natural frequency, and is compensated by a 180° phase lead at every imaginary zero; the phase always oscillates between 0 and , as a result of the interlacing property of the poles and zeros.

30

2.5


Continuous structures

Continuous structures are distributed parameter systems which are governed by partial differential equations. Various discretization techniques, such as the Rayleigh-Ritz method, or finite elements, allow us to approximate the partial differential equation by a finite set of ordinary differential equations. In this section, we illustrate some of the features of distributed parameter systems with continuous beams. This example will be frequently used in the subsequent chapters. The plane transverse vibration of a beam is governed by the following partial differential equation This equation is based on the Euler-Bernouilli assumptions that the neutral axis undergoes no extension and that the cross section remains perpendicular to the neutral axis (no shear deformation). EI is the bending stiffness, is the mass per unit length and the distributed external load per unit length. If the beam is uniform, the free vibration is governed by

The boundary conditions depend on the support configuration: a simple support implies and (no displacement, no bending moment); for a clamped end, we have and (no displacement, no rotation); a free end corresponds to and (no bending moment, no shear), etc... A harmonic solution of the form can be obtained if and satisfy

with the appropriate boundary conditions. This equation defines a eigenvalue problem; the solution consists of the natural frequencies (infinite in number) and the corresponding mode shapes The eigenvalues are tabulated for various boundary conditions in textbooks on mechanical vibrations (e.g. Geradin & Rixen, p. 187). For the pinned-pinned case, the natural frequencies and mode shapes are

Just as for discrete systems, the mode shapes are orthogonal with respect to the mass and stiffness distribution:


31

The generalized mass corresponding to Equ.(2.44) is As with discrete structures, the frequency response function between a point force actuator at and a displacement sensor at is

where the sum extends to infinity. Exactly as for discrete systems, the expansion can be limited to a finite set of modes, the high frequency modes being included in a quasi-static correction as in Equ.(2.34) (Problem P.2.5).

2.6

Guyan reduction

As already mentioned, the size of a discretized model obtained by finite elements is essentially governed by the representation of the stiffness of the structure. For complicated geometries, it may become very large, especially with automated mesh generators. Before solving the eigenvalue problem (2.5), it may be advisable to reduce the size of the model (2.2) by condensing the degrees of freedom with little or no inertia and which are not excited by external forces. The degrees of freedom to be condensed, denoted in what follows, are often referred to as slaves; those kept in the reduced model are called masters and are denoted To begin with, consider the undamped forced vibration of a structure where the slaves are not excited and have no inertia; the governing equation is

or

According to the second equation, the slaves the masters :

are completely determined by

Substituting into Equ.(2.49), we find the reduced equation

which involves only Note that in this case, the reduced equation has been obtained without approximation.

32


The idea in this, so-called, Guyan reduction is to assume that the masterslave relationship (2.51) applies even if have some inertia (sub-matrix not identically 0) or applied forces. We assume the following transformation

The reduced mass and stiffness matrices are obtained by substituting the above transformation into the kinetic and strain energy:

with The second equation produces as in Equ.(2.52). If external loads are applied to , the reduced loads are obtained by equating the virtual work or

Finally, the reduced equation of motion reads

Usually, it is not necessary to consider the damping matrix in the reduction, because it is rarely known explicitly at this stage. The Guyan reduction can be performed automatically in commercial finite element packages, the selection of masters and slaves being made by the user. In the selection process the following should be kept in mind: The degrees of freedom without inertia or applied load can be condensed without affecting the accuracy. Translational degrees of freedom carry more information than rotational ones. In selecting the masters, preference should be given to translations, especially if large modal amplitudes are expected (Problem P.2.7). It can be demonstrated that the error in the mode shape associated with the Guyan reduction is a increasing function of the ratio


33

where is the natural frequency of the mode and is the first natural frequency of the constrained system, where all the degrees of freedom (masters) have been blocked is the smallest solution of det . Therefore, should be kept far above the frequency band where the model is expected to be accurate. Considering the effect of additional constraints on can often be used as a guide in the selection of the degrees of freedom .

2.7

References

K. J. BATHE & E. L. WILSON, Numerical Methods in Finite Element Analysis, Prentice-Hall, 1976. R. E. D. BISHOP & D . C. JOHNSON, The Mechanics of Vibration, Cambridge University Press, 2nd ed., 1979. R. H. CANNON, Dynamics of Physical Systems, McGraw-Hill, 1967. R. W. CLOUGH & J. PENZIEN, Dynamics of Structures, McGraw-Hill, 1975. R. R. CRAIG JR., Structural Dynamics, Wiley, 1981. M. GERADIN & D. RIXEN, Mechanical Vibrations, Theory and Application to Structural Dynamics, Wiley, 1993. P. C. HUGHES, Attitude Dynamics of Three-Axis Stabilized Satellite with a Large Flexible Solar Array, J. Astronautical Sciences, vol. 20, 166–189, Nov.– Dec. 1972. P. C. H UGHES , Dynamics of Flexible Space Vehicles with Activ e Attitude Control, Celestial Mechanics Journal, vol. 9, 21–39, March 1974. T. J. R. HUGHES, The Finite Element Method, Linear Static and Dynamic Finite Element Analysis, Prentice-Hall, 1987b. D. J. INMAN, Vibration, with Control, Measurement, and Stability, PrenticeHall, 1989. L. MEIROVITCH, Computational Methods in Structural Dynamics, Sijthoff & Noordhoff, 1980. V. J. MODI, Attitude Dynamics of Satellites with Flexible Appendages - A Brief Review, J. Spacecraft and Rockets, vol. 11, 743–751, Nov. 1974. O. C. ZIENKIEWICZ & R. L. TAYLOR, The Finite Element Method, McGrawHill, fourth ed., 1989, (2 vol.).


34

2.8

Problems

P.2.1 Using a finite element program, discretize a simply supported uniform beam with an increasing number of elements (4,8,etc...). Compare the natural frequencies with those obtained with the continuous beam theory. Observe that the finite elements tend to overestimate the natural frequencies. Why is that so? P.2.2 Using the same stiffness matrix as in the previous example and a diagonal mass matrix obtained by lumping the mass of every element at the nodes (the entries of the mass matrix for all translational degrees of freedom are where is the number of elements; no inertia is attributed to the rotations), compute the natural frequencies. Compare the results with those obtained with a consistent mass matrix in problem P.2.1. Notice that using a diagonal mass matrix usually tends to underestimate the natural frequencies. P.2.3 Consider the three mass system of section 2.3.3. Show that changing the dummy constraint to mass 2 does not change the pseudo-static flexibility matrix P.2.4 Consider a simply supported beam with the following properties: EI = 10.266 It is excited by a point force at (a) Assuming that a displacement sensor is located at (collocated) and that the system is undamped, plot the transfer function for an increasing number of modes, with and without quasi-static correction for the high-frequency modes. Comment on the variation of the zeros with the number of modes and on the absence of mode 4. Note: To evaluate the quasi-static contribution of the high-frequency modes, it is useful to recall that the static displacement at created by a unit force applied at on a simply supported beam is

The symmetric operator is often called ”flexibility kernel” or Green’s function. (b) Including three modes and the quasi-static correction, draw the Nyquist and Bode plots and locate the poles and zeros in the complex plane for a uniform modal damping of and (c) Do the same as (b) when the sensor location is Notice that the interlacing property of the poles and zeros no longer holds. P.2.5 Consider the modal expansion of the transfer function (2.47) and assume that the low frequency amplitude G(0) is available, either from static

Some concepts of structural dynamics calculations or from experiments at low frequency. approximated by the truncated expansion

35

Show that

can be

P.2.6 Show that the impulse response matrix of a damped structure with rigid body modes reads

where and is the Heaviside step function. P.2.7 Consider a uniform beam clamped at one end and free at the other end; it is discretized with six finite elements of equal size. The twelve degrees of freedom are numbered to starting from the clamped end. We perform various Guyan reductions in which we select according to: (a) all (12 degrees of freedom, no reduction); (b) all (6 d.o.f.); (c) all (6 d.o.f.); (d) (6 d.o.f.); (e) (3 d.o.f.); (f) (3 d.o.f.); For each case, compute the natural frequency of the first three modes and the first natural frequency of the constrained system. Compare the roles of the translations and rotations. P.2.8 Consider a spacecraft consisting of a rigid main body to which one or several flexible appendages are attached. Assume that there is at least one axis about which the attitude motion is uncoupled from the other axes. Let be the (small) angle of rotation about this axis and J be the moment of inertia (of the main body plus the appendages). Show that the equations of motion read

where is the torque applied to the main body, and are the modal masses and the natural frequencies of the constrained modes of the flexible appendages and are the modal participation factors of the flexible modes [i.e. is the work done on mode i of the flexible appendages by the inertia forces associated with a unit angular acceleration of the main body] (Hughes). [Hint: Decompose the motion into the rigid body mode and the components of

36


the constrained flexible modes, express the kinetic energy and the strain energy, write the Lagrangian in the form

and write the Lagrange equations.]

Chapter 3

Actuators‚ piezoelectric materials‚ and active structures 3.1 Introduction The actuators play a critical role in active structures technology and‚ in many cases‚ they constitute the limiting factor. Two broad categories can be distinguished: ”grounded” and ”structure borne” actuators. The former react on a fixed support; they include torque motors‚ force motors (electrodynamic shakers) or tendons. The second category‚ also called ”space realizable”‚ include jets‚ reaction wheels‚ control moment gyros‚ proof-mass actuators‚ active members (capable of both structural functions and generating active control forces)‚ piezo strips‚ etc... Active members and all actuating devices involving only internal‚ self-equilibrating forces‚ cannot influence the rigid body motion of the structure. After a brief review of the principle of the proof-mass actuator and the single axis gyrostabilizer‚ this chapter discusses the constitutive equations of piezoelectric materials and analyses the dynamics of various structures equipped with piezoelectric actuators and sensors‚ namely beams‚ trusses and shells. The duality between the sensor equation and the equivalent piezoelectric loads is emphasized. The danger of using the oversimplified beam theory for predicting the transmission zeros of slender structures with collocated actuator and sensor is pointed out. 37


38

3.2

Proof-mass actuator

A reaction mass is attached to a current-carrying coil moving in a magnetic field created by a permanent magnet; the moving part is attached to the chassis by a spring and damper (Fig.3.1). The current in the coil is governed by

where is the resistance and L the inductance of the coil, and is the back electromotive force (e.m.f.) due to the motion. In the Laplace domain,

where is the impedance of the coil. The equation of motion of the moving part is

where F is the electromagnetic force‚ proportional to the current in the coil

Note that the same constant velocity and the back e.m.f.;

appears in the relationship between the coil is called the voice coil constant. Combining the

Actuators‚ piezoelectric materials‚ and active structures

39

two foregoing equations‚ one gets the relationship between the displacement of the moving mass and the voltage applied

This equation shows that the back e.m.f. appears as an additional damping in the system. On the other hand‚ accelerating the mass with produces a reaction force on the supporting structure equal to

or‚ in Laplace terms‚

Combining with Equ.(3.6)‚ one gets the transfer function between the force applied to the structure and the voltage V:

If one neglects the inductance of the coil‚ function takes the form

is a constant and the transfer

where is the natural frequency associated with the spring-mass system and is the damping ratio (including contributions from the structure and from the back e.m.f.). The Bode plots (amplitude and phase) of the actuator are shown in Fig.3.1. Above some critical frequency the proof-mass actuator behaves like an ideal force generator (constant amplitude, no phase difference). The resonance frequency (and therefore the critical frequency ) can be adapted by changing the mass Because the reaction mass is attached to the structure by a spring, the proof-mass actuator has no authority over the rigid body modes (in contrast to reaction wheels). Operation at low frequency requires a large stroke, which is technically difficult. Medium to high frequency actuators (above 40 Hz) can be obtained with loudspeakers (slightly modified with an added mass ) or commercial electrodynamic shakers provided with an appropriate mass at the tip, to lower the natural frequency to the desired value. An application is discussed later. The proof-mass system can also behave as a passive absorber, if its natural frequency is tuned to one of the structural modes of the structure (e.g. Hagood & Crawley).

40

3.3


Reaction wheels and gyrostabilizers

These devices are torque actuators normally used in attitude control of satellites. They have authority over the rigid body modes as well as the flexible modes. A reaction wheel consists of a rotating wheel whose axis is fixed with respect to the spacecraft; a torque is generated by increasing or decreasing the angular velocity. If the angular velocity exceeds the specification‚ the wheel must be unloaded‚ using another type of actuator (jets or magnetic). In control moment gyros (CMG)‚ the rotating wheel is mounted on gimbals‚ and the gimbal torques are used as control inputs. The principle of a one-axis gyrostabilizer is described in Fig.3.2. Rotating the gimbal about the axis with an angular velocity produces torques:

where is the angular momentum along the axis‚ and is the deviation of the rotor axis with respect to the vertical. The servo motor on the gimbal axis is velocity controlled. The angle is measured also‚ and a small gain feedback maintains the axis of the rotor in the vertical position (for a deeper discussion of the use of CMG in attitude control‚ see Jacot & Liska).

3.4

Piezoelectric actuators

The piezoelectric effect was discovered by Pierre and Jacques Curie in 1880. The direct piezoelectric effect consists of the ability of certain crystalline materials (ceramics) to generate an electrical charge in proportion to an externally


41

applied force; the direct effect is used in force transducers. According to the inverse piezoelectric effect‚ an electric field parallel to the direction of polarization induces an expansion of the ceramic. The piezoelectric effect is anisotropic; it can be exhibited only by materials whose crystal structure has no center of symmetry; this is the case for some ceramics below a certain temperature called the Curie temperature; in this phase‚ the crystal has a built-in electric dipole‚ but the net electric dipole on a macroscopic scale is zero. When the crystal is cooled in the presence of an electric field‚ the dipoles tend to align‚ leading to an electric dipole on a macroscopic scale. After cooling and removal of the poling field‚ the dipoles cannot return to their original position‚ and the ceramic body becomes permanently piezoelectric‚ with the ability to convert mechanical energy to electrical energy and vice versa. The most popular piezoelectric materials are Lead-Zirconate- Titanate (PZT) which is a ceramic‚ and Polyvinylidene fluoride (PVDF) which is a polymer; their main properties are given in Table 3.1. In addition to the piezoelectric effect‚ piezoelectric materials exhibit a pyroelectric effect‚ according to which electric charges are generated when the material is subjected to temperature; this effect is used to produce heat detectors; it will not be discussed here. The direction of expansion with respect to the direction of the electrical field depends on the constants appearing in the constitutive equations. The material can be manufactured in such a way that one of the coefficients dominates the others. Piezoelectric actuators are becoming increasingly important in micropositioning technology. In the modelling of piezoelectric actuators for structural control applications‚ there is a great similarity between the inverse piezoelectric effect and thermal deformations. Table 3.1 gives the properties of typical piezoelectric materials.

3.4.1 Constitutive equations In this section we use the notations of the IEEE standard on piezoelectricity. In an unstressed one-dimensional dielectric medium, the electric displacement D (charge per unit area, expressed in is related to the electric field

E (V/m) by where is the dielectric constant of the material. Similarly, in a one-dimensional elastic body placed in a zero electric field, the strain S and the stress T are related by

where

is the compliance of the material (inverse of the Young’s modulus).

42


For a piezoelectric material‚ the electrical and mechanical constitutive equations are coupled:

In Equ.(3.11), the piezoelectric constant relates the strain to the electric field E in the absence of mechanical stress and refers to the compliance when the electric field is constant. In Equ.(3.12), relates the electric displacement D to the stress under a zero electric field (short-circuited electrodes); is expressed in or Coulomb/Newton. is the dielectric constant under constant stress. The above equations can be transformed into

which are usually rewritten as

where

is the Young’s modulus under constant electric field (in is the constant relating the electric displacement to the strain, for short-circuited electrodes (in is called the coupling coefficient of the piezoelectric material. This name comes from the fact that, at frequencies far below the mechanical resonance frequency of the piezo, can be expressed as

A high value of is desirable for efficient transduction. From Equ.(3.14), one sees that the dielectric constant under constant strain is related to that under constant stress by Equation (3.13) is the starting point for the formulation of the equation of a laminar piezoelectric actuator, while Equ.(3.14) is that for a laminar sensor. In matrix form, Equ.(3.13) and (3.14) read

It is interesting to point out the analogy between Equ.(3.11) and the classical one-dimensional equation of thermoelasticity


43

where is the temperature difference and is the thermal expansion coefficient. In Table 3.1‚ it is important to note that the piezoelectric coefficient [as appears in the actuator equation (3.13)] is 300 times larger for PZT than for PVDF; this indicates the superior actuation capability of piezoceramics as compared to piezopolymers. Figure 3.3 shows the two basic types of piezoelectric actuators (e.g. see Physik Instrumente catalogue): The stacked design (linear actuator) and the laminar design (spatially distributed actuator); they are briefly reviewed in the following sections. For multidimensional media‚ the constitutive equations keep more or less the same form as above‚ except that the quantities involved are tensors (e.g. see IEEE Std‚ Lerch or Piefort). If the direction of polarization coincides with direction 3‚ the constitutive equations can be written in matrix form:

44



45

Examining the actuator equation (3.17)‚ we note that when an electric field is applied parallel to the direction of polarization‚ an extension is observed along the same direction; its amplitude is governed by the piezoelectric coefficient Similarly‚ a shrinkage is observed along the directions 1 and 2 perpendicular to the electric field‚ the amplitude of which are controlled by and respectively. Piezoceramics have an isotropic behaviour in the plane‚ On the contrary‚ when the PVDF is polarized under stress‚ its piezoelectric properties are highly anisotropic‚ with Equation (3.17) also indicates that an electric field normal to the direction of polarization 3 produces a shear deformation controlled by the piezoelectric constant (the same occurs if a field is applied). An interesting feature of this type of actuation is that is the largest of all piezoelectric coefficients ( for PZT). This has led to the interesting torque actuator design of Fig.3.4 (Glazounov et al.).


46

3.4.2

Linear actuator

A linear piezo actuator consists of a stack of thin ceramic layers of thickness between 0.1 and 1 mm separated by electrodes (Fig.3.3). The polarization axis and the applied electric field are perpendicular to the layer and coincide with the direction of expansion of the actuator‚ so that the actuation capability is controlled by the coefficient. The required voltage for layers of 0.1 mm is 100 V (they are called low voltage piezo: LVPZ) while 1 mm layers require 1000 V (high voltage piezo: HVPZ). Both types have similar strain capabilities; the difference is that low voltage piezos have a higher electric capacitance and require a significantly larger current. LVPZ are characterized by low voltage and high current while HVPZ are high voltage and low current devices; this affects the amplifier design. The maximum expansion can be 0.1 to 0.13 %‚ which means that a piezo translator of 100 mm can have a maximum stroke of about The size of the disks can be selected to achieve the appropriate stiffness. For dynamic applications‚ a mechanical preload is introduced to compensate for the small extensional strength of the stack. The stacked design is often referred to as When no external load is applied‚ the change of length is related to the voltage applied‚ by the approximate relationship

where is the number of disks in the stack and V is the applied voltage. Two types of piezotranslators are shown in Fig.3.5. The first one follows exactly the principle of Fig.3.3.a. In the second design‚ the piezoelectric stack is mounted along the main axis of an elliptical structure which acts as a motion amplifier; the expansion of the actuator is in the direction perpendicular to that of the piezo stack; this design leads to more compact actuators.


47

The actual voltage/expansion diagram is nonlinear and exhibits a fairly high hysteresis (Fig.3.6). The maximum width of the hysteresis curve can reach as much as 15 % of the total expansion. For quasi-static applications‚ the hysteresis can be compensated in closedloop by measuring the actual expansion of the stack (for example by gluing a strain gage on the piezo stack). Closed-loop compensation of the hysteresis in dynamic applications such as active damping does not seem to be necessary‚ and may even be dangerous for the stability of the system. Many lightweight space structures consist of trusses. Active members are capable of both structural functions (load transmission) and simultaneously generating active control forces. The actuation system can be based on jackscrews (e.g. see Robertshaw et al.) or on piezotranslators (e.g. Fanson et al.). The former exhibit high rigidity‚ large amplitude‚ low bandwidth‚ low precision and low power efficiency; such actuators are used mainly for reconfiguration. By contrast‚ piezotranslators allow a stroke of up to a few hundreds microns‚ with high bandwidth and high precision (resolution in the nanometer range). Applications using linear piezoelectric actuators will be considered in subsequent chapters.


48

3.4.3

Laminar actuator

In the laminar design‚ the piezoelectric layer is bonded to the structure (Fig.3.3.b). The geometrical arrangement is such that the useful direction of expansion is normal to that of the electric field. Thus‚ the activation capability is governed by the piezoelectric constant or more precisely by note that is only one half of The piezoelectric material may either be ceramics (PZT) or polymers (PVDF); typical thickness of PZT strips is their sizes are generally small and they are covered with silver electrodes. PVDF strips are flexible and may be much thinner and larger; however‚ the actuation capability of PVDF (measured by ) is 300 times smaller than that of PZT (see Table 3.1). PZT is always isotropic in the plane while PVDF may be manufactured both isotropic and anisotropic‚ even almost unidirectional It is the shape of the electrode which determines the active part of the material. Consider the beam of Fig.3.7 with a thin layer of piezoelectric material bonded to one side. We assume that the thickness of the piezoelectric strip is small compared to that of the beam‚ The piezoelectric strip is used as actuator by controlling the voltage V applied to the electrodes‚ creating a constant electric field V/ We assume that the electrodes have a variable width (this can be achieved either by etching the original surface electrode‚ or by cutting the piezoelectric layer). If one rewrites Equ.(3.13) with standard engineering notation‚ one gets‚ within the piezoelectric layer‚


where denotes the Young’s modulus of the piezo‚ and following usual relation applies to the beam

49

The

where E is the Young’s modulus of the beam. The equilibrium equation of the beam is and‚ according to the Euler-Bernouilli assumption‚ the axial deformation and the curvature are related by where is the distance to the neutral axis. Introducing this into Equ.(3.20) and (3.21) and integrating over the cross-section‚ one gets

or

where we have assumed a constant moment arm across the thickness of the film. In this equation‚ the bending stiffness EI refers to the supporting structure and the piezoelectric film together‚ and is the current width of the electrode. Substituting into Equ.(3.22)‚ one gets

Comparing with Equ.(2.40)‚ we see that when the width of the electrodes varies along the beam‚ the piezoelectric actuator produces a distributed load proportional to the second derivative of

50


Similarly‚ for an electrode of constant width‚ the distributed actuator is equivalent to adding concentrated moments at the boundaries of the actuator‚ as represented in Fig.3.8. From Equ.(3.24)‚ the piezoelectric moment is Equivalently‚ a sudden change in the first derivative produces a point force at the discontinuity (Fig.3.8):

at some location

This result is readily obtained by integrating the right hand side of Equ.(3.25) over a small interval about and taking the limit when the size of the interval goes to zero. Equations (3.25) to (3.27) give the equivalent loads associated with various shapes of the electrodes. We note that the actuation authority is strongly dependent on the electrode shape and that it is proportional to the piezoelectric constant as mentioned before‚ is substantially larger for PZT than for PVDF. Equations (3.25) to (3.27) are based on the assumption that the thickness of the piezo strip is much smaller than that of the beam more accurate calculations show that the distance appearing in these equations is actually the distance between the mid-plane of the beam and the mid-plane of the piezo strip.


3.4.4

51

Laminar sensor

Consider a laminar sensor with short-circuited electrodes‚ so that a zero electric field is enforced (E = 0). According to Equ.(3.14)‚ the amount of charge per unit area is related to the strain by or‚ with structural engineerning notations‚

Combining with Equ.(3.23) and assuming that the strain is constant over the thickness of the strip (consequence of we get

The corresponding electric charge is equal to the integral of the electric displacement over the electrode area‚

If the polarization profile is uniform ( current amplifier (Fig.3.9.a) is

constant)‚ the output voltage of a

52


is the constant of the amplifier. Thus‚ the output signal is proportional to the integral of the time derivative of the curvature weighted by the width of the electrode. If a charge amplifier is used (Fig.3.9.b)‚ the output voltage is proportional to the electric charge in the electrode:

If the width of the electrode‚

is constant‚

The sensor output is thus proportional to the difference of slopes at the extremities of the piezo strip. As for the actuator‚ we observe that the sensing capability is proportional to These results assume that a more accurate calculation shows that the distance appearing in Equ.(3.29) to (3.31) is actually the distance between the mid-plane of the beam and the mid-plane of the piezo.

3.4.5 Example: Tip displacement of a cantilever beam

Consider a cantilever beam clamped at and covered with a piezoelectric strip of triangular shape (Fig.3.10). The integral involved in Equ.(3.28) giving the electric charge accumulated on the triangular electrode can be integrated by parts twice according to

The integral in the right hand side vanishes since and taking into account that and we get


53

Thus‚ the output signal is proportional to the tip displacement of the beam.

3.4.6 Spatial modal filters Consider the case of the actuator first. Let

be the modal expansion of the transverse displacements‚ where are the mode shapes and the modal amplitudes. Upon substituting this into the equation of motion (3.25) and using the orthogonality of the mode shapes‚ Equ.(2.45) and (2.46)‚ one gets the following equation which controls the amplitude of mode

This equation shows that the actuator can be designed in such a way that a specific mode is uncontrollable‚ by enforcing that the integral in the right hand side vanishes. Moreover‚ the electrode can be tailored in such a way that only a single mode is excited. This is achieved if the electrode profile is such that

Indeed‚ according to the first orthogonality relation (2.45)‚ this electrode profile can excite only mode without affecting the other modes. As far as the sensor is concerned‚ the output signal of the amplifier‚ Equ.(3.29)‚ can be decomposed into its modal components as

This equation shows that any specific mode can be made unobservable by choosing the electrode profile in such a way that the integral vanishes. Just as for the actuator‚ because of the second orthogonality relation (2.46)‚ the following electrode profile will produce a sensor which is sensitive only to mode In principle‚ if a good structural model is available for the set of modes that one wishes to control‚ it is possible to design a sensor which is sensitive only to these modes‚ and orthogonal to the residual modes. This offers an attractive possibility for spillover alleviation. Note that for a uniform beam‚ the mode shapes are such that as a result of (2.42). It follows that the modal actuator (3.34) and the modal sensor (3.36) have the same shape.

54


Figure 3.11 illustrates the modal filters used for a uniform beam with various boundary conditions; the change of sign indicates a change in the polarity of the strip, which is equivalent to negative values of As an alternative, the part of the sensor with negative polarity can be bonded on the opposite side of the beam. The reader will notice that the electrode shape of the simply supported beam is the same as the mode shape, while for the cantilever beam, the electrode shape is that of the mode shape of a beam clamped at the opposite end.

3.5

Passive damping with piezoceramics


55

A piezoelectric material works as an energy transformer, from mechanical energy to electrical and vice versa. A piezoelectric device subjected to mechanical strain can be regarded as a charge (or current) generator in parallel with a capacitance C (typical values for linear actuators range from to If the structure vibrates, the charge is generated at the frequency of vibration. One way to damp a structure equipped with a piezoelectric device is to connect the piezo to a resistor and an inductor as in Fig.3.12, to obtain a resonant LC circuit tuned on the natural frequency of the structure:

the problem is that‚ for practical values of and C‚ this leads to very large values of L‚ which would not be practical to realize by classical means. However‚ a tunable active inductor can be realized electronically with a gyrator circuit (Edberg et al.). Figure 3.12.b shows experimental results obtained with the truss of Fig.3.17 where one of the piezo translators is connected as indicated above. We see that the damping ratio depends very much on the tuning of to small errors can substantially reduce the efficiency of the system. We shall see later that active damping is often less sensitive to the parameters of the electromechanical system.

3.6

Active cantilever beam

Consider a cantilever beam provided with a pair of rectangular piezoelectric actuator and sensor as in Fig.3.13. The system is modelled by finite elements;

56


the mesh is such that there is a node at both ends of the piezos (each node has two degrees of freedom‚ one translation and one rotation ). We seek the open-loop transfer function between the voltage applied to the actuator and the output voltage of the sensor. According to the foregoing sections‚ the rectangular actuator is equivalent to a pair of torques M with opposite signs and proportional to V:

The gain of the actuator can be computed from the actuator size and the material properties with Equ.(3.26). In the general form (2.2) of the equation of motion‚ the vector force is where the influence vector has the form the only nonzero components correspond to the rotational degrees of freedom of the nodes at and in the model. It follows that the modal excitations are

where is the relative angle (difference of slope) between the ends of the actuator, for mode . Similarly, the output of the charge amplifier, Equ.(3.31), is also proportional to the difference of the rotations at the ends of the sensor, Expanding in modal coordinates, we get

where are the modal components of the slope difference between the ends of the sensor; the gain depends on the sensor size‚ the material properties and the gain of the charge amplifier. Note that if the actuator and sensor are collocated in the sense of the Euler-Bernouilli beam theory‚ the slope differences appearing in the actuator and sensor equations are identical‚

Combining Equ(3.40)‚ (3.41) and the equations governing the structural dynamics‚ (2.19)‚ we obtain the frequency response function (FRF) between the voltage V applied to the actuator and the voltage at the output of the charge amplifier:

The discussion of section 2.3.1 regarding the truncation of the modal expansion applies fully (Problem P.3.4). Equ.(3.42) has the general form (2.36) with no contribution from the rigid body modes‚ because the piezoelectric loads


57

are self equilibrating. This leads to alternating poles and zeros as in Fig.2.6.b. Figure 3.14 shows typical experimental results obtained with a system similar to that shown in Fig.3.13. Observe that the open-loop FRF does not exhibit any roll-off at high frequency. Note also that a small linear phase shift appears in the phase diagram‚ due to the fact that these results have been obtained digitally‚ at a sampling frequency of l0000Hz (the effect of sampling will be discussed briefly in chapter 13). Because of the high-pass behaviour of the charge amplifier‚ discrepancies can be observed also at very low frequency‚ typically below 0.1Hz to 1 H z ‚ depending on the charge amplifier characteristics. If a refined model is required‚ the dynamics of the charge amplifier can be represented by a second order high-pass filter

with the appropriate corner frequency When the modal expansion (3.42) is truncated after mode (quasi-static correction) reads

modes‚ the residual

Its influence on the open-loop transfer function may be quite significant‚ as illustrated in Fig.3.15 which displays the amplitude plots with and without residual mode (truncation after 3 modes). After correction‚ the truncated transfer function follows almost exactly the full expansion (not shown for clarity) for


58

Without correction, the truncated transfer function differs substantially from the full expansion, in particular near the zeros which are strongly dependent on the residual mode this is particularly important because the zeros play a key role in the performance of the control system. We also notice the presence of an extra zero in the corrected diagram; it is due to the feedthrough component which brings back the zero at infinity (due to roll-off) to a finite frequency. The feedthrough component will be reexamined in section 3.8 where the adequacy of the beam model for collocated systems will be questioned.

3.7

Active truss

Figure 3.16 shows a truss structure where some of the bars have been replaced by active struts; each of them consists of a piezoelectric linear actuator colinear with a force transducer. Such an active truss can be used for vibration attenuation‚ or to improve the dimensional stability under thermal gradients. If the stiffness of the active struts matches that of the other bars in the truss‚ the passive behaviour of the truss is basically unchanged. Fig.3.17 shows an example of active truss equipped with two active struts (this truss was built in the late 80 ’s and uses out-dated high-voltage piezotranslators). When a voltage is applied to a unconstrained linear piezoelectric actuator‚ it produces an expansion which‚ if we neglect the hysteresis‚ is proportional


59


60

to the applied voltage where is the gain of the actuator. If the actuator is placed in a truss‚ its effect on the structure can be represented by equivalent piezoelectric loads acting on the passive structure. As for thermal loads‚ the pair of self equilibrating piezoelectric loads applied axially to both ends of the active strut (Fig.3.16) has a magnitude equal to the product of the stiffness of the active strut‚ and the unconstrained piezoelectric expansion

Assuming no damping‚ the equation governing the motion of the structure excited by a single actuator is

where is the influence vector of the active strut in the global coordinate system. The non-zero components of are the direction cosines of the active bar (Problem P.3.6). As for the output signal of the force transducer‚ it is given by

where is the elastic extension of the active strut‚ equal to the difference between the total extension of the strut and its piezoelectric component The total extension is the projection of the displacements of the end nodes on the active strut‚ (Problem P.3.6). Introducing this into Equ.(3.47)‚ we get

Note that‚ because the sensor is located in the same strut as the actuator‚ the same influence vector appears in the sensor equation (3.48) and the equation of motion (3.46). If the force sensor is connected to a charge amplifier of gain the output voltage is given by

Note the presence of a feedthrough component from the piezoelectric extension Proceeding as in the previous section‚ we obtain the modal excitations

and the modal expansion of the output


61

Combining the above equation with that of the dynamics in modal coordinates‚ we get the FRF between the applied voltage V and the output voltage of the charge amplifier:

where we have neglected the damping to simplify the equation. It can be written alternatively

where we have defined

We see that the numerator and the denominator of this expression represent respectively twice the strain energy in the active strut and twice the total strain energy when the structure vibrates according to mode is therefore called the modal fraction of strain energy in the active strut. From Equ.(3.52)‚ we see that determines the residue of mode that is the amplitude of the contribution of mode in the transfer function between the piezo actuator and the force sensor; it can therefore be regarded as a compound index of controllability and observability of mode is readily available from commercial finite element programs; it can be used to select the proper location of the active strut in the structure (Problem P.3.7): the best location is that with the highest for the modes that we wish to control (Preumont et al.). The FRF (3.52) has alternating poles and zeros (Fig.3.18). One can easily see from Equ.(3.52) and (3.48) that the FRF between the piezoelectric displacement and total displacement of the end nodes of the active strut‚ is

At

the total displacement

produced by the piezoelectric loads

is

where is the stiffness of the truss (with the active strut considered as passive) from the end point of the active strut. Writing Equ.(3.54) at and comparing with (3.55)‚ we obtain

62


This ratio can be easily calculated from a static analysis (it is always smaller than 1). As in section 2.3.1‚ the modal expansion of can be truncated after modes provided that a quasi-static correction is applied:

and‚ taking into account Equ.(3.56)‚

This equation involves only the low frequency modes of the truss. It can be written alternatively

The foregoing results apply to an undamped truss with a single piezoelectric linear actuator collocated with a force sensor; the extension to the case with modal damping is straightforward and is left as an exercise (Problem P.3.8).


3.8

63

Piezoelectric shell

The elementary case of a piezoelectric beam has been discussed in section 3.4.3‚ where we introduced the laminar actuators and sensors; analytical expressions for the equivalent piezoelectric loads of a laminar actuator and the output of a laminar sensor were developed with the Euler-Bernouilli beam theory. The open-loop transfer function of an active beam with collocated actuator and sensor was analysed in section 3.6 and the importance of the residual mode (high frequency dynamics) on the location of the open-loop transmission zeros was pointed out. In fact‚ in many situations involving nearly collocated actuators and sensors‚ the Euler-Bernouilli beam theory is inappropriate‚ because the transfer function becomes more sensitive to the detailed strain transmission path within the coupled system (structure-actuator-sensor) which requires more sophisticated modelling techniques accounting for membrane (in-plane) deformations as well as bending ones. This is illustrated by the following experiment. Consider the cantilever plate of Fig.3.19; the steel plate is 0.5 mm thick and 4 piezoceramic strips of thickness are bounded symmetrically as indicated in the figure. The size of the piezos is respectively 55 mm x 25 mm for and and 55 mm × 12.5 mm for and is used as actuator while the sensor is taken successively as and Since they cover the same extension along the beam‚ the various sensor locations cannot be distinguished in the sense of the Euler-Bernouilli beam theory (except for the sign or a constant factor‚ because and are on opposite sides and the size of is twice that of and ) and they should lead to the same transfer function. This is not the case in practice‚ as we can see in the experimental results shown in Fig.3.20. We see that the transmission zeros vary significantly from one configuration to the other (as we shall see later‚ this will dramatically change the performance

64


of the closed-loop system). In fact‚ because of the nearness of the actuator to the sensor‚ and the small thickness of the plate‚ the membrane strains play an important role in the transmission of the strain from the actuator to the sensor; these membrane strains are neglected in the classical beam theory but they can be handled by shell theory. This introductory text on active vibration control is not the place for extensive developments of the theory of piezoelectric shells which can be found in the literature (Lee); we will only write the actuator and sensor equations for piezoelectric shells in the piezoelectric axes and compare them with their equivalent forms for beams. Finally‚ we will write the general form of the dynamic equation of a piezoelectric structure.

3.8.1

Two-dimensional constitutive equations

In this section‚ we will denote vector quantities by {} in order to improve the understanding and to conform with the usual notations of the theory of laminate composites. Consider a two-dimensional piezoelectric lamina in a plane ( ); the poling direction is (normal to the lamina) and the electric field is also applied along In the piezoelectric principal axes‚ the constitutive equations read

where


65

are the stress and strain vector‚ respectively‚ [C] is the matrix of elastic constant‚ E is the component of the electric field along D is the of the electric displacement and the dielectric constant.

3.8.2

Kirchhoff shell

Next‚ consider a piezoelectric lamina bonded to a shell structure (Fig.3.21). If the global axes coincide with the piezoelectric axes of the lamina‚ the constitutive equations can be integrated over the thickness of the shell in the form

where {N} is the vector of in-plane resultant forces and {M} the vector of bending moments;

66


is the deformation vector of the mid-plane and tures:

is the vector of curva-

The matrices A‚B‚D are the classical stiffness matrices of shell theory (e.g. Agarwal & Broutman); is the thickness of the piezoelectric lamina and is the distance between its mid-plane and the mid-plane of the shell. If the piezoelectric lamina is connected to a charge amplifier as in Fig.3.9.b‚ the voltage between the electrodes is set to V = 0 and the sensor equation (3.64) can be integrated over the electrode to produce the sensor output

where the integral extends over the surface of the electrode (the part of the piezo not covered by the electrode does not contribute to the signal). The first part of the integral is the contribution of the membrane strain‚ while the second one is due to bending. If the piezoelectric properties are isotropic the surface integral can be further transformed into a contour integral using one of the Green integrals:

This integral extends over the contour of the electrode (Fig.3.22); the first contribution is the component of the mid-plane in-plane displacement normal to the contour and the second one is associated with the slope along the contour. Similarly‚ for a piezoelectric actuator made of isotropic material‚ the equivalent piezoelectric loads consist of a in-plane force‚ normal to the contour of the electrode‚ and a constant moment‚ acting along the contour of the electrode (Fig.3.23): Comparing these results with the equivalent bending moment (3.26) of the beam theory‚ we see that the beam theory accounts only for the component of the bending moment normal to the beam axis and totally neglects the in-plane force (Fig.3.24). Conversely‚ the sensor signal given by the beam theory accounts only for the component of the rotation along the contour normal to the beam axis.


67

68

3.9


Finite element formulation

The finite element formulation of a piezoelectric continuum can be derived from Hamilton’s principle in which the potential energy density and the virtual work are properly transformed to include the piezoelectric energy (Allik & Hughes). In addition to the classical displacements as in any F.E. formulation‚ the generalized variables include the electric potentials between the electrodes of the piezos. The resulting dynamic equations read

where M is the mass matrix, the stiffness matrix, the coupling matrix connecting the mechanical variables and the electrical variables the matrix of electrical capacitance; is the vector of external mechanical loads applied to the structure and q the electric charges brought to the electrodes. For voltage driven electrodes, the electric potential is controlled and Equ.(3.70) can be rewritten

where represents the equivalent piezoelectric loads. Once the mechanical displacements have been computed, the electric charges appearing on the electrodes can be computed from Equ.(3.71). Prom Equ.(3.72), we see that the dynamics of the system with short-circuited electrodes is the same as if there were no piezoelectric electromechanical coupling. Conversely, open electrodes correspond to a charge condition in this case we can substitute from Equ.(3.71) into Equ.(3.70)


69

This equation shows that the piezoelectric coupling tends to increase the global stiffness of the system if the electrodes are left open (this is because the terms of are negative). The natural frequencies of the system with open electrodes are larger than those with short-circuited electrodes; this difference is usually small but it tends to increase as the amount of embedded piezoelectric material increases. Finally‚ if the electrodes are connected via a passive electrical network of impedance matrix Z‚ Equ.(3.70) and (3.71) must be complemented by the network equation (Fig.3.25). To conclude this section‚ let us reexamine the modeling of the cantilever beams of Fig.3.15 and 3.19. It turns out that‚ when the thickness of the beam is small and the actuator and sensor are collocated‚ the membrane strains contribute significantly to the transfer function. As a result‚ the discrepancy between the beam theory and the shell theory is large‚ particularly near the transmission zeros of the FRF which are strongly affected by the feedthrough from the strain actuator to the strain sensor (Fig.3.26).


70

Figure 3.27 shows computed FRF based on the shell theory for the cantilever plate of Fig.3.19; the model is based on Mindlin-type finite elements (Piefort). The agreement with the experimental results of Fig.3.20 is qualitatively good, although the numerical results still tend to overestimate the spacing between the poles and the zeros. As we shall see in the chapters devoted to active damping, the closed-loop poles are located on branches going from the open-loop poles to the open-loop zeros; overestimating their spacing amounts to overestimating the performance of the control system.

3.10

References

B. D. AGARWAL & L. J. BROUTMAN‚ Analysis and Performance of Fiber Composites‚ Wiley‚ second ed.‚ 1990. H. ALLIK & T. R. J. HUGHES‚ Finite Element Method for Piezoelectric Vibration‚ Int. J. for Numerical Methods in Engineering‚ vol. 2‚ 151–157‚ 1970. E. H. ANDERSON‚ D. M. MOORE‚ J. L. FANSON & M. A. EALEY‚ Development of an Active Member Using Piezoelectric and Electrostrictive Actuation for Control of Precision Structures‚ SDM Conference‚ AIAA paper 90-1085CP‚ 1990. S. E. BURKE & J. E. HUBBARD‚ Active Vibration Control of a Simply Supported Beam Using Spatially Distributed Actuator‚ IEEE Control Systems Magazine‚ pp. 25–30‚ Aug. 1987.


71

G. S. CHEN‚ B. J. LURIE & B. K. WADA‚ Experimental Studies of Adaptive Structures for Precision Performance‚ SDM Conference‚ AIAA paper 89-1327CP‚ 1989. E. F. CRAWLEY & J. DE LUIS‚ Use of Piezoelectric Actuators as Elements of Intelligent Structures‚ AIAA Journal‚ vol. 25‚ no. 10‚ 1373–1385‚ 1987. E. F. CRAWLEY & K. B. LAZARUS‚ Induced Strain Actuation of Isotropic and Anisotropic Plates‚ AIAA Journal‚ vol. 29‚ no. 6‚ 944–951‚ 1991. E. K. DIMITRIADIS‚ C. R. FULLER & C. A. ROGERS‚ Piezoelectric Actuators for Distributed Vibration Excitation of Thin Plates‚ Trans. ASME‚ J. of Vibration and Acoustics‚ vol. 113‚ 100–107‚ Jan. 1991. D. L. EDBERG‚ A. S. Bicos & J. S. FECHTER‚ On Piezoelectric Energy Conversion for Electronic Passive Damping Enhancement‚ Proceedings of Damping’91‚ San Diego‚ 1991. J. L. FANSON‚ G. H. BLACKWOOD & C. C. CHEN‚ Active Member Control of Precision Structures‚ SDM Conference‚ AIAA paper 89-1329-CP‚ 1989. R. L. FORWARD‚ Electronic Damping of Orthogonal Bending Modes in a Cylindrical Mast Experiment‚ AIAA Journal of Spacecraft‚ vol. 18‚ no. 1‚ 11–17‚ Jan.-Feb. 1981. A. E. GLAZOUNOV‚ Q. M. ZHANG & C. KIM‚ A New Torsional Actuator Based on Shear Piezoelectric Response‚ SPIE Conference on Smart Structures and Materials‚ vol. 3324‚ 82–91‚ 1998a‚ san Diego‚ CA. A. E. GLAZOUNOV‚ Q. M. ZHANG & C. KIM‚ Piezoelectric Actuator Generating Torsional Displacement from the D15 Shear Strain‚ Appl. Phys. Lett. 72‚ 2526‚ 1998b. N. W. HAGOOD & E. F. CRAWLEY‚ Experimental Investigation of Passive Enhancement of Damping for Space Structures.‚ AIAA Journal of Guidance‚ vol. 14‚ no. 6‚ 1100–1109‚ Nov.–Dec. 1991. W. L. HALLAUER JR‚ Recent Literature on Experimental Structural Dynamics and Control Research‚ in J. L. JUNKINS‚ editor‚ Mechanics and Control of Large Flexible Structures‚ chap. 18‚ AIAA Progress in Astronautics and Aeronautics‚ Vol. 129‚ 1990. S. HANAGUD‚ M. W. OBAL & A. J. CALISE‚ Optimal Vibration Control by the Use of Piezoceramic Sensors and Actuators‚ AIAA Journal of Guidance‚ vol. 15‚ no. 5‚ 1199–1206‚ Sep.–Oct. 1992. W.-S. HWANG & H. C. PARK‚ Finite Element Modeling of Piezoelectric Sensors and Actuators‚ AIAA Journal‚ vol. 31‚ no. 5‚ 930–937‚ May 1993.

72


IEEE Std‚ Standard on Piezoelectricity‚ ANSI/IEEE Std 176-1987‚ Jan. 1987. A. D. JACOT & D. J. LISKA‚ Control Moment Gyros in Attitude Control‚ AIAA J. of Spacecraft and Rockets‚ vol. 3‚ no. 9‚ 1313–1320‚ Sep. 1966. C.-K. LEE‚ Theory of Laminated Piezoelectric Plates for the Design of Distributed Sensors/Actuators- Part I : Governing Equations and Reciprocal Relationships‚ J. Acoust.-Soc. Am.‚ vol. 87‚ no. 3‚ 1144–1158‚ March 1990. C.-K. LEE‚ W.-W. CHIANG & T. C. O’SULLIVAN‚ Piezoelectric Modal Sensor/Actuator Pairs for Critical Active Damping Vibration Control‚ J. Acoust.Soc. Am.‚ vol. 90‚ no. 1‚ 374–384‚ July 1991. C.-K. LEE & F. C. MOON‚ Modal Sensors/Actuators‚ Trans. ASME‚ J. of Applied Mechanics‚ vol. 57‚ 434–441‚ June 1990. R. LERCH‚ Simulation of Piezoelectric Devices by Two- and Three-Dimensional Finite Elements‚ IEEE Transactions on Ultrasonics‚ Ferroelectrics‚ and Frequency Control‚ vol. 37‚ no. 3‚ May 1990. N. LOIX & A. PREUMONT‚ Remarks on the Modelling of Active Structures with Colocated Piezoelectric Actuators and Sensors‚ ASME Vibration Conference‚ Boston‚ vol. 84-3‚ 335–339‚ Sep. 1995. L. D. PETERSON‚ J. J. ALLEN‚ J. P. LAUFFER & A. K. MILLER‚ An Experimental and Analytical Synthesis of Controlled Structure Design‚ SDM Conference‚ AIAA paper 89-1170-CP‚ 1989. Philips Application Book on Piezoelectric Ceramics‚ (J. Van Randeraat & R. E. Setterington‚ Edts)‚ Mullard Limited‚ London‚ 1974. Physik Instrumente catalogue‚ Products for Micropositioning (PIGmbH‚ FRG). V. PIEFORT‚ Finite Element Modeling of Piezoelectric Active Structures‚ Ph.D. thesis‚ Université Libre de Bruxelles‚ Active Structures Laboratory‚ 2001. A. PREUMONT‚ J. P. DUFOUR & C. MALEKIAN‚ Active Damping by a Local Force Feedback with Piezoelectric Actuators‚ AIAA J. of Guidance‚ vol. 15‚ no. 2‚ 390–395‚ March–April 1992. H. H. ROBERTSHAW‚ R. H. WYNN‚ S. L. HENDRICKS & W. W. CLARK‚ Dynamics and Control of a Spatial Active Truss Actuator‚ SDM Conference‚ AIAA paper 89-1328-CP‚ 1989. K. U CHINO ‚ Ferroelectric Devices‚ Marcel Dekker‚ 2000. J. R. VlNSON‚ The Behavior of Shells Composed of Isotropic and Composite Materials‚ Kluwer Academic Publishers‚ 1993.


73

D. C. ZIMMERMAN‚ G. C. HORNER & D. J. INMAN‚ Microprocessor Controlled Force Actuator‚ AIAA Journal of Guidance‚ vol. 11‚ no. 3‚ 230–236‚ May– June 1988.

3.11

Problems

P.3.1 Consider a beam covered with a piezo patch and an electrode of the shape shown in Fig.3.8.b. Sketch and calculate the equivalent loading given by this electrode configuration. (Note: just like thermal loads‚ piezoelectric loads are internal loads which must be self-equilibrated.) P.3.2 Consider a simply supported beam with collocated piezoelectric rectangular actuator and sensor extending longitudinally from to Using the results of section 2.5‚ show that the expression appearing at the numerator of Equ.(3.42) can be written analytically

P.3.3 From the result of the previous problem‚ sketch the actuator (and the sensor) which maximizes the response of mode 1‚ mode 2 and mode 3‚ respectively. P.3.4 Consider the active cantilever beam of transfer function (3.42). Assuming that G(0) is available from static calculations or from an experiment at low frequency‚ show that the truncated modal expansion including a quasi-static correction can be written

P.3.5 Consider a beam covered with a piezoelectric actuator of triangular shape (the axis of symmetry of the triangle coincides with that of the beam). Show that the equivalent piezoelectric loads consist of two point forces of equal magnitude and opposite signs at each end, and a torque at the larger end. Check that these forces and moment are self equilibrating. P.3.6 Consider the active truss of Fig.3.16. (a) For each active strut, write the influence vector b of the piezoelectric loads in global coordinates. (b) Assuming small displacements, check that the projection of the differential displacements of the end nodes of the active struts are given by where b is the result of (a). P.3.7 Consider a nine bay planar truss similar to that of Fig.3.16. Each bay is square with one diagonal; all the bars have the same cross section. For the following boundary conditions, use a finite element program to calculate the first three flexible modes of the truss. Suggest two reasonable locations of an active strut to control these modes.

74


(a) Free-free boundary conditions. (b) Clamped-free boundary conditions. [Hint: Use the modal fraction of strain energy as index in the selection of the active strut location.] P.3.8 Show that with modal damping Equ.(3.59) is transformed into

Chapter 4

Collocated versus non-collocated control 4.1

Introduction

In the foregoing chapters, we have seen that the use of collocated actuator and sensor pairs, for a lightly damped flexible structure, always leads to alternating poles and zeros near the imaginary axis (Fig.4.1.a). In this chapter, using the root locus technique, we show that this property guarantees the asymptotic stability of a wide class of single-input single-output (SISO) control systems, even if the system parameters are subject to large perturbations. This is because the root locus plot keeps the same general shape, and remains entirely within the left half plane when the system parameters are changed from their nominal values. Such a control system is said to be robust with respect to stability. The

75


76

use of collocated actuator/sensor pairs is recommended whenever it is possible. This interlacing property of the poles and zeros no longer holds for a noncollocated control, and the root locus plot may experience severe alterations for small parameter changes. This is especially true when the sequence of poles and zeros along the imaginary axis is reversed as in Fig.4.1.b. This situation is called a pole-zero flipping. It is responsible for a phase uncertainty of 360° and the only protection against instability is provided by the damping (systems which are prone to such a huge phase uncertainty can only be gain-stabilized).

4.2

Pole-zero flipping

Recall that the root locus shows, in a graphical form, the evolution of the poles of the closed-loop system as a function of the scalar gain applied to the compensator. The open-loop transfer function GH includes the structure, the compensator, and possibly the actuator and sensor dynamics, if necessary. The root locus is the locus of the solution of the closed-loop characteristic equation when the real parameter goes from zero to infinity. If the open-loop transfer function is written

the locus goes from the poles (for to the zeros (as open-loop system, and any point P on the locus is such that

of the

Collocated versus non-collocated control

77

where are the phase angles of the vectors joining the zeros to P and are the phase angles of the vector joining the poles to P (see Fig.4.10). There are branches of the locus going asymptotically to infinity as increases. Consider the departure angle from a pole and the arrival angle at the zero when they experience a pole-zero flipping; since the contribution of the far away poles and zeros remains essentially unchanged, the difference must remain constant after flipping. As a result, a nice stabilizing loop before flipping is converted into a destabilizing one after flipping (Fig.4.2). If the system has some damping, the control system is still able to operate with a small gain after flipping. Since the root locus technique does not distinguish between the system and the compensator, the pole-zero flipping may occur in two different ways: There are compensator zeros near system poles (this is called a notch filter). If the actual poles of the system are different from those assumed in the compensator design, the notch filter may become inefficient (if the pole moves away from the zero), or worse, a pole-zero flipping may occur. This is why notch filters have to be used with extreme care. As we shall see in later chapters, notch filters are generated by optimum feedback compensators and this may lead to serious robustness questions if the parameter uncertainty is large. Some actuator/sensor configurations may produce pole-zero flipping within the system alone, for small parameter changes. These situations are often associated with a pole-zero (near) cancellation due to a deficiency in the controllability or the observability of the system. In some cases, however, especially if the damping is extremely light, instability may occur. No pole-zero flipping can occur within the structure if the actuator and sensor are collocated. The following sections provide examples illustrating these points.


78

4.3

Collocated control

Consider the two-mass problem of Fig.4.3. The system has a rigid body mode along the axis; it is controlled by a force f applied to the main body M. A flexible appendage is connected to the main body by a spring and a damper In this section, a position control system will be designed, using a sensor placed on the main body (collocated); a sensor attached to the flexible appendage will be considered in the next section. With f representing the control torque and and being the attitude angles, this problem is representative of the single-axis attitude control of a satellite, with M representing the main body, and the other inertia representing either a flexible appendage like a solar panel (in which case the sensor can be on the main body, i.e. collocated), or a scientific instrument like a telescope which must be accurately pointed towards a target (now the sensor has to be part of the secondary structure; i.e. non-collocated). A more elaborate single-axis model of a spacecraft is considered in Problem P.2.8. The system equations are :

With the notations

the transfer functions between the input force f and

and

are respectively :

Approximation (4.8) recognizes the fact that, for low damping the far away zero will not influence the closed-loop response. There are two poles near the imaginary axis. In which refers to the collocated sensor, there are two zeros also near the imaginary axis, at As observed earlier, these zeros are identical to the poles of the modified system where the main body has been blocked (i.e. constrained mode of the flexible appendage). When the mass ratio is small, the polynomials in the numerator and denominator are almost equal, and there is a pole-zero cancellation.


79

Let us consider a lead compensator

It includes one pole and one zero located on the negative real axis; the pole is to the left of the zero. Figure 4.4 shows a typical root locus plot for the collocated case when M = 1, and The parameters of the compensator are T = 10 and Since there are two more poles than zeros the root locus has two asymptotes at ±90°. One observes that the system is stable for every value of the gain, and that the bandwidth of the control system can be a substantial part of The lead compensator always increases the damping of the flexible mode (see Problem P.4.5). If there are not one, but several flexible modes, there are as many pole-zero pairs and the number of poles in excess of zeros remains the same in this case), so that the angles of the asymptotes remain ±90° and the root locus never leaves the stable region. The lead compensator increases the damping ratio of all the flexible modes, but especially those having their natural frequency between the pole and the zero of the compensator. Of course, we have assumed that the sensor and the actuator have perfect dynamics; if this is not the case, the foregoing conclusions may be considerably modified, especially for large gains.


80

4.4

Non-collocated control

Figure 4.5 shows the root locus plot for the lead compensator applied to the non-collocated open-loop system characterized by the transfer function Equ.(4.7), with the following numerical data: M = 1, 0.02. The excess number of poles is in this case so that, for large gains, the flexible modes are heading towards the asymptotes at ±60°, in the right half plane. For a gain the closed-loop poles are located at and (these locations are not shown in Fig.4.5 for clarity: the poles of the rigid body mode are close to the origin and those of the flexible mode are located between the open-loop poles and the imaginary axis). The corresponding Bode plots are shown in Fig.4.6; the phase and gain margins are indicated. One observes that even with this small bandwidth (crossover frequency the gain margin is extremely small. A slightly lower value of the damping ratio would make the closed-loop system unstable (Problem P.4.1).

4.5

Notch filter

A classical way of alleviating the effect of the flexible modes in non-collocated control is to supplement the lead compensator with a notch filter with two zeros located near the flexible poles:

The zeros of the notch filter, at are selected right below the flexible poles. The double pole at aims at keeping the compensator proper (i.e. the


81

degree of the numerator not larger than that of the denominator); it can, for example, be selected far enough along the negative real axis. The corresponding root locus is represented in Fig.4.7.a for and This compensator allows a larger bandwidth than the lead compensator alone (Problem P.4.2). To be effective, a notch filter must be closely tuned to the flexible mode that we want to attenuate. However, as we already mentioned, the notch filter suffers from a lack of robustness and should not be used if the uncertainty in the system properties is large. To illustrate this, Fig.4.7.b shows a detail of the root locus near the notch, when the natural frequency of the system is smaller than expected (in the example, is reduced from to the other data are identical to that of Fig.4.5), while the notch filter is kept the same (being implemented in the computer, the notch filter is not subject to parameter uncertainty). The rest of the root locus is only slightly affected by the change. Because the open-loop poles of the flexible mode move from above to below the zero of the notch filter (from to with the zeros at in the example), there is a pole-zero flipping, with the consequence that the branch of the root locus connecting the pole to the zero rapidly becomes unstable. This example emphasizes the fact that notch filters should be used with extreme care, especially for systems where the uncertainty is large (Problem P.4.3).

82

4.6


Pole-zero flipping in the structure

In the previous section, a pole-zero flipping was created by lowering the natural frequency of the flexible mode from above to below the imaginary zero of the notch filter. This led to a sharp alteration of the root locus plot near the flexible poles. This was used as a warning of the danger of using notch filters for systems with large uncertainty. In this section, we use a simple example to demonstrate the occurrence of a pole-zero flipping in a structure for small variations of the parameters. Consider once again the three mass system of Fig.2.1; if there is no damping, the equation of motion is

Transforming into Laplace form (assuming zero initial conditions) one gets

With the same notation as in the previous example, the transfer functions between the applied force f and the displacement of the various masses are respectively


83

The poles of the system are located on the imaginary axis at

has two pairs of zeros, independent of the mass ratio, located at and As usual, they are identical to the natural frequencies of the constrained modes, when is blocked. For the poles and the zeros of cancel each other; for they alternate on the imaginary axis, as we expect for a collocated system. on the other hand, has one pair of zeros at and has no zero at all. The evolution of the poles with the mass ratio is shown in Fig.4.8. The zeros are also shown in the figure; they do not depend on One observes that at there is a pole-zero flipping in In the vicinity of can be written

The pole-zero pattern of the open-loop system is shown in Fig.4.9 for values of close to 1.


84

4.7

Effect on the Bode plots

From Equ.(4.1),

the phase of

for a specific value

is given by

where is the phase angle of the vector joining the zero to and is the phase angle of the vector joining the pole to (Fig.4.10). Accordingly, an imaginary zero at produces a phase lead of 180° for and an imaginary pole produces similarly a phase lag of 180°. Therefore, a pole-zero flipping near the imaginary axis produces a phase uncertainty of 360° in the frequency range between the pole and the zero (Problem P.4.4). It appears that the only way the closed-loop stability can be guaranteed in the vicinity of a pole-zero flipping is to have the open-loop system gain-stabilized (i.e. such that in that frequency range.

4.8

Relation to the mode shapes

Because of their dramatic influence on the phase of the open-loop transfer function, it is essential for the designer to be able to predict all the imaginary zeros within or near the bandwidth of the control system, for a given actuator and sensor location. This fact is not given proper importance by structural engineers involved in identification; in fact, it is still a tradition, in the modal analysis and model up-dating community, to be more concerned with the poles (resonance


85

frequencies) than with the zeros (anti-resonance). It turns out that the zeros are essentially related to the mode shapes of the structure, which unfortunately often tend to be more sensitive to the parameter uncertainty than the natural frequencies. Consider the modal expansion of the undamped transfer function

where and are the modal amplitudes at the actuator and sensor locations, normalized such that An interesting result, due to Martin, states that if two neighbouring modes are such that their residues and have the same sign, there is always an imaginary zero between the two poles (Problem P.4.7). Based on this result, one can anticipate that the pole-zero flipping is associated with a change of sign, that is to a nodal point in the mode shapes. This is illustrated in Fig.4.11, which shows the evolution of the zeros of a simply supported beam with a point force actuator located at when the displacement sensor is moved to the right One sees that, as the sensor moves away from the actuator, the zeros migrate along the imaginary axis. When the sensor reaches which is a nodal point of mode 5, the fourth imaginary zero becomes identical to for Similarly, the third zero crosses at which is a node of mode 4. The same phenomenon occurs for other along the beam. Note that the same observation can be made for the three mass system of section 4.6: It is easy to show that, for the nodal point of mode 1 is located between the mass points 1 and 2; for it coincides with 2, and it moves between 2 and 3 for Extensive studies have been conducted on simple structures, to understand

86


how the zeros migrate when the sensor moves away from the actuator (Miu, Spector & Flashner). The following observations have been made: For structures such as bars in extension, shafts in torsion or simply connected spring-mass systems (non dispersive), when the sensor is displaced from the actuator, the zeros migrate along the imaginary axis towards infinity. The imaginary zeros are the resonance frequencies of the two substructures formed by constraining the structure at the actuator and sensor (Problem P.4.8). For beams with specific boundary conditions, the imaginary zeros still migrate along the imaginary axis, but every pair of zeros that disappears at infinity reappears symmetrically at infinity on the real axis and moves towards the origin (Fig.4.12). Systems with right half plane zeros are called non-minimum phase. Thus, non-collocated control systems are always non-minimum phase, but this does not cause difficulties if the right half plane zeros lie well outside the desired bandwidth of the closed-loop system. When they interfere with the bandwidth, they put severe restrictions on the control system, as we shall see in chapter 8. In addition to purely imaginary and purely real zeros, some conservative systems do possess complex zeros, symmetrically located with respect to


87

the imaginary axis; this was shown numerically on a multiply connected spring-mass system and confirmed experimentally on a beam clamped at one end for a special actuator and sensor configuration (Loix et al.) Besides, simulation results suggest that the presence of complex zeros increases the sensitivity of the closed-loop system with respect to the parameter variations.

4.9

The role of damping

To conclude this chapter, we would like to insist on the role of the damping for non-collocated control systems. We have seen that the imaginary zeros provide the necessary phase lead to compensate the undesirable phase lag caused by the poles. Whenever a flexible pole is not associated with a zero, it produces a net phase lag of 180°. According to the stability criterion, the amplitude of the open-loop transfer function must satisfy whenever the phase lag exceeds 180°. Since the amplitude of in the roll-off region is dominated by the resonant peaks of G, it is clear that the damping of the flexible modes is essential for non-collocated systems (Problem P.4.1). Damping augmentation can be achieved by passive as well as active means. For spacecraft applications, the former often use constrained layers of high damping elastomers placed at appropriate locations in the structure (e.g. Johnson et al. or Ikegami & Johnson). More varied ways are regularly used in civil engineering applications, such as tuned-mass dampers, tuned liquid dampers, chain dampers, etc... Active damping is the subject of next chapter.


88

4.10

References

R. H. CANNON & D. E. ROSENTHAL, Experiment in Control of Flexible Structures with Noncolocated Sensors and Actuators, AIAA Journal of Guidance, vol. 7, no. 5, 546–553, Sep.–Oct. 1984. G. F. FRANKLIN, J. D. POWELL & A. EMANI-NAEMI, Feedback Control of Dynamic Systems, Addison-Wesley, 1986. W. B. GEVARTER, Basic Relations for Control of Flexible Vehicles, AIAA Journal, vol. 8, no. 4, 666–672, April 1970. P. C. HUGHES & T. M. ABDEL-RAHMAN, Stability of Proportional Plus Derivative Plus Integral Control of Flexible Spacecraft, AIAA J. Guidance and Control, vol. 2, no. 6, 499–503, Nov.–Dec. 1979. R. IKEGAMI & D. W. JOHNSON, The Design of Viscoelastic Passive Damping Treatments for Satellite Equipment Support Structures, Proceedings of DAMPING’86, AFWAL-TR-86-3059, 1986. C. D. JOHNSON, D. A. KIENHOLZ & L. C. ROGERS, Finite Element Prediction of Damping in Beams with Constrained Viscoelastic Layers, Shock and Vibration Bulletin, , no. 51, 78–81, May 1981. N. LOIX, J. KOZANEK & E. FOLTETE, On the Complex Zeros of Non-Colocated Systems, Journal of Structural Control, vol. 3, no. 1–2, June 1996. G. D. MARTIN, On the Control of Flexible Mechanical Systems, Ph.D. thesis, Stanford University, 1978. D. K. MIU, Physical Interpretation of Transfer Function Zeros for Simple Control Systems with Mechanical Flexibilities, ASME J. Dynamic Systems Measurement and Control, vol. 113, 419–424, Sep. 1991. V. A. SPECTOR & H. FLASHNER, Sensitivity of Structural Models for Noncollocated Control Systems, ASME J. Dynamic Systems Measurement and Control, vol. 111, no. 4, 646–655, Dec. 1989. V. A. SPECTOR & H. FLASHNER, Modeling and Design Implications of Noncollocated Control in Flexible Systems, ASME J. Dynamic Systems Measurement and Control, vol. 112, 186–193, June 1990.

4.11

Problems

P.4.1 Consider the lead compensator for the non-collocated control of the two-mass system (section 4.4).


89

(a) Determine the value of the damping ratio which would reduce the gain margin to zero. (b) What would be the gain margin if instead of P.4.2 Consider the lead compensator plus notch filter (4.10) for the noncollocated control of the two-mass system (section 4.5). Draw the corresponding Bode plots. Select a reasonable value of the gain and compare the bandwidth, the gain and phase margins with those of the lead compensator of Fig.4.6. P.4.3 (a) Repeat the previous problem when the frequency of the appendage is lower than that of the notch filter compare the Bode plots and comment on the role of the damping. (b) Same as (a) with the frequency of the appendage moving away from the notch filter Comment on the importance of tuning the notch filter. P.4.4 Consider the transfer function (4.14) of the three mass system with (a) Compare the Bode plots for and Comment on the phase diagram and its influence on the robustness. (b) For both values of plot the root locus for the lead compensator (4.9) with T = 10 and P.4.5 Consider the PD regulator

applied to the open-loop structure

Assuming that the modes are well separated, show that, for small gain closed-loop damping ratio of mode is

the

Conclude on the stability condition (Gevarter). [Hint: Use a perturbation method, in the vicinity of and write the closed-loop characteristic equation.] P.4.6 Consider a simply supported uniform beam with a point force actuator and a displacement sensor. Based on the result of the previous problem, sketch a non-collocated actuator and sensor configuration such that a PD regulator is stabilizing for the first three modes. P.4.7 Consider an undamped structure. Show that if two neighbouring modes are such that their residues and have the same sign, there is always an imaginary zero between the two poles (Martin).

90


P.4.8 Consider a system of identical masses M simply connected with springs of stiffness k; assume that a point force is applied on mass and a displacement sensor is connected to mass Show that the zeros of the transfer function are the resonance frequencies of the two substructures (from 1 to and from ), formed by constraining the masses and (Miu). P.4.9 Consider the non-collocated control of the two-mass problem [the system transfer function is given by Equ.(4.7)] with M = 1. For various values of the mass ratio 0.01, 0.001, assuming a lead compensator (4.9), draw a diagram of the bandwidth of the control system, as a function of the damping ratio for the limit of stability (GM = 0).

Chapter 5

Active damping with collocated pairs 5.1

Introduction

The role of damping in the gain stabilization of a control system in the roll-off region has been pointed out in the foregoing chapter. The damping also reduces the settling time of the transient response to impulsive loads. Indeed, since the modal expansion of the impulse response matrix corresponding to (2.19) is

and are a Fourier transform pair, see problem P.2.6], one readily sees that the time constant (the memory) of each modal contribution is proportional to If, for example, and which are common values for spacecrafts, the time to reduce the impulse response by a factor of 10 is longer than comparable to that of one orbit revolution of the spacecraft. If one wants, for example, to maintain a micro-gravity environment or the pointing of a telescope, in spite of the transient loads from the thrusters or the human activity, one easily appreciates the need for damping augmentation. Similarly, the damping reduces the amplitude of the FRF in the vicinity of the resonances and, as a result, the steady state response to wideband disturbances [the variance of the stationary modal response to a white noise excitation is proportional to In this chapter, we concentrate on active means for damping augmentation, in short active damping. We will focuse on control schemes with guaranteed stability (at least to a large extent). As already stressed earlier, the use of collocated actuator and sensor (i.e. physically located at the same place and energetically conjugated, such as 91

92


force and displacement or velocity, or torque and angle) leads to an alternating pole/zero pattern, on the imaginary axis if the structure is undamped, or slightly in the left half plane if the structure is lightly damped. Thanks to this property, a number of active damping schemes with guaranteed stability have been developed and successfully tested with various types of actuators and sensors. They can be implemented in a decentralized manner, with each actuator interacting only with the collocated sensor. In this case, the control system consists of independent SISO loops, whose stability can be readily established from the root locus of where is the structure transfer function between the actuator and the collocated sensor, is the active damping compensator, and is the scalar gain. For practical implementation purposes, however, one should be careful that has enough roll-off at high frequency, to accommodate the actuator and power amplifier dynamics and the inevitable phase lag due to sampling in digital controllers. This implies that some roll-off should appear in if there is a feedthrough component in (which often occurs in collocated systems). Nearly collocated systems may sometimes be preferable to strictly collocated ones, to reduce the feedthrough component in while preserving the interlacing property of the poles and zeros in the frequency band where the active damping is significant. It is also important to note that guaranteed stability does not imply guaranteed performance of the control system. Good performance requires the proper sizing and location of the actuator and sensor, to achieve good controllability and observability. This will be reflected by well separated poles and zeros of the open-loop system, leading to wide loops in the root locus plot. The crucial role played by the actuators in the design of the control system was stressed in earlier chapters. The sensors are also technologically critical, for the following reasons: (i) Often, physical problems have natural outputs, because not all the variables (stress, acceleration, velocity,...) are within the range of normally available sensors. For example, vibration amplitudes below one micron may be difficult to measure, while the corresponding accelerations may be easily measured if the structure is stiff, (ii) Apparently similar control problems may involve displacements, velocities and accelerations varying by several order of magnitudes, which may preclude some measurement systems, (iii) The presence of the sensor may significantly alter the dynamic properties of the system, (iv) Even if, in principle, displacements, velocities and accelerations can be computed from one another, this is rarely the case in practice, for noise reasons; besides, only relative displacement and velocity measurements are possible. It is therefore important that the compensator design be consistent with the actual hardware used in the implementation. In this chapter, we describe stable, local controllers, for various types of

Active damping with collocated pairs

93

collocated actuator/sensor pairs. They are summarized in Table 5.1.

5.2

Direct Velocity Feedback

Let be a set of velocity measurements, and a set of collocated control forces, and are energetically conjugated (in the sense that their dot product has the dimension of power). Dissipation can be achieved by proportional (negative) feedback on The governing equations are

Equation (5.3) is the equation of motion including the perturbation (dimension and the control force (dimension acting through the influence matrix


94

B (Problem P.5.1). The structural damping has been omitted for simplicity. Equation (5.4) is the output equation relating the velocity output to the time derivative of the degrees of freedom (d.o.f) of the structure. The fact that the same matrix B appears in Equ.(5.3) and (5.4) is a consequence of collocation. The control law (5.5) is chosen in such a way that G is positive definite. This guarantees that the power is dissipated, because If G is diagonal, the controller is decentralized because every actuator interacts only with its collocated sensor. Substituting Equ.(5.4) and (5.5) into (5.3), one gets

Therefore, the control forces appear as a viscous damping (electrodynamic damping). The damping matrix is positive semi definite, because the actuators and sensors are collocated. The velocity distributions which belong to the null space of (such that remain undamped; such distributions have vibration nodes at the sensor locations.

5.2.1

Lead compensator

Let us examine the SISO case a little closer. In this case, the matrix B degenerates into a control influence vector From Equ.(2.19), we easily see that, if there is no damping, the open-loop transfer function between the control force and the collocated displacement is

where the sum extends to all the modes. We know that the corresponding poles and zeros alternate on the imaginary axis. It is not desirable to implement the derivative compensator as in Equ.(5.5), because it tends to amplify the noise at high frequency; it is preferable to add


95

a low-pass filter, leading to:

A pole has been added at some distance along the negative real axis. This compensator behaves like a differentiator at low frequency The block diagram of the control system is shown in Fig.5.1; a displacement sensor is now assumed and the structural damping is again omitted for simplicity. Typical root locus plots are shown in Fig.5.2 for two values of the low-pass filter corner frequency The closed-loop pole trajectories go from the open-loop poles to the open-loop zeros following branches which are entirely contained in the left half plane. Since there are two more poles than zeros, there are two asymptotes at ±90°. The system is always stable, and this property is not sensitive to parameter variations, because the alternating pole-zero pattern is preserved under parameter variations. At this point, it is appropriate to explain the physical meaning of the fact that the closed-loop poles are located near the zeros of the open-loop system for large values of the gain We know that the imaginary zeros are in fact antiresonance frequencies (i.e. frequencies at which a harmonic excitation produces no displacement where the force is applied); they are identical to the natural frequencies of the modified system in which a support has been added instead of the actuator (Fig.2.4). In fact, as increases from zero, the compensator brings more and more damping into the structure; for larger the closed-loop


96

poles travel on the loops and go to the open-loop zeros, the damping decreases and finally vanishes when Then, the controller excites the structure at the anti-resonance frequencies and, as a result, the point of application of the control force remains fixed, corresponding to a nodal point being enforced by the control system in the closed-loop modes: the control system acts as a support (Fig.5.3). The very lightly damped closed-loop modes, which are not excitable from the input, are likely to be excited by any perturbation which does not enter at the input. Other implementations of the Direct Velocity Feedback and the Lead compensator are possible (Problem P.5.2). The root locus shows that large damping can be achieved only if the loops going from the poles to the zeros are wide in the left half plane, which occurs only if the poles and zeros are well separated.

5.3

Acceleration feedback

This section addresses the case in which the output of the system is the acceleration and the control input is a force. Acceleration is often easier to measure than displacement or velocity, particularly when the structure is stiff.


5.3.1

97

Direct Velocity Feedback

The easiest way to use the acceleration is to integrate it to obtain the velocity. Thus, Direct Velocity Feedback is obtained with the compensator

The stability properties of this compensator have been examined in the foregoing section. In practice, however, piezoelectric accelerometers use charge amplifiers which behave as high-pass filters [Equ.(3.43)]. This does not significantly affect the results if the corner frequency of the charge amplifier is well below the vibration modes of the structure (Problem P.5.4). Next, we present a compensator involving a second order filter which also enjoys guaranteed stability and exhibits a larger roll-off at high frequency. First, we consider the case of a single degree of freedom oscillator; then, we extend the results to SISO systems with many modes, and to MIMO systems.

5.3.2

Second order filter

The basic idea is to pass the acceleration signal through a second order filter with substantial damping, and generate a force feedback proportional to the output of the filter. For a single degree of freedom oscillator, the equations of the system and the compensator are respectively Structure : Compensator :

The control force per unit mass, is obtained as the product of a scalar gain has the dimension of natural frequency squared) by the output of a second order filter with properties excited by the structure acceleration as input. Although the compensator equation is solved by a computer and does not correspond to a physical system, can be interpreted as the relative displacement of a single d.o.f oscillator with properties seismically excited by the acceleration The block diagram of the control system is shown in Fig.5.4, where represents the accelerometer dynamics (assumed to be perfect). The closed-loop characteristic equation, reads

Using the Routh-Hurwitz criterion (see chapter 11), one can check that the combined system is always stable for (Problem P.5.6). Typical root locus plots are shown in Fig.5.5.

98

5.3.3


SISO system with many modes

As we can see from Fig.5.4, the compensator is


99

Its effect on a SISO system with many modes can be investigated by substituting Equ.(5.7) in the block diagram of Fig.5.4; the block diagram is that of Fig.5.6; the corresponding root locus is shown in Fig.5.7 for two values of in both cases, is used; in Fig.5.7.a, the frequency of the filter, is taken close to mode 2 while in Fig.5.7.b, it is selected close to mode 1. Comparing the figures, one sees that all the modes are positively damped, but the mode with the natural frequency close to is more heavily damped. Thus, the performance of the compensator relies on the tuning of the filter on the mode that we wish to damp (this aspect may become problematic if the system is subject to changes in the parameters). The maximum achievable damping ratio increases with a value of between 0.5 and 0.7 is recommended. For closely spaced modes, stability is still guaranteed, but a large damping ratio cannot be achieved simultaneously for the two modes; besides, small variations of the filter frequency may significantly change the root locus and the modal damping (Preumont et al.). If several modes must be damped, several compensators may be used in parallel:

where the are tuned on the targeted modes (Fig.5.8). The gains can be determined with some experimentation (in the sequential selection of the gains it is generally easier to start with the filter of higher frequency, because the roll-off of the second order filter reduces the influence of the filter on the frequencies above


100

As compared to the Direct Velocity Feedback (5.9), the new compensator has a larger roll-off at high frequency instead of which may decrease the risk of destabilizing high frequency modes. The need for tuning the filter frequency may be a drawback if the natural frequency is not known accurately, or is subject to changes. In the foregoing discussion, we have assumed that the sensors and the actuators are perfect. If the actuator dynamics is taken into account, unconditional stability can no longer be guaranteed. If, for example, a proof-mass actuator is used (with in Equ.(3.8), so that it behaves like a perfect force generator in the frequency range of interest), it can be shown that the closed-loop system becomes unstable for large gains (Problem P.5.5). The effect of the charge amplifier dynamics is left as an exercise (Problem P.5.4).

5.3.4

Multidimensional case

Next, we consider the more general case of a MIMO system with collocated actuator/sensor pairs and an array of second order filters. The structure and controller equations are Structure: Sensors: Controller:


101

with the notations and The structural damping has been omitted for simplicity. As before, B is the control influence matrix, indicating the way the control acts on the structure, is the set of accelerations sensed by the collocated sensors. E is a rectangular matrix, which allows us to use more filters than actuators (so that modes can be damped with actuators). The controller equation consists of independent second order filters fed by linear combinations of the accelerations. The control is constructed by multiplying the controller internal variables by the diagonal positive gain matrix G and then by It can be proved that this system is unconditionally asymptotically stable providing the proof is based on Liapunov’s direct method (Sim & Lee).

5.4

Positive Position Feedback

The Positive Position Feedback (PPF) (Goh & Caughey) is appropriate for a structure equipped with strain actuators and sensors; the objective is, once again, to use a second order filter to improve the roll-off of the control system, allowing high frequency gain stabilization.

5.4.1

SISO system

The block diagram is represented in Fig.5.9. As compared to Fig.5.6, the output is now proportional to the displacements (e.g. strain sensor) and a minus sign appears in the controller block (together with the minus sign in the feedback loop, this produces a positive feedback.). Figure 5.10.a and b show the root locus when the controller is tuned on mode 1 and mode 2, respectively. We see that the tuning property of the controller is very similar to that of Fig.5.7 and, even in presence of a feedthrough component, the open-loop transfer function has a roll-off of However, there is a stability limit which is


102

reached when the open-loop static gain is equal to 1 (Problem P.5.7). Note that this condition is independent of the damping.

5.4.2

Multidimensional case

The extension to SISO systems with many modes is parallel to that of the acceleration feedback. The modal selectivity is also achieved by tuning the filter frequency on the targeted mode. Several modes can be targeted simultaneously with the form (5.15) (with a negative sign in this case, to produce a positive feedback). As for the acceleration feedback, in the design of the compensator involving several modes, it is generally more convenient to start with the component of larger frequency. The MIMO implementation of the PPF is also very similar to that of the acceleration feedback: Structure: Sensors: Controller:

A necessary and sufficient condition for asymptotic stability can be established using Liapunov’s direct method (see Fanson & Caughey). An application of the


103

PPF to the active damping of a plate equipped with piezoelectric actuators and sensors will be examined in chapter 13.

5.5

Integral Force Feedback

Consider the active truss of section 3.7. Each active strut consists of a piezoelectric linear actuator collocated with a force transducer. In this section, we consider the decentralized active damping with local controllers connecting each actuator to its collocated force sensor. The open-loop transfer function between the voltage applied to the piezo and the sensor signal is given by Equ.(3.52) or (3.59). A typical Bode plot of is represented in Fig.3.18; once again, there are alternating poles and zeros along the imaginary axis; there is a feedthrough component and, according to the foregoing discussion, some roll-off is necessary in the active damping compensator D(s) to achieve stability. It is readily established from the root-locus (Fig.5.11) that the positive Integral Force Feedback (IFF):

is unconditionally stable for all values of The negative sign in Equ.(5.24) is combined with the negative sign in the feedback loop (Fig.5.12) to produce a positive feedback. In practice, it is not advisable to implement plain integral control, because it would lead to saturation. A forgetting factor can be introduced by slightly moving the pole of the compensator from the origin to the negative real axis, leading to


104

This does not affect the general shape of the root locus and prevents saturation. Note that piezoelectric force sensors have a built-in high-pass filter.

5.5.1

Modal damping

The behaviour of an active truss equipped with a single active strut has been analysed in section 3.7. Combining with the control law (5.24), we find the equations governing the control system: Structure: Sensor: Controller:

where is the unconstrained piezoelectric displacement (proportional to the voltage applied to the piezo, except for the hysteresis), is the relative displacement of the end nodes of the active element, is the force in the active strut and its stiffness. Note that the stiffness matrix K includes the active strut. is the equivalent piezoelectric load. Combining (5.27) and (5.28), we get

Taking the Laplace transform of (5.26) and substituting δ we get the closed-loop characteristic equation


105

From this equation, we can deduce the open-loop transmission zeros, which coincide with the asymptotic values of the closed-loop poles as Taking the limit, we get which states that the zeros (i.e. the anti-resonance frequencies) coincide with the poles (resonance frequencies) of the structure where the active strut has been removed (corresponding to the stiffness matrix To evaluate the modal damping, Equ.(5.30) must be transformed in modal coordinates with the change of variables Assuming that the mode shapes have been normalized according to and taking into account that we have

The matrix is, in general, fully populated; if we assume that it is diagonally dominant, and if we neglect the off-diagonal terms, it can be rewritten

where

is the fraction of modal strain energy in the active member when the structure vibrates according to mode was defined by Equ.(3.53)]. Substituting Equ.(5.33) into (5.32), we find a set of decoupled equations

For small we can assume a solution of the form into (5.35) and neglecting the terms of the second order in P.5.9)

substituting we get (Problem

This interesting result tells us that, for small gains, the active damping ratio in a given mode is proportional to the fraction of modal strain energy in the active element. This result is very useful for the design of active trusses; the active struts should be located in order to maximize the fraction of modal strain energy in the active members for the critical vibration modes. is readily available from F.E. softwares. Notice that the damping ratio decreases with the frequency of the mode, as we would expect from the integral shape of the controller. A root locus plot of the closed-loop poles can be obtained as follows. If we denote

106


we can write Equ.(5.35) as

or

This equation shows that, for each mode, the closed-loop poles follow the root locus corresponding to the open-loop transfer function

From the comparison with Equ.(5.31), which shows the asymptotic behaviour of the closed-loop system as the zeros at are readily identified as the natural frequencies of the structure when the active member has been removed; the poles are those of the open-loop structure and the integral controller. The root locus is shown in Fig.5.13. It can be shown that the maximum modal damping is given by

and it is achieved for (Problem P.5.10). The preceding result has been established for a single active member; if there are several active members operating with the same control law and the same gain this result can be generalized under similar assumptions. It can be shown that each closed-loop pole follows a root locus governed by Equ.(5.38)


107

where the pole is the natural frequency of the open-loop structure and the zero is the natural frequency of the structure where the active members have been removed. The proof is left as an exercise (Problem P.5.11).

5.6 5.6.1

Remarks Controllability, observability

For all the control laws discussed in this chapter, the guaranteed stability is a consequence of the interlacing of the poles and zeros of the structure along the imaginary axis. This property is guaranteed for collocated systems, for any location of the actuator-sensor pair in the structure. However, if the pair is located where the controllability and observability (they always go together for collocated pairs) are weak, the corresponding loops in the root locus are very small and remain close to the imaginary axis; this severely restricts the achievable active damping. It is important to stress that good performance requires good controllability (Problem P.5.12).

5.6.2

Actuator and sensor dynamics

108


Throughout this chapter, it has been assumed that the actuator and sensor have perfect dynamics. As a result, the active damping algorithms are stable for all gains [except the PPF)]. In practice, however, the open-loop transfer function (5.2) becomes where includes the sensor, actuator and the digital controller dynamics. In chapter 3, we saw that the low frequency behaviour of the proof-mass actuator (3.8) and the charge amplifier (3.43) can both be approximated by a second order high-pass filter. The high frequency behaviour of the actuators and sensors can often be represented by a second order low-pass filter

It is easy to see that the two extra poles in bring the asymptotes inside the right half plane and substantially alter the root locus for Figure 5.14 shows the effect of the second order low-pass filter (5.42) on the root locus plot of the Direct Velocity Feedback (5.8) (we have assumed and The active damping is no longer unconditionally stable and always has some destabilizing influence on the high frequency modes. Fortunately, the modes of the structure are not exactly on the imaginary axis, because of the structural damping (Fig.5.15); this allows us to operate the controller with small gains. The control system becomes insensitive to the actuator dynamics if


109

is far beyond the cross-over frequency of (5.2). The effect of the low frequency dynamics of a proof-mass actuator on the acceleration feedback is examined in Problem P.5.5

5.7

References

J. N. AUBRUN, Theory of the Control of Structures by Low-Authority Controllers, AIAA J. of Guidance, vol. 3, no. 5, 444–451, Sep.–Oct. 1980. M. J. BALAS, Direct Velocity Feedback Control of Large Space Structures, AIAA J. of Guidance, vol. 2, no. 3, 252-253, 1979. A. BAZ, S. POH & J. FEDOR, Independent Modal Space Control with Positive Position Feedback, Trans. ASME, J. of Dynamic Systems, Measurement, vol. 114, no. 1, 96–103, March 1992. R. J. BENHABIB, R. P. IWENS & R. L. JACKSON, Stability of Large Space Structure Control Systems Using Positivity Concepts, AIAA J. of Guidance and Control, vol. 4, no. 5, 487–494, Sep.–Oct. 1981. J. L. FANSON & T. K. CAUGHEY, Positive Position Feedback Control for Large Space Structures, AIAA Journal, vol. 28, no. 4, 717–724, April 1990. R. L. FORWARD, Electronic Damping of Orthogonal Bending Modes in a Cylindrical Mast Experiment, AIAA Journal of Spacecraft, vol. 18, no. 1, 11–17, Jan.–Feb. 1981. W. B. GEVARTER, Basic Relations for Control of Flexible Vehicles, AIAA Journal, vol. 8, no. 4, 666–672, April 1970. C. G OH & T. K. C A U G H E Y , On the Stability Problem Caused by Finite Actuator Dynamics in the Control of Large Space Structures, Int. J. of Control, vol. 41, no. 3, 787–802, 1985. A. PREUMONT, J. P. DUFOUR & C. MALEKIAN, Active Damping by a Local Force Feedback with Piezoelectric Actuators, AIAA J. of Guidance, vol. 15, no. 2, 390–395, March–April 1992. A. PREUMONT, N. Loix, D. MALAISE & O. LECRENIER, Active Damping of Optical Test Benches with Acceleration Feedback, Machine Vibration, vol. 2, 119–124, 1993. D. SCHAECHTER, Optimal Local Control of Flexible Structures, AIAA J. of Guidance, vol. 4, no. 1, 22–26, 1981. E. S IM & S. W. L EE , Active Vibration Control of Flexible Structures with Acceleration or Combined Feedback, AIAA J. of Guidance, vol. 16, no. 2, 413–415, 1993.


110

5.8

Problems

P.5.1 Consider a finite element representation of a, cantilever beam with 5 elements of equal length. Write the influence matrix B [Equ.(5.3)] for (a) a point force actuator at the tip (b) a torque actuator at the tip Check that the sensor output matrix [Equ.(5.4)] for collocated displacement and angular sensor is the transpose P.5.2 Compare the following implementations of the Lead and the Direct Velocity Feedback compensators:

Discuss the conditions under which these compensators would be applicable for active damping. P.5.3 Assuming that the modes are well separated, show that for small gain the Direct Velocity Feedback produces a closed-loop damping ratio

where the mode shapes are normalized in such a way that (Aubrun). [Hint: Use a perturbation method, in the vicinity of and write the closed-loop characteristic equation]. P.5.4 Consider a vibrating structure with a point force actuator collocated with an accelerometer. For the two compensators discussed in connection with the acceleration feedback, analyse the effect of the charge amplifier dynamics [Equ.(3.43)] on the root locus. Show that the stability is no longer guaranteed, but if the sensor dynamics does not affect the control system for small gains. P.5.5 Consider a s.d.o.f. oscillator controlled with a proof-mass actuator; analyse the effect of the actuator dynamics on the active damping by acceleration feedback with a second order filter. Show that if the natural frequency of the actuator is much lower than that of the filter the actuator dynamics has little influence on the root locus for small gains, but that the system becomes unstable for large gains.


111

P.5.6 Using the Routh-Hurwitz criterion, show that the acceleration feedback based on the second order filter is always stable for [Start from the closed loop characteristic equation (5.13)] P.5.7 Consider the second order Positive Position Feedback applied to a s.d.o.f. oscillator; using the Routh-Hurwitz criterion, show that the stability limit is reached when the open-loop static gain is equal to 1. P.5.8 Show that for a collocated SISO system without feedthrough, the Positive Position Feedback can be implemented with the following first order controller

Show that for a s.d.o.f. oscillator, the stability condition is (Baz et al.). P.5.9 Consider a truss with an active strut consisting of a linear piezo actuator collocated with a force sensor. Show that, for small gain the Integral Force Feedback produces a closed-loop damping ratio

(Preumont et al.). [Hint: Start from the closed-loop characteristic equation (5.35) and assume P.5.10 Show that the maximum damping achievable in an active truss with the integral force feedback is

and that it is achieved for P.5.11 Consider a truss structure with several identical active members controlled with the same control law (IFF) and the same gain. Making the proper assumptions, show that each closed-loop pole follows a root locus defined by Equ.(5.39), where the natural frequency is that of the open-loop structure and is that of the structure where the active members have been removed. P.5.12 Compare the transfer functions of Fig.3.26 from the point of view of control authority. Comment on the consequences of the error in predicting the zeros.


Chapter 6

Active vibration isolation 6.1

Introduction

There are two broad classes of problems in which vibration isolation is necessary: 1. Operating equipments generate oscillatory forces which can propagate in the supporting structure (Fig.6.1.a).This situation corresponds to that of an engine in a car, or an attitude control reaction wheel assembly in a spacecraft. 2. Sensitive equipments may be supported by a structure which vibrates appreciably (Fig.6.l.b). This situation corresponds to, for example, a telescope in a spacecraft, a precision machine tool in a workshop, or a passenger seated in a car.

The disturbance may be either deterministic, such as the unbalance of a motor, or random as in a passenger car riding on a rough road. As already mentioned, this book is focused on the feedback strategies for active isolation; they apply to both deterministic and random disturbances. For deterministic sources such as a rotating unbalance, feedforward control can also be very effective (see Fuller et al., chap. 7). In this chapter, we first consider the single-axis isolation problem; the acceleration feedback (also called “sky-hook” damper) and the force feedback are compared. Next, we consider the general purpose 6-axis isolator based on a Stewart platform. The decentralized integral force feedback of a Stewart platform is examined. The joint pointing/isolation control is also briefly examined. Finally, we consider the application of the sky-hook damper to an active vehicle suspension. 113


114

6.2

Passive isolator

First, let us consider the situation of an operating equipment generating a disturbance force (Fig.6.l.a). If the support is fixed, the governing equations are as follows:

In the Laplace domain,

where

and stand for the Laplace transform of respectively and and with the usual notations and The transmissibility of the support is defined in this case as the transfer function between the disturbance force applied to the mass and the force transmitted to the support structure; combining Equ.(6.3) and (6.4), we get

Next, consider the second situation illustrated in Fig.6.1.b; the disturbance is the displacement of the supporting structure (that we call “dirty body”) and the system output is the displacement of the sensitive equipment (referred to as “clean body”). Following a procedure similar to that leading to Equ.(6.5),

Active vibration isolation

115

it is easily established (Problem P.6.1) that the transmissibility of this isolation system, defined in this case as the transfer function between the displacement of the dirty body and that of the clean body, is given by

which is identical to Equ.(6.5); the two isolation problems can therefore be treated in parallel. The amplitude of the corresponding FRF (for is represented in Fig.6.2 for various values of the damping ratio We observe that All the curves are larger than 1 for and become smaller than 1 for Thus the critical frequency separates the domains of amplification and attenuation of the isolator. When the high frequency decay rate is that is -40 dB/decade, while very large amplitudes occur near the corner frequency (the natural frequency of the spring-mass system).


116

The damping reduces the amplitude at resonance, but also tends to reduce the effectiveness at high frequency; the high frequency decay rate becomes

The design of a passive isolator involves a trade-off between the resonance amplification and the high frequency attenuation; the ideal isolator should have a frequency dependent damping, with high damping below the critical frequency to reduce the amplification peak, and low damping above to improve the decay rate. The objective in designing an active isolation system will be to add a force actuator working in parallel with the spring and dashpot (Fig.6.1.c), that will operate in such a way that there is no amplification below and the decay rate is -40dB/decade at high frequency, as represented in dotted line in Fig.6.2.

6.3

The “sky-hook” damper

Consider the single axis isolator of Fig.6.3. It consists of a soft spring acting in parallel with a force actuator generating a control force the objective is to isolate the clean body from the disturbance applied to the dirty body. With the usual Laplace notations, the equations governing the system are as follows: Dirty body:

Clean body:


117

Spring:

In the sky-hook isolator, a feedback link is established between the absolute acceleration of the clean body, and the actuator force with an integral controller (Fig.6.4.a), leading to a

We see that the active control force is proportional to the clean body absolute velocity; this is why this control is called “sky-hook” damper (Fig.6.4.b).

Introducing Equ.(6.10) into (6.8) and using Equ.(6.9), we get

or

Thus, the transmissibility has a corner frequency at its high frequency decay rate is that is -40 dB/decade, and the control gain can be chosen in such a way that the isolator is critically damped the corresponding value of the gain is In this way, we achieve a low-pass filter without overshoot with a roll-off of -40 dB/decade. This transmissibility follows exactly the objective represented in Fig.6.2. Combining Equ.(6.7) to (6.9), the open-loop transfer function of the soft isolator of Fig.6.3 can be written


118

The open-loop poles are the natural frequencies of the system without control. The rigid body modes do not appear in the transfer function (6.13) because they are not controllable from The root locus of the closed-loop poles as the gain of the controller increases is shown in Fig.6.5.

In developing the sky-hook damper, we have assumed that the passive part involves only a spring. If we include also a dash-pot of constant Equ.(6.9) must be replaced by

and Equ.(6.11) becomes

which shows that the high frequency roll-off is reduced to -20dB/decade. Thus, better performances will be achieved without passive damping (Problem P.6.2).

6.4

Force feedback

If we look at Equ.(6.8), we see that the acceleration of the clean body is proportional to the total force transmitted by the interface, As a result, the sky-hook damper can be obtained alternatively with the control configuration of Fig.6.6, where a force sensor has been substituted to the acceleration sensor. Equations (6.7) to (6.9) still hold and we must add the sensor equation


119

It follows that the open-loop transfer function between the actuator force input and the force sensor output F is

It is identical to Equ.(6.13), except for the constant M relating the total force F to the clean body acceleration. As a result, the root locus of Fig.6.5 also applies to the integral force feedback damper which is totally equivalent to the sky-hook damper. Taking into account the constant M relating the clean body acceleration and the total transmitted force, the transmissibility of the active isolator now reads

which is totally equivalent to Equ.(6.12)(except for the slightly different meaning of The control strategies based on acceleration feedback and on force feedback appear as totally equivalent for the isolation of rigid bodies. However, the force feedback has two advantages which may be significant in some circumstances. The first one is related to sensitivity: force sensors with a sensitivity of are commonplace and commercially available; if we consider a clean body with a mass of 1000 kg (e.g. a telescope), the corresponding acceleration is Accelerometers with such a sensitivity are more difficult to find; for example, the most sensitive accelerometer available in the Bruel & Kjaer catalogue is (model 8318). The second advantage is stability when the clean body is flexible; this is discussed below.

6.5

Flexible clean body

When the clean body is flexible, the behaviour of the acceleration feedback and the force feedback are no longer the same, due to different poles/zeros config-

120


urations of the two control strategies. In fact, different sensor configurations correspond to different locations of the zeros in the s-plane. To analyse this situation, consider the clean body with a flexible appendage as represented in Fig.6.7; the nominal numerical values used in the calculations are the mass of the flexible appendage is taken as a parameter to analyse the interaction between the flexible appendage and the isolation system. When is small, the flexible appendage is much more rigid than the isolation system and the situation is not much different from that of a rigid body. Figure 6.8 shows the root locus plots for the acceleration feedback and the force feedback have similar root locus plots, with a new pole/zero pair appearing higher on the imaginary axis; the only difference between the two plots is the distance between the pole and the zero which is larger for the acceleration feedback; as a result, the acceleration feedback produces a larger damping of the higher mode. In contrast,

when

is large, the root locus plots are reorganized as shown in Fig.6.9 for For force feedback (Fig.6.9.a), the poles and zeros still alternate on


121

the imaginary axis, leading to a stable root locus; this property is lost for the acceleration feedback (Fig.6.9.b), leading to an unstable loop for the lower mode. In practice, however, this loop is moved slightly to the left by the structural damping, and the control system can still operate for small gains (conditionally stable).

6.5.1

Free-free beam with isolator

To analyse a little further the situation when the clean body is flexible, consider the vertical isolation of a free-free continuous beam from the disturbance of a dirty body of mass (Fig.6.l0.a). This situation is representative of a large space structure with its attitude control system. Note that the rigid body modes are uncontrollable from the internal force In the numerical example described below, the length of the beam is the mass per unit length is the stiffness of the isolator is and the mass of the dirty body is the stiffness EI of the beam is taken as a parameter. Let be the natural frequencies of the flexible modes of the free-free beam alone (Fig.6.10.b) and be the transmission zeros corresponding to a force excitation and a collocated displacement sensor (or equivalently acceleration). According to section 2.4, are the natural frequencies of the system with an additional restraint at the connecting degree of freedom of the isolator. Because of the collocation (Fig.2.5), the poles and zeros are alternating on the imaginary axis, so that and satisfy

Next, consider the complete system (beam + dirty body) and let be its natural frequencies (flexible mode only, because the rigid body modes are not controllable from the internal force If the control system uses a force sensor

122


(Fig.6.10.d), the transmission zeros, obtained by enforcing a zero force at the connecting d.o.f., are identical to the natural frequencies of the system when the isolator is disconnected from the beam; which are identical to the natural frequencies of the free-free beam. It can be shown that the residues in the modal expansion of the FRF between the force actuator and the force sensor F are all positive (Problem P.6.4). It follows that the poles and zeros alternate and the following relation holds:

This condition guarantees a fixed, interlacing, pole/zero pattern of the open-


123

loop system and the stability of the closed-loop system when a force feedback is used. With an acceleration feedback (sky-hook damper, Fig.6.10.c) the poles are still while the zeros, obtained by enforcing a zero acceleration at the connecting d.o.f. are as for the free-free beam of Fig.6.10.b. This actuator/sensor configuration is no longer collocated, so that no condition similar to (6.19) or (6.20) holds between and When the beam is stiff, the interlacing property is satisfied and the stability is guaranteed, but as the beam becomes more flexible, the values of and decrease at different rates and a pole/zero flipping occurs when they both become equal to the natural frequency of the isolator (Fig.6.11). As a result, the system is no longer unconditionally stable when the flexibility is such that and above. As a conclusion to this section, it seems that the sky-hook damper imple-


124

mentation (acceleration feedback) is preferable when the clean-body is fairly stiff as compared to the isolator corner frequency (e.g. car suspension), to benefit from the better active damping properties of the flexible modes (Fig.6.8). On the contrary, the force feedback implementation looks preferable when the clean body is very flexible (e.g. space structure), to benefit from the guaranteed stability.

6.6

6 d.o.f. isolator

In the foregoing sections, we have analysed a single axis active isolator which combines a -40 dB/decade attenuation rate in the roll-off region with no overshoot at the corner frequency. To fully isolate two rigid bodies with respect to each other (e.g. an engine from the body of a car, or a telescope from the carrier spacecraft) we would need six such isolators judiciously placed, that could be controlled either in a centralized or (more likely) in a decentralized manner. For a number of space applications, generic multi-purpose 6 d.o.f. isolators have been developed with a standard Stewart platform architecture (Fig.6.12). These general purpose isolators can be used to isolate noisy components such as a reaction wheel assembly (RWA) from the rest of the spacecraft (the wheels apply low frequency torques to control the attitude of the spacecraft, but in doing so they also generate undesirable high frequency forces and torques shaking


125

the base of the structure; this situation is depicted in Fig.6.1.a). Alternatively, they can also be used to isolate quiet subsystems such as a telescope or other sensitive payloads from the carrier spacecraft (Fig.6.1.b). The isolation objectives are illustrated in Fig.6.13; the ideal 6 d.o.f. isolation platform should transmit the attitude control torque within the bandwidth of the attitude control system and filter all the high frequency disturbances above

The Stewart platform of Fig.6.12 uses 6 identical active struts arranged in a mutually orthogonal configuration connecting the corners of a cube as indicated in Fig.6.14. The triangular base plate connects the nodes 1,3,5 and the payload plate connect the nodes 2,4,6. This cubic architecture is interesting because it provides a uniform control capability in all directions, a uniform stiffness in all directions, and it minimizes the cross-coupling amongst actuators (being orthogonal to each other). It also tends to minimize the spread of modal frequencies. Figure 6.14 depicts the geometry of the hexapod and the numbering system for the nodes and the struts; the basic frame has its origin at node 0; the reference (or payload) frame has its origin at the geometrical center of the hexapod, noted as node 8, and is perpendicular to the payload plate; the orientation of and is shown in Fig.6.14. If we neglect the flexibility of the struts and the bending stiffness of the flexible joints connecting the struts to the base and payload plates, the dynamic equations of motion can be obtained from rigid body dynamics. If the base is fixed and the payload is an axisymmetrical rigid body of mass with the principal axes of inertia aligned with principal moment

126


of inertia and with the center of mass located at an offset distance from the geometrical center, along the vertical axis the dynamic equation (for small rotations) of the isolator is

where is the vector describing the small displacements and rotations in the payload frame is the vector of active control forces in strut 1 to 6 respectively, M and K are the mass and stiffness matrices of the isolator, given by

where is the stiffness of one strut. B is the influence matrix of the active control force:


127

(B is the transposed of the Jacobian between the small displacements and rotations and the length increments of the legs; L is the length of the struts). The proof of Equ. (6.21) to (6.24) is left as an exercise (Problem P.6.5). The natural frequencies of the isolator and its payload can be obtained by solving the eigenvalue problem on the left side of Equ. (6.21); the or “bounce” mode and the or “torsional” mode are decoupled, with natural frequencies given by

where and is the radius of gyration normalized to the strut length. For most cases, and The remaining four modes are lateral bending coupled with shear; their natural frequencies occur in two identical pairs, solutions of the characteristic equation

where is the radius of gyration normalized to the strut length and is the center of mass offset normalized to the strut length. Note that if the center of mass is at the geometric center and if and the hexapod will have 6 identical natural frequencies, all equal to

6.7

Decentralized control of the 6 d.o.f. isolator

Let us assume that each leg is equipped with a force sensor as in Fig.6.6; since each leg consists of a passive spring of stiffness in parallel with a force actuator, the output equation reads

where is the vector containing the 6 force sensor outputs, is the vector of leg extensions from the equilibrium position, is the strut stiffness and is the vector of active control forces. Taking into account


128

the relationship between the leg extension and the payload frame displacement, we have

Once again, we note that the same matrix B appears in Equ.(6.21) and (6.28) because the sensors and actuators are collocated. Using a decentralized integral force feedback with constant gain we find the controller equation

Combining Equ.(6.21), (6.28) and (6.29), we obtain the closed-loop equation

and, taking into account that

If we transform into modal coordinates, and take into account the orthogonality relationships (2.8) and (2.9), the characteristic equation is reduced to a set of uncoupled equations

Thus, every mode follows the characteristic equation

or

The corresponding root locus is shown in Fig.6.15. It is identical to Fig.6.5 for a single-axis isolator; however, unless the 6 natural frequencies are identical, a given value of the gain will lead to different pole locations for the various modes and it will not be possible to achieve the same damping for all modes. This is why it is recommended to locate the payload in such a way that the spread of the modal frequencies is minimized.

6.7.1

Remarks

(i) The foregoing model is based on the assumption that there is no structural damping and that the only contribution to the stiffness matrix is the axial


129

stiffness of the struts, so that In practice, however, the spherical joints at the connection between the legs and the base plates of the Stewart plateform are replaced by flexible connections with high longitudinal stiffness and low bending stiffness (such flexible tips can be seen in Fig.3.5), in such a way that the stiffness matrix has an additional contribution, from the elastic joints. Thus, and the closed loop equation becomes

According to this equation, the transmission zeros, which are the asymptotic solutions of Equ.(6.34) as are no longer at the origin but are solutions of the eigenvalue problem

This shift of the zeros may have a substantial influence on the practical performances of actual Stewart platforms. (ii) The foregoing model assumes that the legs of the Stewart platform have no mass. In fact, the magnetic circuit of the voice coil actuator is fairly heavy and the local dynamics of the legs may interfere with that of the Stewart platform, resulting in a reduced attenuation of the isolation system.

6.8

Pointing control

At this stage, it is worth pointing out that, in addition to providing vibration isolation, a Stewart platform may also be used for pointing of, for example, a

130


telescope towards a target star (the reference pointing accuracy of the Hubble Space Telescope is but future missions are likely to require better accuracy by at least one order of magnitude). If are the desired pointing angles, the decoupling matrix D relating to the corresponding leg extensions in the nominal position is obtained from the appropriate columns of the Jacobian


131

Figure 6.16 shows a possible control strategy. Note that the six-leg Stewart platform has some built-in redundancy which can be exploited to account for actuator failure. Figure 6.17 shows the block diagram of the combined control strategy for pointing and isolation.

6.9

Vehicle suspension

Figure 6.18 shows a quarter-car model of a vehicle. Although this 2 d.o.f. model is too simple for performing a comprehensive analysis of the ride motion, it is sufficient to gain some insight in the behaviour of passive and active suspensions in terms of vibration isolation (represented by the body acceleration suspension travel and road holding (represented by the tyre deflexion Typical numerical values used in the simulation reported later are also given in the figure (taken from Chalasani, 1984). The stiffness corresponds to the tyre; the suspension consists of a passive part and an active one, assumed to be a perfect force actuator acting as a sky-hook damper in this case (the active control force is applied on both sides of the active device, to the body and to the wheel of the vehicle). Figure 6.19 shows the FRF from the roadway vertical velocity to the car body acceleration for the passive suspension alone; several values of the damping coefficient c are considered. The first peak corresponds to the body resonance (also called sprung mass resonance) and the second one to the wheel resonance (unsprung mass resonance). The passive damping cannot control the body resonance without reducing the isolation at higher frequency. Next, a sky-hook damper is added. Figure 6.20 shows the corresponding FRF from to for various values of the control gain. Note that

132



133

the body resonance can be damped without reducing the isolation at higher frequency but the peak in the FRF corresponding to the wheel resonance cannot be changed by the active control. Figure 6.21 compares the amplitude of the FRF and for two values of the gain. This figure shows that the absolute velocity of the body rolls-off much faster (i.e. has much lower frequency components) than the relative velocity This point will be further discussed when we examine the semi-active suspension devices which try to emulate the sky-hook damper by acting on the flow parameters of the damper acting on the relative velocity.

6.10

References

R. M. C HALASANI , Ride Performance Potential of Active Suspension Systems, Part1: Simplified Analysis Based on a Quarter-Car Model, ASME Symposium on Simulation and Control of Ground vehicles and Transportation systems, Anaheim, CA, Dec. 1984. C. R. F ULLER , S. J. E LLIOTT & P. A. N ELSON , Active Control of Vibration, Academic Press, 1996.


134

Z. J. G ENG & L. S. H AYNES , Six Degree-of-Freedom Active Vibration Control Using the Stewart Platforms, IEEE Transactions on Control Systems Technology, vol. 2, no. 1, 45–53, March 1994. C. E. KAPLOW & J. R. VELMAN, Active Local Vibration Isolation Applied to a Flexible Space Telescope, AIAA J. Guidance and Control, vol. 3, no. 3, 227–233, May–June 1980. D. K ARNOPP , Design Principles for Vibration Control Systems Using SemiActive Dampers, Trans. ASME Journal of Dynamic Systems, Measurement and Control, vol. 112, 448–455, Sep. 1990. D. C. KARNOPP & A. K. TRIKHA, Comparative Study of Optimization Techniques for Shock and Vibration Isolation, Trans. ASME, Journal of Engineering for Industry, series B, vol. 91, no. 4, 1128–1132, 1969. R. A. LASKIN & S. W. SIRLIN, Future Payload Isolation and Pointing System Technology, AIAA J. Guidance and Control, vol. 9, no. 4, 469–477, July–Aug. 1986. J. E. Mc INROY, G. W. NEAT & J. F. O’BRIEN, A Robotic Approach to Fault-Tolerant, Precision Pointing, IEEE Robotics & Automation Magazine, pp. 24–31, Dec. 1999. J. SPANOS, Z. RAHMAN & G. BLACKWOOD, A Soft 6-Axis Active Vibration Isolator, Proceedings of the American Control Conference, Seattle, WA, pp. 412–416, June 1995. D. STEWART, A Platform with Six Degrees of Freedom, Proc. Instn. Mech. Engrs, vol. 180, no. 15, 371–386, 1965–66. D. THAYER, J. VAGNERS, A. VON FLOTOW, C. HARDHAM & K. SCRIBNER, Six-Axis Vibration Isolation System Using Soft Actuators and Multiple Sensors, Proc. of Annual American Astronautical Society Rocky Mountain Guidance and Control Conference (AAS-98-064), pp. 497–506, 1998.

6.11 P.6.1

Problems

Consider the passive isolator of Fig.6.1.b. Find the transmissibility of the isolation system. P.6.2 Analyse the effect of the passive damping on the transmissibility of the sky-hook damper. Is it beneficial or detrimental?


135

P.6.3 Write the differential equations governing the system of Fig.6.7 in state variable form. Using the following values of the parameters: Write the open-loop frequency response for the acceleration feedback and force feedback configurations and draw the corresponding poles/zeros pattern. In both cases draw the root locus for an integral controller. Do the same for investigate the effect of structural damping. P.6.4 Consider the modal expansion of the open-loop FRF of the system of Fig.6.10.d. Show that the residues are all positive and that this results in alternating poles and zeros. P.6.5 Consider the cubic Stewart platform of Fig.6.14: (1) Using energetic considerations, show that the mass matrix M, the stiffness matrix K, and the control force influence matrix B are given by Equ.(6.22) to (6.24). (2) Show that the natural frequencies are given by Equ.(6.25) and (6.26). [Hint: write the Jacobian relating the leg extension to the motion of the payload frame The stiffness matrix is given by where is the stiffness of one strut.] P.6.6 Consider a Stewart platform of cubic architecture supporting an axisymetric payload such that (1) Select the spring stiffness in order to achieve a corner frequency of about 10Hz. (2) Calculate the corresponding natural frequencies. (3) Select the controller gain to achieve a reasonable damping ratio for the 6 modes. P.6.7 Consider the active isolation problem described in the figure below, where is assumed to be a white-noise process. Show that the optimal control minimizing the objective function

is

where the coefficients and depend on the weighting coefficients and This is the “sky-hook” damper (Karnopp & Trikha, 1969). (The solution of this problem requires a prior reading of Chapters 7 and 9). P.6.8 (1) Develop a state space model of the vehicle suspension of Fig.6.18 (use and as state variables). (2) Plot the amplitude of the FRF of the passive suspension for various

136


values of the damping coefficient Compare with Fig.6.19. (3) Consider the active suspension with a sky-hook damper control law. Plot the amplitude of the FRF for the nominal parameters and various values of the gain Compare the frequency content of the absolute velocity of the body with that of the relative velocity P.6.9 Consider the fully active suspension of Fig.6.18 Using the LQR theory, design a state feedback which minimizes the performance index

The terms involved in the state penalty of this performance index represent respectively the vibration isolation (comfort), the suspension travel and the road holding (unsprung mass velocity). The fourth one is the control force. Try various values for the weighting coefficients and Compare the fullstate feedback to the sky-hook damper (Chalasani, 1984). (The solution of this problem requires a prior reading of Chapters 7 and 9).

Chapter 7

State space approach 7.1

Introduction

The methods based on transfer functions are often referred to as Classical Methods; they are quite sufficient for the design of single-input single-output (SISO) control systems, but they become difficult to apply to multi-input multi-output (MIMO) systems. By contrast, the design methods based on the state space approach, often called Modern Methods, start from a system description using first order differential equations governing the state variables. The formalism applies equally to SISO and MIMO systems, which considerably simplifies the design of multivariable control systems. Although it is formally the same for SISO and MIMO systems, we shall first study the state space methods for SISO systems. In this case, it is possible to draw a clear parallel with the frequency domain approach, and it is possible to solve the problems of the optimum state regulator and the optimum state observer on a purely geometric basis, replacing the solution of the Riccati equation by the use of the symmetric root locus. The optimal control of MIMO systems will be considered in chapter 9. In the state variable form, a linear system is described by a set of first order linear differential equations

with the following notations: state vector, input vector, output vector, = system noise, = measurement noise, 137

138


A = system matrix, B = input matrix, C = output matrix, D = feedthrough matrix, E = system noise input matrix. For SISO systems, both and are scalar functions. Equation (7.1) provides a complete description of the internal dynamics of the system, which may be partially hidden in the transfer function (the internal dynamics may include uncontrollable and unobservable modes which do not appear in the input/output relationship). The feedthrough matrix D is often omitted in control textbooks; however, in earlier chapters, we saw several examples where a feedthrough component does occur as a result of the sensor type (e.g. accelerometer) or location (when collocated with the actuator), or as a result of the modal truncation (residual mode); the impact of a feedthrough component on the compensator design was emphasized in chapter 5. The system noise may include environmental loads, modelling errors, unmodelled dynamics (including that of the actuator and sensor), nonlinearities, and the noise in the input vector. The system noise input matrix E is in general different from the input matrix B. The measurement noise includes the sensor noise and the modelling errors. It must be emphasized that the choice of the state variables is not unique. In principle, their number is independent of the specific choice (it is equal to the order of the system), but this is not true in practice, because a model can only be correct over a limited bandwidth; if the state variables can be selected in such a way that some of them do not respond dynamically within the bandwidth of interest, they can be eliminated from the dynamic model and treated as singular perturbations (i.e. quasi-static). In structural control, the modal coordinates often lead to the minimum number of state variables.

State space approach

7.2

139

State space description

In this section we review the state variable form of the dynamic equations for a number of mechanical systems that will be used later in this chapter.

7.2.1

Single degree of freedom oscillator

First, consider the familiar second order differential equation of a single degree of freedom oscillator:

This second order equation implies that there will be two state variables; they can be selected as

With this choice, Equ.(7.3) can be rewritten as a set of two first order equations

or

This equation explicitly shows the system and input matrices A and B. If one measures the displacement, the output matrix is C = (1 0) while if one uses a velocity sensor, and C = (0 1). An accelerometer can only be accounted for by using a feedthrough matrix D in addition to C. In fact, from Equ.(7.5), one gets

The output and feedthrough matrices are respectively

and

The choice (7.4) of the state variables is not unique and it may be convenient to make another choice :

The advantages are that both state variables express a velocity, and that the free response trajectories in the phase plane are slowly decaying spirals (Problem P.7.1). With this choice, Equ.(7.3) can be rewritten

140


This form is sometimes preferred to (7.6) because the system matrix is dimensionally homogeneous.

7.2.2

Flexible structure

Next, consider the multi degree of freedom vibrating system governed by the set of second order differential equations

where is the input force influence matrix, indicating the way the input forces act on the structure. Equation (7.10) is expressed in physical coordinates. If one defines the state vector as it can be rewritten in state space form as

The foregoing state variable form is acceptable if M is invertible. Calculating is straightforward for a lumped mass system. However, Equ.(7.11) may not be practical because the size of the state vector (which is twice the number of degrees of freedom of the system) may be too large. This is why it is customary to transform Equ.(7.10) into modal coordinates before defining the state vector. In this way, the state equation can be restricted to the few structural modes which describe the main part of the system dynamics in the frequency range of interest. Neglecting the high frequency dynamics of the system in the compensator design, however, must be done with extreme care, because the interaction between the neglected modes and the controller may lead to stability problems (spillover instability). This will be of prime concern in the compensator design. Equation (7.10) can be transformed into modal coordinates following the procedure described in chapter 2. If we perform the change of variables the governing equations in modal coordinates read

with the usual notations damping) and

(modal mass), (modal frequencies). In this equation,

(modal

where is the matrix of the mode shapes and is the input force influence matrix. The columns of contain the modal amplitudes at the actuators location, so that the vector represents the work of the input forces on the various mode shapes. Here, exactly as for the single degree of freedom oscillator, one can define the state variable as either

or


141

In the latter case, the state space equation reads

This form is similar to Equ.(7.9), except that diagonal matrices are substituted for scalar quantities. Note that, in most cases, the size of the state vector in Equ(7.14) (which is twice the number of modes included in the model) is considerably smaller than in Equ.(7.11). The output equations for a set of sensors distributed according to the influence matrix are as follows displacement sensors

velocity sensors accelerometers

where the columns of represent the modal amplitudes at the sensor locations. As for the single degree of freedom oscillator, there is a feedthrough component in the output equation for acceleration measurements. For collocated actuators and sensors, the input and output influence matrices are the same : and

7.2.3

Inverted pendulum

As another example of state space description, consider the inverted pendulum with a point mass at a distance from the support, as represented in Fig.7.2.a. The horizontal displacement of the support O is the input of the system. The differential equation governing the motion is

This equation also applies to more complicated situations where the pendulum consists of an arbitrary rigid body (Fig.7.2.b) if is taken as the equivalent length of the pendulum, defined as

where J is the moment of inertia about the center of mass and L is the distance between the center of mass C and the hinge O is the moment of

142


inertia about O). This can be readily established using the Lagrange or the Newton/Euler equations (Problem P.7.2). Equation (7.18) can be linearized about as When the displacement of the hinge is the input, it is convenient to make a change of variable and introduce the absolute position of the point mass, rather than the angle With this transformation, Equ.(7.20) becomes

where is the natural frequency of the pendulum. Note that, in general, the coordinate is different from that of the center of mass C in Fig.7.2.b. Using the state variables we find the state space equation

If the output of the system is the angle of the pendulum, the output and feedthrough matrices are

7.3

System transfer function

In this section, the input-output transfer function is derived from the state space equations. The relationship is formally the same for the scalar and multidimensional cases. The noise terms are deleted in Equ.(7.1) and (7.2), because they have nothing to do with the present discussion.


143

Laplace transforming the state equation

gives and, assuming zero initial conditions,

and, from the output equation (7.2), we have

or with In the scalar (SISO) case, is the transfer function of the system. For MIMO systems, it is the transfer matrix. It is rectangular, with as many rows as there are outputs and as many columns as there are inputs to the system. As an example, consider the single degree of freedom oscillator described by Equ.(7.6), in which the output is the displacement We readily obtain

This result could have been obtained more easily by Laplace transforming Equ.(7.3). Similarly, applying Equ.(7.28) to the inverted pendulum, using (7.23) and (7.24), we find

Once again, this result is straightforward from Equ.(7.20). Before discussing the poles and zeros, an important theorem in Matrix Analysis must be established: If a matrix is partitioned into blocks, the following identity applies

144


The first and third matrices involved in this identity are block triangular, and the determinant of a block triangular matrix is equal to the product of the determinants of the diagonal blocks ; it follows that:

Using this theorem with

instead of A, and – B instead of B, we get

Comparing with Equ.(7.28), we see that the transfer function of a SISO system can be rewritten

Now, we demonstrate that, as in the classical pole-zero representation of transfer functions, the denominator and the numerator of Equ.(7.32) explicitly show the poles and zeros of the system.

7.3.1

Poles and zeros

The poles are the values such that the free response of the system from some initial state has the form Substituting this in the free response equation, one gets

or This equation shows that the poles are the eigenvalues of the system equation A (and the initial conditions are the eigenvectors). The poles are solutions of which is the denominator of (7.32). The zeros of the system are the values of such that an input

applied from appropriate initial conditions The state vector has the form if the following condition is fulfilled:

produces a zero output,


145

that is, if Under this condition, the output is

Therefore, having

requires that

The two conditions (7.37) and (7.39), can be cast in compact form as

The values of for which this system of homogeneous equations has a non-trivial solution are the zeros of the system. They are solution of

which is the numerator of (7.32). From Equ.(7.40), we observe that, to achieve a zero output, the system needs to be excited at the frequency of a transmission zero, and also must start from the appropriate initial conditions other initial conditions would produce a transient output that would gradually disappear if the system is asymptotically stable.

7.4

Pole placement by state feedback

The idea in the state space approach is to synthesize a full state feedback

where the gain vector G (we deal with SISO systems first) is selected to achieve desirable properties of the closed-loop system. The problem is that, in most practical cases, the state vector is not known. Therefore, there must be an additional step of state reconstruction. One of the nice features of the state feedback approach is that the two steps can be performed independently; this property is called the separation principle. We shall address the design of the state feedback first (Fig.7.3), and leave the state reconstruction until later. A system of the order involves state variables. Therefore there are feedback gains that can be adjusted independently. Since there are poles in the system, either real or complex conjugate, they can be assigned arbitrarily by proper choice of the This is true, in principle, if the system is controllable, but it may not always be practical, because the control effort may be too large

146


for the actuators, or the large values of the feedback gains may unduly increase the bandwidth of the control system and lead to noise problems. Substituting the feedback law (7.42) into the system equation (7.1) yields

The closed-loop system matrix is A – BG. Its eigenvalues are the closed-loop poles; they determine the natural behaviour of the closed-loop system and are solutions of the characteristic equation

The state feedback design consists of selecting the gains so that the roots of (7.44) are at desirable locations. If the system is controllable, arbitrary pole locations (either real or complex conjugate pairs) can be enforced by choosing the in such a way that Equ.(7.44) is identical to

The gain values

7.4.1

are achieved by matching the coefficients of (7.44) and (7.45).

Example: oscillator

As an example, consider the single degree of freedom oscillator of Equ.(7.6)

The poles of the open-loop system are represented in Fig.7.4.a. We want to design a state feedback


147

such that the poles of the closed-loop system are moved to the locations indicated in Fig.7.4.b. From Equ.(7.43), the closed-loop system matrix, A – BG, reads

and the characteristic equation

The desired characteristic equation is

Comparing Equ.(7.48) and (7.49), one sees that the desired behaviour will be achieved if the gains are selected according to

These equations indicate that the controller will act as a spring (of stiffness , and damper (of damping constant . This is a general observation about the state feedback which consists of a generalization of the PD control.

148


Matching the coefficients of the characteristic equation to the desired ones rapidly becomes tedious as the size of the system increases. It becomes straightforward if the state equation is written in control canonical form (see section 10.4.1). We shall not expand on these aspects because the techniques for pole placement are automated in most control design softwares. The fact that the poles of the closed-loop system can be located arbitrarily in the complex plane is remarkable. Two questions immediately arise: Can it always be done ? Is it always practical to do it ? The answer to the first question is yes if the system is controllable (controllability is discussed in chapter 10). However, even if the system complies with the controllability condition, it may not be practical, because the control effort involved would be too large, or because the large values of the feedback gains would render the system oversensitive to noise or to modelling errors, when the control law is implemented on the reconstructed state from an observer. These robustness aspects are extremely important and actually dominate the controller design. One reasonable way of selecting the closed-loop poles is discussed in the next section.

7.5

Linear Quadratic Regulator

One effective way of designing a full state feedback in terms of a single scalar parameter is to use the Optimal Linear Quadratic Regulator (LQR). In this section, we analyse this problem for SISO systems; a more general formulation will be examined in chapter 9. We seek a state feedback (7.42) minimizing the performance index

where

is the controlled variable, defined by

The performance index has two contributions; the first one reflects the desire of bringing the controlled variable to zero (minimizing the error) while the second one that of keeping the control input as small as possible. is a scalar weighting factor used as a parameter in the design : large values of correspond to more emphasis being placed on the control cost than on the tracking error. Note that the control variable, , may or may not be the actual output of the system. If this is the case, H = C.


7.5.1

149

Symmetric root locus

The solution of the LQR is independent of the initial conditions; for a SISO system, it can be shown (e.g. Kailath, p.226) that the closed-loop poles are the stable roots (i.e. those in the left half plane) of the characteristic equation

where

is the open-loop transfer function between the input and the controlled variable. Equation (7.54) defines a root locus problem in terms of the scalar parameter the weighting factor of the control cost in the performance index. Note that and affect Equ.(7.54) in an identical manner. As a result, the root locus will be symmetric with respect to the imaginary axis, in addition to being symmetric with respect to the real axis. This is why it is called a symmetric root locus. Note that, since only the left part of the root locus must be considered, the LQR approach is guaranteed to be stable. The regulator design proceeds according to the following steps Select the row vector H which defines the controlled variable appearing in Equ.(7.53). Draw the symmetric root locus in terms of the parameter only the part of the locus in the left half plane.

and consider

Select a value of that provides the appropriate balance between the tracking error and the control effort. Once the closed-loop poles have been selected, the state feedback gains can be computed as for pole placement (a more efficient way of computing the control gains will be discussed in chapter 9). Check with simulations that the corresponding control effort is compatible with the actuator. If not, return to the previous step.

7.5.2 Inverted pendulum To illustrate the procedure, consider the inverted pendulum described by the state space equation (7.23). If we adopt the absolute displacement as the controlled variable, H = (1 0), and the transfer function (7.55) reads

150


The symmetric root locus is represented in Fig.7.5. The closed-loop poles consist of a complex conjugate pair on the left branch of the locus, Note that the transfer function used in the regulator design is different from the open-loop transfer function (7.30) between the input and the output variable (the latter is the tilt angle rather than the absolute position Using instead of in the performance index will not only keep the pendulum vertical, but will also prevent it from drifting away from its initial position. Once appropriate pole locations have been selected, the control gains G can be calculated as indicated in the previous section, by matching the closed-loop characteristic equation (7.44) to the desired one (7.45):

Equating the coefficients of the various powers in s provides

The open-loop system has a stable pole at and an unstable one at The cheapest optimal state feedback which stabilizes the system, obtained for on the root locus, simply relocates the unstable pole at (thus, the closed-loop system has two poles at The corresponding characteristic equation (7.59) is


151

and the optimal gains are

7.6

Observer design

The state feedback (7.42) assumes that the state vector is available at all times. This is not the case in general, because it would require too many sensors, and in many applications, some of the states would simply not be available, for physical reasons. The aim of the observer is to reconstruct the state vector from a model of the system and the output measurement In this way, the state feedback can be implemented on the reconstructed state, rather than on the actual state It must be emphasized that the state reconstruction requires a model of the system. We shall assume that an accurate model is available. To begin with, consider the case where the noise and feedthrough terms are removed from the system and output equations (7.1) and (7.2).

The following form is assumed for the full state observer (also called Luenberger observer).

The first part of Equ.(7.64) simulates the system and the second contribution (innovation) uses the information contained in the sensor signal; is the difference between the actual and the predicted output; the gain matrix K is chosen in such a way that the error between the true state and the reconstructed one, converges to zero. Combining Equ.(7.62) to (7.64), we find the error equation with the initial condition This equation shows that the error goes to zero if the eigenvalues of A – KC (the observer poles) have negative real parts (i.e. if A – KC is asymptotically stable). In a manner parallel to the state feedback design, if the system is observable, the observer poles can be assigned arbitrarily in the complex plane by choosing the values of the independent gains (observability will be discussed in chapter 10). The procedure for pole assignment is identical to that used for the regulator design : If a set of poles has been selected, the gains can be computed by matching the coefficients of the characteristic equation

152


to that of the desired one

(once again, the poles can be only real or complex conjugate pairs). Note that the nice form of the error equation is due to the fact that the same matrices A, B, C have been used in the system equations (7.62), (7.63) and the observer equation (7.64). This amounts to assuming perfect modelling. In practice, however, the actual system will be substituted for Equ.(7.62) and (7.63) while the observer equation (7.64) will be integrated numerically on a computer. As a result, there will always be an error associated with the modelling of the system. If a feedthrough term appears in the output equation, the innovation term in Equ.(7.64) must be taken as instead of this leads to the same error equation (7.65). If noise terms are included as in Equ.(7.1) and (7.2), the error equation becomes

The plant noise appears as an excitation; so does the measurement noise after being amplified by the observer gain K. If one wants the regulator poles to dominate the closed-loop response, the observer poles should be faster than (i.e. to the left of) the regulator poles. This will ensure that the estimation error decays faster than the desired dynamics, and the reconstructed state follows closely the actual one (at least without noise and modelling error). As a rule of thumb, the observer poles should be 2 to 6 times faster than the regulator poles. With noisy measurements, one may wish to decrease the bandwidth of the observer by having the observer poles closer to those of the regulator. This produces some filtering of the measurement noise. In this case, however, the observer poles have a significant influence on the closed-loop response. Notice, from Equ.(7.65), that if the open-loop system is stable (and the model accurate), the error will converge to zero even if the gain K is zero. The observer equation (7.64) is based partly on the system model, and partly on the measurement error, the latter being amplified by the observer gain K. The relative importance which is given to the model and the measurement contributions should depend on their respective quality. If the sensor noise is low, more weight can be placed on the measurement error (large gains , while noisy measurements do require lower gains. The minimum variance observer is that minimizing the variance of the measurement error. If one assumes that the plant noise and measurement noise are white noise processes (i.e. with uniform power distribution over the whole frequency range), the minimum variance observer is known as the Kalman Filter.


7.7

153

Kalman Filter

The Kalman Bucy Filter (KBF) will be studied in detail in chapter 9. Here, exactly as we used the LQR as a sensible pole placement strategy for the regulator design of SISO systems (in terms of a single scalar parameter), we consider the particular case where the plant noise and the measurement noise are scalar white noise processes. In this case, it is, once again, possible to draw a symmetric root locus plot in terms of a single parameter expressing the relative intensity of the plant and measurement noise. Let

be the system and output equations, where w and v are scalar white noise processes. It can be demonstrated that the optimal pole locations corresponding to the KBF are the stable roots of the symmetric root locus defined by the equation where

is the transfer function between the plant noise

and the output

and the parameter is the ratio between the intensity of the plant noise and that of the measurement noise (noisy measurements correspond to small ). The observer design proceeds exactly as for the regulator: Selecting a noise input matrix E, we draw the symmetric root locus as a function of When proper pole locations have been selected on the left half part of the locus, the gains can be computed by matching the coefficients of the characteristic equation (7.66) to that of the desired one (7.67). If none of the values of provides a desirable set of poles, another E matrix can be selected and the procedure repeated. In the LQR approach, the regulator design involves the symmetric root locus of the open-loop transfer function between the input and the controlled variable while in the KBF approach, the observer design is based on the transfer function between the system noise and the output .

If one assumes that the plant noise enters the system at the input (E = B) and the controlled variable is the output variable (H = C), these transfer functions become identical to the open-loop transfer function

154


In this case, the regulator and the observer can both be designed from a single symmetric root locus.

7.7.1

Inverted pendulum

To illustrate the procedure, consider again the inverted pendulum and assume that the noise enters the system at the input (E = B). From Equ.(7.23) and (7.24), the transfer function between the noise an the output Equ.(7.74), is

Thus, although the output variable is different from the controlled variable is proportional to used in the regulator design [Equ.(7.56)]; this is because there is no feedthrough component in as there is in [Equ.(7.30)]. As a result, the root locus used in the regulator design can also be used for the observer design (Fig.7.6). How far the observer poles should be located on the locus really depends on how noisy the measurements are. Note that when the regulator poles are near the asymptotes, the KBF obtained here is very close to the design consisting of assigning the observer poles by scaling the regulator poles to where according to the rule of thumb mentioned before (Problem P.7.5). Finally, assume that the noise enters the system according to From Equ.(7.23) and (7.24), the transfer function between the noise and the


155

output is readily obtained as

The corresponding symmetric root locus is shown in Fig.7.7, assuming One notices that one of the poles goes to the zero at as increases. This is the optimum for the assumed distribution of the plant noise, but the observer obtained in this way may be sluggish, which may not be advisable for reasons mentioned before.

7.8

Reduced order observer

With the full state observer, the complete state vector is reconstructed from the output variable . When the number of state variables is small and the output consists of one of the states, it may be attractive to restrict the state reconstruction to the missing state variables, so that the measured state variable is not affected by the modelling error. This is nice, but on the other hand, by building the state feedback partly on the output measurement without prior filtering, there is a feedthrough component from the sensor noise to the control input, which increases the sensitivity to noise. In this section, we restrict ourselves to the frequent case of reconstructing the velocity from a displacement measurement for a second order mechanical system. This is probably the most frequent situation where a reduced order observer is used. In this case, the reduced observer is of the first order while the full state observer is of the second order. Extensions to more general situations can be found in the literature (e.g. Luenberger).

156

7.8.1


Oscillator

Consider the single degree of freedom oscillator governed by Equ.(7.6). The output measurement is the displacement The velocity equation is

We seek a first order observer governed by the following equations

where is the estimated velocity and and are free parameters; they are selected in such a way that the error equation governing the behaviour of is asymptotically stable, with an appropriate decay rate

Equations (7.78)-(7.80) give the error equation

It can be made identical to Equ.(7.81) if the coefficients are selected according to

From Equ.(7.81),

7.8.2

is the eigenvalue of the observer; it can be chosen arbitrarily.

Inverted pendulum

The same procedure can be applied to the inverted pendulum described by Equ.(7.23) and (7.24). The displacement is obtained from the output measurement and the input as

The velocity equation is

The velocity observer has again the form


and the error equation

157

is readily written as

It becomes identical to (7.81) if the coefficients satisfy

The reduced order observer is therefore

where the only remaining parameter is the eigenvalue of the observer, which can be selected arbitrarily. Note that appears also as the feedthrough component of the measured variable in Equ.(7.93), leading to a direct effect of the measurement noise on the reconstructed velocity (amplified by Because of the existence of a feedthrough component from the measurement to the reconstructed states, and therefore to the control, the bandwidth of a compensator based on a the reduced observer is much wider than that of a compensator based on a full state observer. The simplicity of the observer structure must be weighed against the higher sensitivity to sensor noise. If the latter is significant, the reduced order observer becomes less attractive than the full state observer.

7.9

Separation principle

Figure 7.8 shows the complete picture of the state feedback regulator implemented on the reconstructed states, obtained from a full state observer (the case without feedthrough component is represented for simplicity). The closedloop equations are :

The complete system has state variables. If one uses the reconstruction error, as state variable instead of the reconstructed state vector one can write the closed-loop system equation as

158


It is block triangular and, as a result, the eigenvalues of the closed-loop system are those of the diagonal blocks A – BG and A – KC. Thus, the poles of the closed-loop system consist of the poles of the regulator and those of the observer. This means that the eigenvalues of the regulator and the observer are not changed when the two subsystems are put together. Therefore, the design of the regulator and of the observer can be carried out independently; this is known as the separation principle.

7.10

Transfer function of the compensator

The transfer function of the compensator can be obtained from Equ.(7.94). It reads

The poles of the compensator are solutions of the characteristic equation

Note that they have not been specified anywhere in the design, and that they are not guaranteed to be stable, even though the closed-loop system is. Working with an unstable compensator (which fortunately, is stabilized by the plant!) may bring practical difficulties as, for example, that the open-loop frequency response of the compensator cannot be checked experimentally. Figure 7.9 represents the compensator and the system in the standard unity feedback form used in classical methods (root locus, Bode, Nyquist). There is a major difference between the state feedback design and the classical methods: in classical methods, the structure of the compensator is selected to achieve desired closed-loop properties; in the state feedback design, the structure of the compensator is never directly addressed because the attention is focused on


159

the closed-loop properties; the compensator is always of the same order as the system.

7.10.1

The two-mass problem

Let us illustrate this point on the two-mass problem of Fig.4.3. Defining the state vector we can write Equ.(4.3) and (4.4) as

For a non-collocated displacement sensor,

The open-loop transfer function is given by Equ.(4.7). If the open-loop system has a pair of poles at open-loop transfer function is

where the compact notation

M = 1 and (flexible mode) the

is used for

Using the output as the controlled variable and assuming that the system noise enters at the input, we can design the regulator and the observer using the same symmetric root locus based on It is shown in Fig.7.10. Once the regulator and observer poles have been selected on the left side of the locus, the corresponding gains can be calculated by matching the coefficients of the characteristic equations (7.44) and (7.66) to the desired ones; for the pole locations indicated in the figure, we get

160


The resulting compensator is

It is of the fourth order, like the system. The numerator consists of a PD plus a pair of zeros near the flexible poles of to produce a notch filter. With this nominal compensator, it is possible to draw a conventional root locus plot for (Fig.7.11), which describes the evolution of the poles of the closed-loop system when the scalar gain varies from 0 to The pole locations on this root locus for are, of course, the same as those selected on the symmetric root locus (Fig.7.10). To assess the robustness of the control system, Fig.7.12 shows the root locus plot when the natural frequency is changed from to One observes that the notch filter does not operate properly any more (the pair of zeros of the compensator no longer attracts the flexible poles of the structure) and the closed-loop system soon becomes unstable. The situation is even worse if the natural frequency is reduced to In that case, a pole-zero flipping occurs between the pole of the flexible mode and the zero of the notch filter (Fig.7.13). As we already stressed in chapter 4, the pole-zero flipping changes the departure angles from the poles and the arrival angles at the zeros by 180°, transforming the stabilizing loop of


161

162


Fig.7.11 into a destabilizing one as in Fig.7.13. This lack of robustness is typical of state feedback and notch filters.

7.11

References

A. E. BRYSON JR., Some Connections Between Modern and Classical Control Concepts, ASME, Journal of Dynamic Systems, Measurement, and Control, vol. 101, 91–98, June 1979. R. H. CANNON & D. E. ROSENTHAL, Experiment in Control of Flexible Structures with Noncolocated Sensors and Actuators, AIAA Journal of Guidance, vol. 7, no. 5, 546–553, Sep.–Oct. 1984. G. F. FRANKLIN, J. D. POWELL & A. EMANI-NAEMI, Feedback Control of Dynamic Systems, Addison-Wesley, 1986. T. KAILATH, Linear Systems, Prentice-Hall, 1980. H. KWAKERNAAK & R. SIVAN, Linear Optimal Control Systems, Wiley, 1972. D. LUENBERGER, An Introduction to Observers, IEEE Trans. Autom. Control, AC-16, pp. 596–603, Dec. 1971. D. LUENBERGER, Introduction to Dynamic Systems, Wiley, 1979.


7.12

163

Problems

P.7.1 Consider the oscillator described by Equ.(7.9). For non-zero initial conditions, sketch the free response in the phase plane Show that the image point rotates clockwise along a spiral trajectory. Relate the decay rate of the spiral to the damping ratio. P.7.2 Show that the inverted pendulum of Fig.7.2 is governed by Equ.(7.18) to (7.20), where is the displacement of the support point. P.7.3 Consider an inverted pendulum similar to that of Fig.7.2, but mounted on a cart of mass M and controlled by an horizontal force applied to the cart. If stands for the horizontal displacement of the cart, show that, for small the governing equations can be approximated by

Write the equations in state variable form. P.7.4 For the inverted pendulum of Problem P.7.3, assuming that the full state is available, find a feedback control that balances the stick and keeps the cart stationary near P.7.5 Consider the inverted pendulum of Fig.7.2. Using the same procedure as for the regulator in section 7.5.2, find the analytical expression of the observer gains which locates the observer poles at P.7.6 Consider an inverted pendulum with (Fig.7.2). (a) Compute the transfer function of the compensator such that the regulator and the observer poles are respectively at and (b) Draw a root locus for this compensator. Observe that the closed-loop system is conditionally stable. Find the critical gains corresponding to the limits of stability. Sketch the Nyquist diagram for the nominal system.


Chapter 8

Analysis and synthesis in the frequency domain 8.1

Gain and phase margins

Consider the root locus plot of Fig.4.5. Any point in the locus is solution of the characteristic equation Therefore, we have

The locus crosses the imaginary axis at the point of neutral stability. Since the following relations hold at the point of neutral stability:

Returning to the Bode plot of Fig.4.6, one sees that changing amounts to moving the amplitude plot along the vertical axis; the point of neutral stability is obtained for when the amplitude curve is tangent to the 0 dB line, near the frequency above which the phase exceeds –180°. The system is stable if the algebraic value of the phase of the openloop system is larger than for all frequencies where the amplitude is larger than 1. The system is unstable if this condition is violated. A measure of the degree of stability is provided by the gain and phase margins: The gain margin (GM) indicates the factor by which the gain must be increased to reach the neutral stability. The phase margin (PM) is the amount by which the phase of the open loop transfer function exceeds when The corresponding frequency is called the crossover frequency. 165

166


The situation described above corresponds to a stable open-loop system the gain and phase margins are decreasing functions of the system becomes unstable when the gain exceeds some critical value. There are more complex situations where the system is conditionally stable, when the gain belongs to some interval it becomes unstable when the gain decreases below the threshold (see Problem P.7.6). These cases can be handled with the Nyquist criterion.

8.2

Nyquist criterion

8.2.1

Cauchy’s principle

Consider the feedback system of Fig.8.1, where stands for the combined open-loop transfer function of the system and the compensator (shorthand for in the previous section). The conformal mapping transforms the contour in the plane into a contour in the plane. One of the properties of the conformal mapping is that two intersecting curves with an angle in the plane map into two intersecting curves with the same angle in the plane. Assume is written in the form

Analysis and synthesis in the frequency domain

For any

on

167

the phase angle of G(s) is given by

where and are the phase angles of the vectors connecting respectively the zeros and the poles to From Fig.8.1, it can be seen that, if there are neither poles nor zeros within the contour does not change by when goes clockwise around The contour will encircle the origin only if the contour contains one or more singularities of the function Because the contour is travelled clockwise, one pole of within produces a phase change of that is one counterclockwise encirclement of the origin by Conversely, a zero produces a clockwise encirclement. Thus, the total number of clockwise encirclements of the origin by is equal to the number of zeros in excess of poles of within the contour This is Cauchy 's principle.

8.2.2

Nyquist stability criterion

The foregoing idea provides a simple way to evaluate the number of singularities of the closed-loop system in the right half plane, from the poles and zeros of the open-loop transfer function The contour is selected in such a way that it encircles the whole right half plane as indicated in Fig.8.2. If there are poles on the imaginary axis, indentations are made as indicated on the figure, to leave them outside The closed-loop transfer function is

and the closed-loop poles are solutions of the characteristic equation

168


If

This equation shows that the poles of are the same as those of the open-loop transfer function Let P be the [known] number of unstable poles of the open-loop system. On the other hand, the zeros of are the poles of the closed-loop system; we want to evaluate their number within From Cauchy’s principle, if we consider the mapping the number of clockwise encirclements of the origin by when goes clockwise along

is Now, instead of considering the encirclements of the origin by completely equivalent to consider the encirclements of –1 by number of unstable poles of the closed-loop system is given by

it is Thus, the

where N is the number of clockwise encirclements of –1 by when s follows the contour and P is the number of unstable poles of the open-loop system. All physical systems without feedthrough are such that as (with feedthrough, goes to a constant value). As a result, only the part of the plot corresponding to the imaginary axis must be


169

considered. The polar plot for positive frequencies can be drawn from the Bode plots; that for negative frequencies is the mirror image with respect to the real axis, because If there are poles of on the imaginary axis and if the indentations are made as indicated in Fig.8.2, the poles are outside the contour. According to Equ.(8.2) and (8.3), each pole contributes with an arc at infinity and a rotation of that is 180º clockwise. Returning to the system of Fig.4.5 and 4.6, its Nyquist plot is shown in Fig.8.3: The contribution of the positive frequencies is plotted in full line, and the circle at infinity corresponds to the indentation of the double pole at the origin the contribution of the negative frequencies (not shown) is the mirror image of that of the positive frequencies with respect to the real axis. For small gains (as shown in the figure), the number of encirclements of –1 is zero and the system is stable; for larger gains, there are two encirclements and therefore two unstable poles. This is readily confirmed by the examination of Fig.4.5. Since the instability occurs when encircles –1, the distance from to –1 is a measure of the degree of stability of the system. The relative stability is measured by the gain and phase margins (Fig.8.4). As we shall see, there is a direct relationship between the phase margin and the sensitivity to parameter variations and the disturbance rejection near crossover. Since the closed-loop transfer function is uniquely determined by [Equ.(8.4)], loci of constant magnitude and of constant phase can be drawn in the complex plane they happen to be circles. The M circles are shown in Fig.8.5; the larger magnitudes correspond to smaller circles near – 1. Thus, the maximum amplification corresponds to the smallest circle, tangent to the Nyquist plot of it is reached for a

170


frequency close to the crossover frequency (where the Nyquist plot crosses the unit circle). This is why there is a direct relationship between the maximum amplification and the phase margin PM. In most cases, prevents overshoot of F (Fig.8.6). The of the control system is defined as the frequency corresponding to an attenuation of -3 dB in the closed-loop transfer function. Since the bandwidth can be taken equal to the crossover frequency for design purposes.

8.3

Nichols chart

The Nyquist plot is a convenient tool for evaluating the number of encirclements and the absolute stability of the system. However, because of the linear scale for the magnitude, the Nyquist plot is not always practical in the vicinity of -1. The Nichols chart plot is often more useful for evaluating the relative stability when the open-loop system is stable (Fig.8.7). It consists of a plot of the vs. of the open-loop transfer function There is a one to one relationship between the Nichols chart and the Nyquist plot, but the former brings a considerable amplification to the vicinity of -1, for easy evaluation of the gain and phase margins, and the logarithmic scale allows a much wider range of magnitude in the graph. Unlike the Nyquist plot, the Nichols chart plot can be obtained from the summation of the individual magnitude and phase angle contributions of the poles and zeros, and a change of gain moves


171

the curve along the vertical axis. These two advantages are shared by the Bode plots, but the Nichols chart plot combines the gain and phase information into a single diagram. As for the Nyquist plot, in order to assist in the design, it is customary to draw the loci of constant amplitude M and constant phase N of the closed-loop system (Fig.8.7); they are no longer circles.

8.4

Feedback specification for SISO systems

Consider the SISO feedback system of Fig.8.8, where is the reference input, the output, the disturbance and the sensor noise (unlike the disturbance, the sensor noise does not directly affect the output).

8.4.1

Sensitivity

In this section, we evaluate the sensitivity of the closed-loop transfer function to the variations of the open-loop transfer function; we assume that The closed-loop transfer function is

Its sensitivity to parameter changes is related to that of the open-loop transfer function G by

172


or or

is called the sensitivity function and 1 + G the return difference. Equation (8.10) states that, in the frequency range where the sensitivity of the closed-loop system to parameter changes is much smaller than that of the open-loop system. This is one of the objectives of feedback. On the other hand, if the phase margin is small, goes very near –1 at crossover, and F becomes much more sensitive than G.

8.4.2

Tracking error

Referring to Fig.8.8, the governing equations are


173

or

The tracking error is

From Equ.(8.11) and (8.12), we see that good tracking implies that The sensitivity function range where the command

must be small in the frequency and the disturbance are large.

The closed-loop transfer function frequency range where the sensor noise is large.

must be small in the

Note, however, that Therefore, S and F cannot be small simultaneously; this means that disturbance rejection and noise rejection cannot be achieved simultaneously.

8.4.3

Performance specification

In the previous section, G referred to the open-loop transfer function, including the system and the compensator. In this section, G will refer explicitly to the system and H to the compensator; the open-loop transfer function is GH (Fig.8.9). The general objective of feedback is to achieve sensitivity reduction, good tracking and disturbance rejection at low frequency with a sensor signal which has been contaminated with noise at high frequency From the foregoing section, this is translated into the following design constraints:

174


where is a large positive function (performance specification), defined in the frequency range where good tracking and disturbance rejection is necessary [zero steady-state error requires that and

where defines the frequency above which the sensor noise becomes significant. Of course, in addition to that, the closed-loop system must be stable. Thus minimum values for the gain and phase margins must also be specified:

Equivalently, we may require that the GH curve remains outside a particular M circle from the point We shall see later that stability places constraints on the slope of near crossover (in the frequency range near

8.4.4

Unstructured uncertainty

When the uncertainty on the physical system is known to affect some specific physical parameters like, for example, the natural frequencies and the damping, it is called structured. If little is known about the underlying physical mechanism, it is called unstructured. The unstructured uncertainty will be characterized by an upper bound of the norm of the difference between the transfer function of the actual system, and its model It can be considered either as additive:

where is a positive function defining the upper bound to the additive uncertainty (including the parameter changes and the neglected dynamics), or


175

multiplicative:

where is a positive function defining an upper bound to the multiplicative uncertainty. is usually small at low frequency where the model is accurate, and becomes large at high frequency, due to the neglected dynamics of the system (Fig.8.10). Because defines only the magnitude of the uncertainty, it can be associated with any phase distribution; therefore, it clearly defines a worst case situation. Evaluating or is not a simple task. For large flexible structures, more modes are often available than it would be practical to include in a design model; these extra modes can be used to evaluate the associated uncertainty (Problem P.8.4).

8.4.5

Robust performance and robust stability

The design objectives in the face of uncertainties are that the conditions enumerated in section 8.4.3 must be fulfilled for the perturbed system. Referring to Fig.8.11, the stability of the system will be guaranteed in presence of uncertainty if the distance to instability, is always larger than the uncertainty The stability robustness condition is therefore

This guarantees that the number of encirclements will not be altered by any additive uncertainty bounded by for any phase distribution. Similarly, since condition (8.19) can be written in terms of the multiplicative uncertainty as

or

176


Normally, this condition will never be violated at low frequency, where the uncertainty is small and the loop gain large. At high frequency,

and Equ.(8.21) is reduced to

This condition expresses the gain stability. It is not conservative if the phase is totally unknown, as in the case of pole-zero flipping. Multiplication of (8.20) by gives an alternate form of the stability condition:

Now, if we want to achieve the performances with the perturbed system, Equ(8.14) must be replaced by

This equation will be satisfied if

A sufficient condition is

The design tradeoff for is explained in Fig.8.12. The shaded region at low frequency is excluded for robust performance, and that at high frequency for robust stability.


177

Considering the vicinity of the crossover frequency (Fig.8.13), we can make the following observations: At crossover, becomes

and the stability robustness condition (8.20)

This means that accepting a magnitude error at crossover requires a phase margin of 60º. From this, one can anticipate that the bandwidth of the closed-loop system cannot be much larger than the frequency where At crossover, the return difference is related to the phase margin by

This establishes a direct connection between the phase margin and the sensitivity and the disturbance rejection near crossover. We shall now address the conflict between the quality of the loop near crossover (good phase margin) and the attenuation rate of GH.


178

Bode gain-phase relationships

8.5

Figure 8.12 suggests that good feedback design could be achieved by having a large gain at low frequency and a fast enough decay rate at high frequency. Unfortunately, this is not as simple as that, because the closed-loop stability of the nominal system requires that the phase remains larger than as long as the gain is larger than 1. It turns out that for stable, minimum phase systems (i.e. with neither poles nor zeros in the right half plane), the phase angle and the amplitude are uniquely related. This relationship is expressed by the Bode Integrals. In this section, we shall state the main results without proof; the interested reader can refer to the original work of Bode, or to Horowitz or Lurie.

Integral # 1 Consider the unity feedback with the stable, minimum phase open-loop transfer function If the amplitude diagram has a constant slope corresponding to poles in a log-log diagram the phase is In the general case, the phase at a frequency is given as a weighted average of the gain slope at all frequencies, but with a stronger weight in the vicinity of

where

and the weighting function

is defined by


is strongly peaked near far from that of a Dirac impulse,

179

(Problem P.8.5); its behaviour is not too so that

This relation is very approximate, unless the slope of is nearly constant in the vicinity of it applies almost exactly if the slope is constant over two decades. In this case, one readily sees that

The first integral indicates that a large phase can only be achieved if the gain attenuates slowly. It follows that the rolloff rate in the region near crossover must not exceed and it must often be smaller than this, in order to keep some phase margin.

Integral # 2 Assume one pole

is stable and has an asymptotic rolloff corresponding to more than at infinity). According to section 8.4.2, the sensitivity function represents the fraction of the command or of the disturbance which is transmitted into the tracking error. must be small in the frequency range where or are large. However cannot be small everywhere, because the second Bode integral states that

This relation states that there is as much sensitivity increase as there is decrease sensitivity can only be traded from one frequency band to another, and good disturbance rejection in some frequency range can be achieved only at the expense of making things worse than without feedback outside that frequency range (see also Problem P.8.9). If the open-loop system has unstable poles the second integral becomes

where the sum extends to the unstable poles. This shows that for a system unstable in open-loop, the situation is worse, because there is more sensitivity increase than decrease; fast unstable poles are more harmful than slow ones because they contribute more to the right hand side of Equ.(8.30).

180


Integral # 3 The third integral states that some reshaping of can be performed within the working band (normalized to without affecting the phase outside the working band. This arises from:

If the open-loop transfer function is altered in the working band, in such a way that the integral on the left hand side is unchanged is the same for both transfer functions since only the working band is altered), the weighted phase average is preserved outside the working band. This situation is illustrated in Fig.8.14.

Integral # 4 This integral says that the greater the phase lag, the larger will be the feedback in the working band:

In particular, if two loop shapes and have the same high frequency behaviour, but has a greater phase lag than has a larger magnitude


181

in the working band than

The third and fourth integrals tell us that, in order to achieve a large gain within the working band, the phase lag must be as large as possible outside the working band. The stability limitations on the phase lag are reflected in feedback limitations in the working band. The next section illustrates how these phase-magnitude relationships can be translated into design.

8.6

The Bode Ideal Cutoff

The first Bode integral tells us how the phase at one frequency is affected by the gain slope in the vicinity (about one decade up and down). It also says that local phase increase can be achieved by lowering the gain slope (this situation is illustrated at high frequency in Fig.8.14). Figure 8.15 shows a Nichols chart of the desired behaviour: One wants to keep the feedback constant and as large as possible within the working band in reduced frequency), then reduce it while keeping gain and phase margins of dB and (shaded rectangle in Fig.8.15). It is therefore logical to keep the phase constant and compatible with the phase margin:

The compensator design is thus reduced to that of enforcing an open-loop transfer function (system plus compensator) to map the two segments of

182


Fig.8.15. The result is shown in Fig.8.16: The open-loop transfer function with a constant gain in the working band and a constant phase lag outside the working band is

This transfer function has an ideal behaviour at low frequency. However, the rolloff at high frequency is directly related to the phase margin. Improved high frequency noise attenuation requires higher rolloff. Because of the first integral, we know that if we simply add one segment with a higher slope to the amplitude diagram of Fig.8.16, it will be reflected by an additional phase lag near crossover, which is incompatible with the stability margin. The cure to that is to have first a flat segment dB below 0 dB, for approximately one or two octaves, followed by a segment with a higher slope, taking care of the sensor noise attenuation. The flat segment at dB provides the extra phase lead near crossover which compensates for the extra phase lag associated with the higher rolloff at high frequency. The final design is sketched in Fig.8.17. The reader can compare it to the design tradeoff of Fig.8.12. Once the loop transfer function GH has been obtained, that of the compensator can be deduced and approximated with a finite number of poles and zeros.


8.7

183

Non-minimum phase systems

In the previous sections, we assumed that had no singularity in the right half plane; such systems are called minimum phase (this section will explain why). The effect of right half plane poles was briefly examined in Equ.(8.30). In this section, we consider the effect of right half plane zeros. To begin with, assume that has a single right half plane zero at We can write

where

is the all-pass function

and is the minimum phase transfer function obtained by reflecting the right half plane zero into the left half plane no longer has a singularity in the right half plane]. The Nyquist diagram of is represented in Fig.8.18; follows the unit circle clockwise, from at to at takes its name from the fact that for all and have the same magnitude for all and the same phase at but for their phase differ by that of at and at We have seen in section 4.8 that flexible structures with non-collocated actuators and sensors do have non-minimum phase zeros. If they lie well outside

184


the bandwidth of the system they do no harm, because the corresponding all-pass function brings only very little phase lag near crossover. On the contrary, if the non-minimum phase zero lies at a distance comparable to the bandwidth, and this means that in the design, the phase angle of cannot exceed (without any phase margin) and, consequently, its falling rate cannot exceed instead of for a minimum phase system. The situation can easily be generalized to an arbitrary number of right half plane zeros. If has right half plane zeros at it can be written in the form (8.36) where is minimum phase and is the all-pass function

A typical pole-zero pattern of is shown in Fig.8.19. It is easy to see that for all and that the phase angle decreases by when goes from 0 to or when goes from to This extra clockwise rotation about the origin, associated with the all-pass function, can change the number of encirclements of –1, especially if some zeros are close to the bandwidth; it is therefore decisive in stability considerations. The phase lag associated with the non-minimum phase transfer function is always greater than that of by that of This is why is called minimum phase, because no other stable function with the same amplitude characteristics can have any less phase lag. The design of a compensator for a non-minimum phase system can be done by considering the minimum phase system and modifying the design specifications to account for the extra phase lag generated by the all-pass function. As we have seen, this puts stronger restrictions on the attenuation rate near crossover to avoid instability (Problem P.8.7).


185

A pure time delay T has a transfer function

Just as the all-pass function, it has a constant magnitude and a phase lag which increases linearly with the frequency. A pure delay can always be approximated as closely as needed by an all-pass network approximants), so that the above conclusions also apply (Problem P.8.8).

8.8

Usual compensators

Amongst the desired features of the closed-loop system are the stability properties, expressed by the gain and phase margins, the transient response characteristics, the bandwidth, the high frequency attenuation for sensor noise rejection and the abilty to maintain the output at a desired value with minimum static error. We have seen in the previous sections that these requirements are interdependent and often conflicting. The Ideal Bode Cutoff defines a reasonable compromise. The ease with which it can be implemented, however, depends very much on the system properties. This section briefly reviews the most popular Lead, Lag, PI, PID compensators. Before that, we define the system type which controls the steady state error and tracking ability.

8.8.1

System type

The system type measures the ability of the control system to track polynomials. Consider a unity feedback (Fig.8.8) with the open-loop transfer function

186


The transfer function between the reference input

and the error

According to the final value theorem, the steady state error,

is

is given by

The number of integrators in the open-loop transfer function is called the system type. If the system is type 0: a constant actuating signal maintains a constant output

If

is a step function,

and

The output follows a step input with a steady state error proportional to (1 + If there is one integrator in and the system is type 1: a constant actuating signal maintains a constant rate of change of the output. A step input is tracked with zero steady state error and a ramp input, is tracked with a constant error

Similarly, a system with two integrators in the open-loop transfer function is type a constant actuating signal maintains a constant second derivative of the output (structures with rigid body modes belong to this category); step and ramp inputs are tracked without steady state error and a parabolic input, produces a constant error

is called static error coefficient of the system of type

From Equ.(8.40),

If a polynomial input is applied to a type system, there is a zero steady state error if If there is a constant steady state error defined by the static error coefficient if the reference signal cannot be tracked. From the foregoing discussion, we note that long term errors associated with persistant disturbances can be zero only if the poles of the disturbance (one pole at for a step, one double pole at for a ramp, etc...) are included among the poles of the open-loop transfer function this is called the internal model principle.


8.8.2

187

Lead compensator

The aim of the cascade compensator (Fig.8.20) is to alter the open-loop transfer function from to in order to improve the characteristics of the closed-loop system. The simplest one is the Proportional plus Derivative (PD):

This compensator increases the phase above the breakpoint 1/T. Therefore, selecting 1/T slightly below the crossover frequency of produces an increase in the phase margin. The major drawback of the PD compensator is that the compensation increases with frequency and the high frequency attenuation rate is reduced, which is undesirable if the sensor noise is significant. The Lead compensator eliminates this drawback by adding a pole at higher frequency

The phase lead is still significant but the high frequency amplification is considerably reduced (Fig.8.21). The maximum phase lead occurs at

Normally, in order to maximize the benefit of the phase lead, and T are chosen in such a way that the crossover frequency lies between the zero and the pole of the compensator.

8.8.3

PI compensator

The aim of the Proportional plus Integral (PI) compensator is to increase the gain at low frequency to eliminate the static error. The compensator is

The system type is increased by 1. The amplification at low frequency is obtained at the expense of a phase lag of –90º below the breakpoint In

188



189

order to minimize the phase lag at crossover, the breakpoint must be substantially below the crossover frequency

8.8.4

Lag compensator

The analytical form of the Lag compensator is the same as that of the Lead compensator, except that the pole and zero are placed in reverse order

Normally, the pole and zero are placed at low frequency, to minimize the phase lag at crossover. The lag compensator can be used

1. to reduce the gain at high frequency without affecting that at low frequency, in order to reduce the crossover frequency and increase the phase margin;

2. to increase the gain at low frequency without affecting the high frequency behaviour, in order to improve the steady state error characteristics.

8.8.5

PID compensator

The Proportional plus Integral plus Derivative (PID) compensator consists of cascaded PI and PD compensators (Fig.8.21)

The integral part aims at reducing the static error, while the PD part increases the phase near crossover. Reasonable breakpoint frequencies are

If the increasing behaviour at high frequency is undesirable, the PD part can be replaced by a Lead compensator which achieves some phase lead without changing the asymptotic decay rate of

8.9

References

H. W. BODE, Relations Between Attenuation and Phase in Feedback Amplifier Design, Bell System Technical Journal, July 1940. H. W. BODE, Network Analysis and Feedback Amplifier Design, Van Nostrand, N-Y, 1945.

190


J. J. D’Azzo & C. H. H OUPIS, Feedback Control System Analysis McGraw-Hill, second ed., 1966.

Synthesis,

J. J. DISTEFANO, A. R. STUBBERUD & I. J. WILLIAMS, Feedback and Control Systems, Shaum’s Outline Series, McGraw-Hill, 1967. J. C. DOYLE & G. STEIN, Multivariable Feedback Design: Concepts for a Classical/Modern Synthesis, IEEE Transaction on Automatic Control, vol. AC-26, no. 1, 4–16, Feb. 1981. G. F. FRANKLIN, J. D. POWELL & A. EMANI-NAEMI, Feedback Control of Dynamic Systems, Addison-Wesley, 1986. I. M. HOROWITZ, Synthesis of Feedback Systems, Academic Press, 1963. G. J. KISSEL, The Bode Integrals and Wave Front Tilt Control, AIAA Guidance, Navigation & Control Conference, Portland Oregon, 1990. B. J. LURIE & P. J. ENRIGHT, Classical Feedback Control (with MATLAB), Marcel Dekker Inc., 2000. J. M. MACIEJOWSKI, Multivariable Feedback Design, Addison-Wesley, 1989. R. E. SKELTON, Model Error Concepts in Control Design, Int. J. of Control, vol. 49, no. 5, 1725–1753, 1989.

8.10

Problems

P.8.1 Consider an inverted pendulum with (Fig. 7.2). (a) Compute the transfer function of the state feedback compensator such that the regulator and the observer poles are respectively at and (b) Draw a root locus and the Nyquist diagram for this compensator. Using the Nyquist criterion, show that the closed-loop system is conditionally stable. Find the critical gains corresponding to the limits of stability. P.8.2 Draw the Nichols chart plot for the two-mass problem of section 4.4. P.8.3 Consider a direct velocity feedback applied to a structure with three flexible modes at and the modal damping is uniform, Assume that the actuator and sensor are noncollocated and that the uncertainty on the mode shapes is responsible for some uncertainty on the first pair of zeros which vary from to The second pair of zeros are fixed to Plot the open-loop transfer function on a Nichols chart, for various locations of the uncertain zeros. Comment on the effect of the zero on the stability margins.


191

P.8.4 Consider a simply supported beam excited by a point force at and provided with a displacement sensor at The nominal system is such that and the modal damping is uniform Compute an estimate of the transfer function Next, consider the following modelling errors: (a) Modal truncation: the modal expansion (2.47) is truncated after 4, 6 and 10 modes. (b) Quasi-static correction: same as (a), but start from the result of problem P.2.5. (c)Parametric uncertainty: the model is based on and (d)Errors in the mode shapes are simulated by using in the model. For each situation, compute the transfer function of the model, the additive and the multiplicative uncertainty, respectively and (gain and phase). Comment on the adequacy of the unstructured uncertainty to represent the various situations. P.8.5 Plot the weighting function appearing in the first Bode integral (8.28). P.8.6 Consider a SISO system with the following performance specifications:

Using the first Bode integral, discuss the feasibility of the design. P.8.7 Consider a non-minimum phase system with a real zero at Following the Bode Ideal Cutoff, draw the magnitude and phase diagrams for the corresponding minimum phase system which produce a gain margin of dB and a phase margin of for Do the same for P.8.8 Show that at low frequency, the nonrational function can be approximated by the following all-pass functions of increasing orders:

are called the Padé approximants of (Ts may be substituted for s to allow for any desired delay). For each case, draw the pole-zero pattern and the Bode plots. Compare the phase diagrams and comment on the domain of validity of the various approximations. [Hint: Expand the exponential and the all-pass functions into a McLauren series and match the coefficients.]

192


P.8.9 Consider a stable, minimum phase system with a high frequency attenuation rate larger than Using a Nyquist diagram and geometrical arguments, show that there is always a limit frequency above which there is an amplification of the disturbances.

Chapter 9

Optimal control 9.1

Introduction

The optimal control approach for SISO systems has already been introduced in chapter 7. In this chapter, we extend it to Multi-Input Multi-Output (MIMO) systems. The general results are stated without demonstration and the discussion is focused on the aspects which are important for the control of lightly damped flexible structures.

9.2

Quadratic integral

Consider the free response of an asymptotically stable linear system

The quadratic expression is often used to measure the distance from the equilibrium, Asymptotic stability implies that as and that the quadratic integral

converges, for any semi positive definite weighting matrix Q. Its value depends on the initial state alone. To compute J, consider the decay rate of the positive quadratic form where P is the positive definite solution of the matrix Liapunov equation

193

194


Substituting

and using Equ(9.3), one gets

Since Q is semi positive definite, this result proves that V(t) is a monotonically decaying function [it is indeed a Liapunov function for the system (9.1)]. Integrating Equ.(9.4) from 0 to provides the desired result:

where P is solution of the Liapunov equation (9.3).

9.3

Deterministic LQR

We now formulate the steady state form of the Linear Quadratic Regulator (LQR) problem for MIMO systems. This is the simplest and the most frequently used formulation; it can be extended to a cost functional with finite horizon, and to time varying systems. Consider the system where the system matrix A is not necessarily stable, but it is assumed that the pair (A, B) is controllable (controllability and observability will be discussed in detail in chapter 10). We seek a linear state feedback with constant gain

such that the following quadratic cost functional is minimized

where

is semi positive definite and R is strictly positive definite J has two contributions, one from the states and one from the control The fact that R is strictly positive definite expresses the fact that any control has a cost, while implies that some of the states may be irrelevant for the problem at hand. If a set of controlled variables is defined,

a quadratic functional may be defined as

Optimal control

195

This situation is equivalent to the previous one with In a similar way, one often uses the form where S is a scaling matrix and the scalar is a design parameter. It can be shown (see e.g. Kwakernaak & Sivan) that the solution of the problem (9.6) to (9.8) is where P is the symmetric positive definite solution of the algebraic Riccati equation The Riccati equation is nonlinear in P; the existence and uniqueness of the solution is guaranteed if (A, B) is a controllable pair (stabilizable is in fact sufficient) and is observable (by we mean a matrix H such that Under these conditions, the closed-loop system

is asymptotically stable. From Equ.(9.5), the minimum value of the cost functional is given by

where

is the solution of the Liapunov equation of the closed-loop system

Substituting

we get

This equation is identically satisfied by where P is solution of Equ.(9.12). Thus, the minimum value of the cost functional is related to the solution P of the Riccati equation by

Many techniques are available for solving the Riccati equation; they will not be discussed in this text because they are automated in most control design softwares. In principle, the LQR approach allows the design of multivariable state feedbacks which are asymptotically stable. A major drawback is, of course, that it assumes the knowledge of the full state vector Since the latter is, in general, not directly available, it has to be reconstructed. The poles of the closed-loop system depend on the matrices Q and R. Multiplying both Q and R by a scalar coefficient leads to the same gain matrix G and

196


the same closed-loop poles. In structural control applications, if the controlled variables are not clearly identified, it may be sensible to choose Q in such a way that represents the total (kinetic plus strain) energy in the system. Usually will be chosen where is a constant positive definite matrix and is an adjustable parameter; its value is selected to achieve reasonably fast closed-loop poles without excessive values of the control effort.

9.4

Stochastic response to a white noise

A white noise is a mathematical idealization of a stationary random process for which there is a total lack of correlation between the values of the process at different times. If is a vector white noise process of zero mean, we have

and the covariance matrix reads

where is symmetric semi positive definite and is called the covariance intensity matrix, but this is often abbreviated to covariance matrix. If the components of the random vector are independent, is diagonal. The corresponding power spectral density matrix is

This result shows that the power in the signal is uniformly distributed over the frequency. Although a white noise is not physically realizable [because its variance would not be finite, see the Dirac function in Equ.(9.17)], it is a convenient approximation which is appropriate whenever the correlation time of the signal is small compared to the time constant of the system to which it is applied. In what follows, it is assumed that all processes are Gaussian, so that they are entirely characterized by their second order statistics. Consider the stationary random response of a linear time invariant system to a white noise excitation of covariance intensity matrix

It can be shown that the covariance matrix of the steady state response,

satisfies the Liapunov equation

Optimal control

197

(see e.g. Bryson & Ho). This equation is the dual (adjoint) of Equ.(9.3); it expresses the equilibrium between the damping forces in the system and the random disturbance acting on it. It has a unique positive definite solution if the system is stable. Now, consider the quadratic performance index Since

we can write

where we have used the fact that the expected value applies only to random quantities. The covariance matrix X is the solution of Equ.(9.21). This result can be written alternatively

where P is solution of Equ.(9.3); the proof is left as an exercise (Problem P.9.1).

9.4.1 Remark The notation used in this section is the most commonly used in the literature; however, for the reader who is not trained in stochastic processes, it deserves some clarification: The definition (9.17) of the covariance matrix involves two separate times and for a stationary process, the covariance matrix is a function of the difference (delay) On the contrary, the covariance matrix involved in Equ.(9.20) and (9.21) is not a function of X is the particular value of the covariance function for it is the vector generalization of the mean square value.

9.5 Stochastic LQR The linear quadratic regulator can be formulated in a stochastic environment as follows. Consider the linear time invariant system subjected to a white noise excitation of intensity

Find the control function of all information from the past, that minimizes the performance index

The solution of the problem is a constant gain linear state feedback


198

where P is solution of the Riccati equation

We note that the solution does not depend on the noise intensity matrix and that it is identical to that of the deterministic LQR using the same weighting matrices Q and R. The closed-loop system behaves according to

and the performance index (9.25) can be rewritten

From Equ.(9.23), where

is solution of the Liapunov equation of the closed-loop system

This equation is identical to (9.15). Substituting the gain matrix from (9.27), we can readily establish that it is identically satisfied by the solution P of the Riccati equation (9.28). Thus, and the minimum value of the performance index is

This result is the dual of (9.16) for the deterministic case.

9.6

Asymptotic behaviour of the closed-loop

In the particular case where there is a single controlled variable and a single input, the cost functional (9.10) is reduced to (7.52):

We know from section 7.5 that the closed-loop poles are the stable roots of the characteristic equation

where

Optimal control

199

is the transfer function between the input and the controlled variable Only the part in the left half plane of the symmetric root locus is of interest. If

Equ.(9.34) becomes

For large values of that is when there is a heavy penalty on the control cost, the closed-loop poles are identical to the stable poles of or their mirror image in the left half plane if they are unstable. Thus, the cheapest stabilizing control in terms of control amplitude simply relocates the unstable poles at their mirror image in the left half plane. As the penalty on the control cost decreases, branches go from the poles to the left half plane zeros if the system is minimum phase, or to their mirror image, if the zeros are located in the right half plane. The remaining branches go to infinity. Their asymptotic behaviour can be evaluated by ignoring all but the highest power in in the characteristic equation (9.34). From (9.36), we get

or

The stable solutions lie on a circle of radius in a configuration known as a Butterworth pattern of order (Fig.9.1). For small values of the fact that the system is very fast is reflected by the large distance of the faraway poles to the origin. Some of the poles, however, do not move away but are stuck at the open-loop zeros. Although they


200

remain slow, they will not be noticeable on the controlled variable, because the open-loop zeros are also zeros of the closed-loop system [see Equ.(8.4)]. Extensions to MIMO systems with square open-loop transfer matrix can be found in Kwakernaak & Sivan.

9.7

Prescribed degree of stability

In the previous section, we have seen that as some of the closed-loop poles go to infinity according to Butterworth patterns, while the others go to the open-loop zeros. The latter depend on the choice of the matrix H defining the controlled variables In practice, a single controlled variable cannot always be clearly identified, and the choice of H is often more or less arbitrary: The designer picks up a H matrix and finds a root locus, on which reasonable pole locations can be selected. If one fails to find fast enough poles with acceptable control amplitudes, the process is repeated with another H matrix. It may require several iterations to avoid nearby poles. Anderson and Moore’s procedure allows one to design a regulator with a prescribed degree of stability. The method guarantees that the closed-loop eigenvalues (of A – BG) lie on the left of a vertical line at (Fig.9.2). The idea is to use a modified cost functional

where and

is a positive constant. If one defines the modified variables the system equation in terms of and becomes

and the performance index is

Optimal control

201

Equations (9.39) and (9.40) define a classical LQR problem in terms of the solution is a constant gain feedback

and

with where

is solution of the Riccati equation

The closed-loop system

is asymptotically stable. If one back substitutes

and

in Equ.(9.41), one gets

Thus, the optimal control is a constant gain linear feedback. Besides, since and is asymptotically stable, we know that as approaches infinity, at least as fast as The prescribed degree of stability has been achieved by solving a classical LQR problem for the modified system with the system matrix instead of A.

9.8

Gain and phase margins of the LQR

The LQR has been developed as the solution of an optimization problem. The constant gain feedback assumes the perfect knowledge of the state. We momentarily stay with this assumption and postpone the discussion of the state reconstruction until the next section. In the previous chapter, we emphasized the role played by the gain and phase margins in the good performance of a control system and in the way the performance will be degraded if some change occurs in the open-loop system. A good phase margin improves the disturbance rejection near crossover, and eliminates the oscillations in the closed-loop system; it also protects against instability if some delay is introduced in the loop. This section establishes guaranteed margins for the LQR. Figure 9.3.a shows the classical representation of the state feedback; it is not directly suitable for defining the gain and phase margins. However, the feedback loop can also be represented with a unity feedback as in Fig.9.3.b; the corresponding open-loop transfer function is

It is possible to draw a Nyquist plot of and this allows us to define the gain and phase margins as in the previous chapter.


202

The interesting result is that it can be demonstrated (see Anderson & Moore, p.68) that the LQR satisfies the following inequality.

This implies that the Nyquist plot of always remains outside a disk of unit radius, centered on –1 (Fig.9.4). Simple geometric considerations show that the LQR has a phase margin PM > 60°, and a infinite gain margin. The first result comes from the observation that any open-loop transfer function with a phase margin less than 60° would have to cross the unit circle within the forbidden disk. The second one is the consequence of the fact that the number of encirclements of –1 cannot be changed if one increases the gain by any factor larger than 1. Note that, for curve (2) in Fig.9.4.b, point A could cross -1 if the gain were reduced. However, since A is outside the forbidden circle, this cannot occur if the multiplying factor remains larger than 1/2. This result can be extended to multivariable systems: For each control channel, the LQR has guaranteed margins of

GM = 1/2 to

and

PM > 60°

For the situation depicted in Fig.9.4.a, the phase is close to 90° above crossover; as a result, the gain roll-off rate at high frequency is at most –20dB/decade.

9.9

Full state observer

The state feedback assumes that the states are known at all times. In most practical situations, a direct measurement of all the states would not be feasible. As we already saw earlier, if the system is observable, the states can be reconstructed from a model of the system and the output measurement One should never forget, however, that good state reconstruction requires a good model of the system. If the state feedback is based on the reconstructed states, the separation principle tells us that the design of the regulator and that of the observer can be done independently.

Optimal control

203

Consider the system

where and are uncorrelated white noise processes with zero mean and covariance intensity matrices V and W, respectively. From chapter 7, we know that without noise, a full state observer of the form

converges to the actual states, provided that the eigenvalues of (A – KC) are in the left half plane. In fact, the poles of the observer can be assigned arbitrarily if the system is observable. In presence of plant and measurement noise, and the error equation is the following


204

It shows that both the system noise and the measurement noise act as excitations on the measurement error. Note that the measurement noise is amplified by the gain matrix of the observer, which suggests that noisy measurements will require moderate gains in the observer.

9.9.1

Covariance of the reconstruction error

Comparing Equ.(9.50) with Equ.(9.19) to (9.21), we see that the steady state reconstruction error has zero mean, and a covariance matrix given by the Liapunov equation:

where we have used the fact that

and

are uncorrelated. It can be rewritten

This equation expresses the equilibrium between (as they appear in the equation) the dissipation in the system, the covariance of the disturbance acting on the system, the reduction of the error covariance due to the use of the measurement, and the measurement noise. The latter two contributions depend of the gain matrix K of the observer.

9.10

Kalman-Bucy Filter (KBF)

Since the error covariance matrix depends on the gain matrix K of the observer, one may look for the optimal choice of K which minimizes a quadratic objective function where is a vector of arbitrary coefficients. There is a universal choice of K which makes J minimum for all

where P is the covariance matrix of the optimal observer, solution of the Riccati equation This solution has been obtained as a parametric optimization problem for the assumed structure of the observer given by Equ.(9.49), but it is in fact optimal. The minimum variance observer is also called steady state Kalman-Bucy filter (KBF).

Optimal control

9.11

205

Linear Quadratic Gaussian (LQG)

The so-called Linear Quadratic Gaussian problem is formulated as follows: Consider the completely controllable and observable linear time-invariant system

where

and are uncorrelated white noise processes with intensity matrices and V > 0. Find the control such that the cost functional

is minimized. The solution of this problem is a linear, constant gain feedback

where G is the solution of the LQR problem and is the reconstructed state obtained from the Kalman-Bucy filter. Combining Equ.(9.56), (9.59) and (9.50), we obtain the closed-loop equation

Its block triangular form implies the separation principle: The eigenvalues of the closed-loop system consist of two decoupled sets, corresponding to the regulator and the observer. Note that the separation principle is related to the state feedback and the structure of the observer, rather than to the optimality; it applies to any state feedback and any full state observer of the form (9.49). The compensator equation has exactly the same form as that of Fig.7.9, except that it is no longer restricted to SISO systems. As we already mentioned, the stability of the compensator is not guaranteed, because only the closed-loop poles have been considered in the design.

9.12

Duality

Although no obvious relationship exists between the physical problems of optimal state feedback with a quadratic performance index, and the minimum variance state observer, the algebras of the solution of the two problems are closely related, as summarized in Table 9.1. The duality between the design of the regulator and that of the Kalman filter can be expressed as follows. Consider the fictitious dual control problem: Find that minimizes the performance index


206

for the system The solution is with where P is solution of

It is readily observed that this Riccati equation is that of the Kalman filter of the original problem, and that the gain matrix of the minimum variance observer for the original problem is related to the solution of the fictitious regulator problem by

9.13

Spillover

Flexible structures are distributed parameter systems which, in principle, have an infinite number of degrees of freedom. In practice, they are discretized by a finite number of coordinates (e.g. finite elements) and this is in general quite sufficient to account for the low frequency dynamical behaviour in most practical situations. When it comes to control flexible structures with state feedback and full state observer, the designer cannot deal directly with the finite element model, which is by far too big. Instead, a reduced model must be developed, which includes the few dominant low frequency modes. Due to the inherent low damping of flexible structures, particularly in the space environment, there is a danger that a state feedback based on a reduced model destabilizes the

Optimal control

207

residual modes, which are not included in the model of the structure contained in the observer. The aim of this section is to point out the danger of spillover instability. It is assumed that the state variables are the modal amplitudes and the modal velocities (as in section 7.2.2). In what follows, the subscript refers to the controlled modes, which are included in the control model, and the subscript refers to the residual modes which are ignored in the control design. Although they are not included in the state feedback, the residual modes are excited by the control input and they also contribute to the output measurement (Fig.9.5); it is this closed-loop interaction, together with the low damping of the residual modes, which is the origin of the problem. With the foregoing notation, the dynamics of the open-loop system is

A perfect knowledge of the controlled modes is assumed. The full state observer is

208


and the state feedback The interaction between the control system and the residual modes can be analysed by considering the composite system formed by the state variables where The governing equation is

This equation is the starting point for the analysis of the spillover (Fig.9.5). The key terms are and They arise from the sensor output being contaminated by the residual modes via the term (observation spillover), and the feedback control exciting the residual modes via the term (control spillover). Equ.(9.66) shows that if either or the eigenvalues of the system remain decoupled, that is identical to those of the regulator the observer and the residual modes They are typically located in the complex plane as indicated in Fig.9.6. The poles of the regulator (controlled modes) have a substantial stability margin, and the poles of the observer are located even farther left. On the contrary, the poles corresponding to the residual modes are barely stable, their only stability margin being provided by the natural damping. When both and i.e. when there is both control and observation spillover, the eigenvalues of the system shift away from their decoupled locations. The magnitude of the shift depends on the coupling terms and Since the stability margin of the residual modes is small, even a small

Optimal control

209

shift can make them unstable. This is spillover instability. Not all the residual modes are potentially critical from the point of view of spillover, but only those which are observable, controllable, and are close to the bandwidth of the controller (Problem P.9.6).

9.13.1

Spillover reduction

In the previous section, we have seen that the spillover phenomenon arises from the excitation of the residual modes by the control (control spillover, and the contamination of the sensor output by the residual modes (observation spillover, For MIMO systems, both terms can be reduced by a judicious design of the regulator and the Kalman filter. Control spillover can be alleviated by minimizing the amount of energy fed into the residual modes. This can be achieved by supplementing the cost functional used in the regulator design by a quadratic term in the control spillover:

where the weighting matrix W allows us to penalize some specific modes. This amounts to using the modified control weighting matrix

This control weighting matrix penalizes the excitation whose shape is favorable to the residual modes; this tends to produce a control which is orthogonal to the residual modes. Of course, this is achieved more effectively when there are many actuators, and it cannot be achieved at all with a single actuator. Similarly, a reduction of the observation spillover can be achieved by designing the observer as a Kalman filter with a measurement noise intensity matrix

The extra contribution to the covariance intensity matrix indicates to the filter that the measurement noise has the spatial shape of the residual modes is the covariance matrix of the observation spillover if This tends to desensitize the reconstructed states to the residual modes. Here again, the procedure works better if many sensors are available. The foregoing methodology for spillover reduction was introduced by Sezak et al. under the name of Model Error Sensitivity Suppression (MESS). It is only one of the many methods for spillover reduction, but it is interesting because it pinpoints the role of the matrices R and V in the LQG design. Another interesting situation where the plant noise statistics have a direct impact on the stability margins is discussed in the next section.

210

9.14


Loop Transfer Recovery (LTR)

In section 9.8, we have seen that the LQR has guaranteed stability margins of GM = 1/2 to and PM > 60° for each control channel. This property is lost when the state feedback is based on an observer or a Kalman filter. In that case, the margins can become substantially smaller. The Loop Transfer Recovery (LTR) is a robustness improvement procedure consisting of using a Kalman filter with fictitious noise parameters: If is the nominal plant noise intensity matrix, the KBF is designed with the following plant noise intensity matrix

where is an arbitrary symmetric semi positive definite matrix and q is a scalar adjustment parameter. From the presence of the input matrix B in the second term of (9.70), we see that the extra plant noise is assumed to enter the system at the input. Of course, for q = 0, the resulting KBF is the nominal one. As it can be proved (Doyle & Stein) that, for square, minimum phase open-loop systems the loop transfer function from the control signal to the compensator output (loop broken at the input of the plant, as indicated in Fig.9.7) tends to that of the LQR:

As a result, the LQG/LTR recovers asymptotically the margins of the LQR as

Optimal control

211

Note that The loop breaking point at the input of the plant, as indicated by × in Fig.9.7, is a reasonable one, because this is typically one of the locations where the uncertainty enters the system. The KBF gain matrix, K(q) is a function of the scalar parameter q. For q = 0, K(0) is the optimal filter for the true noise parameters. As q increases, the filter does a poorer job of noise rejection, but the stability margins are improved, with essentially no change in the bandwidth of the closed-loop system. Thus, the designer can select q by trading off noise rejection and stability margins. The margins of the LQG/LTR are indeed substantial; they provide a good protection against delays and nonlinearities in the actuators. They are not sufficient to guarantee against spillover instability, however, because the phase uncertainty associated with a residual mode often exceeds 60° (it may reach 180° if the residual mode belongs to the bandwidth). The LTR procedure is normally applied numerically by solving a set of Riccati equations for increasing values of until the right compromise is achieved. For SISO systems, it can also be applied graphically on a symmetric root locus, by assuming that the noise enters the plant at the input [ in Equ.(7.72)] (Problem P.9.7).

9.15

Integral control with state feedback

Consider a linear time invariant system subject to a constant disturbance

If we use a state feedback to stabilize the system, there will always be a non-zero steady state error in the output Increasing the gain G would reduce the error at the expense of a wider bandwidth and a larger noise sensitivity. An alternative approach consists of introducing an integral action by supplementing Equ.(9.72) by leading to the augmented state vector With the state feedback

the closed-loop equation reads


212

If G and

are chosen in such a way that they stabilize the system, we have

which means that the steady state error will be zero edge of the disturbance

9.16

without knowl-

Frequency shaping

As we saw in earlier chapters, the desirable features of control systems include some integral action at low frequency to compensate for steady state errors and very low frequency disturbances, and enough roll-off at high frequency for noise rejection, and to stabilize the residual dynamics. Moreover, there are special situations where the system is subjected to a narrow-band disturbance at a known frequency. The standard LQG does not give the proper answer to these problems (no integral action, and the roll-off rate of the LQR is only –20 dB/decade). We have seen in the previous section how the state space model can be modified to include some integral action; in this section, we address the more general question of frequency shaping. The weakness of the standard LQG formulation lies in the use of a frequency independent cost functional, and of noise statistics with uniform spectral distribution (white noise). Frequency shaping can be achieved either by considering a frequency dependent cost functional in the LQR formulation, or by using colored (i.e. non-white) noise statistics in the LQG problem.

9.16.1

Frequency-shaped cost functionals

According to Parseval’s theorem, the cost functional of the LQR, Equ.(9.10), can be written in the frequency domain as

where and are the Fourier transforms of and and indicate the complex conjugate transposed (Hermitian). Equ.(9.78) shows clearly that the weighing matrices and S do not depend on meaning that all the frequency components are treated equally. Next, assume that we select frequency dependent weighing matrices

Clearly, if the shaping objectives are to produce a P+I type of controller and to increase the roll-off, we must select to put more weight on low frequency, to achieve some integral action, and to put more weight on high frequency,

Optimal control

213

to attenuate the high frequency contribution of the control. Examples of such functions in the scalar case are

where the corner frequencies and and the exponent are selected in the appropriate manner. Typical penalty functions are represented in Fig.9.8. Likewise, a narrow-band disturbance can be handled by including a lightly damped oscillator at the appropriate frequency in Equation (9.78) can be rewritten

We assume that all the input and output channels are filtered in the same way, so that the weighing matrices are restricted to the form and with and being scalar functions. If we introduce the modified controlled variable

and control we get the frequency independent cost functional


214

or, in the time domain,

This cost functional refers to the augmented system of Fig.9.9, including input filters on all input channels and output filters on all controlled variables. If a state space realization of these filters is available (Problem P.9.9), the complete system is governed by the following equations: Structure

Output filter [state space realization of

Input filter [state space realization of

These equations can be combined together as

with the augmented state vector

Optimal control

215

and the notations

The state feedback is obtained by solving the LQR problem for the augmented system with the quadratic performance index (9.85). Notice that, since the input and output filter equations are solved in the computer, the states and are known; only the states of the structure must be reconstructed with an observer. The overall architecture of the controller in shown in Fig.9.10. It can be shown that the poles of the output filter (eigenvalues of ) appear unchanged in the compensator (Problem P.9.11); this property can be used to introduce a large gain over a narrow frequency range, by introducing a lightly damped pole in (Problem P.9.10).

9.16.2 Noise model As an alternative to the frequency-shaped cost functionals, loop shaping can be achieved by assuming that the plant noise has an appropriate power spectral density, instead of being a white noise. Thus, we assume that is the output of a filter excited by a white noise at the input. If the system is governed by

216


and the plant noise is modelled according to

where is stable and is a white noise (Problem P.9.12). The two sets of equations can be coupled together to form the augmented system

or, with

and the appropriate definitions of

and

Since and are white noise processes, the augmented system fits into the LQG framework and a full state feedback and a full state observer can be constructed by solving the two problems

with the appropriate matrices and V. In Equ.(9.101), note that the filter dynamics is not controllable from the plant input, but this is not a problem provided that is stable, that is if the augmented system is stabilizable (see next chapter). In principle, a large gain over some frequency range can be obtained by proper selection of the poles of and the input and output matrices and However, in contrast to the previous section, the poles of do not appear unchanged in the compensator (Problem P.9.13) and this technique may lead to difficulties for the rejection of narrow-band perturbations (Problem P.9.14).

9.17

References

B. D. O. ANDERSON & J. B. MOORE, Linear Optimal Control, Prentice Hall, Inc. Englewood Cliffs, NJ, 1971. K. J. ÅSTRÖM, Introduction to Stochastic Control Theory, Academic Press, 1970.

Optimal control

217

M. ATHANS, The Role and Use of the Stochastic Linear-Quadratic-Gaussian Problem in Control System Design, IEEE Transactions on Automatic Control, vol. AC-16, no. 6, 529–552, Dec. 1971. M. J. BALAS, Active Control of Flexible Systems, Journal of Optimization Theory and Applications, vol. 25, no. 3, 415–436, 1978. J. C. DOYLE & G. STEIN, Robustness with Observers, IEEE Transactions on Automatic Control, vol. AC-24, no. 4, 607–611, Aug. 1979. J. C. DOYLE & G. STEIN, Multivariable Feedback Design: Concepts for a Classical/Modern Synthesis, IEEE Transaction on Automatic Control, vol. AC-26, no. 1, 4–16, Feb. 1981. K. GUPTA, Frequency-Shaped Cost Functionals: Extension of Linear Quadratic Gaussian Methods, AIAA J. of Guidance and Control, vol. 3, no. 6, 529–535, Nov.-Dec. 1980. H. KWAKERNAAK & R. SIVAN, Linear Optimal Control Systems, Wiley, 1972. J. M. MACIEJOWSKI, Multivariable Feedback Design, Addison-Wesley, 1989. J. R. SEZAC, P. LIKINS & T. CORADETTI, Flexible Spacecraft Control by Model Error Sensitivity Suppression, Proceedings of the VPI&SU/AIAA Symposium on Dynamics & Control of Large Flexible Spacecrafts, Blacksburg, VA, 1979. D. M. WIBERG, State Space and Linear Systems, McGraw-Hill Schaum's Outline Series in Engineering, 1971.

9.18

Problems

P.9.1 Consider the linear system (9.19) subjected to a white noise excitation with covariance intensity matrix Show that the quadratic performance index can be written alternatively

where P is the solution of the Liapunov equation (9.3). P.9.2 Consider the inverted pendulum of section 7.2.3. Using the absolute displacement as control variable, design a LQR by solving the Riccati equation, for various values of the control weight Compare the result to that obtained in section 7.5.2 with the symmetric root locus. P.9.3 Same as problem P.9.2 but with the procedure of section 9.7. Check that for all values of the closed-loop poles lie to the left of the vertical line at (select to the left of ). Compare the state feedback gains to those of the previous problem.

218


P.9.4 For one of the LQR designed at Problem P.9.2, draw the Nyquist plot of Evaluate the gain and phase margins. P.9.5 Consider the state space equation (7.14) of a flexible structure in modal coordinates and assume that the mode shapes have been normalized in such a way that Show that the total energy (kinetic + strain) can be written in the form with where is the state vector defined as P.9.6 Consider a simply supported uniform beam with a point force actuator at and a displacement sensor at Assume that the system is undamped and that and (a) Write the equations in state variable form using the state variable defined as (b) Design a LQR for a model truncated after the first three modes, using Q = I (see Problem P.9.5); select the control weight in such a way that the closed-loop poles are and (c) Check that a full state Luenberger observer with poles located at

shifts the residual mode from to (this example was used by Balas to demonstrate the spillover phenomenon). (d) Using a model with 3 modes and assuming that the plant noise intensity matrix has the form design a KBF and plot the evolution of the residual modes 4 and 5 (in closed-loop) as the noise intensity ratio increases (and the observer becomes faster). (e) For the compensator designed in (d), assuming that all the modes have a structural damping of plot the evolution with the parameter q of the open-loop transfer function corresponding to 5 structural modes (including 2 residual modes). [Hint: Use the result of Problem P.2.5 to compute P.9.7 Reconsider the inverted pendulum of Problem P.9.4. Assume that the output is the absolute position of the pendulum. Design a Kalman filter assuming that the plant noise enters the system at the input (E = B). Apply the LTR procedure and check that, as increases, the open-loop transfer function tends to that of the LQR (Problem P.9.4). Check the effect of the procedure on the bandwidth of the control system. [Note: The assumption that the output of the system is the absolute position rather than the tilt angle may appear as a practical restriction, but it is not, because can always be obtained indirectly from and by Equ.(7.21). It is necessary to remove the feedthrough component from the output before applying the LTR procedure.] [Hint: The KBF/LTR is the limit as of the symmetric root locus (7.71) based on E = B.]

Optimal control

219

P.9.8 Consider the two-mass problem of section 7.10.1. (a) Design a LQR by solving the Riccati equation for various values of the control weight Show that for some we obtain the same gains as those obtained with the symmetric root locus in section 7.10.1. (b) For these gains, draw the Nyquist plot of the LQR, evaluate the gain and phase margins. (c) Assuming that the plant noise enters at the input, design a KBF by solving the Riccati equation for various values of the noise intensity ratio Show that for some q, we obtain the same gains as those obtained with the symmetric root locus. Calculate the gain and phase margins. (d) Apply the LTR technique with increasing q; draw a set of Nyquist plots of showing the evolution of the gain and phase margins. Check that as P.9.9 Find a state space realization of the input and output filters and corresponding to the weighting matrices (9.80):

The latter is known as Butterworth filter of order its poles are located on a circle of radius according to Fig.9.1. P.9.10 Consider the two-mass problem of section 7.10.1. Assume that the system is subjected to a sinusoïdal disturbance at acting on the main body. Using a frequency-shaped cost functional, design a LQG controller with good disturbance rejection capability. Compare the performance of the new design to the nominal one (time response, sensitivity function,...). [Hint: use a lightly damped oscillator as output filter

where is kept as design parameter.] P.9.11 Show that the compensator obtained by the frequency-shaped cost functional has the following state space realization:

where is the gain of the observer for and is the gain of the state feedback. Note that, as a result of the structure of the system matrix, the poles of the compensator include those of the output filter,

220


P.9.12 Find a state space realization of the noise model (9.99) (9.100) achieving the following power spectral density:

(this filter combines in cascade the two filters used in problem P.9.9). P.9.13 Show that the compensator obtained by using a noise model in the loop shaping has the following state space realization:

where is the gain matrix of the regulator of the augmented system and is the corresponding observer gain matrix. Note that the system matrix is no longer block triangular, so that the poles of the compensator differ from those of P.9.14 Repeat Problem P.9.10 using a noise model ( is the output of a second order filter). Compare with the result of Problem P.9.10.

Chapter 10

Controllability and Observability 10.1

Introduction

Controllability measures the ability of a particular actuator configuration to control all the states of the system; conversely‚ observability measures the abilty of the particular sensor configuration to supply all the information necessary to estimate all the states of the system. Classically‚ control theory offers controllability and observability tests which are based on the rank deficiency of the controllability and observability matrices: The system is controllable if the controllability matrix is full rank‚ and observable if the observability matrix is full rank. This answer is often not enough for practical engineering problems where we need more quantitative information. Consider for example a simply supported uniform beam; the mode shapes are given by (2.44). If the structure is subject to a point force acting at the center of the beam‚ it is obvious that the modes of even orders are not controllable because they have a nodal point at the center. Similarly‚ a displacement sensor will be insensitive to the modes having a nodal point where it is located. According to the rank tests‚ as soon as the actuator or the sensor are slightly moved away from the nodal point‚ the rank deficiency disappears‚ indicating that the corresponding mode becomes controllable or observable. This is too good to be true‚ and any attempt to control a mode with an actuator located close to a nodal point would inevitably lead to difficulties‚ because this mode is only weakly controllable or observable. In this chapter‚ after having discussed the basic concepts‚ we shall turn our attention to the quantitative measures of controllability and observability‚ and apply the concept to model reduction. 221


222

10.1.1

Definitions

Consider the linear time-invariant system

The system is completely controllable if the state of the system can be transferred from zero to any final state within a finite time. The system is stabilizable if all the unstable eigenvalues are controllable or‚ in other words‚ if the non controllable subspace is stable. The system is completely observable if the state can be determined from the knowledge of and over a finite time segment. In the specialized literature‚ observability refers to the determination of the current state from future output‚ while the determination of the state from past output is called reconstructibility. For linear‚ time-invariant systems‚ these concepts are equivalent and do not have to be distinguished. The system is detectable if all the unstable eigenvalues are observable‚ or equivalently‚ if the unobservable subspace is stable.

10.2

Controllability and observability matrices

The simplest way to introduce the controllability matrix is to consider the single input discrete-time system governed by the difference equation

where A is the system matrix and Assuming that the system starts from rest‚ state vector resulting from the scalar input

or

the are

input vector. the successive values of the

223

Controllability and Observability where

is equal to the order of the system. The

matrix

is called the controllability matrix; its columns span the state space which can be reached after exactly samples. If is full rank‚ the state vector can be transferred to any final value after only samples. By solving Equ.(10.4)‚ we find

Next‚ consider the values of

for

Once again‚

It turns out that the rank of the rectangular matrix

is the same as that of and that the columns of the two matrices span the same space. This is a consequence of the Cayley-Hamilton theorem‚ which states that every matrix A satisties its own characteristic equation. Thus‚ if the characteristic equation of A is

A satisfies the matrix equation

It follows that for any is linearly dependent on the columns of the controllability matrix as a result‚ increasing the number of columns does not enlarge the space which is spanned (Problem P.10.1). In conclusion‚ the system (10.3) if completely controllable if and only if (iff) the rank of the controllability matrix is This result has been established for a single-input discrete-time linear system‚ but it also applies to multi-input discrete as well as continuous time linear systems. The linear time-invariant system (10.1) with inputs is completely controllable iff the controllability matrix


224

is such that We then say that the pair (A‚ B) is controllable. If is not full rank‚ the subspace spanned by its columns defines the controllable subspace of the system. In a similar manner‚ the system (10.1) (10.2) is observable iff the observability matrix

is such that In this case‚ we say that the pair (A‚ C) is observable. From the fact that

we conclude that the pair (A‚ C) is observable iff the dual system is controllable. Conversely‚ the pair (A‚ B) is controllable iff the dual system is observable. The duality between observability and controllability has already been stressed in section 9.12.

10.3

Examples

10.3.1

A cart with two inverted pendulums

Consider two inverted pendulums with the same mass and lengths and placed on a cart of mass M (Fig.10.1.b). Assume that the input variable is the force applied to the cart (in contrast to section 7.2.3‚ where the input was the displacement of the support). Using the state variables we can write the linearized equations near as

where and controllability matrix is

(Problem P. 10.3). The

Controllability and Observability

225

It can be checked easily that this matrix is full rank provided If the rank of is reduced to 2. Thus‚ when the time constants of the two pendulums are the same‚ the system is not controllable (in practical applications‚ it is likely that the difficulties in controlling the system will appear long before reaching ). Next‚ consider the observability of the system from the measurement of We have C = (1‚0‚0‚0) and the observability matrix is

Since we conclude that the system is always observable from a single angle measurement; this result is somewhat surprising‚ but true.


226

10.3.2

Double inverted pendulum

Next‚ consider a double inverted pendulum on a cart as in Fig.10.1.c. To simplify the equations without losing any generality in the discussion‚ we assume that the two arms have the same length‚ and that the two masses are the same. The equations of motion can be written more conveniently by using the absolute tilt angles of the two arms (Problem P.10.4). Using the state vector we can write the linearized equations about the vertical position as

where

and

Since det the observability matrix from

We have det able from alone.

10.3.3

The controllability matrix reads

the system is always controllable. Similarly‚ reads

which indicates that the system is indeed observ-

Two d.o.f. oscillator

Consider the mechanical system of Fig.10.2. It consists of two identical undamped single d.o.f. oscillators connected with a spring of stiffness The input of the system is the point force applied to mass 1. The mass and stiffness matrices are respectively

Defining the state vector state space equation

and using Equ.(7.11)‚ we find the


where

227

The controllability matrix reads

det indicates that the system is no longer controllable as approaches 0. Indeed‚ when the stiffness of the coupling spring vanishes‚ the two masses become uncoupled and mass 2 is uncontrollable from the force acting on mass 1.

10.4

State transformation

Consider a SISO system

Since A is and and are both the system has If we consider the non singular transformation of the state‚

parameters.

the transformed state equation is

or

with the proper definition of and The non singular transformation matrix T contains free parameters which can be chosen to achieve special

228


properties for the transformed system; we shall discuss an example in detail in the next section. It can be shown ( Problem P.10.5) that the controllability matrix of the transformed system‚ is related to that of the original system by For any non singular transformation T, the rank of is the same as that of Thus, the property of controllability is preserved by any non singular transformation.

10.4.1

Control canonical form

We have seen in the previous section that the transformation matrix T can be selected in such a way that the transformed system has special properties. A form which is especially attractive from the state feedback point of view is the control canonical form‚ where the transformed system is expressed in terms of the coefficients and appearing in the system transfer function

The transformed matrices are (Problem P.10.6)

Besides the fact that the transformation between the state space model in control canonical form and the input-output model is straightforward‚ it is easy to compute the state feedback gains to achieve a desired closed-loop characteristic equation. Indeed‚ if the state feedback is applied‚ the closed-loop system matrix becomes

The corresponding characteristic equation is


229

Thus‚ in control canonical form‚ the state feedback gains can be obtained directly from the coefficients of the closed-loop characteristic equation‚ making pole placement very simple. The state feedback gains in the original state space system are slightly more difficult to compute‚ as we now examine. In principle‚ the linear transformation matrix leading from the original state space representation to the control canonical form can be obtained from Equ.(10.27): where and are the controllability matrices of the original system and of the control canonical form (Problem P.10.7)‚ respectively. From

it follows that the state feedback gains those in control canonical form‚ by

of the original model are related to

This formula is not very practical‚ because it requires the inverse of the controllability matrix. However‚ it can be expressed alternatively by Ackermann’s formula where and is the closed-loop characteristic polynomial‚ expressed in terms of the open-loop system matrix A. Equation (10.34) states that the gain vector is in fact the last row of The demonstration uses the Cayley-Hamilton theorem; it is left to the reader (Problem P.10.8). Note that does not have to be calculated explicitly; instead‚ it is more convenient to proceed in two steps‚ by first solving the equation

for

and then computing

10.4.2

Left and right eigenvectors

If the non-symmetric system matrix A has distinct eigenvalues‚ its eigenvectors will be linearly independent and can be taken as the columns of a regular matrix P: where that

is a diagonal matrix with the eigenvalues of A. It follows


230

If we define

and right multiply the foregoing equation by

The columns of P and of Q (i.e. the rows of left eigenvectors of A‚ respectively‚ because

From the definition of

we get

) are called the right and

the left and right eigenvectors are orthogonal

From Equ.(10.36)‚ we have

10.4.3

Diagonal form

Let us use the right eigenvector matrix P as state transformation matrix

Following the procedure described earlier in this section‚ we can write the transformed state equation as

Since is a diagonal matrix with entries equal to the poles of the system, Equ.(10.42) shows that the transformed system behaves like a set of independent first order systems. The diagonal form is also called the modal form, and the states are the modes of the system. Note that this concept of mode is related only to the matrix A and is different from the vibration modes as defined in section 2.2 (for an undamped structure, the entries of are identical to the natural frequencies of the structure, as illustrated in the example of section 10.7). For MIMO systems, Equ.(10.42) and (10.43) become

10.5

PBH test

It is easy to show (Problem P.10.10) that the controllability matrix in diagonal form reads


231

The second matrix in this expression is called a Vandermonde matrix; it is nonsingular if the eigenvalues are distinct. In this case‚ the rank of is the same as that of diag As a result‚ the system is controllable iff

Thus‚ any left eigenvector orthogonal to the input vector is uncontrollable. From Equ.(10.42)‚ we see that is in fact a measure of the effective input of the control in mode and can therefore be regarded as a measure of controllability of mode From the duality between controllability and observability‚ the foregoing results can readily be extended to observability; the observability matrix reads (Problem P.10.10)

Once again‚ a system with distinct eigenvalues is controllable iff

any right eigenvector orthogonal to the output vector is unobservable. From Equ.(10.43)‚ we see that is a measure of the contribution of mode to the output From Equ.(10.46) and (10.48)‚ we conclude that a system with multiple eigenvalues cannot be controlled from a single input‚ nor observed from a single output. The tests (10.47) and (10.49) are often called the Popov-Belevitch-Hautus (in short PBH) eigenvector tests of controllability and observability. For a MIMO system‚ is a row vector; its entry measures the controllability of mode from the input Similarly‚ the component of measures the observability of mode from the component of the output vector.

10.6

Residues

Next‚ consider the open-loop transfer function of the system‚

From Equ.(10.42) and (10.43)‚ it can be written alternatively


232

Since position

is diagonal‚ we easily obtain the following partial fraction decom-

where the residue of mode

is the product of the observability and controllability measures of mode For MIMO systems‚ the partial fraction decomposition becomes

with the residue matrix Its entry combines the observability of mode controllability from input

10.7

from output

and the

Example

In order to dissipate any confusion about the eigenvectors of A and the mode shapes of the structure (section 2.2)‚ let us consider a flexible structure with one input and one output; we assume that the dynamic equations are written in state variable form (7.14) and‚ to make things even clearer‚ we further assume that the system is undamped and that the mode shapes are normalized according to We use the notation and to emphasize the fact that and contain the amplitude of the mode shapes at the actuator and sensor locations‚ respectively. With these notations‚ the state space equation reads

In this equation‚ the state vector is

where is the vector of the amplitudes of the structural modes. The non-diagonal system matrix can be brought to diagonal form according to Equ.(10.40); we get


233

We see that the natural frequencies of the system appear with positive and negative signs on the diagonal of ‚ but the eigenvectors of A have nothing to do with the mode shapes of the structure. The PBH eigenvector tests read

Thus‚ the controllability and observability measures and are proportional to the modal amplitudes (a) and respectively. Introducing this in Equ.(10.52) and combining the complex conjugate eigenvalues‚ the partial fraction decomposition can be reduced to

where the sum extends to all the structural modes This result is identical to Equ.(2.47). To conclude this example‚ we see that when the state equation is written in modal coordinates as in Equ.(10.56)‚ the PBH tests and the associated controllability and observability measures provide no more information than the amplitude of the mode shapes‚ and This fact could have been anticipated from Equ.(10.61).

10.8

Sensitivity

The ultimate goal of the control system is to relocate the closed-loop poles at desirable locations in the complex plane; this should be done‚ preferably‚ with moderate values of the gain‚ in order to limit the control effort and the detrimental effects of noise and modelling errors. The closed-loop poles of a SISO system are solutions of the characteristic equation they start from the open-loop poles for and move gradually away as increases‚ in a direction which is dictated by the compensator The rate of change of near is a direct measure of the capability of the control system; it can be evaluated as follows: for a small gain if the open-loop poles are distinct‚ we can approximate The partial fraction decomposition (10.52) becomes

and the characteristic equation


234

or

Upon taking the limit

we get

This result shows that the rate of change of the closed-loop poles near is proportional to the corresponding residue and to the magnitude of the transfer function of the compensator The latter observation explains why the poles located in the roll-off region of the compensator move only very slowly for small

10.9

Controllability and observability Gramians

Consider the linear time-invariant system (10.1); the controllability measures the ability of the controller to control all the system states from the particular actuator configuration‚ or equivalently‚ the abilty to excite all the states from the input Consider the response of the system to a set of independent white noises of unit intensity:

If the system is asymptotically stable (i.e. if all the poles of A have negative real parts)‚ the response of the system is bounded‚ and the steady state covariance matrix is finite; it reads (Problem P.10.12.a)

is called the Controllability Gramian. According to section 9.4‚ it is solution of the Liapunov equation

The system is controllable if all the states of the system can be excited; this condition is fulfilled iff is positive definite. From the duality between the observability and controllability‚ we know that the pair (A‚ C) is observable iff the pair is controllable. It follows that the system is observable iff the observability Gramian


235

is positive definite. Substituting to (A‚ B) in Equ.(10.67)‚ we see that‚ if A is asymptotically stable‚ is solution of

Just as the controllability Gramian reflects the ability of the input to perturb the states of the system‚ the observability Gramian reflects the ability of nonzero initial conditions of the state vector to affect the output of the system. This can be seen from the following result (Problem P.10.12.b):

If we perform a coordinate transformation

the Gramians are transformed according to

where the notation refers to the controllability Gramian after the coordinate transformation (10.71). The proof is left to the reader (Problem P.10.13).

10.10

Relative controllability and observability

As we have just seen‚ the Gramians depend on the choice of state variables. Since‚ in most cases‚ the latter are not dimensionally homogeneous‚ nor normalized in an appropriate manner‚ the magnitude of the entries of the Gramians are not physically meaningful for identifying the least controllable or least observable part of a system. This information would be especially useful for model reduction. It is possible to perform a coordinate transformation such that the controllability and observability Gramians are diagonal and equal; this unique set of coordinates is called internally balanced (Moore)

10.10.1 Internally balanced coordinates Let and be the controllability and observability Gramians of an asymptotically stable time-invariant linear system. We perform a spectral decomposition of according to


236

where is a unitary matrix and is the diagonal matrix of eigenvalues (all positive if is positive definite). If we define we can write equivalently (when is a lower triangular matrix‚ this decomposition is called a Choleski factorization). From Equ.(10.73) and (10.74)‚ if we perform a change of coordinates with

the Gramians become

Next‚ we perform the spectral decomposition of

(with and use the transformation matrix another change of coordinates

according to

to perform

Equ.(10.72) and (10.73) show that the Gramians in the new coordinate system read

Thus‚ in the new coordinate system‚ the controllability and observability Gramians are equal and diagonal

For this reason‚ the new coordinate system is called internally balanced; it is denotated The global coordinate transformation is

and the internally balanced model is readily obtained from Equ.(10.23) and (10.24). From Equ.(10.72) and (10.73)‚ we see that‚ for any transformation T‚

It follows that Thus‚ the eigenvalues of are the entries of matrix contains the right eigenvectors of

and the transformation


10.11

237

Model reduction

Consider the partition of a state space model according to

If‚ in some coordinate system‚ the subsystem has the same impulse response as the full order system‚ it constitutes an exact lower order model of the system; the model of minimum order is called the minimum realization. Model reduction is concerned with approximate models‚ and involves a tradeoff between the order of the model and its ability to duplicate the behaviour of the full order model within a given frequency range.

10.11.1

Transfer equivalent realization

If we consider the partial fraction decomposition (10.54)‚ one reduction strategy consists of truncating all the modes with poles far away from the frequency domain of interest (and possibly including their contribution to the static gains) and also those with small residues which are only weakly controllable or observable (or both). This procedure produces a realization which approximates the transfer function within the frequency band. However‚ since the uncontrollable part of the system is deleted‚ even if it is observable‚ the reduced model cannot reproduce the response to disturbances that may excite the system. This may lead to problems in the state reconstruction. To understand this‚ recall that the transfer function which is the relevant one for the observer design‚ is that between the plant noise and the output (section 7.7). If the plant noise does not enter at the input‚ does have contributions from all observable modes‚ including those which are uncontrollable from the input. The procedure can be improved by including all the modes which have a significant contribution to too.

10.11.2

Internally balanced realization

Internally balanced coordinates can be used to extend the concept of minimum realization. The idea consists of using the entries of the joint Gramian to partition the original system into a dominant subsystem‚ with large entries and a weak one‚ with small The reduction is achieved by cutting the weak subsystem from the dominant one. The following result guarantees that the reduced system remains asymptotically stable. If the internally balanced system is partitioned according to (10.87) and if the joint Gramian is


238

the two subsystems and and internally balanced‚ such that

are asymptotically stable

The proof is left to the reader (Problem P.10.14). Thus‚ if we order the internally balanced coordinates by decreasing magnitude of and if the subsystems 1 and 2 are selected in such a way that the global system is clearly dominated by subsystem 1. The model reduction consists of severing subsystem 2‚ as indicated in Fig. 10.3‚ which produces the reduced system

10.11.3

Example

Consider a simply supported uniform beam with a point force actuator at and a displacement sensor at We assume that and The natural frequencies and the mode shapes are given by (2.43) and (2.44); we find etc... The system can be written in state variable form according to (7.14). In a second step‚ the system can be transformed into internally balanced coordinates following the procedure of section 10.10.1. Two kinds of reduced models have been obtained as follows: Transform into internally balanced coordinates and delete the subsystem corresponding to the smallest entries of the joint Gramian. Delete the modal coordinates corresponding to the smallest static gains


239

in the modal expansion of the transfer function

Figure 10.4 compares the amplitude plots of the input-output frequency response function of the reduced models with 8 and 12 states‚ with that of the full model; the internally balanced realization and the modal truncation based on the static gains are almost identical (they cannot be distinguished on the plot). Figure 10.5 compares the results obtained with the same reduced models‚ for the frequency response function between a disturbance applied at and the output sensor. Once again‚ the results obtained with the internally balanced realization and the modal truncation based on the static gains are nearly the same (we can notice a slight difference near for the reduced models with 12 states); the reduced models with 8 states are substantially in error in the vicinity of because mode 3‚ which has been eliminated during the reduction process (it is almost not controllable from the input)‚ is excited by


240

the disturbance.

10.12

References

J. ACKERMANN‚ Sampled-Data Control Systems‚ Springer-Verlag‚ 1985. G. F. FRANKLIN‚ J. D. POWELL & A. EMANI-NAEMI‚ Feedback Control of Dynamic Systems‚ Addison-Wesley‚ 1986. A. M. A. HAMDAN & A. H. NAYFEH‚ Measure of Modal Controllability and Observability for First and Second Order Linear Systems‚ AIAA J. of Guidance‚ Control‚ and Dynamics‚ vol. 12‚ no. 5‚ 768‚ 1989. P. C. HUGHES‚ Space Structure Vibration Modes: How Many Exist? Which Ones are Important?‚ IEEE Control Systems Magazine‚ February‚ 1987a. J. L. JUNKINS & Y. KIM‚ Introduction to Dynamics and Control of Flexible Structures‚ AIAA Education Series‚ 1993.

241

Controllability and Observability T. KAILATH‚ Linear Systems‚ Prentice-Hall‚ 1980.

Y. KIM & J. L. JUNKINS‚ Measure of Controllability for Actuator Placement‚ AIAA J. of Guidance‚ Control‚ and Dynamics‚ vol. 14‚ no. 5‚ 895–902‚ Sep.– Oct. 1991. H. KWAKERNAAK & R. SIVAN‚ Linear Optimal Control Systems‚ Wiley‚ 1972. B.C. MOORE‚ Principal Component Analysis in Linear Systems: Controllability‚ Observability and Model Reduction‚ IEEE Trans‚ on Automatic Control‚ vol. AC-26‚ no. 1‚ 17–32‚ 1981. R‚ E. SKELTON‚ Dynamic System Control - Linear System Analysis and Synthesis‚ Wiley‚ 1988. R. E. SKELTON & P. C. HUGHES‚ Modal Cost Analysis for Linear Matrix Second-Order Systems‚ ASME J. of Dynamic Systems‚ Measurement‚ and Control‚ vol. 102‚ 151–158‚ Sep. 1980. D. M. WIBERG‚ State Space and Linear Systems‚ McGraw-Hill Schaum’s Outline Series in Engineering‚ 1971.

10.13 P.10.1

Problems Show that for a

the rank of the matrix

is the same as that of the controllability matrix for any P.10.2 Consider the inverted pendulum of Fig.10.1.a‚ where the input variable is the force applied to the cart. Show th at the equation of motion near is

Write the equation in state variable form using Compute the controllability matrix. [Hint: use Lagrange’s equations] P.10.3 Consider two inverted pendulums on a cart as in Fig.10.1.b. Show that the equations of motion near and are


242

P.10.4 Consider the double inverted pendulum of Fig.10.1.c. Show that the equations of motion near and are

where and are the absolute angles of the two arms‚ and P.10.5 Show that for two sets of state variables related by the non singular transformation the controllability matrices are related by

P.10.6 Show that the control canonical form (10.29) is a state space realization of the transfer function (10.28). P.10.7 Show that for the controllability matrix of the control canonical form reads

P.10.8 Demonstrate Ackermann’s formula (10.34) for SISO systems. [Hint: Proceed according to the following steps: (1) Show that (2) Using the Cayley-Hamilton theorem‚ show that (3) Show that (4) Using the result of Problem P.10.7‚ show that P.10.9 Consider the single degree of freedom oscillator of section 7.4.1. Calculate the state feedback gains leading to the characteristic equation (7.49) using Ackermann’s formula. Compare with (7.50) and (7.51). P.10.10 Show that for a system in diagonal form‚ the controllability and observability matrices are given by Equ.(10.46) and (10.48). P.10.11 The PBH rank tests state that The pair (A‚ b) is controllable iff

for all The pair

is observable iff

for all Show that these tests are equivalent to the eigenvector tests (10.47) and (10.49).


243

P.10.12 Consider an asymptotically stable linear time-invariant system. Show that (a) The steady state covariance matrix due to independent white noise inputs of unit intensity is equal to the controllability Gramian:

(b) The free response from initial conditions

satisfies

where is the observability Gramian. [Hint: the state impulse response is and the free output response from non-zero initial conditions is P.10.13 Show that for the coordinate transformation the Gramians are transformed according to

P.10.14 Show that if an internally balanced system is partitioned according to (10.87)‚ the two subsystems and are internally balanced with joint Gramians and [Hint: Partition the Liapunov equations governing and


Chapter 11

Stability 11.1

Introduction

A basic knowledge of stability of linear systems has been assumed throughout the previous chapters. Stability was associated with the location of the poles of the system in the left half plane. In chapter 7, we saw that the poles are the eigenvalues of the system matrix A when the system is written in state variable form. In chapter 8, we examined the Nyquist criterion for closedloop stability of a SISO system; we concluded on the stability of the closedloop system from the number of encirclements of – 1 by the openloop transfer function In this chapter, we examine the salient results of Liapunov’s theory of stability; it is attractive for mechanical systems because of its exceptional physical meaning and its wide ranging applicability, especially for the analysis of nonlinear systems, and also in controller design. We will conclude this chapter with a class of collocated controls that are especially useful in practice, because of their guaranteed stability, even for nonlinear systems; we will call them energy absorbing controls. The following discussion will be restricted to time invariant systems (also called autonomous), but most of the results can be extended to time varying systems. As in the previous chapters, most of the general results are stated without demonstration and the discussion is focussed on vibrating mechanical systems; a deeper discussion can be found in the references. Consider a time-invariant system, linear or not The equilibrium state there is some

is stable in the sense of Liapunov if, for every (depending on such that, if then

In this statement, stands for a norm, measuring the distance to the equilibrium; the Euclidean norm is defined as . States which are not 245


246

stable in the sense of Liapunov are unstable. Stability is a local property; if it is independent of the size of the initial perturbation it is global. The equilibrium state is asymptotically stable if it is stable in the sense of Liapunov and if, for any close to 0, as Thus, for a mechanical system, asymptotic stability implies some damping, unlike Liapunov stability. For a linear time-invariant system, since asymptotic stability is always global; nonlinear systems exhibit more complicated behaviours and they can have more than one equilibrium point (Problem P. 11.1). The stability of an equilibrium point is related to the behaviour of the free trajectories starting in its neighbourhood; if all the trajectories eventually converge towards the equilibrium point, it is asymptotically stable; if the trajectories converge towards a limit cycle, the system is unstable (Problem P.11.2). The above definitions of internal stability refer to the free response from nonzero initial conditions. In some cases, we are more interested in the input-output response: A system is externally stable if every bounded input produces a bounded output. For obvious reasons, this is also called BIBO stability. External stability has no relation to internal (zero-input) stability in general, except for linear time-invariant systems, where it is equivalent to asymptotic stability (if the system is both controllable and observable).

11.1.1

Phase portrait

As we have already mentioned, the stability of an equilibrium point is related to the behaviour of the trajectories in its vicinity. If we can always find a small domain containing the equilibrium point, such that all trajectories starting within this domain remain arbitrarily close to the origin, the equilibrium is stable; if all trajectories starting in a small domain eventually converge towards the origin, the equilibrium is asymptotically stable, and if this occurs for any initial condition, we have global asymptotic stability. The complete set of trajectories is called the phase portrait; to visualize it, consider the second order system

Defining the state variables and we can easily represent the trajectories in the phase plane various situations are considered in Fig. 11.1. Figure 11.1.a corresponds to a linear oscillator with viscous damping; all the trajectories consist of spirals converging towards the origin (the decay rate is governed by the damping); the system is globally asymptotically stable. Figure 11.1.b shows the phase portrait of an unstable linear system (poles at -2 and +1); all the trajectories are unbounded. The situation depicted on

Stability

247

Fig.11.1.c is that of a Van der Pol oscillator (Problem P.11.2), all the trajectories converge towards a limit cycle; the system is unstable, although all the trajectories are bounded.

11.2

Linear systems

Since the stability of a system is independent of the state space coordinates, it is convenient to consider the diagonal form (10.44), where is the diagonal matrix with the eigenvalues of A. The free response from non-zero initial conditions reads

Each state coordinate follows an exponential The system is stable in the sense of Liapunov if If (strictly negative), the system is globally asymptotically stable (and also externally stable).


248

If the characteristic equation is available in the form

it is not necessary to compute all the eigenvalues to assess the asymptotic stability of the system; this can be done directly from the coefficients of the characteristic polynomial by the Routh-Hurwitz criterion.

11.2.1

Routh-Hurwitz criterion

Assume that the characteristic polynomial is written in the form (11.3) with all coefficients a necessary and sufficient condition for all the roots to have negative real parts is that all the determinants defined below must be positive. The determinants are constructed as follows 1. Form the array:

where

are the coefficients of the characteristic polynomial, and

2. Compute the determinants:

All the eigenvalues have negative real parts iff for all If not all the coefficients are positive, that is if for some it is not necessary to compute all the determinants to conclude that the system is not asymptotically stable (it may still be stable in the sense of Liapunov if some

Stability

249

11.3 Liapunov’s direct method 11.3.1

Introductory example

Consider the linear oscillator

We know that it is asymptotically stable for positive damping ; its phase portrait is represented in Fig. 11.1.a; for any disturbed state the system returns to the equilibrium The total energy of the system is the sum of the kinetic energy of the mass and the strain energy in the spring:

is positive definite because it satisfies the two conditions

The time derivative of the total energy during the free response is

and, upon substituting

and

from Equ.(11.5),

We see that is always negative for a structure with positive damping. Since E is positive and decreases along all trajectories, it must eventually go to E = 0 which, from (11.6), corresponds to the equilibrium state This implies that the system is asymptotically stable. Here, we have proved asymptotic stability by showing that the total energy decreases along all trajectories; Liapunov’s direct method (also called second method, for chronological reasons), generalizes this concept. Unlike other techniques (Eigenvalues, Routh-Hurwitz, Nyquist,...), the method is also applicable to nonlinear and time varying systems. In what follows, we shall restrict our attention to time-invariant systems for which the theorems have a simpler form; more general results can be found in the literature (e.g. Vidyasagar).

11.3.2

Stability theorem

A time-invariant Liapunov function candidate is a continuously differentiable, locally positive definite function, i.e. satisfying


250

where is a certain domain containing the origin. Theorem: Consider a system governed by the vector differential equation

such that The equilibrium state is stable (in the sense of Liapunov) if one can find a Liapunov function candidate such that

for all trajectories in the neighbourhood of the origin. If condition (11.11) is satisfied, is called a Liapunov function for the system (11.10). The Liapunov function is a generalization of the total energy of the linear oscillator considered in the introductory example. The foregoing theorem is only a sufficient condition; the fact that no Liapunov function can be found does not mean that the system is not stable. There is no general procedure for constructing a Liapunov function, and this is the main weakness of the method. As an example, consider the simple pendulum, governed by the equation

where is the length of the pendulum, gravity. Introducing the state variables

the angle and and

the acceleration of we rewrite it

Let us again use the total energy (kinetic plus potential) as Liapunov function candidate: It is indeed positive definite in the vicinity of

and, substituting and along the trajectories:

We have

from Equ.(11.13), we obtain the time derivative

which simply expresses the conservation of energy. Thus, satisfies condition (11.11), is a Liapunov function for the pendulum and the equilibrium point is stable. We now examine a stronger statement for asymptotic stability.

Stability

11.3.3

251

Asymptotic stability theorem

Theorem: The state is asymptotically stable if one can find a continuously differentiable, positive definite function such that

for all trajectories in some neighbourhood of the origin. Besides, if is such that there exists a nondecreasing scalar function of the distance to the equilibrium (Fig.11.2), such that and

then the system is globally asymptotically stable under condition (11.15).

11.3.4 Lasalle’s theorem Going back to the linear oscillator of section 11.3.1, we see that Equ.(11.8) does not comply with the requirement (11.15) to be stricly negative; indeed, whenever even if (i.e. whenever the trajectory crosses the axis in Fig.11.1.a). The answer to that is given by Lasalle’s theorem, which extends the asymptotic stability even if is not strictly negative. Theorem: The state is asymptotically stable if one can find a differentiable positive definite function such that

for all trajectories, provided that the set of points where

contains no trajectories other than the trivial one


252

As an example, consider the nonlinear spring with friction governed by

where is the nonlinear restoring force and that and are continuous functions such that

is the friction. We assume

and are entirely contained in the first and third quadrant). It is easy to see that the linear oscillator is the particular case with

and that the simple pendulum corresponds to

The total energy is taken as Liapunov function candidate

where the first term is the kinetic energy, and the second one, the potential energy stored in the spring. The time derivative is

Since the set of points where is globally asymptotically stable.

11.3.5

does not contain trajectories, the system

Geometric interpretation

To visualize the concept, it is useful to consider, once again, a second order system for which the phase space is a plane. In this case, can be visualized by its contours (Fig. 11.3). The stability is associated with the behaviour of the trajectories with respect to the contours of V. If we can find a locally positive definite function such that all the trajectories cross the contours downwards (curve 1), the system is asymptotically stable; if some trajectories follow the contours, the system is stable in the sense of Liapunov (curve 2). The trajectories crossing the contours upwards (curve 3) correspond to instability, as we now examine.

253

Stability

11.3.6

Instability theorem

In the previous sections, we examined sufficient conditions for stability. We now consider a sufficient condition for instability. Let us start with the well known example of the Van der Pol oscillator

Taking the Liapunov function candidate

we have We see that, whenever Thus, applies everywhere in a small set containing the origin; this allows us to conclude that the system is unstable. In this example, is positive definite; in fact, instability can be concluded with a weaker statement: Theorem: If there exists a function continuously differentiable such that along every trajectory and for arbitrarily small values of the equilibrium is unstable. It can be further generalized as follows: Theorem: If there is a continuously differentiable function such that in an arbitrary small neighbourhood of the origin, there is a region where V > 0 and V = 0 on its boundaries; (ii) at all points of along every


254

trajectory and (iii) the origin is on the boundary of then, the system is unstable. The visual interpretation is shown in Fig. 11.4: A trajectory starting at within will intersect the contours in the direction of increasing values of V, increasing the distance to the origin; it will never cross the lines OA and OB because this would require

11.4

Liapunov functions for linear systems

Consider the linear time-invariant system

We select the Liapunov function candidate

where the matrix P is symmetric positive definite. Its time derivative is

if P and Q satisfy the matrix equation

This is the Liapunov equation, that we already met in chapter 9. Thus, if we can find a pair of positive definite matrices P and satisfying Equ.(11.30), both V and are positive definite functions and the system is asymptotically stable. Theorem: The following statements are equivalent for expressing asymptotic stability:

255

Stability

1. All the eigenvalues of A have negative real parts. 2. For some positive definite matrix Q, the Liapunov equation has a unique solution P which is positive definite.

3. For every positive definite matrix Q, the Liapunov equation has a unique solution P which is positive definite. Note that, in view of Lasalle’s theorem, Q can be semi-positive definite, provided that on all nontrivial trajectories. The foregoing theorem states that if the system is asymptotically stable, for every one can find a solution P > 0 to the Liapunov equation. Note that the converse statement (for every P > 0, the corresponding Q is positive definite) is, in general, not true; this means that not every Liapunov candidate is a Liapunov function. The existence of a positive definite solution of the Liapunov equation can be compared with the Routh-Hurwitz criterion, which allows us to determine whether or not all the eigenvalues of A have negative real parts without computing them.

11.5 Liapunov’s indirect method This method (also known as the first method), allows us to draw conclusions about the local stability of a nonlinear system from the analysis of its linearization about the equilibrium point. Consider the time invariant nonlinear system

Assume that is continuously differentiable and that so that is an equilibrium point of the system. The Taylor’s series expansion of near reads

where Taking into account that and neglecting the second order term, we obtain the linearization around the equilibrium point

where A denotes the Jacobian matrix of

at

Liapunov’s indirect method assesses the local stability of the nonlinear system (11.31) from the eigenvalues of its linearization (11.33). Theorem: The nonlinear system (11.31) is asymptotically stable if the eigenvalues of A have negative real parts. Conversely, the nonlinear system is unstable if


256

at least one eigenvalue of A has a positive real part. The method is inconclusive if some eigenvalues of A are purely imaginary. We shall restrict ourselves to the proof of the first part of the theorem. Assume that all the eigenvalues of A have negative real parts; then, we can find a symmetric positive definite matrix P solution of the Liapunov equation

Using have

as Liapunov function candidate for the nonlinear system, we

Using the Taylor’s series expansion

we find

Taking into account Equ.(11.35) and the fact that

, we obtain

Sufficiently near is dominated by the quadratic term which is negative; is therefore a Liapunov function for the system (11.31) which is asymptotically stable. We emphasize the fact that the conclusions based on the linearization are of a purely local nature; the global asymptotic stability of the nonlinear system can only be established by finding a global Liapunov function.

11.6

An application to controller design

Consider the asymptotically stable linear system

with the scalar input

subject to the saturation constraint

If P is solution of the Liapunov equation

with

is a Liapunov function of the system without control . With control, we have

Any control

257

Stability

where the scalar function is such that because The following choice of makes

will stabilize the system, as negative as possible:

This discontinuous control is often called bang-bang; it is likely to produce chattering near the equilibrium. The discontinuity can be removed by

where the saturation function is defined as

However, the practical implementation of this controller requires the knowledge of the full state which is usually not available in practice; asymptotic stability is no longer guaranteed if is reconstructed from a state observer.

11.7

Energy absorbing controls

Consider a vibrating mechanical system, linear or not, stable in open-loop, and such that the conservation of the total energy (kinetic + potential) applies when the damping is neglected. Because there is always some natural damping in practice, the total energy E of the system actually decreases with time during its free response, Suppose that we add a control system using a collocated actuator/sensor pair; if we denote by the power dissipated by damping 0) and by the power flow from the control system to the structure, we have

If we can develop a control strategy such that the power actually flows from the structure to the control system, (the control system behaves like an energy sink), the closed-loop system will be asymptotically stable. Now, consider the situation depicted in Fig.11.5.a, where we use a point force actuator and a collocated velocity sensor. If a velocity feedback is used,

with

we have

The stability of this Direct Velocity Feedback was already pointed out for linear system, in section 5.2. Here, it is generalized to nonlinear structures. Even more generally, any nonlinear control


258

where is such that will be stabilizing. Next, consider the dual situation (Fig. 11.5.b) where the actuator controls the relative position of two points inside the structure, and the sensor output is the dynamic force T in the active member (T is collocated with ; this situation is that of the active truss considered in section 3.7. Again, the power flow into the structure is

It is easy to verify that the positive Integral Force Feedback

with

will be stabilizing, because

The stability of the control law (11.50) was established for linear structures in section 5.5, using the root locus technique. Here, we extend this result to nonlinear structures, assuming perfect actuator and sensor dynamics. In chapter 14, we will apply this control law to the active damping of cable-structure systems. Because of their global asymptotic stability for arbitrary nonlinear structures, we shall refer to the controllers (11.48) and (11.50) as energy absorbing controllers. Note that, unlike those discussed in the previous section, these controllers do not require the knowledge of the states, and are ready for implementation; the stability of the closed-loop system relies very strongly on the collocation of the sensor and the actuator. Once again, we emphasize that we have assumed perfect sensor and actuator dynamics; finite actuator and sensor dynamics always have a detrimental effect on stability.

Stability

11.8

259

References

L. P. GRAYSON, The Status of Synthesis Using Lyapunov’s Method, Automatica, vol. 3, 91–121, 1965. E. A. GUILLEMIN, The Mathematics of Circuit Analysis, Wiley, 1949. R. E. KALMAN & J. E. BERTRAM, Control System Analysis and Design Via the Second Method of Lyapunov (1. Continuous-Time Systems), ASME J. of Basic Engineering, pp. 371-393, June 1960. L. MEIROVITCH, Methods of Analytical Dynamics, McGraw-Hill, 1970. A. H. NAYFEH & D. T. MOOK, Nonlinear Oscillations, Wiley, 1979. P. C. PARKS & V. HAHN, Stability Theory, Prentice Hall, 1993. M. VIDYASAGAR, Nonlinear Systems Analysis, Prentice-Hall, 1978. D. M. WIBERG, State Space and Linear Systems, McGraw-Hill Schaum’s Outline Series in Engineering, 1971.

11.9

Problems

P.11.1

Show that the nonlinear oscillator

with P.11.2

has three equilibrium points. Check them for stability. Consider the Van der Pol oscillator

with Show that the trajectories converge towards a limit cycle (Fig.11.1.c) and that the system is unstable. P. 11.3 Plot the phase portrait of the simple pendulum

P.11.4 Show that a linear system is externally (BIBO) stable if its impulse response satisfies the following inequality

for all


260

P.11.5 Show that a linear time-invariant system is asymptotically stable if its characteristic polynomial can be expanded into elementary polynomials and with all the coefficients positive. P.11.6 Examine the asymptotic stability of the systems with the following characteristic polynomials: (i) (ii) (iii) P.11.7 Examine the stability of the Rayleigh equation

with the direct method of Liapunov. P.11.8 Examine the stability of the following equations:

P.11.9

(a) Show that, if A is asymptotically stable,

where M is a real symmetric matrix, satisfies the matrix differential equation

(b) Show that the steady state value

satisfies the Liapunov equation

P.11.10 Consider the free response of the asymptotically stable system from the initial state Show that, for any the quadratic integral

is equal to where P is the solution of the Liapunov equation

Stability

P.11.11

261

Consider the linear time invariant system

Assume that the pair (A, B) is controllable and that the state feedback has been obtained according to the LQR methodology:

where P is the positive definite solution of the Riccati equation

with and Prove that the closed-loop system is asymptotically stable by showing that is a Liapunov function for the closed-loop system. Note: From section 9.3, we readily see that is in fact the remaining cost to equilibrium:

P.11.12

Consider the bilinear single-input system

where A is asymptotically stable (the system is linear in and in but it is not jointly linear in and because of the presence of the bilinear matrix N). Show that the closed-loop system is globally asymptotically stable for the nonlinear state feedback where P is the solution of the Liapunov equation

P.11.13

Consider the free response of a damped vibrating system

The total energy is

(a) Show that its decay rate is

(b) Show that if we normalize the mode shapes according to use the state space representation (7.14), the total energy reads

and if we

262


P.11.14 Consider a linear structure with a point force actuator collocated with a velocity sensor. Using the state space representation (7.14) and taking the total energy as Liapunov function, show that the controller (11.41) is equivalent to (11.48).

Chapter 12

Semi-active control 12.1

Introduction

Active control systems rely entirely on external power to operate the actuators and supply the control forces. In many applications, such systems require a large power source, which makes them vulnerable to power failure. This is why the civil engineering community is reluctant to use active control devices for earthquake protection. Semi-active devices require a lot less energy than active devices; and the energy can often be stored locally, in a battery, thus rendering the semi-active device independent of any external power supply. Another critical issue with active control is the stability robustness with respect to sensor failure; this problem is especially difficult when centralized controllers are used. On the contrary, semi-active control devices are essentially passive devices where properties (stiffness, damping,...) can be adjusted in real time, but they cannot input energy directly in the system being controlled. Semi-active control devices have the capability to vary the resistance law in some way in order to achieve a strong dependence of the control law on the relative velocity. The variable resistance law can be achieved in a wide variety of forms, as for example position controlled valves, rheological fluids, or piezoelectrically actuated friction joints. Magneto-rheological fluids exhibit very fast switching (of the order of millisecond) with a substantial yield strength; this makes them excellent contenders for semi-active devices, particularly for small and medium-size devices, and justifies their extensive discussion in this chapter. It is likely that semi-active control devices will find their way in many vibration control applications, for large and medium amplitudes (vehicle suspension, earthquake, washing machines,...). However, it should be kept in mind that, in most cases, semi-active devices are designed to operate in the ”post yield” region, when the stress exceeds some controllable threshold; this makes them 263

264


inappropriate for vibrations of small amplitude where the stress remains below the minimum controllable threshold in the device. It should also be pointed out that, in many applications (e.g. the washing machine), the cost of the control system is a critical issue (it is much more important than the optimality of the performances); this often leads to simplified control architectures with extremely simple sensing devices. This chapter begins with a review of magneto-rheological (MR) fluids and a brief overview of their applications to date. Next, some semi-active control strategies are discussed.

12.2

Magneto-rheological (MR) fluids

In 1947, W.Winslow observed a large rheological effect (apparent change of viscosity) induced by the application of an electric field to colloidal fluids (insulating oil) containing micron-sized particles; such fluids are called electrorheological (ER) fluids. The discovery of MR fluid was made in 1951 by J.Rabinow, who observed similar rheological effects by application of a magnetic field to a fluid containing magnetizable particles. In both cases, the particles create columnar structures parallel to the applied field (Fig.12.1) and these chain-like structures restrict the flow of the fluid, requiring a minimum shear stress for the flow to be initiated. This phenomenon is reversible, very fast (response time of the order of millisecond) and consumes very little energy. When no field is applied, the rheological fluids exhibit a Newtonian behaviour. Typical values of the maximum achievable yield strength are given in Table 12.1. ER fluids performances are generally limited by the electric field breakdown strength of the fluid while MR fluids performances are limited by the magnetic saturation of the particles. Iron particles have the highest saturation magnetisation. In Table 12.1, we note that the yield stress of MR fluids is 20 to 50 times larger than that of ER fluids. This justifies why most practical applications use MR fluids. Typical particle sizes are 0.1 to and typical particle fractions are between 0.1 and 0.5; the carrier fluids are selected on

Semi-active control

265

the basis of their tribology properties and thermal stability; they also include additives that inhibit sedimentation and aggregation.

The behaviour of MR fluids is often represented as a Bingham plastic model with a variable yield strength depending on the applied magnetic field H (Fig.12.2.a and 12.2.b). The flow is governed by the equation


266

where is the shear stress, is the shear strain and is the viscosity of the fluid. The operating range is the shaded area in Fig.12.2.c. Below the yield stress (at strains of order the material behaves viscoelastically:

where G is the complex material modulus. This model is also a good approximation for MR devices (with appropriate definitions for and ). However, the actual behaviour is more complicated and includes stiction and hysteresis such as shown in Fig.12.2.d; more elaborate models attempting to account for the hysteresis are available in the literature.

12.3

MR devices

Figure 12.3 shows the three operating modes of controllable fluids: valve mode, direct shear mode and squeeze mode; the valve mode is the normal operating mode of MR dampers and shock absorbers (Fig.12.4); the direct shear mode is that of clutches and brakes.

Semi-active control

267

The cost of the MR fluid contributes significantly to the total cost of a MR device. In order to bring this cost down by reducing the amount of fluid encapsulated, foam devices have been introduced, where the MR fluid is constrained in an absorbent matrix by capillarity, without seal. Lord corporation has developed such a variable friction device for application in washing machines; it uses only 3ml of MR fluid and produces a friction force up to 100N; the target price of this damper is of the order of 1 to 3 USD.

12.4

Semi-active control

Semi-active devices are essentially passive and cannot generate arbitrary control forces; they are restricted to control forces opposing the relative velocity which are exclusively dissipative With this restriction, the control element allows one to vary the control force arbitrarily within the operating range. Figure 12.5 shows the operating range of a classical viscous damper with a variable damping coefficient, and that of a MR device; in these plots, represents the force applied to the device, equal and opposite to that applied to the structure. In the variable damper of Fig.l2.5.a, the control variable is typically the voltage applied to an electromagnet which commands the opening of the valve between the two chambers of the damper, thus controlling the damping of the device [details regarding the design of semi-active dampers to achieve fast switching are discussed by Karnopp]. The operating range is the shaded area between the two lines of minimum and maximum damping coefficients. For the MR device of Fig.l2.5.b, the control variable is the current through the coil generating the magnetic field which commands the variable yield force of the device; the operating range shown in the figure assumes a Bingham model for the device.


268

12.5

Open-loop control

In this mode of operation, no feedback is used, the variable damper works in a on/off configuration. A desired level of attenuation is obtained by selecting the control variable according to the operating conditions of the system. As an example, consider the passive isolator problem of section 6.2 during the acceleration or the deceleration of a rotating machinery. When the rotation velocity is below and close to the natural frequency of the suspension, it is beneficial to have a large damping value (Fig.6.2) and the damping coefficient is set to its maximum value. On the contrary, when the rotation velocity is above the natural frequency of the suspension, a better isolation is achieved with a lower damping coefficient and the control variable is set to zero, leading to the minimum damping coefficient.

12.6

Feedback control

12.6.1

Continuous control

We will use an active suspension with a sky-hook damper as an example to illustrate various feedback implementations of the semi-active control with a

Semi-active control

269

variable damping device. The active sky-hook damper was analysed in chapter 6; the virtual sky-hook damper connects the inertial frame (sky) to the sprung mass M while the semi-active damper is placed between the sprung and the unsprung mass (Fig. 12.6). The desired control force is but the semiactive device produces a control force it is able to generate this force only if the sign of is the same as that of and if the magnitude of the desired force is within the operating range of the damper, that is if

where and are the lower and upper bounds of the damping coefficient achievable with the device. When and have opposite signs, the device can only supply a force with a sign opposite to the desired force and the best it can do is to set the damping coefficient to its minimum value. The variable damping coefficient which best matches the demanded control force is given by

where 1(.) is the Heaviside step function (which is equal to 1 when its argument is positive, and equal to zero when its argument is negative or zero). Note that the foregoing control law involves a fast switching to whenever the relative velocity switches signs; it also assumes the capability of controlling the

270


Semi-active control

271

magnitude of the damping coefficient in the device; this requires a calibrated proportional valve. Figure 12.7 show the results of a simulation of a semi-active suspension such as that of Fig.12.6 subjected to a band limited white noise disturbance load (180Nrms in 0.2H z – 10H z); the nominal values of the parameters are given in Fig.12.6. Figure 12.7.(a) to (c) show the disturbance force, the absolute velocity of the clean body and the relative velocity Figure 12.7.(d) shows the demanded control force (dashed line) and the actual control force Whenever possible, tracks and this corresponds to a damping constant within the bounds (Fig. 12.7.(e)). When the demanded force exceeds the damper capabilities or when the velocities and have opposite signs, the damping saturates at either the maximum or the minimum limit. We know from Fig.6.21 that the signal contains higher frequency components than Figure 12.8 illustrates what happens when one tries to

272


Semi-active control

273

emulate a low frequency signal with segments of the higher frequency signal The sign changes of the product produce steps and introduce higher harmonics in the actual control force which are likely to produce a parasitic excitation of the sprung mass, if it is not strictly rigid. The semi-active isolation strategy based on emulating the sky-hook damper works well for narrow-band disturbances, but tends to be less effective when subjected to wide-band disturbances.

12.6.2

On-off control

Fast responding servo valves are expensive; an alternative on/off strategy consists of switching the control variable between 0 and its maximum value, thus switching the damping coefficient from to according to the sign of This provides a maximum damping whenever and have the same sign and a minimum force when and have opposite signs. Figure 12.9 illustrates the foregoing control law for the same example as in Fig.12.7; we note that the time-history response of the clean body velocity is very similar to the previous case and that the rms level of control force is generally higher, but, unlike the previous one, this implementation requires only a on/off valve.


274

12.6.3

Force feedback

A more general strategy which applies to any type of semi-active device is described in Fig.12.10. First, an active control algorithm computes a desired control force (any of the active damping strategies discussed in the foregoing chapters can be used at this stage). Next, a secondary clipping controller tries to make the semi-active device to replicate the force that the active device would apply on the structure. This requires the measurement of the actual force applied to the structure. Once again, there are two options at this stage: on/off control and continuous control. The on/off clipping strategy is the simplest (Fig.l2.10.b): when the two forces have the same sign and the magnitude of the actual force is smaller than the magnitude of the desired control force the control variable is set to the maximum otherwise the control variable is set to 0.

This strategy is often called ” clipped on/off ” .

As an alternative to the foregoing on/off clipping control, a continuous controller can be implemented as shown in Fig.12.11. The control variable of the semi-active element is now the output of a proportional controller which tries to minimize the error between the actual control force and the desired one The bias input aims at increasing the working band where a proportional feedback is achieved; if and the proportional gain is taken very large, this control strategy is equivalent to the clipped on/off. This continuous controller applies to all kind of semi-active devices provided the magnitude of the control force in the device is an increasing function of the control variable

Semi-active control

275

12.7 References J. D. CARLSON & M. R. JOLLY, MR Fluid, Foam and Elastomer Devices, Mechatronics, vol. 10, 555–569, 2000. J. D. CARLSON & J. L. SPROSTON, Controllable Fluids in 2000 - Status of ER and MR Fluid Technology, Proc. of Actuator 2000, Bremen, pp. 126–130, June 2000. J. P. COULTER, K. D. WEISS & J. D. CARLSON, Engineering Applications of Electrorheological Materials, Journal of Intelligent Material Systems and Structures, vol. 4, 248–259, April 1993. L. GAUL & R. NITSCHE, Friction Control for Vibration Suppression, (DETC99/ VIB-8191) Movie Symposium, ASME Design Engineering Technical Conferences, Las Vegas, 1999. D. KARNOPP, Design Principles for Vibration Control Systems Using SemiActive Dampers, Trans. ASME Journal of Dynamic Systems, Measurement and Control, vol. 112, 448–455, Sep. 1990. D. KARNOPP, M. CROSBY & R. A. HARWOOD, Vibration Control Using SemiActive Suspension Control, Journal of Engineering for Industry, vol. 96, 619– 626, 1974. N. D. SIMS, R. STANWAY & A. R. JOHNSON, Vibration Control Using SMART Fluids: A State of the Art Review, The Shock and Vibration Digest, vol. 31, no. 3, 195–203, May 1999. B. F. SPENCER, S. J. DYKE, M. K. SAIN & J. D. CARLSON, Phenomenological Model of a Magnetorheological Damper, ASCE Journal of Engineering Mechanics, vol. 123, no. 3, 230–238, 1997.

12.8

Problems

P.12.1 Consider a MR device operating according to the direct shear mode (Fig.12.3.b); the electrodes move with respect to each other with a relative velocity U. If A is the area of the electrode, the distance between the electrodes, the viscous components and the field-induced yield stress components are respectively

From these equations, show that the minimum volume of active fluid to achieve a given control ratio for a specified maximum controlled force and a

276


maximum relative velocity U reads

where is the viscosity and is the maximum yield strength induced by the magnetic field. From this result, can be regarded as a figure of merit of a controllable fluid (Coulter et al). This explains the superiority of the MR fluids over ER fluids (see Table 12.1).

Chapter 13

Applications After a brief overview of some critical aspects of digital control, this chapter applies the concepts developed in the foregoing chapters to a few applications; it is based on the work done at the Active Structures Laboratory of ULB over the past ten years. The active tendon control of cable structures will be considered in the next chapter.

13.1

Digital implementation

In recent years, low cost microprocessors have become widely available, and digital has tended to replace analog implementation. There are many reasons for this: digital controllers are more flexible (it is easy to change the coefficients of a programmable digital filter), they have good accuracy and a far better stability than analog devices which are prone to drift due to temperature and ageing. Digital controllers are available with several hardware architectures, including microcontrollers, PC boards, and digital signal processors (DSP). It appears that digital signal processors are especially efficient for structural control applications. Although most controller implementation is digital, current microprocessors are so fast that it is always more convenient, and sometimes wise, to perform a continuous design of the compensator and transform it into a digital controller as a second step, once a good continuous design has been achieved. This does not mean that the control designer may ignore digital control theory, because even though the conversion from continuous to digital is greatly facilitated by software tools for computer aided control engineering, there are a number of fundamental issues that have to be considered with care; they will be briefly mentioned below. For a deeper discussion, the reader may refer to the literature on digital control (e.g. Åström & Wittenmark; Franklin & Powell). 277

278

13.1.1


Sampling, aliasing and prefiltering

Since digital controllers operate on values of the process variables at discrete times, it is important to know under what conditions a continuous signal can be recovered from its discrete values only. The answer to this question is given by Shannon’s theorem (also called the sampling theorem), which states that, to recover a band-limited signal with frequency content from its sampled values, it is necessary to sample at least at If a signal is sampled at any frequency component above the limit frequency will appear as a component at a frequency lower than This phenomenon is called aliasing, and the limit frequency that can be theoretically recovered from a digital signal is often called Nyquist frequency, by reference to the exploratory work of Nyquist. Aliasing is of course not acceptable and it is therefore essential to place an analog low-pass filter at a frequency before the analog to digital converter (ADC) (Fig.13.1). However analog prefilters have dynamics and, as we know from the first Bode integral, a sharp cut-off of the magnitude is always associated with a substantial phase lag at the cut-off frequency As is related to it is always a good idea to sample at a high rate and to make sure that the cut-off frequency of the prefilter is substantially higher than the crossover frequency of the control system. If the phase lag of the prefilter at crossover is significant, it is necessary to include the prefilter dynamics in the design (as a rule of thumb, the prefilter dynamics should be included in the design if the crossover frequency is higher than

Applications

279

A simple solution to prefiltering is to introduce an analog second order filter

which can be built fairly easily with an operational amplifier and a few passive components. A second order Butterworth filter corresponds to Higher order filters are obtained by cascading first and second order systems; for example, a fourth order Butterworth filter is obtained by cascading two second order filters with the same cut-off frequency and and respectively (Problem P.13.1).

13.1.2

Zero-order hold, computational delay

Sampling can be viewed as an impulse modulation converting the continuous signal into the impulse train

where T is the sampling period The construction of a process which holds the sampled values constant during a sampling period is made by passing through a zero-order hold which consists of a filter with impulse response (Fig.13.2). It is easy to show that the corresponding transfer function is (Problem P.13.2)


280

and that it introduces a linear phase lag Another effect of the sampling is the computational delay which is always present between the access to the computer through the ADC and the output of the control law at the digital to analog converter DAC. This delay depends on the way the control algorithm is implemented; it may be fixed, equal to the sampling period T, or variable, depending on the length of the computations within the sampling period. A time delay T is characterized by the transfer function it introduces a linear phase lag this effect is clearly seen in the phase diagram of Fig.3.14. Rational approximations of the exponential by all-pass functions (Padé approximants) were discussed in Problem P.8.8. The output of the DAC is also a staircase function; in some applications, it may be interesting to smooth the control output, to remove the high frequency components of the signal, which could possibly excite high frequency mechanical resonances. The use of such output filters, however, should be considered with care because they have the same detrimental effect on the phase of the control system as the prefilter at the input. In applications, it is advisable to use a sampling frequency at least 20 times, and preferably 100 times the crossover frequency of the continuous design, to preserve the behaviour of the continuous system to a reasonable degree.

13.1.3

Quantization

After prefiltering at a frequency below the Nyquist frequency the signal is passed into the ADC for sampling and conversion into a digital signal of finite word length (typically N=14 or 16 bits) representing the total range of the analog signal. Because of the finite number of quantization levels, there is always a roundoff error which represents times the full range of the signal; the quantization error can be regarded as a random noise. The signal to noise ratio is of the order provided that the signal is properly scaled to use the full range of the ADC. Near the equilibrium point, only a small part of the dynamic range is used by the signal, and the signal to noise ratio drops substantially. The quantization error is also present at the output of the DA C; the finite word length of the digital output is responsible for a finite resolution in the analog output signal; the resolution of the output is where R is the dynamic range of the output and M the number of bits of the DA C. To appreciate the limitations associated with this formula, consider a positioning problem with a range of and a DAC of 16 bits; the resolution on the output will be limited to Quantization errors may be responsible for limit cycle oscillations. Let us briefly mention that the finite word length arithmetic in the digital controller is another source of error, because finite word length operations are no longer associative or distributive, due to rounding. We shall not pursue this matter which is closer to the hardware (e.g. see Jackson).

Applications

13.1.4

281

Discretization of a continuous controller

Although all the design methods exist in discrete form, it is quite common to perform a continuous design, and to discretize it in a second step. This procedure works quite well if the sampling rate is much higher than the crossover frequency of the control system (in structural control, it is quite customary to have Assume that the compensator transfer function has been obtained in the form

For digital implementation, it must be transformed to the form of a difference equation

The corresponding

transfer function is

where is the delay operator. The coefficients and obtained from those of following Tustin’s method: related by the bilinear transform

of and

can be are

where T is the sampling period. This transformation maps the left half into the interior of the unit circle in the and the imaginary axis from into the upper half of the unit circle from to (e.g. see Franklin & Powell or Oppenheim & Schafer) Tustin’s method can be applied to multivariable systems written in state variables; for the continuous system described by

the corresponding discrete system resulting from the bilinear transform (13.7) reads

with

282


The proof is left as an exercise (Problem P.13.3).

13.2

Active damping of a truss structure

One of the earliest active damping experiments that we performed at ULB is that with the truss of Fig.3.17, built in 1989. It consists of 12 bays of each, made of steel bars of diameter; it is clamped at the bottom, and two active struts are located in the lower bay as shown in Fig.3.17 and 13.3. Similar studies were performed at other places at about the same time (Fanson et al.; Chen et al.; Peterson et al.). The distinctive feature of this work was that the active strut was built with low cost commercial components (Philips linear piezoelectric actuator and Bruel & Kjaer piezoelectric force sensor). The design was such that the length and the stiffness of the active strut almost exactly matched that of one bar; in this way, the insertion of the active element did not change the stiffness of the structure. Because of the high-pass nature of the piezoelectric sensors, only the dynamic component of the force is measured by the force sensor. Other types of active members with built-in viscous damping have recently been developed (Hyde & Anderson). The mathematical modelling of an active truss was examined in section 3.7 and the active damping with integral force feedback was investigated in section 5.5. It was found that the closed-loop poles of an active truss provided with a single active element follow the root locus defined by Equ.(5.38).

13.2.1

Actuator placement

More than any specific control law, the location of the active member is the most important factor affecting the performance of the control system. The active element should be placed where its authority over the modes it is intended to control is the largest. According to Equ.(5.36), the control authority is proportional to the fraction of modal strain energy in the active element, It follows that the active struts should be located in order to maximize in the active members for the critical vibration modes. The search for candidate locations where active struts can be placed is greatly assisted by the examination of the map of the fraction of strain energy in the structural elements, which is directly available in commercial finite element packages. Such a map is presented in Fig. 13.3; one sees that substituting the active member for the bar No 98 provides a strong control on mode (in-plane bending mode), but no control on mode 2 (out-of-plane bending mode), which is almost uncontrollable from an active member placed in bar No 98 By contrast, an active member substituted for the bar No 2 offers a reasonable control on mode 1 and excellent control on mode 2 these two locations

Applications

283

were selected in the design. The fraction of modal strain energy is well adapted to optimization techniques for actuator placement.


284

13.2.2

Implementation, experimental results

Using the bilinear transform (13.7), we can readily transform the integral control law (IFF)

into the difference equation

which we recognize as the trapezoid rule for integration. In order to avoid saturation, it is wise to slightly modify this relation according to

where is a forgetting factor slightly lower than 1. depends on the sampling frequency; it can either be tuned experimentally or obtained from a modified compensator

where the breakpoint frequency is such that to produce a phase of 90° for the first mode and above (Problem P.13.5). Note that, for a fast sampling rate, the backward difference rule

works just as well as (13.16). In our experiment, the two active members operated independently in a decentralized manner; the control law was implemented initially on a PC at Hz and later on a DSP at Figure 13.4 shows the force signal in the active members during the free response after an impulsive load, first without, and then with control. Figure 13.5 shows the frequency response between a point force applied at A along the truss and an accelerometer located at B, at the top of the truss. A damping ratio larger than 0.1 was obtained for the first two modes. Finally, it is worth pointing out that : The dynamics of the charge amplifier does not influence the result appreciably, provided that the corner frequency of the high-pass filter is significantly lower than the natural frequency of the targeted mode. In this application as in all applications involving active damping with piezo struts, no attempt was made to correct for the large hysteresis of the piezotranslator; it was found that the hysteresis does not deteriorate the closed-loop response significantly, as compared to the linear predictions.

Applications

285

286

13.3


Active damping generic interface

The active strut discussed in the previous section can be integrated into a generic 6 d.o.f. interface connecting arbitrary substructures. Such an interface is shown

Applications

287

in Fig.l3.6.a and b (the diameter of the base plates is it consists of a Stewart platform with a cubic architecture [this provides a uniform control capability and uniform stiffness in all directions, and minimizes the cross-coupling thanks to mutually orthogonal actuators (Geng & Haynes)]. However, unlike in section 6.6 where each leg consists of a single d.o.f. soft isolator, each leg consists of an active strut including a piezoelectric actuator, a force sensor and two flexible tips (Fig.3.5).

13.3.1

Active damping

The control is a decentralized IFF with the same gain for all loops. Let

be the dynamic equation of the passive structure (including the interface). According to section 3.7, the dynamics of the active structure is governed by

where the right hand side represents the equivalent piezoelectric loads : is the vector of piezoelectric extensions, is the stiffness of one strut and B is the influence matrix of the interface in global coordinates. The output consists of the six force sensor signals which are proportional to the elastic extension of the active struts

where coordinates by

is the vector of global leg extensions, related to the global

The same matrix appears in Equ.(13.20) and (13.22) because the actuators and sensors are collocated. Using a decentralized IFF with constant gain on the elastic extension,

the closed-loop characteristic equation is obtained by combining Equ.(13.20) to (13.23):

In this equation, the stiffness matrix K refers to the complete structure, including the full contribution of the Stewart platform. The open-loop poles are where are the natural frequencies of the complete structure. The open-loop zeros are the asymptotic values of the eigenvalues of Equ.(13.24) when they are solution of


288

The corresponding stiffness matrix is where the axial stiffness of the legs of the Stewart platform has been removed from K. Without bending stiffness in the legs, this matrix is singular and the transmission zeros include the rigid body modes (at of the structure where the piezo actuators have been removed. However, the flexible tips are responsible for a non-zero bending stiffness of the legs and the eigenvalues of Equ.(13.25) are located at at some distance from the origin along the imaginary axis. Upon transforming into modal coordinates, and assuming that the normal modes are normalized according to we get

where As in section 5.5.1, the matrix is, in general, fully populated; assuming it is diagonally dominant and neglecting the off diagonal terms, we can rewrite it where is the fraction of modal strain energy in the active damping interface, that is the fraction of the strain energy concentrated in the legs of the Stewart platform when the structure vibrates according to the global mode i. From the definition of the open-loop transmission zeros, we also have

and the characteristic equation (13.26) can be rewritten as a set of uncoupled equations

or

This equation is identical to Equ.(5.38) and all the results of section 5.5.1 apply. Note that, in this section, the results of section 5.5.1 have been extended to a multi-loop decentralized controller with the same gain for all loops.

13.3.2

Experiment

The test set-up is shown in Fig.l3.6.c; the interface is used as a support for the truss discussed in the previous section (used as a passive truss in this case). The six independent controllers have been implemented on a DSP board; the feedback gain is the same for all the loops. Figure 13.7 shows some typical experimental results; the time response shows the signal from one of the force

Applications

289

290


sensors of the Stewart platform when the truss is subjected to an impulse at mid height from the base, first without, and then with control. The FRFs (with and without control) are obtained between a disturbance applied to the piezoactuator in one leg and its collocated force sensor. One sees that fairly high damping ratios can be achieved for the low frequency modes but also significant damping in the high frequency modes The experimental root locus of the first two modes is shown in Fig.13.8; it is compared to the analytical prediction of Equ.(13.31). In drawing Fig.13.8, the transmission zeros are taken as the asymptotic natural frequencies of the system as

13.3.3

Pointing and position control

As a closing remark, we wish to emphasize the potential of the stiff Stewart platform described here for precision pointing and precision control. With piezoceramic actuators of 50 stroke, the overall stroke of the plateform is 90,103 and along the and directions (in the payload plate axis of Fig.6.14) and 1300, 1050 and around the and axes, respectively. Embedding active damping in a precision pointing or position control loop can be done with the HAC/LAC strategy discussed in section 13.6.

13.4

Active damping of a plate

In 1993, at the request of ESA, we developed a laboratory demonstration model of an active plate controlled by PZT piezoceramics; it was later transformed into a flight model (to be flown in a canister) by our industrial partner SPACEBEL and the experiment (named CFIE: Control-Flexibility Interaction Experiment), was successfully flown by NASA in the space shuttle in September 1995. According to the specifications, the experiment should fit into a ”GAS” canister (cylinder of diameter and high), demonstrate significant gravity effects, and use the piezoelectric technology. We settled on a very flexible steel plate of thickness hanging from a support as shown in Fig. 13.9; two additional masses were mounted, as indicated in the figure, to lower the natural frequencies of the system. The first mode is in bending and the second one is in torsion. Because of the additional masses, the structure has a significant geometric stiffness due to the gravity loads, which is responsible for a rise of the first natural frequency from Hz in zero gravity to 0.9 Hz with gravity. The finite element model of the structure in the gravity field could be updated to match the experimental results on the ground, but the in-orbit natural frequencies could only be predicted numerically and were therefore subject to uncertainties. As we know from the previous chapters, in order to achieve active damping, it is preferable to adopt a collocated actuator/sensor configuration. In principle, a strictly collocated configuration can be achieved with self-sensing actuators

Applications

291

(Dosch et al.), but from our own experience, these systems do not work well, mainly because the piezoceramic does not behave exactly like a capacitance as assumed in the self-sensing electronics. As a result, self-sensing was ruled out and we decided to adopt a nearly collocated configuration, which is quite sufficient to guarantee alternating poles and zeros at low frequencies. However, as we saw in section 3.8, nearly collocated piezoelectric shells are not trivial to model, because of the importance of the membrane strains in the input-output relationship; this project was at the origin of our work on the finite element modelling of piezoelectric shells (Piefort).

13.4.1

Control design

According to section 5.4, achieving a large active damping with a Positive Position Feedback (PPF) and strain actuator and sensor pairs relies on two conditions: obtaining a precise tuning of the controller natural frequency on the targeted mode and using an actuator/sensor configuration leading to sufficient spacing between the poles and the zeros, so that wide loops can be obtained. We will discuss the tuning issue a little later; for nearly collocated systems, the distance between the poles and the zeros depends strongly on local effects in the strain transmission. In the CFIE experiment, the control system consists of two independent control loops with actuator/sensor pairs placed as indicated in Fig.13.9; finite element calculations confirmed that the spacing between the poles and the zeros was acceptable. The controller consists of two independent PPF loops, each of them targeted at modes 1 and 2 of the structure, respectively at

292


and with gravity and and in zero gravity (predicted from finite element calculations). The compensator reads

The determination of the gains and requires some trial and error; as already mentioned, it is generally simpler to adjust the gain of the filter of higher frequency first, because the roll-off of the second order filter reduces the influence from the filter tuned on a lower frequency. Note that, although its stability is guaranteed for moderate values of and the performance of the PPF depends heavily on the tuning of the filter frequencies and on the targeted modes and It is therefore essential to predict the natural frequencies accurately. To illustrate the degradation of the performance when the controller is not tuned properly, Fig.13.10 shows the sensitivity of the performance, taken as the maximum closed-loop modal damping of the first mode, as a function of the relative error in the frequency of the PPF filter; corresponds to the optimally tuned filter, leading to a modal damping over 0.13. We see that even small tuning errors can significantly reduce the performance, and that a 20 % error makes the control system almost ineffective. This problem was particularly important in this experiment where the first natural frequency could not be checked from tests. Fig.13.11 illustrates the performance of the control system on the laboratory demonstration model; it shows the free response measured by laser of one of the additional mass after some disturbance, with and without control, when the tuning is optimal. The laboratory demonstration model shown in Fig.13.9 is very flimsy and would not withstand the environmental loads (static and dynamics) during the launch of the spacecraft; the test plate would even buckle under its own weight

Applications

293

if turned upside down. As a result, the flight model was substantially reinforced with a strong supporting structure, and a latching mechanism was introduced to hold the plate during the launch. The flight model successfully underwent the vibration tests before launch, but the charge amplifiers were destroyed because the amount of electric charge generated during the qualification tests was several orders of magnitude larger than the level expected during the in-orbit experiment; the problem was solved by changing the electronic design, to include low leakage diodes with appropriate threshold at the input of the charge amplifiers. No problem occurred during the flight.

13.5

Active damping of a stiff beam

We begin with a few words about the background in which this problem was brought to our attention in the early 90’s. Optical instruments for space applications require an accuracy on the wave front in the range of to The ultimate performance of the instruments must be evaluated on earth, before launch, in a simulated space environment. This is done on sophisticated test benches resting on huge seismically isolated slabs and placed in a thermal vacuum chamber. Because of the constraint on accuracy, the amplitude of the microvibrations must remain below or, equivalently, if the first natural frequency of the supporting structure is around 60 Hz, the acceleration must remain below This limit is fairly easy to exceed, even under such apparently harmless excitations as the noise generated by the air conditioning of the clean room. Beyond the specific problem that we have just mentioned, the damping of microvibrations is a fairly generic problem which has many appli-

294


cations in other fields of precision engineering, such as machine tools, electronic circuit lithography, etc...

13.5.1

System design

A simple active damping device has been developed, based on the following premises: (i) The control system should use an accelerometer which is more appropriate than a displacement or a velocity sensor for this problem (an acceleration of can be measured with a commercial accelerometer, while a displacement of requires a sophisticated laser interferometer). (ii) The structures considered here are fairly stiff and well suited to the use of a proofmass actuator without excessive stroke (section 3.2). (iii) The sensor and the actuator should be collocated, in order to benefit from guaranteed stability. The test structure is represented in Fig. 13.12; it consists of a steel beam of mounted on three supports located at the nodes of the second free-free mode, to minimize the natural damping. The first natural frequency of the beam is The proof-mass actuator consists of a standard electrodynamic

Applications

295

shaker (Bruel & Kjaer 4810) fitted with an extra mass of to lower its natural frequency to about 20 Hz. In this way, the amplitude diagram of the frequency response is nearly constant for indicating that the proof-mass actuator behaves nearly as an ideal force generator (section 3.2). The phase diagram is also nearly flat above 40 Hz, but contains a linear phase due to the digital controller, and also some phase lag due to the inductance of the shaker, which was neglected to obtain Equ.(3.8). For a sampling frequency of 10000 Hz, the phase lag due to sampling does not exceed 3.6° at 100 Hz and, for the particular case of the B&K electrodynamic shaker, the impedance of the coil [Equ.(3.3)] reads

where the breakpoint frequency is 6300 Hz, leading to a phase lag of about 1° at 100 Hz and 10° at l000Hz. According to Table 5.1, the control law can be either leading to a Direct Velocity Feedback, or the set of second order filters (5.15) that we called acceleration feedback in section 5.3. Both have guaranteed stability (assuming perfect actuator and sensor dynamics). In choosing between the two alternatives, we must take the following aspects into account: (i) Since the transfer function of the structure does not have any roll-off, the roll-off of the open-loop system is entirely controlled by the compensator. (ii) The Direct Velocity Feedback is wide band, while the acceleration feedback, based on second order filters, must be tuned on the targeted modes. (iii) In theory, the phase margin of the Direct Velocity Feedback is 90° for all modes, but its roll-off is only The acceleration feedback has a roll-off of but the phase margin gradually vanishes for the modes which are above the frequency appearing in the filter of the compensator (Problem P. 13.6). Based on the foregoing facts and depending on the structure considered, one alternative may be more effective than the other in not destabilizing the high frequency dynamics, which is more sensitive to the finite dynamics of the actuator and sensor, delays, etc... For the test structure of Fig. 13.12, which is fairly simple and does not involve closely spaced modes, both compensators have been found very effective; the damping ratio of the first mode has been increased from to

13.6

The HAC/LAC strategy

In active structures for precision engineering applications, the control system is used to reduce the effect of transient and steady state disturbances on the controlled variables. Active damping is very effective in reducing the settling time of transient disturbances and the effect of steady state disturbances near the resonance frequencies of the system; however, away from the resonances, the active damping is completely ineffective and leaves the closed-loop response

296


essentially unchanged. Such low gain controllers are often called Low Authority Controllers (LAC), because they modify the poles of the system only slightly (Aubrun). To attenuate wide-band disturbances, the controller needs larger gains, in order to cause more substantial modifications to the poles of the open-loop system; this is the reason why they are often called High Authority Controllers (HAC). Their design requires a model of the structure and, as we saw in chapter 8, there is a trade-off between the conflicting requirements of performance-bandwidth and stability in the face of parametric uncertainty and unmodelled dynamics. The parametric uncertainty results from a lack of knowledge of the structure (which could be reduced by identification) or from changing environmental conditions, such as the exposure of a spacecraft to the sun. Unmodelled dynamics include all the high frequency modes which cannot be predicted properly, but are candidates for spillover instability. When collocated actuator/sensor pairs can be used, stability can be achieved using positivity concepts (Benhabib et al.), but in many situations, collocated pairs are not feasible for HAC; we know from chapter 4 that such configurations do not have a fixed pole-zero pattern and are much more sensitive to parametric uncertainty. LQG controllers are an example of HAC; their lack of robustness with respect to the parametric uncertainty was pointed out in section 7.10. The situation is even worse for the unmodelled dynamics, particularly for very flexible structures which have a high modal density, because there are always flexible modes near the crossover frequency. Without frequency shaping, LQG methods often require an accurate modelling for approximately two decades beyond the bandwidth of the closed-loop system, which is unrealistic in most practical situations.

Applications

297

The HAC/LAC approach originated at Lockheed in the early 80’s; it consists of combining the two approaches in a dual loop control as shown in Fig. 13.13. The inner loop uses a set of collocated actuator/sensor pairs for decentralized active damping with guaranteed stability; the outer loop consists of a non-collocated HAC based on a model of the actively damped structure. This approach ha s the following advantages: The active damping extends outside the bandwidth of the HA C and reduces the settling time of the modes which are outside the bandwidth. The active damping makes it easier to gain-stabilize the modes outside the bandwidth of the outer loop (improved gain margin). The larger damping of the modes within the controller bandwidth makes them more robust to the parametric uncertainty (improved phase margin). Singular value robustness measures generalize the phase and gain margin for MIMO systems (e.g. Kosut et al., or Mukhopadhyay & Newsom).

13.6.1

Wide-band position control

In order to illustrate the HAC/LAC strategy for a non-collocated system,let us consider once again the active truss of Fig.3.17. The objective is to design a wide-band controller using one of the piezo actuators (#1) to control the tip displacement along one coordinate axis (Fig. 13.14), measured by a laser interferometer. The compensator should have some integral action at low frequency, to compensate the thermal perturbations and avoid steady state errors; the targeted bandwidth of includes the first two vibration modes. Note that the actuator and the displacement sensor are located at opposite ends of the structure, so that the actuator action cannot be transmitted to the sensor without exciting the entire truss. The LAC consists of the active damping discussed in section 13.2; the transfer function between the input voltage of the actuator and the tip displacement is shown in Fig.13.15 for various values of the gain of the active damping. One observes that the active damping works very much like passive damping, affecting only the frequency range near the natural frequencies. Below the behaviour of the system is dominated by the first mode; the second mode does not substantially affect the amplitude of and the phase lag associated with the second mode is compensated by the phase lead of a zero at a frequency slightly lower than (although not shown, the general shape of the phase diagram can be easily drawn from the amplitude plot). From these observations, we conclude that mode 2, which is close to mode 1, will be phase-stabilized with mode 1 and, as a result, the compensator design can be based on a model including a single vibration mode; the active damping can be closely approximated by passive damping. Thus, the compensator design is based on the very simple model of a damped oscillator.

298


Applications

299

13.6.2 Compensator design The compensator should be designed to achieve integral action at low frequency and to have enough roll off at high frequency to avoid spillover instability. The standard LQG is not well suited to these requirements, because the quadratic performance index puts an equal weight on all frequencies; the design objectives require larger weights on the control at high frequency to avoid spillover, and larger weights on the states at low frequency to achieve integral action; both can be achieved by the frequency-shaped LQG as explained in section 9.16. The penalty on the high frequency components of the control is obtained by passing the control through a low-pass filter (selected as a second order Butterworth filter in this case) and the P+I action is achieved by passing the output (which is also the control variable through a first order system as indicated in Fig.9.9. The state feedback is obtained by solving the LQR problem for the augmented system with the quadratic perfomance index

The structure of the compensator is that of Fig.9.10; the frequency distribution of the weights for the original problem is shown in Fig 9.8; the large weights on the states at low frequency correspond to the integral action, and the large penalty on the control at high frequency aims at reducing the spillover. The states of the structure (only two in this case) must be reconstructed with an observer; in this case, a Kalman filter is used; the noise intensity matrices have been selected to achieve the appropriate dynamics.

13.6.3

Results

The Bode plots of the compensator are shown in Fig.13.16; it behaves like an integrator at low frequency, provides some phase lead near the flexible mode and crossover, and roll-off at high frequency. The open-loop transfer function of the control system, GH, is shown in Fig.13.17 (G corresponds to the model); the bandwidth is and the phase margin is PM = 38°. The effect of this compensator on the actual structure can be assessed from Fig.13.18. As expected, the second flexible mode is phase stabilized and does not cause any trouble. On the other hand, we observe several peaks corresponding to higher frequency modes in the roll-off region; some of these peaks exceed and their stability must be assessed from the Nyquist plot, which is also represented in Fig.13.18. We see that the first peak exceeding 1 in the roll-off region (noted 1 in Fig. 13.18) is indeed stable (it corresponds to the wide loop in the right side of the Nyquist plot). The second peak in the roll-off region (noted 2) is slightly unstable for the nominal gain of the compensator; some reduction of the gain is necessary to achieve stability (small loop near -1 in the Nyquist plot); this reduces the bandwidth to about A detailed examination showed that the potentially unstable mode corresponds to a local mode of the support

300


Applications

301

of the mirror of the displacement measurement system. This local mode is not influenced by the active damping; the situation could be improved by a redesign of the support for more stiffness and more damping (e.g. passive damping locally applied). This controller has been implemented digitally on a DSP processor


302

with a sampling frequency of 1000 Hz. Figure 13.19 compares the predicted step response with the experimental one. The settling time is reduced to about 10 times faster than what would be achievable with a PID compensator.

13.7

Volume displacement sensors

The volume velocity

of a vibrating plate is defined as

where is the transverse displacement of the plate and the integral extends over the entire plate area. It is a fairly important quantity in vibroacoustics, because it is strongly correlated to the sound power radiated by the plate (Johnson & Elliott), and the modes which do not contribute to the net volume velocity (anti-symmetric modes for a symmetric plate) are poor contributors to the sound power radiations at low frequency (Fahy). In this section, we discuss the sensing of the volume velocity with an arrangement of piezoelectric sensors; note that the same sensor arrangement can be used to measure the volume displacement V by using a charge amplifier instead of a current amplifier as we discussed in Fig.3.9, so that the two quantities are fully equivalent from a sensor design viewpoint. This section examines three totally different concepts for sensing the volume displacement with piezoelectric sensors; the first one is based on a distributed

Applications

303

sensor initially developed for beams, and extended to plates by discretizing them into narrow strips; the second is based on a discrete array sensor connected to a linear combiner; it is subjected to spatial aliasing. The third concept is based on a new electrode design which allows to tailor the effective piezoelectric properties of PVDF films.

13.7.1

QWSIS sensor

The Quadratically Weighted Strain Integrator Sensor (QWSIS) is a distributed sensor which applies to any plate without rigid body mode (Rex & Elliott). Beam

Consider a beam fixed at both ends: it is covered with a PVDF sensor with a parabolic electrode, as indicated in Fig.13.20; the profile of the electrode is defined by

According to Equ.(3.30), if the electrode is connected to a charge amplifier, the sensor output is

Upon integrating by parts twice, and taking into account the boundary conditions and we get

Since the width of an electrode of parabolic shape has a constant second derivative with respect to the space coordinate, the output of the sensor is proportional to the volume displacement:

304


Plate In the QWSIS, the plate is discretized in a set of narrow strips (Fig.13.21) which are provided with parabolic electrodes connected in series; if we consider the elementary strips as beams, the total amount of electric charge is proportional to the volume displacement of the plate. The QWSIS sensor is based on the beam theory, but the actual behaviour of the plate produces curvatures in two directions; assuming that the piezoelectric orthotropy axis 1 of the sensor coincides with the axis of the strip, the amount of electric charges generated by the sensor can be obtained by integrating Equ.(3.61) over the electrode area with the electrical boundary condition E = 0 enforced by the charge amplifier:

where and are the strain components along the orthotropy axes in the mid-plane of the sensor. If the membrane strains in the plate are small as compared to the bending strains,

where is the distance between the mid-plane of the sensor and the midplane of the baffle plate [see Equ.(3.67)]. If this equation is reduced to that of a beam, which means that the convergence of the sensor is guaranteed

Applications

305

when the number of strips increases. However, although strongly anisotropic, PVDF exhibits a piezoelectric coefficient which is at least 20% of which introduces a bias in the sensor output. Dual actuator A piezoelectric strip can be used either as a sensor or as an actuator; in the latter case, according to the beam theory [see Equ.(3.25)], a distributed actuator of width produces a distributed load proportional to the second derivative of the width of the electrode, Accordingly, if the QWSIS is used as an actuator, it is equivalent to a uniform pressure actuator. This led to the idea of building a collocated active structural acoustic (ASAC) plate with one side covered with a QWSIS volume displacement sensor and the opposite side covered with the dual actuator (Gardonio et al.). Unfortunately, such an arrangement performs poorly, because the input-output relationship between the strain actuator and the strain sensor is dominated by the membrane strains in the plate, which have been ignored in the theory, and are not related to the transverse displacements is the useful output of the system). The anisotropy of PVDF can be exploited to improve the situation, by placing the strips of the actuator and the sensor orthogonal to each other (Piefort).

13.7.2

Discrete array sensor

In this section, we discuss an alternative set-up using a discrete array of strain sensors bonded on the plate according to a regular mesh (Fig. 13.22). The strain sensors consist of piezo patches connected to individual charge amplifiers with output they are connected to a linear combiner, the output of which is

The coefficients of the linear combiner can be adjusted by software in order that the sensor output be as close as possible (in some sense) of a desired quantity such as a modal amplitude, or, in this case, the volume displacement. The electric charges generated by each strain sensor is a linear combination of the modal amplitude

where is the electric charge generated on sensor by a unit amplitude of mode The volume displacement V is also a linear combination of the modal amplitude,

306


where is the modal volume displacement of mode At low frequency, V is dominated by the contribution of the first few modes and therefore only these modal amplitudes, have to be reconstructed from the electric charges produced by a redundant set of piezoelectric strain sensors leading to

(where the coefficients Equ.(13.44), we find

are unknown at this stage). Combining with

where

Equation (13.46) has the form of a linear combiner with constant coefficients it can be rewritten in the frequency domain

Applications

307

where is the FRF between a disturbance applied to the baffled plate and the volume displacement, measured with a laser scanner vibrometer, is the FRF between the same disturbance and the electric charge on sensor i in the array. If this equation is written at a set of discrete frequencies regularly distributed over the frequency band of interest, it can be transformed into a redundant system of linear equations,

or, in matrix form, where Q is a complex valued rectangular matrix is a complex-valued vector and is the vector of linear combiner coefficients (real). Since the FRFs Q and V are determined experimentally, the solution of this redundant system of equations requires some care to eliminate the effect of noise; the coefficients resulting from the pseudo-inverse in the mean-square sense (see problem P.1.1)

are highly irregular and highly sensitive to the disturbance source. This difficulty can be overcome by using a singular value decomposition of Q (e.g. Strang), where

and are unitary matrices containing the eigenvectors of and respectively (the superscript H stands for the Hermitian, that is the conjugate transpose), and is the rectangular matrix of dimension with the singular values on the diagonal (equal to the square root of the eigenvalues of and If are the column vectors of and are the column vectors of Equ.(13.52) reads

and the pseudo-inverse is

This equation shows clearly that, because of the presence of the lowest singular values tend to dominate the pseudo-inverse; this is responsible for the high variability of the coefficients resulting from Equ.(13.51). The problem can be solved by truncating the singular value expansion (13.54) and deleting

308


the contribution relative to smaller singular values which are dominated by the noise. Without noise, the number of singular values which are significant (i.e. the rank of the system) is equal to the number of modes responding significantly in the frequency band of interest (assuming this number smaller than the number of sensors in the array); with noise, the selection is slightly more difficult, because the gap in magnitude between significant and insignificant singular values disappears; some trial and error is needed to identify the optimum number of singular values in the truncated expansion (François et al.). Figure 13.23 shows typical results obtained with a glass plate covered with an array of 4 × 8 PZT patches.

13.7.3

Spatial aliasing

The volume displacement sensor of Fig.13.23 is intended to be part of a control system to reduce the sound transmission through a baffled plate in the low frequency range (below 250 Hz), where the correlation between the volume velocity and the sound power radiation is high. Figure 13.23.c shows that the output of the array sensor follows closely the volume displacement below 400 Hz. However, in order to be included in a feedback control loop, the quality of the sensor must

Applications

309

be guaranteed at least one decade above the intended bandwidth of the control system. Figure 13.24 shows a numerical simulation of the open-loop FRF of a SISO system where the input consists of 4 point force actuators controlled with the same input current and the output is the volume displacement of the 4 × 8 array sensor. The comparison of the sensor output with the actual volume displacement reveals substantial differences at higher frequency, the amplitude of the sensor output being much larger than the actual volume displacement, which is inacceptable from a control point of view, for obvious reasons. This is due to spatial aliasing, as explained in Fig.13.25. The left part of Fig.13.25 shows the shape of mode (1,1) and mode (1,15); the diagrams on the right show the electric charges generated by the corresponding mode shape on the PZT patches. We observe that the electric charges generated by mode (1,15) have the same shape as those generated by mode (1,1). Thus, at the frequency 1494.4 Hz, the plate vibrates according to mode (1,15) which contributes only little to the volume displacement; however the output of the array sensor is the same as that of mode (1,1) which contributes a lot to the volume displacement; this explains why the high frequency amplitude of the FRF are much larger than expected. Note that it is a typical property of aliasing that a higher frequency component is aliased into a lower frequency component symmetrical with respect to the Nyquist frequency. In this case, the number of patches in the array being 8 along the length of the plate, mode (1,15) is aliased into the symmetrical one with respect to 8, that is into mode (1,1); similarly, mode (1,13) would be aliased into mode (1,3). The most obvious way to alleviate aliasing is to increase the sampling rate, that is, in this case, to increase the size of the array; this is illustrated in Fig.13.26 where we see that an array of 16 × 32 gives a good agreement up to 5000 Hz. However, dealing with such big arrays brings practical problems with the need for independent conditioning electronics (charge amplifier) prior to the linear combiner. If we give up the programmability of the linear combiner, the coefficients can be incorporated into the size of the electrodes, leading to the design of Fig.13.27, which requires only a single charge amplifier. The shape of this sensor has some similarity with the QWSIS (in two dimensions).

13.7.4

Distributed sensor

For the design of Fig.13.27 involving an electrode connecting 16 × 32 patches of variable size, the spatial aliasing still occurs above 10000 Hz; it can be pushed even further by increasing the number of patches. This suggests the distributed sensor with a single “porous” electrode shown in Fig.13.28 (Preumont et al.). The electrode is full in the center of the plate and becomes gradually porous as one moves towards the edge of the plate, to achieve an electrode density which produces the desired weighting coefficient This motif can be placed on

310


Applications

311

one side or on the two sides of the piezo material; for manufacturing, it seems simpler to apply the motif on one side only, with a continuous electrode on the other side. The amount of electrical charges generated on the electrode is given by Equ.(13.40) where the integral extends over the area of the electrode; it follows that tailoring the porosity of the electrode (i.e. is equivalent to tailoring

312


Applications

313

314


Applications

315

the piezoelectric constants of the material, and Equation (13.40) assumes that the size of the electrode is much larger than its thickness. However, when the motif of the electrode becomes small, tridimensional (edge) effects start to appear and the relationship between the porosity and the equivalent piezoelectric property is no longer linear. The exact relationship between the porosity and the equivalent piezoelectric coefficients can be explored with a tridimensional finite element analysis software (Piefort). Figure 13.29 shows the isopotential surfaces for the two electrode configurations when a small sample is subjected to a strain along the axis and a potential difference V = 0 is enforced between the electrodes; the material assumed in this study is isotropic PVDF polarized in the direction perpendicular to the electrodes; the edge effects appear clearly in the figures. For this sample, Fig.13.30 shows the relationship between the effective piezoelectric coefficient and the fraction of electrode area; the two electrode configurations are considered for two sample thicknesses we observe that for a very thin sensor, the two electrode configurations produce

316


the same results and the relationship is almost linear. The potential of this concept of “porous” electrode for shaping the effective piezoelectric properties of the material for bidimensional structures is far beyond the design of a volume displacement sensor. Modal filtering is another obvious application.

13.8

References

Digital control

K. J. ASTROM & B. WITTENMARK, Computer-Controlled Systems, Theory and Design, Prentice-Hall, 2nd ed., 1990. G. F. FRANKLIN & J. D. POWELL, Digital Control of Addison-Wesley, 1980.

Dynamic Systems,

H. HANSELMANN, Implementation of Digital Controllers - A Survey, Automatica, vol. 23, no. 1, 7-32, 1987. L. B. JACKSON, Digital Filters and Signal Processing, Kluwer, 1986. B. C. KUO, Digital Control Systems, SRL Pub. Co., Champaign, Illinois, 1977. A. V. OPPENHEIM & R. W. SCHAFER, Digital Signal Processing, PrenticeHall, 1975.

Active damping of a truss

E. H. ANDERSON, D. M. MOORE, J. L. FANSON & M. A. EALEY, Development of an Active Member Using Piezoelectric and Electrostrictive Actuation for Control of Precision Structures, SDM Conference, AIAA paper 90-1085CP, 1990. G. S. CHEN, B. J. LURIE & B. K. WADA, Experimental Studies of Adaptive Structures for Precision Performance, SDM Conference, AIAA paper 89-1327CP, 1989. J. L. FANSON, G. H. BLACKWOOD & C. C. CHEN, Active Member Control of Precision Structures, SDM Conference, AIAA paper 89-1329-CP, 1989. T. T. HYDE & E. H. ANDERSON, Actuator with Built-in Viscous Damping for Isolation and Structural Control, AIAA Journal, vol. 34, no. 1, 129-135, Jan. 1996.

Applications

317

L. D. PETERSON, J. J. ALLEN, J. P. LAUFFER & A. K. MILLER, An Experimental and Analytical Synthesis of Controlled Structure Design, SDM Conference, AIAA paper 89–1170-CP, 1989. A. PREUMONT, J. P. DUFOUR & C. MALEKIAN, Active Damping by a Local Force Feedback with Piezoelectric Actuators, AIAA J. of Guidance, vol. 15, no. 2, 390–395, March–April 1992.

Stewart platform

A. ABU HANIEH, A. PREUMONT & N. LOIX, Piezoelectric Stewart Platform for General Purpose, Active Damping Interface and Precision Control, 9th European Space Mechanisms and Tribology Symposium, Liege, Belgium, 2001. Z. J. GENG & L. S. HAYNES, Six Degree-of-Freedom Active Vibration Control Using the Stewart Platforms, IEEE Transactions on Control Systems Technology, vol. 2, no. 1, 45–53, March 1994. J. E. Mc INROY, G. W. NEAT & J. F. O’BRIEN, A Robotic Approach to Fault-Tolerant, Precision Pointing, IEEE Robotics & Automation Magazine, pp. 24-31, Dec. 1999. D. STEWART, A Platform with Six Degrees of Freedom, Proc. Instn. Mech. Engrs, vol. 180, no. 15, 371–386, 1965–66.

Active damping of a plate

J. J. DOSCH, D. J. INMAN & E. GARCIA, A Self-Sensing Piezoelectric Actuator for Collocated Control, J.of Intelligent Materials, Sytems and Structures, vol. 3, 166–185, Jan. 1992. J. L. FANSON & T. K. CAUGHEY, Positive Position Feedback Control for Large Space Structures, AIAA Journal, vol. 28, no. 4, 717–724, April 1990. N. LOIX, A. CONDE REIS, P. BRAZZALE, J. DETTMANN & A. PREUMONT, CFIE: In-Orbit Active Damping Experiment Using Strain Actuators, Space Microdynamics and Accurate Control Symposium, Toulouse, May 1997. V. PlEFORT, Finite Element Modeling of Piezoelectric Active Structures, Ph.D. thesis, Universite Libre de Bruxelles, Active Structures Laboratory, 2001.

318


Active damping of a stiff beam

A. PREUMONT, N. LOIX D. MALAISE & O. LECRENIER, Active Damping of Optical Test Benches with Acceleration Feedback, Machine Vibration, vol. 2, 119-124, 1993.

HAC/LAC

J. N. AUBRUN, Theory of the Control of Structures by Low-Authority Controllers, AIAA J. of Guidance, vol. 3, no. 5, 444–451, Sep.–Oct. 1980. R. J. BENHABIB, R. P. IWENS & R. L. JACKSON, Stability of Large Space Structure Control Systems Using Positivity Concepts, AIAA J. of Guidance and Control, vol. 4, no. 5, 487–494, Sep.–Oct. 1981. K. GUPTA, Frequency-Shaped Cost Functionals: Extension of Linear Quadratic Gaussian Methods, AIAA J. of Guidance and Control, vol. 3, no. 6, 529–535, Nov.-Dec. 1980. R. L. KOSUT, H. SALZWEDEL & A. EMANI-NAEMI, Robust Control of Flexible Spacecraft, AIAA J. of Guidance, vol. 6, no. 2, 104–111, March-April 1983. V. MUKHOPADHYAY & J. R. NEWSOM, A Multiloop System Stability Margin Study Using Matrix Singular Values, AIAA J. of Guidance, vol. 7, no. 5, 582–587, Sep.–Oct. 1984. E. K. PARSONS, An Experiment Demonstrating Pointing Control on a Flexible Structure, IEEE Control Systems Magazine, pp. 79–86, April 1989. A. PREUMONT, Active Structures for Vibration Suppression and Precision Pointing, Journal of Structural Control, vol. 2, no. 1, 49–63, June 1995. Volume displacement sensors

F. FAHY, Sound and Structural Vibration, Academic Press, 1987. P. DE MAN & A. PREUMONT, Piezoelectric Array Sensing of Volume Displacement: A Hardware Demonstration, J. of Sound and Vibration, vol. 244, no. 3, 395–405, July 2001. P. GARDONIO, Y. LEE, S. ELLIOTT & S. DEBOST, Active Control of Sound Transmission Through a Panel with a Matched PVDF Sensor and Actuator Pair, Active 99, Fort Lauderdale, Florida, USA, Dec. 1999.

Applications

319

M. E. JOHNSON & S. J. ELLIOTT, Active Control of Sound Radiation Using Volume Velocity Cancellation, J. of the Acoustical Society of America, vol. 98, 2174-2186, 1995. V. PlEFORT, Finite Element Modeling of Piezoelectric Active Structures, Ph.D. thesis, Universite Libre de Bruxelles, Active Structures Laboratory, 2001. A. PREUMONT, Method for Shaping Laminar Piezoelectric Actuators and Sensors and Related Devices, European patent application No 01870068.2, March 2001. A. PREUMONT, P. DE MAN & A. FRANQOIS, On the Feedback Control of a Baffled Plate, submitted to the Journal of Sound and Vibration, 2001. J. REX & S. J. ELLIOTT, The QWSIS - A New Sensor for Structural Radiation Control, MOVIC-1, Yokohama, September 1992. G. STRANG, Linear Algebra and its Applications, New York: Harcourt Brace Jovanovich, third edition, 1988.

13.9

Problems

P.13.1 An antialiasing filter of bandwidth second order filters of the form

can be obtained by cascading

The Butterworth filters correspond to order 2: order 4: order 6: The Bessel filters correspond to order 2: order 4: order 6: (a) Compare the Bode plots of the various filters and, for each of them, evaluate the phase lag for and (b) Show that the poles of the Butterworth filter are located on a circle of radius according to the configurations depicted in Fig.9.1. (c) Show that, at low frequency, the Bessel filter has a linear phase, and can be approximated with a time delay (Åström & Wittenmark).

320

P.13.2


Show t hat the transfer function of the zero-order hold is

Show that the frequency response function is

Draw the amplitude and phase plots. P.13.3 Using the bilinear transform, show that the discrete equivalent of Equ.(13.8) (13.9) is given by Equ.(13.10)-(13.13). P.13.4 Consider a truss structure with several identical active members controlled with the same control law (IFF) and the same gain. Making the proper assumptions, show that each closed-loop pole follows a root locus defined by Equ.(5.39), where the natural frequency is that of the open-loop structure and is that of the structure where the active members have been removed. P.13.5 For the active truss of section 13.2, show that the compensator

is equivalent to provided that the breakpoint frequency a is such that Show that its digital counterpart is

P.13.6 Consider a simply supported beam with a point force actuator and a collocated accelerometer at Assume that EI = 1 and Design a compensator to achieve a closed-loop modal damping for and 2, using the Direct Velocity Feedback and the second order filter (see Table 5.1). Draw the Bode plots for the two compensators and compare the phase margins. For both cases, check the effect of the delay corresponding to a sampling frequency 100 times larger than the first natural frequency of the system and that of the actuator dynamics, assuming that the force actuator is a proof-mass with a natural frequency (assume

Chapter 14

Tendon Control of Cable Structures 14.1

Introduction

Cable structures are used extensively in civil engineering: suspended bridges, cable-stayed bridges, guyed towers, roofs in large public buildings and stadiums. The main span of current cable-stayed bridges (Fig.14.1) can reach more than (e.g. Normandy bridge, near Le Havre, in France). These structures are very flexible, because the strength of high performance materials increases faster than their stiffness; as a result, they become more sensitive to wind and traffic induced vibrations. Large bridges are also sensitive to flutter (Fig. 14.2) which, in most cases, is associated with the aeroelastic damping coefficient in torsion becoming negative above a critical velocity (Scanlan & Tomko). The situation can be improved either by changing the aerodynamic shape of the deck, or by increasing the stiffness and damping in the system; the difficulty in active damping of cable structures lies in the strongly nonlinear behaviour of the cables, particularly when the gravity loads introduce some sag (typical sag to length ratio is 0.5% for a cable-stayed bridge). The structure and the cables interact with linear terms (at the natural frequency of the cable and quadratic terms resulting from stretching (at the latter may produce parametric excitation if some tuning conditions are satisfied (parametric excitation has indeed been identified as the source of vibration in several existing cable-stayed bridges). Cable structures are not restricted to civil engineering applications; the use of cables to achieve lightweight spacecrafts was recommended in Herman Oberth’s early books on astronautics. Tension truss structures have already been used for large deployable mesh antennas. The use of guy cables is probably the most efficient way to stiffen a structure in terms of weight; in addition, if the structure is deployable and if the guy cables have been properly designed, they may be 321

322


Tendon Control of Cable Structures

323

used to prestress the structure, to eliminate the geometric uncertainty due to the gaps. This chapter examines the possibility of connecting guy cables to active tendons to bring active damping into cable structures; the same strategy applies to large space structures and to cable-stayed bridges and other civil engineering structures; however, the technology used to implement the control strategy is vastly different (piezoelectric actuators for space and hydraulic actuators for bridges).

14.2

Tendon control of strings and cables

The mechanism by which an active tendon can extract energy from a string or a cable is explained in Fig. 14.3 with a simplified model assuming only one mode (Rayleigh-Ritz) and for situations of increasing complexity. The simplest case is that of a linear string with constant tension (Fig. 14.3.a); the equation becomes nonlinear when the effect of stretching is added (cubic nonlinearity). In Fig.14.3.b, a moving support is added; the input of this active tendon

324


produces a parametric excitation, which is the only way one can control a string with this type of actuator. The difference between a string and a cable is the effect of gravity, which produces sag (Fig.14.3.c). In this case, the equations of motion in the gravity plane and in the plane orthogonal to it are no longer the same, and they are coupled. In the gravity plane coordinate), the active tendon control still appears explicitely as a parametric excitation, but also as an inertia term whose coefficient depends on the sag of the cable; even for cables with moderate sag (say sag to length ratio of 1% or more), this contribution becomes significant and constitutes the dominant control term of the equation. On the contrary, in the out-of-plane equation coordinate), the tendon control appears explicitely only through the parametric excitation, as for the string.

14.3

Active damping strategy

Figure 14.4 shows a schematic view of a cable-structure system, where the control is the support displacement, T is the tension in the cable, the transverse vibration of the cable and the vibration of the structure; we seek a control strategy for moving the active tendon to achieve active damping in the structure and the cable. Any control law based on the non-collocated measurements of the cable and structure vibration

must, at some stage, rely on a simplified model of the system; as a result, it is sensitive to parametric variations and to spillover. Such control laws have been investigated by Chen and Fujino & coworkers; they work in specific conditions, when the vibration is dominated by a single mode, but they become unstable when the interaction between the structure and the cable is strong.


325

By contrast, we saw in section 11.7 that, if a force sensor measuring the tension T in the cable is collocated with the active tendon, the positive Integral Force Feedback

produces an energy absorbing controller. A high-pass filter is necessary to eliminate the static tension in the cable. To establish the vibration absorbing properties of Equ.(14.2) when T is the dynamic component of the tension in the cable, one can show that the dynamic contribution to the total energy, resulting from the vibration around the static equilibrium position, is a Liapunov function. Thus, the stability is guaranteed if we assume perfect sensor and actuator dynamics. Note that the fact that the global stability is guaranteed does not imply that all the vibration modes are effectively damped. In fact, from a detailed examination of the dynamic equations (e.g. Fujino et al. or Achkire), it appears that not all the cable modes are controllable with this actuator and sensor configuration. The odd numbered in-plane modes (in the gravity plane) can be damped substantially because they are linearly controllable by the active tendon (inertia term in Fig. 14.3) and linearly observable from the tension in the cable; all the other cable modes are controllable only through active stiffness variation (parametric excitation in Fig. 14.3), and observable from quadratic terms due to cable stretching. The nice thing is that these weakly controllable modes are never destabilized by the control system, even at the parametric resonance, when the natural frequency of the structure is twice that of the cable. The reader will observe that the Integral Force Feedback (14.2) is exactly the same as that used for damping the truss in section 13.2; the discussion of section 13.2.2 regarding the implementation also applies here.

14.4

Basic Experiment

Figure 14.5 shows the test structure that was built to represent the ideal situation of Fig.14.4; the cable is a long stainless steel wire of cross section provided with additional lumped masses at regular intervals, to achieve a sag to span ratio comparable to actual bridges; the active tendon is materialized by a piezoelectric linear actuator acting on the support point with a lever arm, to amplify the actuator displacement by a factor 3.4; this produces a maximum axial displacement of for the moving support. A piezoelectric force sensor is colinear with the actuator; because of the high-pass behaviour of this type of sensor, it measures only the dynamic component of the tension in the cable. The spring-mass system (in black on the figure) has an adjustable mass which allows us to tune its natural frequency; a shaker and an accelerometer are attached to it, to evaluate the performance of the control

326



327

system. In addition, a non-contact laser measurement system was developed to measure the cable vibration. Figure 14.6 shows the effect of the control system on the structure; we see that the controller brings a substantial amount of damping to the system. As far as the cable modes are concerned, the out-of-plane modes and the antisymmetric in-plane modes are not affected by the controller (except for large amplitudes where the cable stretching becomes significant); the amount of active damping brought into the symmetric in-plane modes depends very much on the sag to span ratio. The control system behaves nicely, even at the parametric resonance, when the natural frequency of the structure is exactly twice that of the cable.

14.5

Approximate linear theory

In this section, we follow an approach similar to that of sections 5.5.1 and 13.3.1 to predict the closed-loop poles of the cable-structure system. To do that, we assume that the dynamics of the active cables can be neglected and that their interaction with the structure is restricted to the tension in the active cables. Accordingly, the governing equation is

where M and K refer to the passive structure (including a linear model of the passive cables but excluding the active cables), T is the vector of tension in the active cables and B is their influence matrix. If we neglect the cable dynamics, they behave like bars and the tension in the cables is given by

where is the stiffness matrix of the cables, are the relative displacements of the end points of the cables projected along the chord lines, and is the vector of active displacements of the tendons. Combining Equ.(14.3) and (14.4), we get This equation indicates that is the stiffness matrix of the structure including all the guy cables (passive + active). Next, we assume that all the active cables are controlled according to the feedback law:

(IFF with equal gain on the elastic extension of the cables). Equ.(14.4) to (14.6), we obtain the closed-loop equation

Combining

328


It is readily observed that the open-loop poles, solutions of the characteristic equation for satisfy

while the zeros, solutions of Equ.(14.7) for

satisfy

which is the eigenvalue problem for the open-loop structure where the active cables have been removed. Next, let us project the characteristic equation on the normal modes of the structure with all the cables, which are normalized according to According to the orthogonality condition,

where are the natural frequencies of the complete structure. In order to derive a simple and powerful result about the way each mode evolves with let us assume that the mode shapes are little changed by the active cables, so that we can write where are the natural frequencies of the structure where the active cables have been removed. Comparing Equ.(14.10) and (14.11), we get

where

is the fraction of modal strain energy in the active cables. Upon projecting the characteristic equation (14.7) into modal coordinates and taking into account Equ.(14.10) and (14.11), we get

or

which is identical to (5.38). This result indicates that the closed-loop poles can be predicted by performing two modal analyses (Fig.14.7), one with all the cables, leading to the open-loop poles and one with only the passive cables, leading to the open-loop zeros and drawing the independent root loci (14.14). As in section 5.5.1, the maximum modal damping is given by


329

and it is achieved for As for the active truss, for small gains, the modal damping resulting from the active tendon control is proportional to the fraction of modal strain energy:

Equ.(14.15) and (14.16) can be used very conveniently in the design of actively controlled cable structures. Recall that the foregoing results are based on the assumption that (i) the dynamics of the active cables can be neglected, that (ii) the passive one, behave linearly, and that (iii) the mode shapes are not changed (substantially) when the active cables are removed. A comparison of this result with experiments is given below.

14.6

Application to space structures

14.6.1

Guyed truss experiment

This experiment aims at comparing the closed-loop predictions of the linear model with experiments. The test structure consists of the active truss of

330



331

Fig.3.17 equipped with three identical cables made of synthetic fiber “Dynema” of 1 diameter (Fig.14.8.a); the tension in the cables is not important provided that the effective Young modulus (due to sag) is close to the actual one; in this case, the tension is such that the cable frequency is above The design of the active tendon is shown in Fig.14.8.b; the amplification ratio of the lever arm is 3, leading to a maximum stroke of The natural frequencies with and without the active cables (respectively and are given in Table 14.1. Figure 14.9 shows the root locus predicted by the linear model together with the experimental results for various values of the gain; only the upper part of the loops is available experimentally because the control gain is limited by the saturation due to the finite stroke of the actuators. The agreement between the experimental results and the linear predictions [Equ.(14.14)] is quite good.

14.6.2

JPL-MPI testbed

To illustrate further the application of the control strategy to the damping of large space trusses, let us consider a numerical model of the microprecision interferometer testbed used at JPL to develop the technology of precision structures for future interferometric missions (Neat et al.). The first three flexible modes are displayed in Fig.14.10. We investigate the possibility of stiffness augmentation and active damping of these modes with a set of three active tendons acting on Kevlar cables of diameter, connected as indicated in Fig.14.11 (Kevlar properties : tensile strength The global added mass for the three cables is only (not including the active tendons and the control system). The natural frequencies of the first three modes, with and without the cables, are reported in Table 14.2; the root locus of the three global flexible modes as functions of the control gain are represented in Fig.14.12; for the modal damping ratios are

14.6.3

Free floating truss experiment

In order to confirm the spectacular analytical predictions obtained with the numerical model of the JPL-MPI testbed, a similar structure (although smaller) was built and tested (Fig.14.13); the free-floating condition was simulated by hanging the structure with soft springs. The active tendon consists of a APA 100 M amplified actuator from CEDRAT Recherche together with a B&K 8200

332



333

334


force sensor and flexible tips. The stroke is and the total weight of the tendon is the cable is made of Dynema with axial stiffness EA = 19000 N. The natural frequencies of the first flexible modes, with and without cables, are reported in Table 14.3. Figure 14.14 shows the comparison between the analytical predictions of the linear model and the experiments.

14.6.4

Microvibrations

All the results discussed above have been obtained for vibrations in a range going from millimeter to micron; in order to apply this technology to future large space platforms for interferometric missions, it is essential that these results be confirmed for microvibrations. In fact, it could well be that, for very small amplitudes, the behaviour of the control system be dominated by the nonlinearity of the actuator (hysteresis of the piezo) or the noise in the sensor or in the voltage amplifier. Tests have been conducted for vibrations of decreasing amplitudes, and the influence of the various hardware components has been


335

analysed (Bossens), these tests indicate that active damping is feasible at the nanometric level, provided that adequately sensitive components are used.

14.7

Application to cable-stayed bridges

In this final section, we summarize some of the findings of a research project called “ACE” which was funded by the EC in the framework of the Brite-Euram program, between 1997 and 2000, and involved several academic and industrial partners: DERA, UK; Newland Technology Ltd, UK; Bouygues, France; VSL, France; Johs Holt A.S, Norway; Mannesmann Rexroth AG, Germany; Joint Research Center of the EC, Italy; Technische Universität Dresden, Germany; Université Libre de Bruxelles, Belgium. The overall objective of the project was to demonstrate the use of the active control in civil engineering.

14.7.1

Laboratory experiment

The test structure is a laboratory model of a cable-stayed bridge during its construction phase, which is amongst the most critical from the point of view of the wind response. The structure consists of two half decks mounted symmetrically

336


with respect to a central column of about high (Fig. 14.15); each side is supported by 4 cables, two of which are equipped with active piezoelectric tendons identical to those of Fig.14.8.b. The cables are provided with lumped masses at regular intervals, so as to match the sag to length ratio of actual stay cables (a discussion of the similarity aspects can be found in (Warnitchai et al.). Figure 14.16 compares the evolution of the first bending and torsion closed-loop poles of the deck with the analytical predictions of the linear theory. The agreement is good for small gains, when the modal damping is smaller than 20%.

14.7.2

Control of parametric resonance

In this experiment, the bridge deck is excited with an electrodynamic shaker and the tension in the two passive cables on one side is chosen in such a way that the first in-plane mode of one of them is tuned on the global bending mode of the deck while the other is tuned on to experience the parametric resonance when the deck vibrates (Fig.14.17). This tuning is achieved by monitoring the cable vibration with a specially developed non-contact optical measurement system (Achkire). Figure 14.18 shows the vibration amplitude of the deck and the transverse amplitude of the in-plane mode of the two passive cables when


337

338


the deck is excited at resonance; the excitation starts at and the control is turned on after We note that: 1. The amplitude of the cable vibration is hundred times larger than that of the deck vibration. 2. The parametric resonance is established after some transient period in which the cable vibration changes from frequency to The detail of the transition to parametric resonance is shown clearly in the central part of Fig.14.19. 3. The control brings a rapid reduction of the deck amplitude (due to active damping) and a slower reduction of the amplitude of the cable at resonance (due only to the reduced excitation from the deck, since no active damping is applied to this cable). 4. The control suppresses entirely the vibration of the cable at parametric resonance This confirms that a minimum deck amplitude is necessary to trigger the parametric resonance.

14.7.3

Large scale experiment

Although appropriate to demonstrate control concepts in labs, the piezoelectric actuators are inadequate for large scale applications. For cable-stayed bridges, the active tendon must simultaneously sustain the high static load (up to and produce the dynamic load which is at least one order of magnitude lower than the static one (< ±10%). This has led to an active tendon design consisting of two cylinders working together: one cylinder pressurized by an accumulator compensates for the static load, and a smaller double rod cylinder drives the cable dynamically to achieve the control law. The two functions are integrated in a single cylinder, as illustrated in Fig.14.20; the double rod part of the cylinder is achieved by a “rod in rod” design; this solution saves hydraulic energy and reduces the size of the hydraulic components. The cylinder is position controlled; the long term changes of the static loads as well as the temperature differences require adaptation of the hydraulic conditions of the accumulator. The actuators used in this project were designed and manufactured by Mannesmann-Rexroth in collaboration with the Technical University of Dresden. The mock-up (Fig.14.21 and 14.22) was designed and manufactured by Bouygues in the framework of the ACE project; it has been installed on the reaction wall of the ELSA facility at the Joint Research Center in Ispra. It consists of a cantilever beam supported by 8 stay cables the stay cables are provided with additional masses to achieve a representative sag-to-length ratio (the overall mass per unit length is 15 An intermediate support can also be placed along the deck to tune the first global mode and the cable frequencies. Because of the actuator dynamics and the presence of a static load, the implementation of the control requires some alterations from the basic idea of Fig.14.4: (i) A high-pass filter must be included after the force sensor to eliminate the static load in the active cables.


339

340



341

(ii) In hydraulics, the flowrate is directly related to the valve position which is the control element; it is therefore more natural to control the actuator velocity than its position. In addition, a proportional controller acting on the actuator velocity is equivalent to an integral controller acting on the actuator displacement. The actual implementation of the control is shown in Fig. 14.23. The overall controller includes a high-pass filter with a corner frequency at (to eliminate the static load), an integrator and a low-pass filter with corner at to eliminate the internal resonance of the hydraulic actuator. The overall FRF of the active control device is represented in Fig.14.24. The dotted line refers to the digital controller alone (between 1 and 3 in Fig. 14.23) while the full line includes the actuator dynamics (between 1 and 2 in Fig. 14.23). Notice that (i) the controller behaviour follows closely a pure integrator in the frequency range of interest and (ii) the actuator dynamics introduces a significant phase lag above the dominant modes of the bridge. Figure 14.25 shows the envelope of the time response of the bridge deck displacement near the tip when a sweep sine input is applied to a proof-mass

342


actuator (MOOG, max. inertial force 40 located off axis near the end of the deck (Fig. 14.22). The three curves correspond to various values of the gain of the decentralized controller when the two active tendons are in operation corresponds to the open-loop response). The instantaneous frequency of the input signal is also indicated on the time axis, to allow the identification of the main contributions to the response. One sees that the active tendon control brings a substantial reduction in the vibration amplitude, especially for the first global bending mode. Using a band-limited white noise excitation and a specially developed identification technique based on the spectral moments of the power spectral density of the bridge response, M. Auperin succeeded in isolating the first global mode of the bridge. Figure 14.26 compares the experimental root-locus with the predictions of the linear approximation; the agreement is surprisingly good, especially if one thinks of the simplifying assumptions leading to Equation (14.14) . The marks on the experimental and theoretical curves indicate the fraction of optimum gain where corresponds to the largest modal damping ratio (theoretical value Note that the maximum damping ratio is close to 17%.


343

344

14.8


References

Y. ACHKIRE, Active Tendon Control of Cable-Stayed Bridges, Ph.D. thesis, Active Structures Laboratory, Université Libre de Bruxelles, Belgium, May 1997. Y. ACHKIRE & A. PREUMONT, Active Tendon Control of Cable-Stayed Bridges, Earthquake Engineering and Structural Dynamics, vol. 25, no. 6, 585–597, June 1996. Y. ACHKIRE &; A. PREUMONT, Optical Measurement of Cable and String Vibration, Shock and Vibration, vol. 5, 171–179, 1998. F. BOSSENS, Contrôle Actif Des Structures Câblées: De la Théorie À L’implémentation, Ph.D. thesis, Active Structures Laboratory, Université Libre de Bruxelles, Belgium, Oct. 2001. F. BOSSENS & A. PREUMONT, Active Tendon Control of Cable-Stayed Bridges: A Large-Scale Demonstration, Earthquake Engineering and Structural Dynamics, 30, pp. 961–979, 2001. J.-C. CHEN, Response of Large Space Structures with Stiffness Control, AIAA J. of Spacecraft, vol. 21, no. 5, 463–467, 1984.


345

Y. FUJINO & T. SUSUMPOW, An experimental study on active control of planar cable vibration by axial support motion, Earthquake Engineering and Structural Dynamics, vol. 23, 1283–1297, 1994. Y. FUJINO, P. WARNITCHAI & B. M. PACHECO, Active Stiffness Control of Cable Vibration, ASME J. of Applied Mechanics, 60, pp. 948–953, 1993. Y. C. FUNG, An Introduction to the Theory of Aeroelasticity, Dover, 1969. A. H. NAYFEH & D. T. MOOK, Nonlinear Oscillations, Wiley, 1979. G. W. NEAT, A. ABRAMOVICI, J. M. MELODY, R. J. CALVET, N. M. NERHEIM & J. F. O’BRIEN, Control Technology Readiness for Spaceborne Optical Interferometer Missions, Proceedings SMACS-2, Toulouse, pp. 13–32, 1997. A. PREUMONT & Y. ACHKIRE, Active Damping of Structures with Guy Cables, AIAA, J. of Guidance, Control, and Dynamics, vol. 20, no. 2, 320–326, March-April 1997. A. PREUMONT, Y. ACHKIRE & F. BOSSENS, Active Tendon Control of Large Trusses, AIAA Journal, vol. 38, no. 3, 493–498, March 2000. A. PREUMONT & F. BOSSENS, Active Tendon Control of Vibration of Truss Structures: Theory and Experiments, J, of Intelligent Material Systems and Structures, vol. 2, no. 11, 91–99, Feb. 2000. R. H. SCANLAN & J. TOMKO, Airfoil and Bridge Deck Flutter Derivatives, ASCE J. Eng. Mech. div., 100, pp. 657–672, Aug. 1974. P. WARNITCHAI, Y. FUJINO, B. M. PACHECO & R. AGRET, An Experimental Study on Active Tendon Control of Cable-Stayed Bridges, Earthquake Engineering and Structural Dynamics, vol. 22, no. 2, 93–111, 1993. J. N. YANG & F. GIANNOPOULOS, Active Control and Stability of Cable-Stayed Bridge, ASCE J. Eng. Mech. div., 105, pp. 677–694, Aug. 1979a. J. N. YANG & F. GIANNOPOULOS, Active Control of Two-Cable-Stayed Bridge, ASCE J. Eng. Mech. div., 105, pp. 795–810, Oct. 1979b.


Bibliography A. ABU HANIEH, A. PREUMONT & N. LOIX, Piezoelectric Stewart Platform for General Purpose, Active Damping Interface and Precision Control, 9th European Space Mechanisms and Tribology Symposium, Liége, Belgium, 2001. Y. ACHKIRE, Active Tendon Control of Cable-Stayed Bridges, Ph.D. thesis, Active Structures Laboratory, Université Libre de Bruxelles, Belgium, May 1997. Y. ACHKIRE & A. PREUMONT, Active Tendon Control of Cable-Stayed Bridges, Earthquake Engineering and Structural Dynamics, vol. 25, no. 6, 585–597, June 1996. Y. ACHKIRE & A. PREUMONT, Optical Measurement of Cable and String Vibration, Shock and Vibration, vol. 5, 171–179, 1998. J. ACKERMANN, Sampled-Data Control Systems, Springer-Verlag, 1985. B. D. AGARWAL & L. J. BROUTMAN, Analysis and Performance of Fiber Composites, Wiley, second ed., 1990. H. ALLIK & T. R. J. HUGHES, Finite Element Method for Piezoelectric Vibration, Int. J. for Numerical Methods in Engineering, vol. 2, 151–157, 1970. B. D. O. ANDERSON & J. B. MOORE, Linear Optimal Control, Prentice Hall, Inc. Englewood Cliffs, NJ, 1971. E. H. ANDERSON, D. M. MOORE, J. L. FANSON & M. A. EALEY, Development of an Active Member Using Piezoelectric and Electrostrictive Actuation for Control of Precision Structures, SDM Conference, AIAA paper 90-1085CP, 1990. K. J. ÅSTRÖM, Introduction to Stochastic Control Theory, Academic Press, 1970. K. J. ÅSTRÖM & B. WITTENMARK, Computer-Controlled Systems, Theory and Design, Prentice-Hall, 2nd ed., 1990. 347

348


M. ATHANS, The Role and Use of the Stochastic Linear-Quadratic-Gaussian Problem in Control System Design, IEEE Transactions on Automatic Control, vol. AC-16, no. 6, 529–552, Dec. 1971. J. N. AUBRUN, Theory of the Control of Structures by Low-Authority Controllers, AIAA J. of Guidance, vol. 3, no. 5, 444–451, Sep.-Oct. 1980. J. N. AUBRUN, K. R. LORELL, T. W. HAVAS & W. C. HENNINGER, Performance Analysis of the Segment Alignment Control System for the Ten-Meter Telescope, Automatica, vol. 24, no. 4, 437–453, 1988. M. J. BALAS, Active Control of Flexible Systems, Journal of Optimization Theory and Applications, vol. 25, no. 3, 415–436, 1978. M. J. BALAS, Direct Velocity Feedback Control of Large Space Structures, AIAA J. of Guidance, vol. 2, no. 3, 252–253, 1979. K. J. BATHE & E. L. WILSON, Numerical Methods in Finite Element Analysis, Prentice-Hall, 1976. A. BAZ, S. POH & J. FEDOR, Independent Modal Space Control with Positive Position Feedback, Trans. ASME, J. of Dynamic Systems, Measurement, vol. 114, no. 1, 96–103, March 1992. R. J. BENHABIB, R. P. IWENS & R. L. JACKSON, Stability of Large Space Structure Control Systems Using Positivity Concepts, AIAA J. of Guidance and Control, vol. 4, no. 5, 487–494, Sep.–Oct. 1981. R. E. D. BISHOP & D. C. JOHNSON, The Mechanics of Vibration, Cambridge University Press, 2nd ed., 1979. H. W. BODE, Relations Between Attenuation and Phase in Feedback Amplifier Design, Bell System Technical Journal, July 1940. H. W. BODE, Network Analysis and Feedback Amplifier Design, Van Nostrand, N-Y, 1945. F. BOSSENS, Contrôle Actif Des Structures Câblées: De la Theorie A L’implémentation, Ph.D. thesis, Active Structures Laboratory, Université Libre de Bruxelles, Belgium, Oct. 2001. F. BOSSENS & A. PREUMONT, Active Tendon Control of Cable-Stayed Bridges: A Large-Scale Demonstration, Earthquake Engineering and Structural Dynamics, 30, pp. 961–979, 2001. A. E. BRYSON JR., Some Connections Between Modern and Classical Control Concepts, ASME, Journal of Dynamic Systems, Measurement, and Control, vol. 101, 91–98, June 1979.

Bibliography

349

S. E. BURKE & J. E. HUBBARD, Active Vibration Control of a Simply Supported Beam Using Spatially Distributed Actuator, IEEE Control Systems Magazine, pp. 25–30, Aug. 1987. R. H. CANNON, Dynamics of Physical Systems, McGraw-Hill, 1967. R. H. CANNON & D. E. ROSENTHAL, Experiment in Control of Flexible Structures with Noncolocated Sensors and Actuators, AIAA Journal of Guidance, vol. 7, no. 5, 546–553, Sep.–Oct. 1984. J. D. CARLSON & M. R. JOLLY, MR Fluid, Foam and Elastomer Devices, Mechatronics, vol. 10, 555–569, 2000. J. D. CARLSON & J. L. SPROSTON, Controllable Fluids in 2000 - Status of ER and MR Fluid Technology, Proc. of Actuator 2000, Bremen, pp. 126–130, June 2000. R. M. CHALASANI, Ride Performance Potential of Active Suspension Systems, Part1: Simplified Analysis Based on a Quarter-Car Model, ASME Symposium on Simulation and Control of Ground vehicles and Transportation systems, Anaheim, CA, Dec. 1984. G. S. CHEN, B. J. LURIE & B. K. WADA, Experimental Studies of Adaptive Structures for Precision Performance, SDM Conference, AIAA paper 89-1327CP, 1989. J.-C. CHEN, Response of Large Space Structures with Stiffness Control, AIAA J. of Spacecraft, vol. 21, no. 5, 463–467, 1984. R. L. CLARK, W. R. SAUNDERS & G. P. GIBBS, Adaptive Structures, Dynamics and Control, Wiley, New York, NY, 1998. R. W. CLOUGH & J. PENZIEN, Dynamics of Structures, McGraw-Hill, 1975. J. P. COULTER, K. D. WEISS & J. D. CARLSON, Engineering Applications of Electrorheological Materials, Journal of Intelligent Material Systems and Structures, vol. 4, 248–259, April 1993. R. R. CRAIG JR, Structural Dynamics, Wiley, 1981. E. F. CRAWLEY & J. DE LUIS, Use of Piezoelectric Actuators as Elements of Intelligent Structures, AIAA Journal, vol. 25, no. 10, 1373–1385, 1987. E. F. CRAWLEY & K. B. LAZARUS, Induced Strain Actuation of Isotropic and Anisotropic Plates, AIAA Journal, vol. 29, no. 6, 944–951, 1991. J. J. D’AZZO & C. H. HOUPIS, Feedback Control System Analysis & Synthesis, McGraw-Hill, second ed., 1966.

350


E. K. DIMITRIADIS, C. R. FULLER & C. A. ROGERS, Piezoelectric Actuators for Distributed Vibration Excitation of Thin Plates, Trans. ASME, J. of Vibration and Acoustics, vol. 113, 100–107, Jan. 1991. J. J. DISTEFANO, A. R. STUBBERUD & I. J. WILLIAMS, Feedback and Control Systems, Shaum’s Outline Series, McGraw-Hill, 1967. J. J. DOSCH, D. J. INMAN & E. GARCIA, A Self-Sensing Piezoelectric Actuator for Collocated Control, J. of Intelligent Materials, Sytems and Structures, vol. 3, 166–185, Jan. 1992. J. C. DOYLE & G. STEIN, Robustness with Observers, IEEE Transactions on Automatic Control, vol. AC-24, no. 4, 607–611, Aug. 1979. J. C. DOYLE & G. STEIN, Multivariable Feedback Design: Concepts for a Classical/Modern Synthesis, IEEE Transaction on Automatic Control, vol. AC-26, no. 1, 4–16, Feb. 1981. D. L. EDBERG, A. S. BICOS & J. S. FECHTER, On Piezoelectric Energy Conversion for Electronic Passive Damping Enhancement, Proceedings of Damping’91, San Diego, 1991. F. FAHY, Sound and Structural Vibration, Academic Press, 1987. J. L. FANSON, G. H. BLACKWOOD & C. C. CHEN, Active Member Control of Precision Structures, SDM Conference, AIAA paper 89-1329-CP, 1989. J. L. FANSON & T. K. CAUGHEY, Positive Position Feedback Control for Large Space Structures, AIAA Journal, vol. 28, no. 4, 717–724, April 1990. R. L. FORWARD, Electronic Damping of Orthogonal Bending Modes in a Cylindrical Mast Experiment, AIAA Journal of Spacecraft, vol. 18, no. 1, 11–17, Jan.–Feb. 1981. A. FRANÇOIS, P. DE MAN & A. PREUMONT, Piezoelectric Array Sensing of Volume Displacement: A Hardware Demonstration, J. of Sound and Vibration, vol. 244, no. 3, 395–405, July 2001. G. F. FRANKLIN & J. D. POWELL, Digital Control of Dynamic Systems, Addison-Wesley, 1980. G. F. FRANKLIN, J. D. POWELL & A. EMANI-NAEMI, Feedback Control of Dynamic Systems, Addison-Wesley, 1986. Y. FUJINO & T. SUSUMPOW, An experimental study on active control of planar cable vibration by axial support motion, Earthquake Engineering and Structural Dynamics, vol. 23, 1283–1297, 1994.

Bibliography

351

Y. FUJINO, P. WARNITCHAI & B. M. PACHECO, Active Stiffness Control of Cable Vibration, ASME J. of Applied Mechanics, 60, pp. 948–953, 1993. C. R. FULLER, S. J. ELLIOTT & P. A. NELSON, Active Control of Vibration, Academic Press, 1996. Y. C. FUNG, An Introduction to the Theory of Aeroelasticity, Dover, 1969. M. V. GANDHI & B. S. THOMPSON, Smart Materials and Structures, Chapman & Hall, 1992. P. GARDONIO, Y. LEE, S. ELLIOTT & S. DEBOST, Active Control of Sound Transmission Through a Panel with a Matched PVDF Sensor and Actuator Pair, Active 99, Fort Lauderdale, Florida, USA, Dec. 1999. L. GAUL & R. NITSCHE, Friction Control for Vibration Suppression, (DETC99/ VIB-8191) Movic Symposium, ASME Design Engineering Technical Conferences, Las Vegas, 1999. W. K. GAWRONSKI, Dynamics and Control of Structures - A Modal Approach, Springer, 1998. Z. J. GENG & L. S. HAYNES, Six Degree-of-Freedom Active Vibration Control Using the Stewart Platforms, IEEE Transactions on Control Systems Technology, vol. 2, no. 1, 45–53, March 1994. M. GERADIN & D. RIXEN, Mechanical Vibrations, Theory and Application to Structural Dynamics, Wiley, 1993. W. B. GEVARTER, Basic Relations for Control of Flexible Vehicles, AIAA Journal, vol. 8, no. 4, 666–672, April 1970. A. E. GLAZOUNOV, Q. M. ZHANG & C. KIM, A New Torsional Actuator Based on Shear Piezoelectric Response, SPIE Conference on Smart Structures and Materials, vol. 3324, 82–91, 1998a, san Diego, CA. A. E. GLAZOUNOV, Q. M. ZHANG & C. KIM, Piezoelectric Actuator Generating Torsional Displacement from the D15 Shear Strain, Appl. Phys. Lett. 72, 2526, 1998b. C. GOH & T. K. CAUGHEY, On the Stability Problem Caused by Finite Actuator Dynamics in the Control of Large Space Structures, Int. J. of Control, vol. 41, no. 3, 787–802, 1985. L. P. GRAYSON, The Status of Synthesis Using Lyapunov’s Method, Automatica, vol. 3, 91–121, 1965. E. A. GUILLEMIN, The Mathematics of Circuit Analysis, Wiley, 1949.

352


K. GUPTA, Frequency-Shaped Cost Functionals: Extension of Linear Quadratic Gaussian Methods, AIAA J. of Guidance and Control, vol. 3, no. 6, 529–535, Nov.–Dec. 1980. N. W. HAGOOD & E. F. CRAWLEY, Experimental Investigation of Passive Enhancement of Damping for Space Structures., AIAA Journal of Guidance, vol. 14, no. 6, 1100–1109, Nov.–Dec. 1991. W. L. HALLAUER JR, Recent Literature on Experimental Structural Dynamics and Control Research, in J. L. JUNKINS, editor, Mechanics and Control of Large Flexible Structures, chap. 18, AIAA Progress in Astronautics and Aeronautics, Vol. 129, 1990. A. M. A. HAMDAN & A. H. NAYFEH, Measure of Modal Controllability and Observability for First and Second Order Linear Systems, AIAA J. of Guidance, Control, and Dynamics, vol. 12, no. 5, 768, 1989. S. HANAGUD, M. W. OBAL & A. J. CALISE, Optimal Vibration Control by the Use of Piezoceramic Sensors and Actuators, AIAA Journal of Guidance, vol. 15, no. 5, 1199–1206, Sep.–Oct. 1992. H. HANSELMANN, Implementation of Digital Controllers - A Survey, Automatica, vol. 23, no. 1, 7–32, 1987. C. H. HANSEN & S. D. SNYDER, Active Control of Sound and Vibration, E&FN Spon, London, 1996. I. M. HOROWITZ, Synthesis of Feedback Systems, Academic Press, 1963. P. C. HUGHES, Attitude Dynamics of Three-Axis Stabilized Satellite with a Large Flexible Solar Array, J. Astronautical Sciences, vol. 20, 166–189, Nov.– Dec. 1972. P. C. HUGHES, Dynamics of Flexible Space Vehicles with Active Attitude Control, Celestial Mechanics Journal, vol. 9, 21–39, March 1974. P.O. HUGHES, Space Structure Vibration Modes: How Many Exist? Which Ones are Important?, IEEE Control Systems Magazine, February, 1987a. P. C. HUGHES & T. M. ABDEL-RAHMAN, Stability of Proportional Plus Derivative Plus Integral Control of Flexible Spacecraft, AIAA J. Guidance and Control, vol. 2, no. 6, 499–503, Nov.–Dec. 1979. T. J. R. HUGHES, The Finite Element Method, Linear Static and Dynamic Finite Element Analysis, Prentice-Hall, 1987b. W.-S. HWANG & H. C. PARK, Finite Element Modeling of Piezoelectric Sensors and Actuators, AIAA Journal, vol. 31, no. 5, 930–937, May 1993.

Bibliography

353

T. T. HYDE & E. H. ANDERSON, Actuator with Built-in Viscous Damping for Isolation and Structural Control, AIAA Journal, vol. 34, no. 1, 129–135, Jan. 1996. D. C. HYLAND, J. L. JUNKINS & R. W. LONGMAN, Active Control Technology for Large Space Structures, J. of Guidance, vol. 16, no. 5, 801–821, Sep.–Oct. 1993. IEEE Std, Standard on Piezoelectricity, ANSI/IEEE Std 176–1987, Jan. 1987. R. IKEGAMI & D. W. JOHNSON, The Design of Viscoelastic Passive Damping Treatments for Satellite Equipment Support Structures, Proceedings of DAMPING’86, AFWAL-TR-86-3059, 1986. D. J. INMAN, Vibration, with Control, Measurement, and Stability, PrenticeHall, 1989. L. B. JACKSON, Digital Filters and Signal Processing, Kluwer, 1986. A. D. JACOT & D. J. LISKA, Control Moment Gyros in Attitude Control, AIAA J. of Spacecraft and Rockets, vol. 3, no. 9, 1313–1320, Sep. 1966. H. JANOCHA, editor, Adaptronics and Smart Structures (Basics, Materials, Design and Applications), Springer, 1999. C. D. JOHNSON, D. A. KIENHOLZ & L. C. ROGERS, Finite Element Prediction of Damping in Beams with Constrained Viscoelastic Layers, Shock and Vibration Bulletin, , no. 51, 78–81, May 1981. M. E. JOHNSON & S. J. ELLIOTT, Active Control of Sound Radiation Using Volume Velocity Cancellation, J. of the Acoustical Society of America, vol. 98, 2174–2186, 1995. S. M. JOHSI, Control of Large Flexible Space Structures, Lecture Notes in Control and Information Sciences, Springer-Verlag, vol. 131, 1989. J. L. JUNKINS, editor, Mechanics and Control of Large Flexible Structures, vol. 129, 1990. J. L. JUNKINS & Y. KIM, Introduction to Dynamics and Control of Flexible Structures, AIAA Education Series, 1993. T. KAILATH, Linear Systems, Prentice-Hall, 1980. R. E. KALMAN & J. E. BERTRAM, Control System Analysis and Design Via the Second Method of Lyapunov (1. Continuous-Time Systems), ASME J. of Basic Engineering, pp. 371–393, June 1960.

354


C. E. KAPLOW & J. R. VELMAN, Active Local Vibration Isolation Applied to a Flexible Space Telescope, AIAA J. Guidance and Control, vol. 3, no. 3, 227–233, May–June 1980. D. KARNOPP, Design Principles for Vibration Control Systems Using SemiActive Dampers, Trans. ASME Journal of Dynamic Systems, Measurement and Control, vol. 112, 448–455, Sep. 1990. D. KARNOPP, M. CROSBY & R. A. HARWOOD, Vibration Control Using SemiActive Suspension Control, Journal of Engineering for Industry, vol. 96, 619– 626, 1974. D. C. KARNOPP & A. K. TRIKHA, Comparative Study of Optimization Techniques for Shock and Vibration Isolation, Trans. ASME, Journal of Engineering for Industry, series B, vol. 91, no. 4, 1128–1132, 1969. Y. KIM & J. L. JUNKINS, Measure of Controllability for Actuator Placement, AIAA J. of Guidance, Control, and Dynamics, vol. 14, no. 5, 895–902, Sep.– Oct. 1991.

G. J. KISSEL, The Bode Integrals and Wave Front Tilt Control, AIAA Guidance, Navigation & Control Conference, Portland Oregon, 1990. R. L. KOSUT, H. SALZWEDEL & A. EMANI-NAEMI, Robust Control of Flexible Spacecraft, AIAA J. of Guidance, vol. 6, no. 2, 104-111, March-April 1983. B. C. KUO, Digital Control Systems, SRL Pub. Co., Champaign, Illinois, 1977. H. KWAKERNAAK & R. SIVAN, Linear Optimal Control Systems, Wiley, 1972. R. A. LASKIN & S. W. SIRLIN, Future Payload Isolation and Pointing System Technology, AIAA J. Guidance and Control, vol. 9, no. 4, 469–477, July–Aug. 1986. C.-K. LEE, Theory of Laminated Piezoelectric Plates for the Design of Distributed Sensors/Actuators- Part I : Governing Equations and Reciprocal Relationships, J. Acoust.-Soc. Am., vol. 87, no. 3, 1144–1158, March 1990. C.-K. LEE, W.-W. CHIANG & T. C. O’SULLIVAN, Piezoelectric Modal Sensor/Actuator Pairs for Critical Active Damping Vibration Control, J. Acoust.Soc. Am., vol. 90, no. 1, 374–384, July 1991.

C.-K. LEE & F. C. MOON, Modal Sensors/Actuators, Trans. ASME, J. of Applied Mechanics, vol. 57, 434–441, June 1990. R. LERCH, Simulation of Piezoelectric Devices by Two- and Three-Dimensional Finite Elements, IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 37, no. 3, May 1990.

Bibliography

355

N. LOIX, A. CONDE REIS, P. BRAZZALE, J. DETTMANN & A. PREUMONT, CFIE: In-Orbit Active Damping Experiment Using Strain Actuators, Space Microdynamics and Accurate Control Symposium, Toulouse, May 1997. N. LOIX, J. KOZANEK & E. FOLTETE, On the Complex Zeros of Non-Colocated Systems, Journal of Structural Control, vol. 3, no. 1–2, June 1996. N. LOIX & A. PREUMONT, Remarks on the Modelling of Active Structures with Colocated Piezoelectric Actuators and Sensors, ASME Vibration Conference, Boston, vol. 84-3, 335–339, Sep. 1995. D. LUENBERGER, An Introduction to Observers, IEEE Trans. Autom. Control, AC-16, pp. 596–603, Dec. 1971. D. LUENBERGER, Introduction to Dynamic Systems, Wiley, 1979. B. J. LURIE & P. J. ENRIGHT, Classical Feedback Control (with MATLAB), Marcel Dekker Inc., 2000. J. M. MACIEJOWSKI, Multivariable Feedback Design, Addison-Wesley, 1989. G. D. MARTIN, On the Control of Flexible Mechanical Systems, Ph.D. thesis, Stanford University, 1978. J. E. MC INROY, G. W. NEAT & J. F. O’BRIEN, A Robotic Approach to Fault-Tolerant, Precision Pointing, IEEE Robotics Automation Magazine, pp. 24–31, Dec. 1999. L. MEIROVITCH, Methods of Analytical Dynamics, McGraw-Hill, 1970. L. MEIROVITCH, Computational Methods in Structural Dynamics, Sijthoff & Noordhoff, 1980. L. MEIROVITCH, Dynamics and Control of Structures, Wiley, 1990. D. K. MIU, Physical Interpretation of Transfer Function Zeros for Simple Control Systems with Mechanical Flexibilities, ASME J. Dynamic Systems Measurement and Control, vol. 113, 419–424, Sep. 1991. D. K. MIU, Mechatronics - Electromechanics and Contromechanics, SpringerVerlag, 1993. V. J. MODI, Attitude Dynamics of Satellites with Flexible Appendages - A Brief Review, J. Spacecraft and Rockets, vol. 11, 743–751, Nov. 1974. B.C. MOORE, Principal Component Analysis in Linear Systems: Controllability, Observability and Model Reduction, IEEE Trans. on Automatic Control, vol. AC-26, no. 1, 17–32, 1981.

356


V. MUKHOPADHYAY & J. R. NEWSOM, A Multiloop System Stability Margin Study Using Matrix Singular Values, AIAA J. of Guidance, vol. 7, no. 5, 582–587, Sep.–Oct. 1984. A. H. NAYFEH & D. T. MOOK, Nonlinear Oscillations, Wiley, 1979. G. W. N EAT, A. A BRAMOVICI, J. M. M ELODY, R. J. C ALVET, N. M. N ERHEIM & J. F. O’BRIEN, Control Technology Readiness for Spaceborne Optical Interferometer Missions, Proceedings SMACS-2, Toulouse, pp. 13–32, 1997. P. A. NELSON & S. J. ELLIOTT, Active Control of Sound, Academic Press, 1992. A. V. OPPENHEIM & R. W. SCHAFER, Digital Signal Processing, PrenticeHall, 1975. P. C. PARKS & V. HAHN, Stability Theory, Prentice Hall, 1993.

E. K. PARSONS, An Experiment Demonstrating Pointing Control on a Flexible Structure, IEEE Control Systems Magazine, pp. 79–86, April 1989. L. D. PETERSON, J. J. ALLEN, J. P. LAUFFER & A. K. MILLER, An Experimental and Analytical Synthesis of Controlled Structure Design, SDM Conference, AIAA paper 89-1170-CP, 1989. Philips Application Book on Piezoelectric Ceramics, (J. Van Randeraat & R. E. Setterington, Edts), Mullard Limited, London, 1974. Physik Instrumente catalogue, Products for Micropositioning (PI GmbH, FRG). V. PIEFORT, Finite Element Modeling of Piezoelectric Active Structures, Ph.D. thesis, Université Libre de Bruxelles, Active Structures Laboratory, 2001. A. PREUMONT, Active Structures for Vibration Suppression and Precision Pointing, Journal of Structural Control, vol. 2, no. 1, 49–63, June 1995. A. PREUMONT, Method for Shaping Laminar Piezoelectric Actuators and Sensors and Related Devices, European patent application No 01870068.2, March 2001. A. PREUMONT & Y. ACHKIRE, Active Damping of Cable Structures, ESA Conference on Spacecraft Structures Materials and Mechanical Testing, March, 1996. A. PREUMONT & Y. ACHKIRE, Active Damping of Structures with Guy Cables, AIAA, J. of Guidance, Control, and Dynamics, vol. 20, no. 2, 320–326, March-April 1997.

Bibliography

357

A. PREUMONT, Y. ACHKIRE & F. BOSSENS, Active Tendon Control of Large Trusses, AIAA Journal, vol. 38, no. 3, 493–498, March 2000. A. PREUMONT & F. BOSSENS, Active Tendon Control of Vibration of Truss Structures: Theory and Experiments, J. of Intelligent Material Systems and Structures, vol. 2, no. 11, 91–99, Feb. 2000. A. PREUMONT, P. DE MAN & A. FRANÇOIS, On the Feedback Control of a Baffled Plate, submitted to the Journal of Sound and Vibration, 2001. A. PREUMONT, J. P. DUFOUR & C. MALEKIAN, Active Damping by a Local Force Feedback with Piezoelectric Actuators, AIAA J. of Guidance, vol. 15, no. 2, 390–395, March–April 1992. A. PREUMONT, N. LOIX, D. MALAISE & O. LECRENIER, Active Damping of Optical Test Benches with Acceleration Feedback, Machine Vibration, vol. 2, 119–124, 1993. J. REX & S. J. ELLIOTT, The QWSIS - A New Sensor for Structural Radiation Control, MOVIC-1, Yokohama, September 1992. H. H. ROBERTSHAW, R. H. WYNN, S. L. HENDRICKS & W. W. CLARK, Dynamics and Control of a Spatial Active Truss Actuator, SDM Conference, AIAA paper 89-1328-CP,1989. R. H. SCANLAN & J. TOMKO, Airfoil and Bridge Deck Flutter Derivatives, ASCE J. Eng. Mech. div., 100, pp. 657–672, Aug. 1974. D. SCHAECHTER, Optimal Local Control of Flexible Structures, AIAA J. of Guidance, vol. 4, no. 1, 22–26, 1981. K. SETO, Trends in Active Control in Japan, MOVIC-1, Yokohama, Sep. 1992. J. R. SEZAC, P. LIKINS & T. CORADETTI, Flexible Spacecraft Control by Model Error Sensitivity Suppression, Proceedings of the VPI&SU/AIAA Symposium on Dynamics & Control of Large Flexible Spacecrafts, Blacksburg, VA, 1979. E. SIM & S. W. LEE, Active Vibration Control of Flexible Structures with Acceleration or Combined Feedback, AIAA J. of Guidance, vol. 16, no. 2, 413–415, 1993. N. D. SIMS, R. STANWAY & A. R. JOHNSON, Vibration Control Using SMART Fluids: A State of the Art Review, The Shock and Vibration Digest, vol. 31, no. 3, 195–203, May 1999. R. E. SKELTON, Dynamic System Control - Linear System Analysis and Synthesis, Wiley, 1988.

358


R. E. SKELTON, Model Error Concepts in Control Design, Int. J. of Control, vol. 49, no. 5, 1725–1753, 1989. R. E. SKELTON & P. C. HUGHES, Modal Cost Analysis for Linear Matrix Second-Order Systems, ASME J. of Dynamic Systems, Measurement, and Control, vol. 102, 151–158, Sep. 1980. J. SPANOS, Z. RAHMAN & G. BLACKWOOD, A Soft 6-Axis Active Vibration Isolator, Proceedings of the American Control Conference, Seattle, WA, pp. 412–416, June 1995. D. W. SPARKS JR & J. N. JUANG, Survey of Experiments and Experimental Facilities for Control of Flexible Structures, AIAAJ of Guidance, vol. 15, no. 4, 801–816, July–Aug. 1992. V. A. SPECTOR & H. FLASHNER, Sensitivity of Structural Models for Noncollocated Control Systems, ASME J. Dynamic Systems Measurement and Control, vol. 111, no. 4, 646–655, Dec. 1989. V. A. SPECTOR & H. FLASHNER, Modeling and Design Implications of Noncollocated Control in Flexible Systems, ASME J. Dynamic Systems Measurement and Control, vol. 112, 186–193, June 1990. B. F. SPENCER, S. J. DYKE, M. K. SAIN & J. D. CARLSON, Phenomenological Model of a Magnetorheological Damper, ASCE Journal of Engineering Mechanics, vol. 123, no. 3, 230–238, 1997. D. STEWART, A Platform with Six Degrees of Freedom, Proc. Instn. Mech. Engrs, vol. 180, no. 15, 371–386, 1965–66. G. STRANG, Linear Algebra and its Applications, New York: Harcourt Brace Jovanovich, third edition, 1988. D. THAYER, J. VAGNERS, A. VON FLOTOW, C. HARDHAM & K. SCRIBNER, Six-Axis Vibration Isolation System Using Soft Actuators and Multiple Sensors, Proc. of Annual American Astronautical Society Rocky Mountain Guidance and Control Conference (AAS-98-064), pp. 497–506, 1998. K. UCHINO, Ferroelectric Devices, Marcel Dekker, 2000. M. VIDYASAGAR, Nonlinear Systems Analysis, Prentice-Hall, 1978. J. R. VINSON, The Behavior of Shells Composed of Isotropic and Composite Materials, Kluwer Academic Publishers, 1993. P. WARNITCHAI, Y. FUJINO, B. M. PACHECO & R. AGRET, An Experimental Study on Active Tendon Control of Cable-Stayed Bridges, Earthquake Engineering and Structural Dynamics, vol. 22, no. 2, 93–111, 1993.

Bibliography

359

D. M. WIBERG, State Space and Linear Systems, McGraw-Hill Schaum's Outline Series in Engineering, 1971. R. N. WILSON, F. FRANZA & L. NOETHE, Active Optics. 1. A System for Optimizing the Optical Quality and Reducing the Costs of Large Telescopes, Journal of Modern Optics, vol. 34, no. 4, 485–509, 1987. J. N. YANG & F. GIANNOPOULOS, Active Control and Stability of Cable-Stayed Bridge, ASCE J. Eng. Mech. div., 105, pp. 677–694, Aug. 1979a. J. N. YANG & F. GIANNOPOULOS, Active Control of Two-Cable-Stayed Bridge, ASCE J. Eng. Mech. div., 105, pp. 795–810, Oct. 1979b. O. C. ZIENKIEWICZ &: R. L. TAYLOR, The Finite Element Method, McGrawHill, fourth ed., 1989, (2 vol.). D. C. ZIMMERMAN, G. C. HORNER & D. J. INMAN, Microprocessor Controlled Force Actuator, AIAA Journal of Guidance, vol. 11, no. 3, 230–236, MayJune 1988.


Index Clipped on/off, 274 Collocated (actuator/sensor). 26, 75, 91 Control canonical form, 228 Controllability, 221 matrix, 223 Gramian, 234, 243 Coupling coefficient (piezo), 42 Covariance (intensity) matrix, 196 Crossover frequency, 170, 177

Acceleration feedback, 96 ADC, 278, 280 Ackermann’s formula, 229, 242 Actuator/sensor dynamics, 107 Aliasing in time, 278 in space, 308 All-pass function, 183, 191 Anti-resonance, 27, 85, 95 Array sensor, 305 AS AC, 305 Asymptotic stability, 246, 251

DAC, 280 Damping active,8, 91, 93, 282 normal (classical), 20 passive (with piezos), 54 Rayleigh, 18, 20 typical values, 20 role of, 87, 91 Delay, 185, 191 Decentralized control, 94, 103 Diagonal form, 230 Direct Velocity Feedback, 93, 97 Distributed sensor, 302, 309 Duality, 205 Dynamic flexibility matrix, 21 DSP, 277

Bandwidth, 5, 9, 170 Bang-bang control, 257 Beam (Euler-Bernouilli), 30, 48, 55 Bessel filter, 319 BIBO (external) stability, 246 Bilinear system, 261 transform (Tustin), 281 Bingham model, 265 Bode Integrals, 178 Ideal Cutoff, 181 Butterworth pattern, 199 filter, 219, 279, 319

Eigenvectors (left and right), 229 Electro-rheological (ER) fluid, 7, 264, 265 ELSA facility, 338 Energy absorbing controller, 257

Cable-stayed bridge, 321 Cable structures, 321 Cauchy’s principle, 166 Cayley-Hamilton theorem, 223 CFIE experiment, 290, 317 Charge amplifier (dynamics), 57

Feedback control, 7, 171

361

362

Feedforward control, 9 Feedthrough, 22, 138 Finite element, 18, 68 Flipping (pole/zero), 76, 82, 160 Flutter, 321 Forgetting factor, 103, 284 Fraction of modal strain energy 61, 283 Frequency shaping, 212, 299 Gain margin, 165, 174 of the LQR, 201 Gain-phase relationships, 178 Gain stability, 84, 176 Global stability, 246 Gramians, 234, 243 Guyan reduction, 31 Guyed truss, 329 Gyrostabilizer, 40 HAC/LAC, 295 Instability theorem, 253 Integral Force Feedback, 103, 258, 325 Integral control(state feedback), 211 Interferometer, 3, 331 Internally balanced coordinates, 235 realization, 237 Isolation, 113 Kalman-Bucy filter (KBF), 153, 204 Kirchhoff shell, 65 Lag compensator, 189 Lasalle’s theorem, 251 Lead compensator, 79, 110, 187 Liapunov direct method, 249 indirect method, 255 function, 249, 254 equation, 193, 196, 234, 254 Limit cycle, 247, 259, 280


Linear combiner, 305 LQR, 148, 194, 197 LQG, 205, 299 Loop Transfer Recovery (LTR), 210

MATLAB, 13 Mechatronics, 12 Magneto-rheological (MR) fluid, 7, 263-266 Magnetostrictive materials, 7 Margin (gain & phase), 165, 174, 177, 201 Microprecision interferometer (JPL-MPI), 331 Mindlin shell, 70 Model reduction, 237 Mode shapes, 18, 30, 84 rigid body mode, 19 Modal filter, 53, 316 Modal fraction of strain energy 61, 283 Nichols chart, 170, 172

NITINOL, 6 Noise model, 215 Non-minimum phase, 86, 183 Notch filter, 77, 80, 160 Normandy bridge, 321 Nyquist stability criterion, 167 frequency, 278 Observability, 221 matrix, 224 Gramian, 234, 243 Observer, 151, 202 Luenberger, 151 minimum variance, 152 reduced order, 155 On-off control, 273 Padé approximants, 185, 191 Parametric excitation, 321, 324

Index

resonance, 336 Parseval’s theorem, 212 Passive isolator, 114 Popov-Belevitch-Hautus(PBH) test, 230, 233, 242 PD compensator, 187 Performance specification, 173 robust, 175 PI compensator, 187 PID compensator, 189 Phase change materials, 7 Phase margin, 165, 174, 177, 201 of the LQR, 201 Phase portrait, 246 Piezoelectric materials, 6 constitutive equations, 41, 44 properties, 45 6, 45 PZT, 6, 45, 48 Piezoceramics, 6 Piezopolymers, 6 Piezoelectric beam, 48, 52, 55, 303 Piezoelectric shell, 63, 291 Poles, 144 Pole placement, 145 Pole-zero flipping, 76, 82, 160 Polyvinylidene fluoride (PVDF6, 45 Porous electrode, 309, 313 Positive Position Feedback, 101, 291 Power spectral density matrix, 196 Proof-mass actuator, 38, 294 Quantization, 280 QWSIS, 303 Rayleigh damping, 18,20 Reaction wheel (RWA), 40, 124 Residual mode, 22, 57 Residues, 231 Return difference, 172, 177 Riccati equation, 195, 204 Robust (stability, performance), 175

363

Roll-off, 100, 101, 179, 295 Routh-Hurwitz criterion, 97, 248 Sampling, 278 Shannon’s (sampling) theorem, 278 Saturation function, 257 Sensitivity, 171, 233, 292 Sensitivity function, 172 Sensor dynamics, 107 Semi-active, 263, 267 Shape Memory Alloys (SMA), 6 Separation principle, 157, 205 SIMULINK, 13 Singular value decomposition, 307 Sky-hook, 116, 131 Smart materials, 6 Smart structures, 5 Sound power radiation, 302 Spillover, 9, 206 reduction, 53, 209 State transformation, 227 Stability asymptotic, 246, 251 external, 246 gain, 84, 176 global, 246 internal, 246 in the sense of Liapunov, 245 Nyquist criterion, 167 robust, 175 Routh-Hurwitz criterion, 248 theorem, 249 Stewart platform, 124, 286 Symmetric root locus, 149, 153 System type, 185 Takoma Narrows bridge, 322 Telescope, 2,3 Tendon control, 321 TERFENOL-D, 7 Truss (active), 58, 103, 282 Transmissibility, 114 Transmission zeros, 27, 58, 86, 129, 144

364

Tustin’s method, 281 Unstructured uncertainty, 174 Vandermonde matrix, 231 Van der Pol oscillator, 247, 253, 259 Volume displacement, 302 Volume velocity, 302 White noise, 196 Zeros, 27, 58, 86, 129, 144 Zero-order hold, 279, 320


Mechanics SOLID MECHANICS AND ITS APPLICATIONS

Series Editor: G.M.L. Gladwell 69.

70. 71. 72. 73.

74. 75. 76.

77.

78. 79. 80. 81. 82. 83. 84. 85. 86. 87. 88. 89.

P. Pedersen and M.P. Bendsøe (eds.): IUTAM Symposium on Synthesis in Bio Solid Mechanics. Proceedings of the IUTAM Symposium held in Copenhagen, Denmark. 1999 ISBN 0-7923-5615-2 S.K. Agrawal and B.C. Fabien: Optimization of Dynamic Systems. 1999 ISBN 0-7923-5681-0 A. Carpinteri: Nonlinear Crack Models for Nonmetallic Materials. 1999 ISBN 0-7923-5750-7 F. Pfeifer (ed.): IUTAM Symposium on Unilateral Multibody Contacts. Proceedings of the IUTAM Symposium held in Munich, Germany. 1999 ISBN 0-7923-6030-3 E. Lavendelis and M. Zakrzhevsky (eds.): IUTAM/IFToMM Symposium on Synthesis of Nonlinear Dynamical Systems. Proceedings of the IUTAM/IFToMM Symposium held in Riga, Latvia. 2000 ISBN 0-7923-6106-7 J.-P. Merlet: Parallel Robots. 2000 ISBN 0-7923-6308-6 J.T. Pindera: Techniques of Tomographic Isodyne Stress Analysis. 2000 ISBN 0-7923-6388-4 G.A. Maugin, R. Drouot and F. Sidoroff (eds.): Continuum Thermomechanics. The Art and Science of Modelling Material Behaviour. 2000 ISBN 0-7923-6407-4 N. Van Dao and E.J. Kreuzer (eds.): IUTAM Symposium on Recent Developments in Non-linear ISBN 0-7923-6470-8 Oscillations of Mechanical Systems. 2000 S.D. Akbarov and A.N. Guz: Mechanics of Curved Composites. 2000 ISBN 0-7923-6477-5 M.B. Rubin: Cosserat Theories: Shells, Rods and Points. 2000 ISBN 0-7923-6489-9 S. Pellegrino and S.D. Guest (eds.): IUTAM-IASS Symposium on Deployable Structures: Theory and Applications. Proceedings of the IUTAM-IASS Symposium held in Cambridge, U.K., 6–9 September 1998. 2000 ISBN 0-7923-6516-X A.D. Rosato and D.L. Blackmore (eds.): IUTAM Symposium on Segregation in Granular Flows. Proceedings of the IUTAM Symposium held in Cape May, NJ, U.S.A., June 5–10, 1999. 2000 ISBN 0-7923-6547-X A. Lagarde (ed.): IUTAM Symposium on Advanced Optical Methods and Applications in Solid Mechanics. Proceedings of the IUTAM Symposium held in Futuroscope, Poitiers, France, August 31–September 4, 1998. 2000 ISBN 0-7923-6604-2 D. Weichert and G. Maier (eds.): Inelastic Analysis of Structures under Variable Loads. Theory and Engineering Applications. 2000 ISBN 0-7923-6645-X T.-J. Chuang and J.W. Rudnicki (eds.): Multiscale Deformation and Fracture in Materials and Structures. The James R. Rice 60th Anniversary Volume. 2001 ISBN 0-7923-6718-9 S. Narayanan and R.N. lyengar (eds.): IUTAM Symposium on Nonlinearity and Stochastic Structural Dynamics. Proceedings of the IUTAM Symposium held in Madras, Chennai, India, 4–8 January 1999 ISBN 0-7923-6733-2 S. Murakami and N. Ohno (eds.): IUTAM Symposium on Creep in Structures. Proceedings of the IUTAM Symposium held in Nagoya, Japan, 3-7 April 2000. 2001 ISBN 0-7923-6737-5 W. Ehlers (ed.): IUTAM Symposium on Theoretical and Numerical Methods in Continuum Mechanics of Porous Materials. Proceedings of the IUTAM Symposium held at the University of Stuttgart, Germany, September 5-10,1999. 2001 ISBN 0-7923-6766-9 D. Durban, D. Givoli and J.G. Simmonds (eds.): Advances in the Mechanis of Plates and Shells The Avinoam Libai Anniversary Volume. 2001 ISBN 0-7923-6785-5 U. Gabbert and H.-S. Tzou (eds.): IUTAM Symposium on Smart Structures and Structonic Systems. Proceedings of the IUTAM Symposium held in Magdeburg, Germany, 26–29 September 2000. 2001 ISBN 0-7923-6968-8

Mechanics SOLID MECHANICS AND ITS APPLICATIONS Series Editor: G.M.L. Gladwell 90. 91.

92.

93. 94.

Y. Ivanov, V. Cheshkov and M. Natova: Polymer Composite Materials – Interface Phenomena & Processes. 2001 ISBN 0-7923-7008-2 R.C. McPhedran, L.C. Botten and N.A. Nicorovici (eds.): IUTAM Symposium on Mechanical and Electromagnetic Waves in Structured Media held in Sydney, NSW, Australia, 18-22 Januari 1999. 2001 ISBN 0-7923-7038-4 D.A. Sotiropoulos (ed.): IUTAM Symposium on Mechanical Waves for Composite Structures Characterization. Proceedings of the IUTAM Symposium held in Chania, Crete, Greece, June 14-17, 2000. 2001 ISBN 0-7923-7164-X V.M. Alexandrov and D.A. Pozharskii: Three-Dimensional Contact Problems. 2001 ISBN 0-7923-7165-8 J.P. Dempsey and H.H. Shen (eds.): IUTAM Symposium on Scaling Laws in Ice Mechanics and Ice Dynamics. held in Fairbanks, Alaska, U.S.A., 13-16 June 2000. 2001 ISBN 1-4020-0171-1

Kluwer Academic Publishers – Dordrecht / Boston / London

Vibration Control of Active Structures: An Introduction (Solid Mechanics and Its Applications)

Vibration Control of Active Structures (Solid Mechanics and Its Applications)

Vibration Control of Active Structures: An Introduction, 3rd Edition

Fracture Mechanics: An Introduction (Solid Mechanics and Its Applications)

Active Control of Vibration

An introduction to structural optimization (Solid Mechanics and Its Applications)

Active Control of Structures

Elasticity of Transversely Isotropic Materials (Solid Mechanics and Its Applications) (Solid Mechanics and Its Applications)

Design-Oriented Analysis of Structures: A Unified Approach (Solid Mechanics and Its Applications, Volume 95) (Solid Mechanics and Its Applications)

Elasticity (Solid Mechanics and Its Applications)

Elasticity (Solid Mechanics and Its Applications)

Non-Linear Mechanics of Materials (Solid Mechanics and Its Applications)

Elasticity of Transversely Isotropic Materials (Solid Mechanics and Its Applications)

Optimal Control An Introduction to the Theory and Its Applications

The Behavior of Structures Composed of Composite Materials (Solid Mechanics and Its Applications)

Structural Analysis: With Applications to Aerospace Structures (Solid Mechanics and Its Applications)

IUTAM Symposium on Transformation Problems in Composite and Active Materials (Solid Mechanics and Its Applications)

Signal Processing for Active Control (Signal Processing and its Applications)

Random Vibration of Structures

Flexible Multibody Dynamics (Solid Mechanics and Its Applications 176)

Structural Optimization with Uncertainties (Solid Mechanics and Its Applications)

A Modern Course in Aeroelasticity (Solid Mechanics and Its Applications)

Random Vibration of Structures

Wave Mechanics and Its Applications

MATLAB Codes for Finite Element Analysis: Solids and Structures (Solid Mechanics and Its Applications)

Functional Analysis: Applications in Mechanics and Inverse Problems (Solid Mechanics and Its Applications)

Continuum and solid mechanics : concepts and applications

An Introduction to Fluid Mechanics and Transport Phenomena (Fluid Mechanics and Its Applications)

An Introduction to Fluid Mechanics and Transport Phenomena (Fluid Mechanics and Its Applications)

Advanced Structural Dynamics and Active Control of Structures