Adaptive Multi Mode Vibration Control of Dynamically Loaded Flexible Structures PDF

Adaptive Multi Mode Vibration Control
of Dynamically Loaded Flexible

Structures
by
Hendra Tjahyadi, B.Eng.(Electonics), M.Eng.(Instrumentation and
Control)
School of Informatics and Engineering,
Faculty of Science and Engineering
July , 2006
A thesis presented to the
Flinders University of South Australia
in total fulllment of the requirements for the degree of
Doctor of Philosophy
Adelaide, South Australia, 2007
c (Hendra Tjahyadi, 2007)
Contents
Abstract xv
List of Abbreviations xix
Certication xx
Acknowledgements xxi
Publication in Support of the Thesis xxiii
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research Methodology . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Control of Vibration . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Active Vibration Control . . . . . . . . . . . . . . . . . . . . . . . 8
1.4.1 Feedforward and Feedback Control . . . . . . . . . . . . . 8
1.4.2 Wave Control and Modal Control . . . . . . . . . . . . . . 9
1.5 Modal Based Controllers for Multi-mode Vibration Control . . . . 10
1.6 Control Methods for Systems with Varying Parameters . . . . . . 13
1.7 Control Methods for Systems with Uncertainties . . . . . . . . . . 16
ii
CONTENTS iii
1.7.1 Robust Control . . . . . . . . . . . . . . . . . . . . . . . . 16
1.7.2 Adaptive Control . . . . . . . . . . . . . . . . . . . . . . . 17
1.8 Natural Frequency Estimator . . . . . . . . . . . . . . . . . . . . 20
1.9 Multiple Model Adaptive Control . . . . . . . . . . . . . . . . . . 22
1.10 Aim of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.11 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.12 Original Contributions to the Thesis . . . . . . . . . . . . . . . . 27
2 Modelling of Flexible Structures 30
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2 Description of Experimental Plant (Experimental Model) . . . . . 32
2.3 Analytical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3.1 Flexural Vibration of Beams . . . . . . . . . . . . . . . . . 37
2.3.2 Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . . 41
2.4 Modal Analysis Using ANSYS (Numerical Models) . . . . . . . . 60
2.5 Simulation Models . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3 Multiple Model Resonant Control 74
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.2 Structure of a Resonant Controller . . . . . . . . . . . . . . . . . 76
3.3 Discrete-time Resonant Control . . . . . . . . . . . . . . . . . . . 82
3.3.1 Input-output Stability . . . . . . . . . . . . . . . . . . . . 84
3.3.2 Stability of a Discrete-time Resonant Control System . . . 86
CONTENTS iv
3.4 Multiple Model Control . . . . . . . . . . . . . . . . . . . . . . . . 88
3.5 Multi-model Multi-mode Resonant Control (M
4
RC) . . . . . . . . 94
3.5.1 Case 1: All the Possible Loading Condition are a priori
Known . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.5.2 Case 2: Only the upper and Lower Bounds of Operating
Region are a priori Known . . . . . . . . . . . . . . . . . . 100
3.6 Simulation Studies . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.6.1 Resonant Controller . . . . . . . . . . . . . . . . . . . . . 103
3.6.2 M
4
RC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.7 Experimental Studies . . . . . . . . . . . . . . . . . . . . . . . . . 118
3.7.1 Resonant Controller . . . . . . . . . . . . . . . . . . . . . 119
3.7.2 M
4
RC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
3.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4 Natural Frequency Estimator 133
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.2 On-line Parameter Estimation Using RLS . . . . . . . . . . . . . 141
4.3 Stability of the RLS Algorithm . . . . . . . . . . . . . . . . . . . 148
4.4 Factors Which Inuence the Accuracy of RLS . . . . . . . . . . . 150
4.4.1 RLS Characteristics . . . . . . . . . . . . . . . . . . . . . . 151
4.4.2 Parameter Tracking Resolution . . . . . . . . . . . . . . . 155
4.4.3 Parameter Perturbation Error . . . . . . . . . . . . . . . . 158
4.5 Design of the Natural Frequency Estimator for Flexible Structures 159
4.6 Simulation Studies of the Proposed Natural Frequency Estimator 164
CONTENTS v
4.6.1 Eects of Preltering and T Selection . . . . . . . . . . . . 165
4.6.2 Natural Frequency Estimator for Cantilever Beam Models 168
4.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
5 Adaptive Resonant Control 180
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
5.2 Adaptive Resonant Control (ARC) . . . . . . . . . . . . . . . . . 183
5.3 Multi-model Multi-mode Adaptive Resonant Control (M
4
ARC) . 187
5.4 Simulation Studies of ARC and M
4
ARC . . . . . . . . . . . . . . 192
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
6 Summary, Conclusion and Future Work 207
6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
6.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
6.3 Recommendations for Future Work . . . . . . . . . . . . . . . . . 215
A Simulation Models 218
B Passivity 227
C Simulink
TM
Models 229
C.1 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
C.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
Bibliography 245
List of Figures
1.1 The twisted roadway of the Tacoma Narrows bridge before its failure. 2
1.2 Block diagram of gain scheduling control . . . . . . . . . . . . . . 14
1.3 Block diagram of the Self Tuning Regulator scheme . . . . . . . . 18
1.4 Block diagram of Model Reference Adaptive System scheme . . . 19
2.1 Plant with a collocated sensor-actuator pair . . . . . . . . . . . . 34
2.2 Loading model congurations . . . . . . . . . . . . . . . . . . . . 35
2.3 Change of frequency responses for various loading conditions . . . 37
2.4 A beam in exural vibration . . . . . . . . . . . . . . . . . . . . . 39
2.5 A small element of the beam . . . . . . . . . . . . . . . . . . . . 39
2.6 A cantilever beam . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.7 A cantilever beam with n attached masses . . . . . . . . . . . . . 51
2.8 Frequency responses of simulation models . . . . . . . . . . . . . 63
2.9 Alternating pole-zero pattern of exible structures with collocated
sensor and actuator . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.10 Comparative responses of the three-mode model and the ten-mode
model of Model 1 with white noise input . . . . . . . . . . . . . . 66
model of Model 1 with pulse signal input . . . . . . . . . . . . . . 67
vi
LIST OF FIGURES vii
model of Model 4 with white noise input . . . . . . . . . . . . . . 68
model of Model 4 with pulse signal input . . . . . . . . . . . . . . 69
2.14 Comparative responses of real plant and simulation model for Model
1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.18 Mode shapes for Model 1 . . . . . . . . . . . . . . . . . . . . . . . 71
3.1 Frequency response of a exible cantilever beam . . . . . . . . . . 75
3.2 Flexible structure control system . . . . . . . . . . . . . . . . . . 77
3.3 Block diagram of resonant control . . . . . . . . . . . . . . . . . . 77
3.4 Frequency response of a dual mode resonant controller . . . . . . 78
3.5 Closed-loop responses for variations in a plant natural frequencies 81
3.6 Closed-loop responses for variations in a plant damping factors . . 81
3.7 A canonical feedback system . . . . . . . . . . . . . . . . . . . . . 85
3.8 Multiple model control method using weighting function scheme . 90
LIST OF FIGURES viii
3.9 Multiple model control method using supervisor scheme . . . . . . 91
3.10 Filter bank system for the m
th
mode . . . . . . . . . . . . . . . . 96
3.11 Switching system for the m
th
mode . . . . . . . . . . . . . . . . . 97
3.12 Block diagram of M
4
RC for controlling M modes . . . . . . . . . 98
3.13 Comparative responses of BPF with resonant controller structure
with Butterworth BPF . . . . . . . . . . . . . . . . . . . . . . . . 99
3.14 Frequency response of resonant controller with
c
= 0.01 and k
d
=
10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.15 Model array in the M
4
RC model bank . . . . . . . . . . . . . . . 102
3.16 Response of Model 1 and the corresponding control signal . . . . 106
3.17 Response of Model 2 and the corresponding control signal . . . . . 106
3.20 Frequency response of Model 1 . . . . . . . . . . . . . . . . . . . . 108
3.24 Frequency response of Model 1 with only the 2
nd
mode controller
active . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.25 Open-loop system response for the 1 3 4 model sequence . . 111
3.26 Closed-loop system response for the 1 3 4 model sequence . 112
3.27 Schematic diagram of the implemented M
4
RC . . . . . . . . . . . 113
3.28 Schematic diagram of the multi-model control with MMSE super-
visor scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
LIST OF FIGURES ix
3.29 M
4
RC switching behaviour for the 1 3 4 model sequence . . 115
3.30 Closed-loop multiple model resonant control responses for the 1
3 4 model sequence . . . . . . . . . . . . . . . . . . . . . . . . 115
3.31 M
4
3.32 Control signals generated by the multiple model resonant control
for the 1 2 4 model sequence . . . . . . . . . . . . . . . . . 117
3.33 Closed-loop responses of multiple model resonant control for the 1
2 4 model sequence . . . . . . . . . . . . . . . . . . . . . . . 117
3.34 The experimental set-up . . . . . . . . . . . . . . . . . . . . . . . 119
3.43 Frequency response of Model 1 with only the 2
nd
mode controller
active . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3.44 Open-loop system response for the 1 3 4 model sequence . . 125
3.45 Closed-loop system response for the 1 3 4 model sequence
with the controller designed based on Model 3 . . . . . . . . . . . 126
3.46 M
4
RC closed-loop response for the 1 3 4 model sequence . . 127
3.47 M
4
LIST OF FIGURES x
3.48 M
4
RC closed-loop response for the 1 2 4 model sequence . . 129
3.49 M
4
3.50 M
4
RC closed-loop response for the 1 4 model sequence . . . . . 130
3.51 M
4
RC switching behaviour for the 1 4 model sequence . . . . . 131
4.1 Block diagram of an on-line parameter estimator . . . . . . . . . 142
4.2 Recursive structure of parameter adaptation algorithm . . . . . . 142
4.3 Equivalent feedback representation of RLS . . . . . . . . . . . . . 150
4.4 Frequency responses of

A(e
j
) for dierent order . . . . . . . . . . 154
4.5 Eect of sampling period on pole location . . . . . . . . . . . . . 157
4.6 Estimation result without preltering . . . . . . . . . . . . . . . 166
4.7 Estimation result with preltering using single LPF . . . . . . . . 167
4.8 Estimation result with preltering using BPFs . . . . . . . . . . . 167
4.9 Estimation result with preltering using BPFs and dierent sam-
pling period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
4.10 Schematic diagram for natural frequency estimator . . . . . . . . 169
4.11 Estimation results for Model 1 . . . . . . . . . . . . . . . . . . . . 171
4.12 Magnied steady-state results for Model 1 estimation . . . . . . . 171
4.13 Estimation results for Model 4 . . . . . . . . . . . . . . . . . . . . 172
4.15 Estimation results for the 1 3 4 load sequence . . . . . . . . 173
4.16 Estimation result for the 1 2 4 load sequence . . . . . . . . 173
LIST OF FIGURES xi
5.1 Block diagram of the MMAC . . . . . . . . . . . . . . . . . . . . 182
5.2 Block diagram of the ARC for controlling three modes. . . . . . . 183
5.3 Block diagram of the M
4
ARC for controlling three modes . . . . . 188
5.4 Frequency response of resonant controller with
c
= 0.05 and k
d
=
10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
5.5 Model array in the M
4
ARC model bank . . . . . . . . . . . . . . 191
5.6 Simulation responses of the (a) M
4
RC, (b) ARC and (c) M
4
ARC
5.7 M
4
ARC switching behaviour for the 1 2 4 model sequence . 195
4
4
ARC
4
4
ARC
5.10 M
4
5.11 M
4
5.12 Responses of the (a) M
4
4
ARC for the 1
2 4 model sequence . . . . . . . . . . . . . . . . . . . . . . . . 201
5.13 M
4
4
4
ARC for the 1
3 4 model sequence . . . . . . . . . . . . . . . . . . . . . . . . 203
LIST OF FIGURES xii
4
4
ARC for the 1
4 model sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
5.16 M
4
5.17 M
4
ARC switching behaviour for the 1 4 model sequence . . . . 204
C.1 Schematic diagram of adap sim model . . . . . . . . . . . . . . . . 232
C.2 Detail schematic diagram of the frequency estimator block . . . . 233
C.3 Detail schematic diagram of the controller block . . . . . . . . . . 234
C.4 Schematic diagram of m4rc sim model . . . . . . . . . . . . . . . 236
C.5 Schematic diagram of lter bank and switching system block . . . 237
C.6 Schematic diagram of narendra sim model . . . . . . . . . . . . . 238
C.7 Schematic diagram of m4arc sim model . . . . . . . . . . . . . . . 239
C.8 Detail of lter bank and switching systems block of m4arc sim model240
C.9 Schematic diagram of adapt exp model . . . . . . . . . . . . . . . 242
List of Tables
2.1 Properties of the beam . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2 Model parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3 The rst three natural frequencies of the experimental models . . 36
2.4 The rst three frequency equations roots of the models . . . . . . 58
2.5 Comparison of natural frequencies obtained from experimental and
analytical results . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.6 Comparison of natural frequency results from ANSYS with results
from analytical method . . . . . . . . . . . . . . . . . . . . . . . . 61
3.1 Models parameters . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.2 Plant and controller congurations for resonant controller simula-
tion study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
3.3 Attenuation level for the range of models . . . . . . . . . . . . . . 105
3.4 Plant and controller congurations for M
4
RC simulation study . . 113
3.5 Attenuation level for the range of models . . . . . . . . . . . . . . 120
5.1 Plant and controller conguration for three dierent simulation
study cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
5.2 Maximum overshoot percentage and settling time of ARC and
M
4
ARC for dierent loading changes . . . . . . . . . . . . . . . . 196
xiii
LIST OF TABLES xiv
5.3 Maximum overshoot percentage and settling time of ARC and
M
4
ARC for dierent loading changes . . . . . . . . . . . . . . . . 202
A.1 The rst ten natural frequencies, eigenfunctions and damping ra-
tios of Model 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
tios of Model 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
tios of Model 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
tios of Model 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
A.5 The transfer function coecients and pole positions of Model 1 . . 223
C.1 C S-functions that are used for simulation and experimental im-
plementations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
C.2 Simulink model for simulations. . . . . . . . . . . . . . . . . . . . 231
C.3 Simulink model for experimental implementation . . . . . . . . . 241
Abstract
In this thesis, three control methodologies are proposed for suppressing multi-
mode vibration in exible structures. Controllers developed using these methods
are designed to (i) be able to cope with large and sudden changes in the systems
parameters, (ii) be robust to unmodelled dynamics, and (iii) have a fast transient
response. In addition, the controllers are designed to employ a minimum number
of sensor-actuator pairs, and yet pose a minimum computational demand so as
to allow real-time implementation.
A cantilever beam with magnetically clamped loads is designed and con-
structed as the research vehicle for evaluation of the proposed controllers. Using
this set-up, sudden and large dynamic variations of the beam loading can be
tested, and the corresponding changes in the plants parameters can be observed.
Modal testing reveals that the rst three modes of the plant are the most sig-
nicant and need to be suppressed. It is also identied that the rst and third
modes are spaced more than a decade apart in frequency. The latter characteris-
tic increases the diculty of eectively controlling all three modes simultaneously
using one controller. To overcome this problem, the resonant control method is
chosen as the basis for the control methodologies discussed in this thesis.
The key advantage of resonant control is that it can be tuned to provide
specic attenuation only at and immediately close to the resonant frequency
of concern. Consequently, it does not cause control spillover to other modes
owing to unmodeled dynamics. Because of these properties, a resonant controller
xv
LIST OF TABLES xvi
can be congured to form a parallel structure with the objective of targeting
and cancelling multiple modes individually. This is possible regardless of the
mode spacing. In addition, resonant control requires only a minimum number
of collocated sensor-actuator pairs for multi-mode vibration cancellation. All
these characteristics make resonant control a suitable candidate for multi-mode
vibration cancellation of exible structures.
Since a resonant controller provides negligible attenuation away from the nat-
ural frequencies that it has been specically designed for, it is very sensitive to
changes of a systems natural frequencies and becomes ineective when these
mode frequencies change. Hence, for the case of a dynamically loaded structure
with consequent variations in mode frequencies, the resonant control method must
be modied to allow tracking of system parameter changes. This consideration
forms the theme of this thesis, which is to allow adaptive multi-mode vibra-
tion control of dynamically-loaded exible structures. Three controller design
methodologies based on the resonant control principle are consequently proposed
and evaluated.
In the rst approach, all possible loading conditions are assumed to be a priori
known. Based on this assumption, a multi-model multi-mode resonant control
(M
4
RC) method is proposed. The basis of the M
4
RC approach is that it comprises
a bank of known loading models that are designed such that each model gives op-
timum attenuation for a particular loading condition. Conceptually, each model is
implemented as a set of xed-parameter controllers, one for each mode of concern.
In reality, each mode controller is implemented as an adjustable resonant con-
troller that is loaded with the xed-model parameters of the corresponding mode.
The M
4
RC method takes advantage of the highly frequency-sensitive nature of
resonant control to allow simple and rapid selection of the optimum controller.
Identication of the set of resonant frequencies is implemented using a bank of
LIST OF TABLES xvii
band-pass lters that correspond to the mode frequencies of the known mod-
els. At each time interval a supervisor scheme determines for each mode which
model has the closest frequency to the observed vibration frequency and switches
the corresponding model controller output to attenuate the mode. Selection is
handled on a mode-by-mode basis, such that for each mode the closest model
is selected. The proposed M
4
RC is relatively simple and less computationally
complex compared to other multi-model methods reported in the literature. In
particular, the M
4
RC uses a simple supervisor scheme and requires only a single
controller per mode. Other multi-model methods use more complex supervision
schemes and require one controller per model. The M
4
RC method is evaluated
through both simulation and experimental studies. The results reveal that the
proposed M
4
RC is very eective for controlling multi-mode vibration of a exible
structure with known loading conditions, but is ineective for unmodeled loading
conditions.
In the second approach, the assumption that all loading conditions are a priori
known is relaxed. An adaptive multi-mode resonant control (ARC) method is
proposed to control the exible structure for all possible (including unknown)
loading conditions. On-line estimation of the structures natural frequencies is
used to update the adaptive resonant controllers parameters. The estimation of
the natural frequencies is achieved using a parallel set of second-order recursive
least-squares estimators, each of which is designed for a specic mode of concern.
To optimise the estimation accuracy for each mode frequency, a dierent sampling
rate suitable for that mode is used for the corresponding estimator. Simulation
and experiment results show that the proposed adaptive method can achieve
better performance, as measured by attenuation level, over its xed-parameter
counterpart for a range of unmodeled dynamics. The results also reveal that,
for the same sequences of known loading changes, the transient responses of the
ARC are slower than those of the M
4
RC.
LIST OF TABLES xviii
In the third approach, a hybrid multi-model and adaptive resonant control is
utilized to improve the transient response of the ARC. The proposed multi-model
multi-mode adaptive resonant control (M
4
ARC) method is designed as a combi-
nation of the M
4
RC and ARC methods. The basis of the proposed method is to
use the M
4
RC xed-parameter model scheme to deal with transient conditions
while the ARC adaptive parameter estimator is still in a state of uctuation.
Then, once the estimator has reached the vicinity of its steady-state, the adap-
tive model is switched in place of the xed model to achieve optimum control
of the unforeseen loading condition. Whenever a loading change is experienced,
the simple M
4
RC supervisor scheme is used to identify the closest model and
to load the adjustable resonant controllers with the xed parameters for that
model. Meanwhile, the mode estimators developed for the ARC method are used
to identify the exact plant parameters for the modes of concern. As soon as these
parameters stop rapidly evolving and reach their steady-state, they are loaded
into the respective adjustable controllers. The same process is repeated whenever
a loading change occurs. Given the simplicity of the M
4
ARC method and its min-
imal computation demand, it is easily applicable for real-time implementation.
Simulation and experiment results show that the proposed M
4
ARC outperforms
both the ARC with respect to transient performance, and the M
4
RC with respect
to unmodeled loading conditions.
The outcomes of this thesis provide a basis for further development of the
theory and application of active control for exible structures with unforeseen
conguration variations. Moreover, the basis for the proposed multi-model adap-
tive control can be used in other areas of control (not limited to vibration can-
cellation) where fast dynamic reconguration of the controller is necessary to
accommodate structural changes and uctuating external disturbances.
List of Abbreviations
ADC Analog to Digital Converter
AIS Adaptive Input Shaping
ARC Adaptive Resonant Control
ARMAX Auto Regressive Moving Average with eXternal input
ARX Auto Regressive with eXternal input
BIBO Bounded Input Bounded Output
BPF Band Pass Filter
DAC Digital to Analog Converter
FEM Finite Element Method
FFT Fast Fourier Transform
FRF Frequency Response Function
HPF High Pass Filter
IIR Innite Impulse Response
IMSC Independent Modal Space Control
LHP Left Half Plane
LPF Low Pass Filter
M
4
ARC Multi-Model Multi-Mode Adaptive Resonant Control
M
4
RC Multi-Model Multi-Mode Resonant Control
MIMO Multi Input Multi Output
MMAC Multiple Model Adaptive Control
MMC Multiple Model Control
MMSE Minimum Mean Squares Error
MRAS Model Reference Adaptive System
ODE Ordinary Dierential Equation
PDE Partial Dierential Equation
PLL Phase Locked Loop
PPF Positive Position Feedback
RLS Recursive Least Squares
SDOF Single Degree of Freedom
SISO Single Input Single Output
STR Self Tuning Regulator
xix
Certication
I certify that this thesis does not incorporate without acknowledgement any ma-
terial previously submitted for a degree or diploma in any university; and that to
the best of my knowledge and belief it does not contain any material previously
published or written by another person except where due reference is made in the
text.
As requested under Clause 14 of Appendix D of the Flinders University Re-
search Higher Degree Student Information Manual I hereby agree to waive the
conditions referred to in Clause 13(b) and (c), and thus
Flinders University may lend this thesis to other institutions or individuals
for the purpose of scholarly research;
Flinders University may reproduce this thesis by photocopying or by other
means, in total or in part, at the request of other institutions or individuals
for the purpose of scholarly research.
Adelaide, 26 July 2006
Hendra Tjahyadi
xx
Acknowledgements
It is a real privilege for me to work under the supervision of Associate Prof.
Fangpo He and Associate Prof. Karl Sammut. Their valuable inputs and com-
ments always encourage me to work better, think deeper and broader, and to
write clearer. Without their constructive criticism and guidance, I believe my
thesis will not be in a good shape. To Fangpo, thank you for long hours dis-
cussions and debates, and your intriguing parables. To Karl thank you for your
encouragement and patience, and for help me dealing with the bureaucracy.
I want to give thanks to all my friends who help me and my family during my
study in Adelaide. My deep gratitude goes to the member of Ascot Community
Uniting Church for your supports and prayers. To Alex Yates who thoroughly
read my thesis draft and give a very helpful recommendations. Thanks for your
willingness to help and your availability for long discussions. My special thanks
to reverend John Blanskby, who become my mentor and my wonderful friend
during my stay in Adelaide. For my friends in Engineering school Ning, May and
Tae Hwan thanks for our friendships.
This research has been undertaken with the support of the AusAid scholarship
and Flinders University Postgraduate Scholarship that is gratefully acknowledged.
My sincere appreciation for Elaine Kane, the AusAid liaison ocer in Flinders
University, for her helps and supports.
To my wife Mega and my daughters Clarice and Christy to whom I mostly in
debt, thank you so much for your constant love, prayer, support, patience, and
xxi
LIST OF TABLES xxii
encouragement during the preparation of this thesis.
Finally, I want to thanks Jesus Christ my saviour, to be with me wherever I
go and even carry me when I pass the desert part of my life journey.
Publications in Support of the
Thesis
The major contributions presented in this thesis have been peer reviewed and
reported in the following refereed publications:
1. Tjahyadi, H., He, F., and Sammut, K. Vibration control of a cantilever
beam using multiple model adaptive control. In Proceedings of the Ameri-
can Control Conference ACC2004 (Boston, Massachusetts, July 2004), pp.
2907-2908.
2. Tjahyadi, H., He, F., and Sammut, K. Vibration control of a cantilever beam
using adaptive resonant control. In Proceedings of the 5th Asian Control
Conference ASCC2004 (Melbourne, Australia, July 2004), pp. 1786-1790.
3. Tjahyadi, H., He, F., and Sammut, K. Vibration control of a cantilever beam
using multiple model adaptive resonant control. In Proceedings of IFAC
Workshop on Adaptation and Learning in Control and Signal Processing
(ALCOSP04)(Yokohama, Japan, September 2004), pp. 409-413.
4. Tjahyadi, H., He, F., and Sammut, K. Multi-mode vibration control of a
exible cantilever beam using adaptive resonant control. Smart Materials
and Structures, 15:270-278, 2006.
5. Tjahyadi, H., He, F., and Sammut, K. M
4
ARC: Multi-Model-Multi-Mode
Adaptive Resonant Control for Dynamically-Loaded Flexible Beam Struc-
xxiii
LIST OF TABLES xxiv
tures. In preparation to be submitted to IEEE Trans. Control Systems
Technology
Chapter 1
Introduction
This chapter discusses the motivation for the research described in this thesis.
The research methodology is presented, followed by a literature review. Finally,
an outline of the thesis is given along with a list of original contributions.
1.1 Motivation
The quest to improve the dynamic performance, the operating eciency, and the
amount of material that is used in systems that employ mechanical structures, has
prompted many designers to employ lightweight materials and to reduce the cross
sectional dimensions of those structures. The benets of using lightweight mate-
rials and reduced cross sectional dimensions include: reduced weight, decreased
wind resistance, reduced energy requirements, increased acceleration, and a de-
creased in the physical space requirements. Some practical outcomes resulting
from these benets include: the enhanced manoeuvrability of robotic arms, the
increased range of aircraft, and the reduced fuel consumption of vehicles.
However, one side-eect of employing lightweight materials and reducing the
cross sectional dimensions is that the structures become more exible. Flexible
structures are more susceptible to the detrimental eects of unwanted vibration,
particularly when they operate at or near their natural frequencies or when they
1
CHAPTER 1. INTRODUCTION 2
are excited by disturbances that coincide with their natural frequencies.
The enhanced vibration of a structure subjected to an oscillating force close to
a natural frequency of the structure is known as resonance. For many structures,
operating at or near resonance is undesirable because it can degrade performance,
reduce lifetime, and even lead to rapid destruction. A paradigm for this behaviour
is the failure of the Tacoma Narrows Bridge in 1940. The shallow and narrow span
of the bridge made it exible and the deck of the bridge acted like an aircraft wing
in uncontrolled turbulence when it was subjected to an unpredicted wind pattern
[109]. This one example shows how important it is to attenuate or control the
vibration in exible structures, especially those that occur at or near resonance.
Figure 1.1: The twisted roadway of the Tacoma Narrows Bridge before
its failure.
Rather than sacrice the gains to be made by reducing the mass of structures,
interest has centred on means of reducing vibration when it occurs. Particular
emphasis has been placed on active control, in which anti-phase excitation is
introduced in order to destructively interfere with vibration frequencies.
The design and implementation of a high performance vibration controller for
a exible structure can be a dicult task. The diculty is due to the following
factors:
1. The order of a exible structure can be very large and mostly unknown [52].
Structures with large orders have many modes of vibration and controlling
all of the modes is, in practice, impossible. Therefore, controllers are usu-
ally designed to attenuate only the most signicant or dominant vibration
modes. In the design of a system that consists of a controller and a exible
structure, it is common practice to approximate the actual exible struc-
ture with a low order approximation model [96]. This approximation is
achieved by removing all the modes that lie outside the bandwidth of inter-
est. However, the use of a low order approximation model can cause system
uncertainties due to unmodeled dynamics. These unmodeled dynamics can
lead to system instability due to spillover, which is any unwanted vibra-
tion associated with other modes. There are two types of spillover: control
spillover and observation spillover. Control spillover occurs when the control
force excites unmodeled dynamics. The excitation of these unmodeled dy-
namics can degrade the performance of the system. Observation spillover
refers to the measurement error caused by the contribution of excluded
modes to the sensor measurements. While control spillover leads to poor
performance, observation spillover can lead to instability, as shown in the
numerical simulation by Balas [8].
2. To control a exible structure that has widely separated multi-mode vi-
bration a wide-band controller is needed. However, the design of a high
performance wide-band controller is dicult. The diculty is due to the
design trade-o between the error reduction in one frequency band and the
increase of sensitivity at other frequencies, as explained in Bodes theorem
[94]. That is, a wide-band controller is necessary to cover all the modes
of interest, but a narrow-band controller is required to achieve good at-
tenuation of a certain mode, without having a negative eect on the other
modes.
3. In many applications, the parameters of systems with exible structures
are not xed. The parameters can be time varying with dierent variations
in characteristics. The variations can be relatively small and continuous,
such as in the reduction of the total mass of aeroplane due to its fuel con-
sumption, or sudden and large, such as when an aircraft releases a missile,
or when a robot arm picks up or drops a heavy object. For some applica-
tions, the variations are not a priori known , e.g., a robot arm collecting
samples of unknown mass or being subjected to unforeseen environmental
disturbances. For these kinds of applications the diculty associated with
the design of controllers is even greater. A xed-parameter controller might
be able to achieve stability, but the capability of this type of controller
in dealing with system uncertainty is limited [152]. Fixed-parameter con-
trollers usually give poor performance or even become unstable when the
parameters are extended to values beyond the limits of the initial design.
These factors assist designers in determining the essential requirements of a
vibration controller for exible structures. To retain the high-speed performance
of the exible structures the transient response of the closed-loop system needs
to be fast with respect to the rate of change of the systems states. Furthermore,
a controller must be able to give optimum attenuation to unwanted vibration
regardless of any system uncertainties, such as those caused by variations to
the system parameters or by the unmodeled dynamics due to mode truncation.
Moreover, the controllers must be able to eectively attenuate vibration modes
of interest without introducing spillover. Other desirable requirements include,
employing a minimum number of sensor-actuator pairs, using a simple design
structure, and imposing a minimal computational load when processing real-time
applications.
Given these requirements, the question is how to design and implement a
suitable controller. Seeking the answer to this questions is the motivation for the
research described in this thesis.
1.2 Research Methodology
The research methodology consists of four steps. In the rst step, a literature re-
view overviews the relevant existing methods for controlling the vibration of ex-
ible structures, and discusses the relevant shortcomings or gaps in those methods.
Based on the gaps that are found in the existing methods, new control methods
are proposed. The rst method that is proposed, is relatively simple that works
well under very tight assumptions. This is followed by proposals for relatively
complex methods that allow the assumptions to be relaxed, and therefore with
wider applications.
In the second step, an experimental plant that can be used as a tool for the
design and evaluation of the eectiveness of the proposed control methods, is
designed and implemented. The plant chosen must represent a real application
and have the essential characteristics of a exible structure. A exible cantilever
beam with magnetically clamped loads located at arbitrary positions along the
beam is chosen as the experimental plant. The reasons for choosing this structure
are:
1. A cantilever beam with relatively heavy loads mounted along the length of
the beam is widely used as a basic model for a number of exible advanced
engineering structures such as robot arms or aircraft wings [137, 138].
2. This structure has well- separated multi-mode vibration. The frequency of
the rst mode is separated by more than one decade from its corresponding
third mode frequency.
3. This structure can exhibit large and sudden changes to its natural frequen-
cies by dropping the load(s).
From the above reasons it can be seen that a cantilever beam can be used to
represent real applications and that the chosen cantilever beam has the required
characteristics of a testbed that can be used for the design and evaluation of the
proposed control methods.
In the third step, simulation models of the experimental plant are imple-
mented, and computer simulations are exercised. These simulations reduce the
design time, increase the success rate of the real-time implementation, and help
in the evaluation of the performances of the proposed controllers prior to their
use with the experimental plant.
In the fourth step, the proposed controllers are used with the experimental
plant and the eectiveness of the proposed control methods are evaluated.
1.3 Control of Vibration
Vibration control can be categorized into two major techniques: passive control
and active control.
Passive Control
With passive control, vibration is attenuated or absorbed by passive components
such as vibration dampers and dynamic absorbers that use mass-spring-damper
decoupling. This technique is conventional and well developed [120]. However, it
has two major drawbacks. Firstly, it is ineective at low frequencies. The natural
frequency is inversely proportional to the square root of the spring compliance
and to the mass of the damper. Hence, at low frequencies, the volume and mass
requirements are often impractically large for many applications where physical
space and mass loading are critical [53, 93]. Secondly, the passive technique only
works eectively for a narrow band of frequencies and is not easy to modify.
The passive components are sensitive to the characteristics of the structures (i.e.,
damping factors and natural frequencies) and to the sources of vibration. A
redesign of the components is usually needed when the characteristics of the
structures or the sources of vibration are changed.
Active Control
With active control an electronic controller sends command signals to actua-
tor(s) such as electromagnetic shakers, piezoelectric ceramics and lms, or mag-
netostrictive devices to generate a secondary vibrational response. This secondary
response can reduce the overall response by destructively interfering with the re-
sponse of the system due to the primary source of vibration. In contrast with
passive control, active control works eectively over a wide bandwidth where the
working band does not depend on the characteristics of the structure, and is lim-
ited only by the bandwidth of the actuators. Furthermore, the actuators are less
sensitive to the characteristics of the structures and the vibration sources. There-
fore, the same actuators can be used even if the characteristics of the structures
or the vibration sources are changed. To maintain the system performance, the
electronic controller might need to be modied, but this modication is relatively
easy, especially with digital controllers.
From this discussion, it is clear that passive control is not suitable for control-
ling exible structures with varying parameters. On the other hand, active control
shows good potential. This thesis focuses on the design of active controllers.
1.4 Active Vibration Control
1.4.1 Feedforward and Feedback Control
Active control can be classied as feedforward or feedback control depending on
the derivation of the error signal.
Feedforward Control
The design of feedforward control is based on the assumption that a reference sig-
nal representing the primary excitation is available or that the primary excitation
to the structure can be directly observed. This approach has a very attractive
feature because if the reference signal and the primary excitation are exactly the
same, it is theoretically possible to achieve exact cancellation of the vibration
[45]. Another advantage of this approach is that the overall system is always sta-
ble since this approach does not modify the characteristics of the original plant.
However, in many cases such as in the attenuation of the resonant response of
an impulsively excited structure or in the attenuation of vibration in a structure
with sudden parameter changes, a suitable reference signal may not be available.
Therefore, the only possible way to control vibration in such cases is to employ
feedback control [45].
Feedback Control
In feedback control, the error signal, which is the dierence between the desired
response and the controlled output, is fed to the controller. The controller then
generates control signals to drive the error signal to zero. With feedback control,
stability becomes a major concern because the feedback modies the character-
istic of the original plant.
Due to the variations of the system parameters and the excitation signal in the
exible cantilever beam with varying load conditions, feedforward control is not
suitable for application with this system. Therefore, the design control method
discussed in this thesis focuses on feedback control.
1.4.2 Wave Control and Modal Control
Active control can also be classied according to the model descriptions upon
which the control design is based. The most common descriptions of the vibration
of continuous systems are in terms of waves and modes of motion [35]. These two
descriptions lead to two dierent approaches for active control: wave control
and modal control. The rst concentrates on controlling the ow of vibrational
energy through a structure, while the second focuses on controlling the modes of
the structure.
Wave Control
In a structure where the ow of vibrational energy from one part to another is
signicant and needs to be reduced, wave control is normally used. Wave control
design makes use of the wave equation of a structure and the local properties at
and around the control region. Since inherently the local properties of the struc-
ture are less sensitive to system properties wave control has a good robustness.
However, because it does not take into account global motion, global behaviour
can adversely aect the amount of control achieved.
When implementing wave control, to achieve good performance over a large
area of a structure, large numbers of sensors in the form of a sensor array, are
required [44, 57, 131]. Therefore, one of the desirable design requirements, stated
earlier, namely employing a minimum number of sensor-actuator pairs, cannot be
achieved if wave control is used. Moreover, most implementations of wave control
are of the feedforward type [19, 24, 44, 57, 89, 131]. For these reasons, no further
discussion or investigation of wave control is considered.
Modal Control
For a control system where the objective is to attenuate the overall structural
vibration, modal control is usually employed [45]. In modal control theory, the
structures dynamics are broken up into a series of decoupled second-order or-
dinary dierential equations (ODE) through coordinate transformation. Each
second-order ODE is similar to a single-degree-of-freedom (SDOF) system, thus
making it easy to analyse. This transformation also draws upon physical intuition
and insight gained from the experience of modal analysis which is well developed.
Since the modal properties (i.e., damping factors, natural frequencies and mode
shapes) depend on the global properties of the structure, and since by reducing
the amplitudes of the structural modes the space average mean square velocity
over the whole structure is reduced, modal control is said to be a global control
method [89, 35]. As a natural consequence of its global nature, modal control
may suer from spillover.
1.5 Modal Based Controllers for Multi-mode Vi-
bration Control
The main modal control methods that can be found in the literature for control-
ling multi-mode vibration in exible structures include: independent modal space
control (IMSC) [90], positive position feedback (PPF) control [37], and resonant
control [43, 113, 114].
The IMSC method controls each mode separately in order to evade the spillover
problem. The modal control for each mode uses only the respective displacement
and velocity feedback for the corresponding mode to avoid the coupling of modal
equations. Independent displacement and velocity gains are selected as modal
gains for each mode. Therefore, an independent controller can be designed for
each mode of vibration. The control gain for each mode can be found from solv-
ing a second-order Riccati equation from optimal control theory by minimizing
a quadratic performance index J which is the sum of the potential energy and
kinetic energy of the vibrating system. IMSC, however, has a signicant disad-
vantage in that each mode requires its own sensor-actuator pair [54]. In order
to control multi-mode vibration using a single sensor-actuator pair, Baz et al.
[12, 13] developed a time-sharing technique, referred to as modied independent
modal space control, whereby each mode of vibration is separately controlled ac-
cording to the modes energy. At periodic intervals, the energy in each mode is
calculated and the mode with the highest energy is controlled. Once the con-
trolled mode has settled and the energy in another mode becomes dominant,
the control eort is redirected to that mode. This method, however, has a high
computational load, which is caused by the need to calculate and compare the
energies in all modes of interest at every time interval.
The PPF method was proposed by Goh and Caughey [37] to avoid the spillover
problem. A special second-order compensator is designed such that its gain drops
sharply at high frequencies so as to avoid exciting residual modes. An example
of a PPF application used to control multi-mode vibration with only a single
sensor-actuator pair is shown in [132]. Here the PPF controls slewing and vibra-
tion of a single-link exible manipulator with two dominant modes. Two parallel
PPF controllers are employed to cancel the rst two modes. Each controller pa-
rameters are based on the corresponding natural frequency of the mode for which
it is designed. To test the performance of the controlled system subjected to
parameter changes, the actuator position is shifted from the original position to
make the natural frequencies change from the original frequencies. Experimental
results show that PPF is insensitive to parameter changes but still gives reason-
able attenuation. Although PPF does not introduce spillover, it does, however,
possess a drawback that could aect its suitability for multi-mode applications.
The phase of the PPF compensator tends to 0 for vibration frequencies below
the targeted mode and for frequencies above. Hence, for a single-mode con-
troller the eective structural exibility below the targeted mode is increased.
This will in turn lead to larger steady-state errors [136]. This problem becomes
compounded when using parallel PPF compensators to handle multiple modes
since the modal gain of each compensator is not independent and the higher mode
compensators may adversely aect the attenuation of the lower modes [34]. Con-
sequently, the attenuation of the lower modes with a multi-mode compensator is
compromised thus rendering the PPF compensator structure unsuitable for the
overall attenuation of multiple modes.
The Resonant Control method proposed by Pota et al. [113] is based on the
resonant characteristic of exible structures. The controller applies high gain at
the natural frequency and rolls o quickly away from the natural frequency thus
avoiding spillover. It is also described as having a decentralized characteristic
from a modal control perspective [43], thereby making it possible to treat each of
the systems modes in isolation. Hence, in contrast to PPF, the gain selection for
each mode of the resonant controller is independent, which enables the design of
the multi-mode controller to be as simple as that of the single-mode controller.
Other advantages of this controller are that it is able to control multi-mode vi-
bration using only a single collocated sensor-actuator pair, and that its design
is based only on the structure natural frequencies thus resulting in a minimal
computational load suitable for real-time implementation.
The characteristics of resonant control meet the design requirements set at
the beginning of this chapter. The control method is simple with only one design
parameter. It is robust to unmodeled dynamics that can cause spillover, and
it is capable of suppressing multi-mode vibration using a single sensor-actuator
pair. Based on these characteristics, the resonant control will be used as a basis
for controller design. However, the frequency-sensitivity characteristic of the
resonant controller is a limitation if the structure natural frequencies are altered
by changes in its conguration and/or loading. Therefore, to enable the resonant
controller to cope with the system uncertainties, modications or extensions of
the controller are also investigated.
1.6 Control Methods for Systems with Varying
Parameters
Two basic methods that can be found in the literature for controlling systems
with varying parameters are: gain-scheduling control [14, 25, 60, 72, 116, 143]
and multiple model control (MMC) [5, 6, 22, 38, 40, 48, 110, 75].
Gain-Scheduling Control
Gain-scheduling control is one of the most commonly applied techniques for con-
trolling variable parameter systems, especially where the dynamic behaviour of
the systems change with the operating conditions [1, 14, 25, 72]. The reason
that gain-scheduling is so commonly adopted is that this method oers some
practical advantages such as simple design and low computational load. The
gain-scheduling method is based on the assumption that there exists a rigid re-
lationship between the measurable variables, known as auxiliary variables, that
characterize the systems environment and the operating conditions, and the pa-
rameters of the system. The principle of gain-scheduling is to change or schedule
the parameters of the controller as functions of the auxiliary or scheduling vari-
ables, as shown in Fig. 1.2 [1]. From the gure, gain-scheduling can be viewed as
a feedback control system in which the feedback gains are adjusted using feedfor-
ward compensation [1]. The adjustment of the controller gains is precomputed
o-line and, therefore, provides no feedback to compensate for incorrect schedules
[56]. This is one of the disadvantages of the gain-scheduling control, where unpre-
dictable changes in the plants dynamics may lead to deterioration of performance
or even to complete failure.
Another drawback of gain-scheduling is assuming the availability of the aux-
Controller Plant
Scheduler
controller
parameters
Auxiliary
input
r
y
u
Figure 1.2. Block diagram of gain scheduling control.
iliary variables. Gain-scheduling control is not appropriate for a system with no
direct correlation between the measurable variables and the system parameters.
For the multi-mode vibration of a exible beam with varying load conditions,
simple correlations between measurable variables such as position, velocity, dis-
placement or frequency with the current loading conditions or parameters of the
system do not exist. All the measurable variables are dependent not only on the
current loading condition but also on the external disturbance. Therefore, gain-
scheduling control method is not suitable for controlling multi-mode vibration in
a exible structure with varying loading conditions.
Multiple Model Control
In contrast to gain-scheduling control, multiple model control (MMC) does not
need auxiliary variables for its operation. In MMC, a set of model-controller pairs,
referred to as model bank and controller bank, is designed. The design of the
models in the model bank is based on the a priori knowledge of the plant. Several
models are provided to cover the various plant conditions that may exist between
the upper and lower bounds of the plant. It is assumed that at least one model or
a weighted combination of models in the model bank will be close enough to the
current plant condition. Corresponding to each model, a controller is designed
to achieve desired performance. A supervisor or a weighting function is used
to choose a single model or a combination of models as the best representation
of the current plant condition. Two methods to determine the criteria for the
best representation of the current plant condition are found in the literature
[5, 6, 22, 38, 40, 48, 64, 75, 101]. The rst method is based on the similarity of
the models to the current plant. In this method, a probability estimate using
Bayes rule is commonly used [5, 22, 40, 48]. The second method is based on the
error tracking on the time horizon. In this method, the minimum mean-squares
error algorithm is commonly used [38, 64, 75, 101].
A shortcoming of the MMC is the high computational load requirement. In-
tensive computations are necessary to solve Bayes rule or the minimum mean-
squares algorithm in the supervisor or weighting function scheme. The computa-
tional load also increases with the number of models in the model bank. MMC
with its existing supervisor or weighting function scheme requires a powerful pro-
cessor for real-time implementation and it may be impractical to implement if
the number of models is large. To overcome this shortcoming, a simple supervi-
sor scheme can be designed based on the selective attenuation characteristic of
resonant control so that a less powerful processor can be used.
The prior knowledge assumption used in the MMC method is satised if all
the loading conditions are a priori known. If, however, loading and environmen-
tal conditions cannot be previously predicted then alternative methods, which
specically deal with uncertain systems, are required.
1.7 Control Methods for Systems with Uncer-
tainties
The exible beam with its varying loading conditions being used as the experi-
mental plant in this research has large uncertainties. The uncertainties are due
to the variation in the loading conditions and the unmodeled dynamics due to
truncation in the modeling process. There are two major control methods for
dealing with systems with uncertainties [41, 108, 154, 158]: robust control and
adaptive control.
1.7.1 Robust Control
Robust control uses a design method which focuses on the robustness of the con-
trol algorithm. In this approach, the controller is designed to withstand certain
degrees of bounded uncertainties caused by non-linearity, modeling errors or ex-
ogenous disturbances. The design typically requires that the uncertainties be
bounded in a specic region [118]. Therefore, it is important to know the pa-
rameters boundaries before designing a robust control. Once the controller is
designed, its parameters do not change and control performance is guaranteed
within the designed region.
There are two main design approaches in robust control: H
control and
Lyapunovs stability direct method. In a H
controller, robustness is achieved

using a state feedback control law that minimizes the supremum norm of the
transfer function from the disturbance input to the systems output [79, 97]. In
Lyapunovs stability direct method, robustness is maintained via the use of a
control law in which the Lyapunov function decreases along a trajectory. This is
based on the principle that a dynamic system will settle down at its equilibrium
point if the systems total energy decreases [67].
In general, robust control is suitable for dealing with small uncertainties
[152, 154, 158]. However robust control cannot give satisfactory performance
when applied to a control system with large uncertainties. For this kind of a
system, the amount of uncertainty is too large to compensate for using a xed-
parameter controller. Furthermore, the response at some operating points may
have to be overly conservative in order to satisfy specications at other operat-
ing points, while the controlled process itself varies signicantly during operation
[157]. Because robust control is not suitable for the control of exible structures
with large uncertainty it will not be considered any further.
1.7.2 Adaptive Control
In contrast to the robust control approach, the adaptive control approach uses
an on-line identication technique to estimate the current plant parameters. Its
distinctive feature is that it consists of a tuneable controller and an identifying
mechanism (an estimator). Sastry and Bodson [129] dene adaptive control as
a direct aggregation of a (non-adaptive) control methodology with some form of
recursive system identication.
Based on how the estimator is combined with the controller, adaptive control
is divided into two dierent approaches: indirect adaptive control and direct
adaptive control.
Indirect Adaptive Control
In indirect adaptive control it is assumed that a model of the plant is available
through on-line estimation, so that the controller parameters can be up-dated on-
line. In principle, one can combine any parameter estimation scheme with any
control method to form an indirect adaptive control. The algorithm used in this
approach can be described in two steps. The rst step is to estimate the param-
eters of the plant, and the second step is to determine the controller parameters
based on the estimated parameters from the rst step. This approach is based
on the certainty equivalence principle [1]. Using this principle, the controller
parameters are computed by assuming that the current plant model parame-
ter estimations are equal to the real one. The most popular implementation of
Parameter
estimator
Controller
design
r
y Adjustable
controller
u
Plant
Figure 1.3. Block diagram of the Self Tuning Regulator scheme.
indirect adaptive control is the Self Tuning Regulator (STR) scheme which is
illustrated in Fig. 1.3.
Direct Adaptive Control
In direct adaptive control, it is assumed that there exists a parameterization of
the controller such that the closed-loop system behaves in the desired fashion.
Therefore, the parameters which need to be identied are the controller parame-
ters. Direct adaptive control usually has the disadvantage that all process zeros
are cancelled [150]. This implies that direct adaptive control is intended only for
plants with a stable inverse or minimum phase systems, where the cancellation
of all the process zeros is guaranteed. Direct adaptive control is usually imple-
mented in a Model Reference Adaptive System (MRAS) scheme, as shown in Fig.
1.4.
Adaptation
mechanism
Reference
model
r
Adjustable
controller
y
m y
u
Plant
-
+
Figure 1.4: Block diagram of Model Reference Adaptive System
scheme.
Since the on-line parameter estimator used in adaptive control will track the
changes to the plant parameters and produce new parameters for the controller,
the adaptive control approach is applicable to systems with large parameter vari-
ations.
In the eld of vibration control, numerous applications of adaptive control
are found in the literature. Wherein the adaptive control is applied to sys-
tems subjected to varying disturbances [4, 16, 49, 69, 87, 112, 133, 134, 144]
and to systems with varying parameters [7, 11, 50, 51, 121, 125, 141]. For
systems subjected to varying disturbances, the estimator is used to estimate
and/or track the variations of disturbances. The estimation results are then used
to update the adaptive controller parameters so that optimum attenuation is
achieved. Several estimation methods such as zero crossing [4], gradient descent
[49, 55], phase-locked-loop (PLL) [16], least-mean-squares [112], and recursive-
least-squares (RLS) [69, 87, 133, 134, 144] are commonly used to estimate the
disturbances. For systems with varying parameters, the estimator is used to esti-
mate or track parameters changes. The most commonly used estimation method
for estimating the parameters of a system is the RLS [7, 51, 121, 125, 141]. Com-
pared to robust control, adaptive control possesses two distinctive advantageous
features. Firstly, adaptive control is capable of dealing with large and sudden
changes to a system parameters. Secondly, adaptive control is compatible with
any standard controller design method. It is relative easy to modify or convert
any standard control method into an adaptive controller. Because of these two
features adaptive resonant control methods are included in this research.
1.8 Natural Frequency Estimator
One simple technique that is used to make a resonant controller adaptive is
to use a zero-crossing method to measure the vibrating structures frequency
[3, 4]. However, this method only works eectively when the structure is sub-
jected to a single frequency excitation. For a broadband disturbance, which
includes multiple frequencies that match the structure natural frequencies, the
zero-crossing method is unable to identify any of the natural frequencies. Con-
sequently, the controller will be unable to attenuate the corresponding resonant
vibration. Thus, for multiple-frequency excitation, an eective natural frequency
estimation method is required.
With adaptive control, the computational eciency and convergence rate of
the natural frequency estimator are as critical as the estimators ability to produce
accurate estimations of the system natural frequencies. The non-parametric fre-
quency estimation approach that is based on the computationally intensive Fast
Fourier Transform [145, 124] is therefore not suitable for adaptive control meth-
ods. Rew et al. [122] surveyed the parametric frequency estimation approach
and developed a real-time natural frequency estimator based on the Recursive
Least-Squares (RLS) method combined with the Bairstow method. The results
from that paper show that for multi-mode frequency estimation, the estimator
is sensitive to sampling rate selection and to unmodeled high frequency modes.
The estimator will fail to give reasonable estimations for the lower modes if the
sampling rate is too high relative to the natural frequencies of those modes. For
a system with resonant modes that are spaced more than a decade apart in
frequency, a sampling rate that suitable for one mode will be too high for the
lower modes, thus the estimator cannot give reasonable results. The word length
limitation in the digital implementation also contributes to inaccuracy in the
estimation, especially for high sampling rates and high order implementations,
as shown in [9, 117]. Furthermore, as discussed by Wahlberg and Ljung [148],
frequency domain analysis shows that the least-squares method has a tendency
to emphasize high frequencies, especially at high sampling rates. This analysis
led designers to use preltering to remove frequencies above the highest mode
of interest in order to improve the estimation accuracy [20, 125, 126]. However,
although Rovner and Franklin [126] show that preltering improves the estima-
tion accuracy, particularly for the highest modes, the estimators high-pass lter
characteristic still produces inaccurate results for the lower modes.
As part of this research, an adaptive resonant controller for the control of
multi-mode vibration in a exible beam with varying load conditions is designed.
To achieve the desired requirements set out at the beginning of the thesis, a natu-
ral frequency estimator, that can accurately estimate system natural frequencies
when the resonant modes are spaced more than a decade apart in frequency,
is required. To design an accurate estimator, the sampling rate selection, the
preltering components, and the determination of the estimator order must be
included in the design considerations.
1.9 Multiple Model Adaptive Control
Due to the on-line estimation process, an adaptive controller inherently has a
relatively slow transient performance compared to a xed-parameter controller
[30, 41, 50, 157]. For a large and sudden change in the system parameters, adap-
tive control might produce a large transient performance. For certain systems
with strict operational requirements such as an aircraft, a large transient per-
formance may be unacceptable. Consequently, it is necessary to improve the
transient performances of adaptive controllers used in this type of application.
An approach that has been used to improve the transient performances of
an adaptive control is to combine adaptive control with MMC, as proposed by
Narendra and Balakrishnan [100]. The method is known as multiple model adap-
tive control (MMAC). In this method, a bank of a priori known xed-parameter
models, which represent the possibilities of dierent conditions, is provided. In
addition one or more adaptive models are included in the model bank to add an
adaptation capability to the system. A bank of controllers, with each controller
associated with a corresponding model in the model bank, is designed to satisfy
the control objective. The paired model-controller banks work in parallel. At
every instant a supervisor selects the best model based on a certain criterion, and
assigns the corresponding controller to be applied to the plant. Concurrently,
the adaptive model(s) is ne tuned to bring its parameters closer to the current
parameters of the plant so that optimum performance can be achieved. In prin-
ciple this method uses xed-parameter models when the adaptive model is still
in the transient condition, and switches to the adaptive model once the model
has reached its steady-state condition. With the assumption that there is at least
one model close enough to the current plant condition each time the system is
changed from one condition to another, the transient response of the system is
improved.
Along with the MMC method, the main disadvantage of the MMAC method
is the high computational demand of the supervisor scheme. The computational
demand of the MMAC also increases due to the estimation process of the adaptive
models. Therefore, to practically apply the MMAC method a simple supervisor
scheme is required.
1.10 Aim of the Thesis
Motivated by the need for a simple high performance controller for suppressing
vibration in a dynamically loaded exible structure, this thesis describes research
that is devoted to the investigation and development of active control methods
for minimizing multi-mode vibration with varying natural frequencies. The im-
plemented controllers are able to give good performance when subjected to large
and sudden changes of the system parameters, but simple enough to implement in
a real-time system. A cantilever beam with magnetically clamped loads is chosen
as the research vehicle. This cantilever beam is specically designed to exhibit
large variations in natural frequencies and has wide ranging natural frequencies,
where the rst and the third natural frequencies are separated by more than a
decade.
Investigations into three control methods: multi-model multi-mode resonant
control (M
4
RC), multi-mode adaptive resonant control (ARC) and multi-model
multi-mode adaptive resonant control (M
4
ARC) are proposed, discussed and eval-
uated in this thesis.
M
4
RC
M
4
RC is a multi-model control approach. In this approach it is assumed that all
the possible loading conditions for the exible beam are a priori known through a
modeling process. From the a priori knowledge of the plant, a bank of models is
designed such that each model gives optimum attenuation for a particular loading
condition. A simple supervisor scheme, based on the utilization of a lter bank
system, is used to determine for each mode which model has the closest frequency
to the observed vibration frequency.
ARC
If the assumption, that not all the possible loading conditions are a priori known,
is relaxed, then it is possible for unknown loading conditions to be encountered
by the system. To cope with unknown loading conditions, ARC is proposed to
control the system. The proposed ARC uses only the natural frequencies of the
system as the design parameter and an on-line natural frequency estimator is
proposed to estimate the natural frequencies.
M
4
ARC
Finally, to improve the transient response of the proposed ARC, M
4
ARC is pro-
posed. The M
4
ARC is a combination of the ARC and the M
4
RC. The principle
of the M
4
ARC is to use the M
4
RC to deal with the transient conditions and then
switch to the ARC once the systems steady-state is reached in order to achieve
optimum control during unforeseen loading changes.
1.11 Outline of the Thesis
This thesis presents the design and implementation of the resonant-based con-
trollers for attenuating multi-mode vibration in dynamically loaded exible struc-
tures. A detailed outline of the thesis structure is given below.
In Chapter 2, the experimental plant for validating the proposed controllers
is rst discussed. The derivation of the mathematical models and the implemen-
tation of the corresponding simulation model of the experimental plant are then
presented. Since all the proposed controllers are designed using the modal control
method, the mathematical models of the beam for dierent loading conditions
are derived using modal analysis. In many exible structures, such as beams, the
torsional and axial vibration are very small compared to the exural vibration.
Therefore, only the derivations of equations of motions for exural vibration of a
beam with arbitrary loadings along the beam are considered. The mathematical
models obtained from the modeling processes are validated by comparing these
models with valid models cited in the literature. The models obtained are then
compared with the experimental plant and with the numerical models obtained
using Finite Element Method software ANSYS
TM
. The comparisons are used
to nd out whether the numerical models obtained from ANSYS are accurate
enough to be used as models for the design of the controllers. Finally, in or-
der to implement the simulation models of the experimental plant, the transfer
function form of the mathematical models are derived using modal analysis in
ANSYS. The simulation models are then implemented in Simulink
TM
. Details on
the derivation of the simulation models and the transfer functions of the models
are given in Appendix A.
In Chapter 3, the M
4
RC is proposed to enable the resonant control to cope
with the natural frequency variations in a multi-mode system. A brief introduc-
tion to resonant control is presented at the beginning of the chapter, followed by
a presentation of the controller structure and an analysis of its characteristics. To
implement the resonant controller in a real-time micro controller platform, the
continuous-time resonant controller is transformed to its corresponding discrete-
time version. To preserve the passivity characteristic of the continuous resonant
controller, a bilinear transformation method is used as the discretization method.
The stability analysis of the discrete time resonant control is then obtained using
the passivity theorem. Denitions of passivity in continuous and discrete systems
are given in Appendix B. A discussion of the multiple model control approach as
the basis of the M
4
RC design is presented prior to describing the M
4
RC design.
To test the eectiveness of the proposed method, a series of simulation studies
are conducted, and the results then veried through experimental studies.
In Chapter 4, an on-line natural frequency estimator is proposed to give an
adaptation capability to the resonant controller which uses natural frequency as
the design parameter. A literature review of the estimation methods used to
determine the natural frequencies of exible structures is given at the beginning
of this chapter. The literature review is focused on the shortcomings of the
extant methods and the reasons for their shortcomings. A review of the on-
line parameter estimation concept and the derivation of the RLS algorithm are
given followed by a stability analysis of the RLS algorithm. The eects of the
high-pass characteristics of the RLS, the sampling period selection, and the nite
word-length limitation on the estimation accuracy of the RLS-based estimator
are then analysed. This analysis leads to the proposed method for designing
an accurate natural frequency estimator for a exible structure. Simulation and
experimental studies are then used to evaluate the performance of the proposed
estimator.
In Chapter 5, the ARC and the M
4
ARC are proposed. The chapter starts with
a brief review of adaptive control transient response and the MMAC method. Due
to the simple characteristics of resonant control, the proposed ARC uses only the
natural frequencies as the adaptation parameter. Since the controller is formed
in an indirect method, the stability analysis of the controller can be given by
analysing the stability of the estimator and the stability of the controller sep-
arately. To improve the transient response of the adaptive resonant controller,
a modication of the MMAC concept is applied to form M
4
ARC. A simple su-
pervisor scheme to reduce the computational burden and avoid rapid switching
between controllers in the system is proposed. The performances of the pro-
posed ARC and M
4
ARC are compared with the xed-parameter resonant control
method in simulation and experimental studies.
In Chapter 6, a summary and a conclusion obtained from the research are
presented and recommendations for further continuation of the research are given.
Simulink models for simulation and experimental implementations are given
in Appendix C.
1.12 Original Contributions to the Thesis
To the authors knowledge, the design and implementation of the three control
methods for multi-mode vibration control of a exible structure, namely: M
4
RC
(in Chapter 3), ARC, and M
4
ARC (in Chapter 5), and the multi-mode natural
frequency estimator (in Chapter4) as presented in this thesis, are original. More
specically, the original contributions are highlighted as follows:
1. A new approach to the multiple model control method using a simple su-
pervisor scheme, M
4
RC, is proposed in Chapter 3. The M
4
RC is proposed
to enable the resonant control to perform optimum vibration attenuation
in systems with varying natural frequencies. The M
4
RC design is an adap-
tation of the multiple model method with a new supervisor scheme. The
approach is new in the sense that the supervisor scheme in the M
4
RC uses
a measurement of how close the systems output vibration frequencies are
to the natural frequencies of the known models as the index performance
to choose the optimum controller. This approach is relatively simple and
less computationally complex compared to existing strategies reported in
the literature. The M
4
RC also only uses a single adjustable controller per
mode, while the existing multi-model methods use one controller per model.
Therefore, it is more feasible to implement the M
4
RC in a real-time system.
2. A new real-time natural frequency estimator for exible structures is pro-
posed in Chapter 4. The proposed approach is new in the sense that the
estimator is a set of parallel second-order estimators that use dierent sam-
pling periods for dierent modes. It is dierent to the classical RLS esti-
mator, which uses a single estimator where the estimators order is equal
to the order of the system to be estimated, and the estimator uses a single-
rate sampling period. Parallel structure and dierent sampling periods for
dierent modes are employed to improve the accuracy estimation for wide-
band multi-mode exible structures where the frequency of the rst mode
is separated by more than one decade from its corresponding third mode
frequency. The parallel low-order structure of the estimator has the advan-
tages that it is more accurate, more robust and requires less computational
power than its classical counterpart.
3. A new simple and robust adaptive control method, ARC, and a new ap-
proach to the multiple model adaptive control, M
4
ARC, are developed in
Chapter 4. Compared to extant adaptive control methods for multi-mode
vibration, the ARC has two advantages. Firstly, the ARC only uses natural
frequencies as adaptation parameters. This results in a simple controller.
Secondly, because the ARC only uses the natural frequencies of the system
as the adaptation parameters, it is inherently robust to unmodeled dynam-
ics caused by mode truncation. This is because the natural frequencies of
the system do not change in the presence of mode truncation. Mode trun-
cation changes the zeros of the system, but not the poles, which correspond
to the natural frequencies of the system. Compared to the MMAC, M
4
ARC
has two principal dierences. Firstly, M
4
ARC uses a simple supervisor that
utilizes information from a lter bank system and a natural frequency esti-
mator, compared to MMAC which uses a supervisor utilising the minimum
mean squares algorithm. The use of a simple supervisor in M
4
ARC, reduces
the computational requirements and also avoids rapid switching. Secondly,
in the M
4
ARC only one adjustable controller is used in the system regardless
of the number of models (xed and adaptive), while in MMAC the number
of controllers is equal to the number of models used in the system. Overall,
M
4
ARC is simpler than MMAC, and is thus more feasibly implemented in
a real-time system.
Apart from the above original contributions, this thesis also provides a general
formulation of the frequency equation for a cantilever beam with two lumped
masses at arbitrary positions along the beam (in Chapter 2). This ready-to-use
formula can be used as a basic tool to form simulation models. These simulation
models can assist in the design of vibration controllers used for exible structures
with varying parameters.
The outcomes of these three original contributions provide a basis for further
research into the implementation of adaptive control applicable to a large class
of exible structures.
Chapter 2
Modelling of Flexible Structures
In this chapter, the design and implementation of the experimental plant and
the models used to simulate the experimental plant are discussed. The purpose
for the modeling is outlined and the reasons for choosing the experimental plant,
and the modeling steps, are given. The description of the cantilever beam used
in the experimental plant is followed by a description of the analytical method
used to obtain the mathematical models of the plant. The implementation of the
simulation models is then discussed.
2.1 Introduction
As mentioned in Chapter 1, an experimental plant together with its mathemat-
ical representation are required as a design and evaluation tool for the proposed
controllers.
Once the experimental plant is selected, modeling methods can then be ap-
plied to nd models that represent the experimental plant. Using these models
the dynamics of the plant can be studied. As all of the proposed control methods
employ natural frequency as the controller parameter, the modeling methods are
focused on nding models with accurate representations of the natural frequencies
of the systems studied.
30
CHAPTER 2. MODELLING OF FLEXIBLE STRUCTURES 31
The mathematical model of a exible structure is derived using fundamental
theories of physics and mathematics such as Newtons laws, Hookes laws, La-
granges equations, Hamiltons principle, etc. The mathematical model that is
obtained is known as the equation of motion, which is usually given in the form of
a Partial Dierential Equation (PDE). To determine the dynamics of the model,
the solution of the PDE needs to be found. There are two common methods
used to nd the solution of the PDE: the analytical method and the numerical
method.
The analytical method gives an exact solution of the PDE. The solution is in
a closed form and is expressed in terms of known functions. Although analytical
methods can be used to very accurately describe the dynamics of structures, the
types of applications where this method can be applied are limited. The analyt-
ical method is only applicable for systems that are characterized by uniformly
distributed parameters and simple boundaries [92]. In many cases, even though
closed-form solutions may be possible, great eort and time are required to obtain
them. Therefore in practice the analytical method has fewer application areas
than the numerical method.
In the numerical method, a discrete version of the model is produced. The spa-
tial dependence in the solution of the PDE is eliminated by applying spatial dis-
cretization and the dierential eigenvalue problems are transformed into an alge-
braic form [91]. Several methods exist for constructing the discrete model [27, 91,
92]: Rayleighs method, Rayleigh-Ritzs method, Galerkins method, assumed-
modes method, collocation method, Holzers method, Myklestads method and
the nite element method (FEM). FEM is currently the most widely used method
for representing discrete models [92]. The FEM package ANSYS
TM
is used here
to study the dynamics of the structures. Due to the use of approximation, numer-
ical methods do not give the same exact results as analytical methods. However,
improvements in the computational power of computers and improvements to
approximation algorithms have led to accuracy improvements for the numerical
method.
Computer simulation is an important step in the control design process. It
provides a exible and relatively inexpensive means by which to study the dynam-
ics of a plant, design controllers, and evaluate the performances of the controllers
prior to their implementation in an actual system. The computer based simu-
lation tool Simulink
TM
is used for the simulations. To use Simulink simulation
models derived from the modication of the mathematical models obtained from
the analytical method or from the modication of the numerical models obtained
from the numerical method, need to be implemented.
The implementation of the models is undertaken in four steps. In the rst step,
the experimental plant (experimental models), which consists of the cantilevered
beam and associated loads is built. In the second step, analytical models of
the experimental models are derived. In the third step, numerical models of
the experimental models are built using ANSYS. The numerical models are then
compared with the analytical models and the experimental models to determine
their accuracy. In the fourth step, the numerical models are used to construct
the simulation models in Simulink.
2.2 Description of Experimental Plant (Experi-
mental Model)
A cantilever beam carrying relatively heavy loads mounted along the beam can
serve as a basic representative model for a number of exible structures such as
robot arms and aircraft wings [137, 138]. If the loading on the structure is changed
the natural frequencies of the system will change. Depending on the locations
and mass of the loads, relatively drastic changes in the natural frequencies can
be experienced.
For a cantilever beam, the sensitivity of a sensor to dierent modes of vi-
bration, and the authority of an actuator in attenuating the dierent modes of
vibration are determined by the position of the sensor-actuator pair along the
beam. For an actuator to have maximum authority over a specic mode it must
be placed at the position where the curvature is maximum for that particular
mode [27]. The same also applies for the sensitivity of the sensor. Because there
is a high average strain at the root of a cantilever structure, maximum sensitivity
and control authority can be achieved by placing the sensor-actuator pair close
to the root of the structure [42, 63, 153].
The schematic of the experimental plant is shown in Fig. 2.1. A uniform mild
steel beam with two identical bonded piezoceramic patches, and two magnetically
clamped loads, M
1
and M
2
are mounted in a cantilever arrangement to a concrete
block. The piezoceramics are 50mm50mm PSI-5A4E, 1.02 mm thick patches,
placed 1 cm from the root of the beam in order to obtain the maximum sensitivity
and control authority. The top patch is used as an actuator and the bottom patch
is used as a sensor. The properties of the beam are shown in Table 2.1.
Property Notation Value
Length L 500 (mm)
Width w 50 (mm)
Thickness h 3 (mm)
Mass m 589.5 (gram)
Cross section area A 150 (mm
2
)
Mass density 7860 (kg/m
3
)
Elastic modulus E 210 (GPa)
Table 2.1. Properties of the beam.
From this conguration, four dierent experimental models with dierent
loading conditions are tested. The models referred to as Model 1, Model 2,
Model 3 and Model 4 are illustrated in Figs. 2.2(a), (b), (c) and (d), respectively.
From piezo
amplifier
To signal conditioning
From switch box
L
y
x
x
2
x
1
M
1
M
2
Piezo ceramic patches
Figure 2.1. Plant with a collocated sensor-actuator pair.
The parameters of the models are shown in Table 2.2. In Model 1, two masses M
1
and M
2
are attached at x
1
and x
2
, respectively. In Model 2 only M
2
is attached
to the beam. In Model 3 only M
1
is attached to the beam. Model 4 is a cantilever
beam with no attached masses. The positions of x
1
and x
2
are chosen arbitrarily
within the broad consideration that the closer the attached masses are to the free
end of the beam the lower the natural frequencies of the structures.
Parameter Notation Value
Position of load 1 x
1
350 (mm)
Mass of load 1 M
1
283.3 (gram)
Position of load 2 x
2
475 (mm)
Mass of load 2 M
2
193.4 (gram)
Table 2.2. Model parameters.
The natural frequencies of the experimental models are the most important
parameters to be considered for the proposed control methods. Therefore, a series
of experiments are undertaken to measure the natural frequencies of the four
experimental models. The natural frequencies are obtained by applying a sweep
M
1
M
2
x
2
M
2
L
x
1
x
2
(a) Model 1 (b) Model 2
M
1
x
1
(c) Model 3 (d) Model 4
Figure 2.2. Loading model congurations.
signal to each model. The amplitude of the vibration is then measured against the
frequency of the signal. The frequencies where the amplitude of vibration forms
a peak are the natural frequencies of the models. In the experiments, a sweep
signal from a signal generator is amplied using an in-house built piezo power
amplier, and the output from the amplier is applied to the piezoactuator.
The maximum output from the piezosensor is of the order of 500 volts. The
piezosensors output is attenuated by 100 using a dierential probe, HZ115 from
Hameg Instruments
TM
. The attenuation is necessary in order to make the signal
level suitable for input to the analog-to-digital converter. Experiments show that
the dominant modes for all the experimental models are the rst three modes.
The frequencies for the rst three modes are shown in Table 2.3.
The table shows that frequencies for the rst mode of all four experimental
models are separated by more than one decade from their corresponding third
mode frequencies. This conrms that the experimental plant is wide-band, which
Model Natural frequency (Hz)
1 6
44.8
132.7
2 7.1
53.7
155.1
3 8.1
59.4
147.5
4 10.5
64
174
Table 2.3: The rst three natural frequencies of the experimental mod-
els.
is one of the requirements for the plant used in this research. It is also observed
that the added loads make the rst three natural frequencies decrease by as much
as 43.4%, 30.0%, and 27.0%, respectively. The changes to the frequency response
as the model changes are shown in Fig. 2.3. From the gure it can be seen
that the added loads not only change the natural frequencies but also change the
magnitudes of the vibration at each natural frequency. The gure also reveals
the highly resonant characteristic of the exible structure.
In the next section analytical models of the experimental model will be de-
rived.
2.3 Analytical Model
Two steps are used in building an analytical model of vibrating structures: (i)
formulating the physical dynamics of the plant, and (ii) nding the solution of the
dynamic formulation. The dynamics of the plant are formulated in the language
of mathematics to form equations of motion, which for distributed parameter
systems such as exible beams are typically given in the form of a PDE. To solve
Figure 2.3: Change of frequency responses for various loading condi-
tions.
this PDE, two analytical methods, the modal approach and the wave approach,
are commonly used [27]. All the proposed control methods considered in this
research are based on modal control. Therefore, to obtain suitable models for the
proposed modal-based controllers, the modal approach is used to solve the PDE
for the experimental models.
This section is divided into two subsections. In the rst subsection, an equa-
tion of motion for exural vibration in a cantilever beam is derived. In the second
subsection, modal analysis is used to nd the solution of the equation of motion.
2.3.1 Flexural Vibration of Beams
Flexural vibration can occur in many structures such as aircraft wings, robot ma-
nipulators, bridges and buildings. In structures where the thickness is relatively
small compared to the length and width, exural vibration is usually more critical
to structural integrity than other vibration such as axial vibration and torsional
vibration [35]. In the experimental plant, the cantilever beam that is used is
thin and relatively long. Hence the torsional and axial vibration in this structure
is negligible compared to the exural vibration. The research here is concerned
with controlling the exural vibration of the structure. Hence, the equation of
motion derived here is only for the exural vibration.
Consider a thin beam of length L with cross-sectional area A, as shown in Fig.
2.4, whose density and exural rigidity at point x are (x) and EI(x), respectively.
The distributed transverse force is f(x, t). The following assumptions are valid
for the beam:
1. The material follows Hookes Law.
2. The shear deformation is negligible compared to the bending deformation.
3. The rotation of the element is negligible compared to the vertical/transverse
translation.
The free-body diagram of an element dx is shown in Fig. 2.5, where Q denotes
the shearing force and M the bending moment. Applying Newtons second law
to vertical force components (y direction) gives
_
Q(x, t) +
Q(x, t)
x
dx
_
Q(x, t) +f(x, t)dx = A(x)dx
2
y(x, t)
t
2
(2.1)
which is equivalent to
Q(x, t)
x
+f(x, t) = A(x)
2
y(x, t)
t
2
. (2.2)
The moment equation of motion about the axis normal to x and y (out-of-the
page direction) is
_
M(x, t) +
M(x, t)
x
dx
_
M(x, t) +
_
Q(x, t) +
Q(x, t)
x
dx
_
dx
+f(x, t)dx
dx
2
= 0. (2.3)
The terms involving the second power in dx are negligible. Cancelling M(x, t) in
(2.3) and disregarding the terms involving the second power in dx, the shearing
L
y
x
x x+dx
f(x,t)
(x),EI(x)
Figure 2.4. A beam in exural vibration.
y(x,t)
M
Q
f(x,t)dx
dx
x
M
M
+
dx
x
Q
Q
+
x x+dx
y
x
Figure 2.5. A small element of the beam.
force in terms of the bending moment is
Q(x, t) =
M(x, t)
x
. (2.4)
Using the relation from (2.4), (2.2) becomes
2
M(x, t)
x
2
+f(x, t) = A(x)
2
y(x, t)
t
2
. (2.5)
From the theory of beam bending, the bending moment can be related to the
deection as [120]
M(x, t) = EI(x)
2
y(x, t)
x
2
. (2.6)
By inserting (2.6) into (2.5), the equation of motion for the forced exural vibra-
tion could be described as
2
x
2
_
EI(x)
2
y(x, t)
x
2
_
+A(x)
2
y(x, t)
t
2
= f(x, t). (2.7)
The experimental plant is a uniform beam. Therefore, (2.7) can be simplied to
EI
4
y(x, t)
x
4
+A
2
y(x, t)
t
2
= f(x, t). (2.8)
The PDE (2.8) describes the deection of the beam y(x, t) for dierent positions
x at dierent times t. To nd a unique solution for y(x, t) from (2.8), two initial
conditions and four boundary conditions are needed. The boundary conditions
for the models of the experimental plant in this research are [120]
Fixed at the left end (x = 0): The deection and the slope of the deection
at the left end are zero.
y(0, t) = 0,
y(0, t)
x
= 0. (2.9)
Free at the right end (x = L): The shearing force and the bending moment
at the right end are zero.
EI
2
y(L, t)
x
2
= 0,
x
_
EI
2
y(x, t)
x
2
_
(L,t)
= 0. (2.10)
A mass (M) attached at the right end (x = L): The shearing force at the
right end is zero, the bending moment at the right end is equal to the tensile
force caused by the lumped mass (M).
EI
2
y(L, t)
x
2
= 0,
x
_
EI
2
y(x, t)
x
2
_
(L,t)
= M
2
y(L, t)
t
2
. (2.11)
To nd the solution of PDE (2.8) for all the experimental models, modal analysis
is used.
2.3.2 Modal Analysis
Modal analysis can be used to solve the PDE that describes the equation of
motion of a structure [27, 45, 66, 91, 92]. The solution is obtained by rst solving
the eigenvalue problem for the PDE to produce a set of orthogonal functions
known as eigenfunctions. The eigenfunctions are the basis functions used to
describe the dynamics of the structure. The orthogonality of the eigenfunctions
allows one to replace the PDE with the superposition of an innite number of
decoupled second-order ordinary dierential equations (ODE). Each decoupled
ODE represents the motion of the structure for a specic mode of vibration, which
is characterized in terms of a natural frequency, mode shape, and damping. To
nd the natural frequency of the structure, a frequency equation is generated from
the eigenfunctions and the boundary conditions. Two methods that are widely
used to nd the frequency equation are the frequency determinant method and the
Laplace transforms method [85, 86]. While the mode shape and natural frequency
can be found either analytically or experimentally, the damping parameter can
only be found experimentally [45]. Modal analysis is used here to nd analytical
models for the experimental plant. To nd the frequency equation, the frequency
determinant method is used.
Consider the typical PDE for exible structures
L[y(x, t)] +C
_
y(x, t)
t
_
+M
_
2
y(x, t)
t
2
_
= f(x, t). (2.12)
L and M are linear homogeneous dierential operators of order 2p and 2q re-
spectively and q p. Here, x is the spatial coordinate, which is dened over a
domain D. The general arbitrary input is denoted by f, which is distributed over
D. The boundary conditions corresponding to PDE (2.12) can be expressed as
B
i
[y(x, t)] = 0, i = 1, 2, ..., p (2.13)
where B
i
is a linear homogeneous dierential operator of order less than or equal
to 2p 1.
In many practical structures, it is suciently accurate to assume proportional
damping for the purposes of estimating the natural frequencies and mode shapes
of the structures [45]. Proportional damping is a special type of damping where
the damping matrix is proportional to the mass and stiness matrices. By as-
suming proportional damping, C can be dened as a linear combination of the
mass and stiness operators, L and M
C = c
1
L +c
2
M, (2.14)
where c
1
and c
2
are non-negative constants. The advantage of using proportional
damping is that the mode shapes (eigenfunctions) for both the damped and un-
damped cases are the same and the natural frequencies (eigenvalues) are also very
similar [45]. Therefore, the eigenfunctions y
m
(x) can be obtained by solving the
eigenvalue problem associated with undamped version of (2.12)
L[y
m
(x)] =
m
M[y
m
(x)] , m = 1, 2, ... (2.15)
and its associated boundary conditions (2.13). The natural frequencies
m
are
determined from the eigenvalues
m
=
2
m
. (2.16)
To be used as the basis function to nd the solution of PDE (2.12), the eigen-
functions need to be orthogonal. To develop the general principle of orthogonality,
dene the self-adjoint eigenvalue problem. For any two arbitrary eigenfunctions
y
m
(x) and y
n
(x), the eigenvalue problem is self-adjoint if [27]
_
D
y
m
(x)L[y
n
(x)] dD(x) =
_
D
y
n
(x)L[y
m
(x)] dD(x) (2.17)
and
_
D
y
m
(x)M[y
n
(x)] dD(x) =
_
D
y
n
(x)M[y
m
(x)] dD(x). (2.18)
Now, let
2
m
and
2
n
be two distinct eigenvalues with corresponding eigenfunctions
y
m
(x) and y
n
(x) resulting from the solution of the self-adjoint eigenvalue problem.
Then the generalized condition of the orthogonality equation can be obtained as
_
D
y
m
(x)M[y
n
(x)] dD(x) = 0, for
2
m
=
2
n
. (2.19)
From the eigenvalue problem, it follows that
_
D
y
m
(x)L[y
n
(x)] dD(x) = 0, for
2
m
=
2
n
. (2.20)
For simplicity, the eigenfunctions are typically normalized with respect to M
such that
_
D
y
m
(x)M[y
n
(x)] dD(x) =
mn
, (2.21)
where
mn
is the Kronecker delta function, where
mn
= 1 for m = n, and zero
otherwise. Thus, (2.20) can be written as
_
D
y
m
(x)L[y
n
(x)] dD(x) =
mn
2
m
. (2.22)
Since proportional damping is assumed,
_
D
y
m
(x)C [w
n
(x)] dD(x) = c
1
mn
2
m
+c
2
mn
(2.23)
which can be written as
_
D
y
m
(x)C [w
n
(x)] dD(x) =
mn
2
m
m
(2.24)
with
m
=
1
2
m
(c
1
2
m
+c
2
) (2.25)
where
m
is damping ratio of the mode m. Equations (2.21), (2.22), and (2.24)
are orthonormal eigenfunctions which form a basis that span the solution space
of the eigenvalue problem. Thus, the response of the system can be computed at
any arbitrary point in the domain D as a superposition of these basis functions
[27]
y(x, t) =
m=1
y
m
(x)q
m
(t), (2.26)
where q
m
(t) are the generalized coordinates.
To solve PDE (2.12) using modal analysis, substituting (2.26) into (2.12), one
obtains
L
_

m=1
y
m
(x)q
m
(t)
_
+C
_

t
m=1
y
m
(x)q
m
(t)
_
+M
_

2
t
2
m=1
y
m
(x)q
m
(t)
_
= f(x, t). (2.27)
Multiplying (2.27) by y
n
(x) and integrating over the domain D, gives
m=1
q
m
(t)
mn
2
m
+
m=1
q
m
(t)
mn
2
m
m
+
m=1
q
m
(t)
mn
= Q
n
(t) (2.28)
where
Q
n
(t) =
_
D
y
n
(x)f(x, t)dD(x) (2.29)
is dened as the n
th
generalized force. Taking advantage of the orthogonality
conditions (2.22), (2.21) and (2.24), the summation in (2.28) only holds when
m = n. Thus, one obtains an innite set of uncoupled ordinary dierential
equations
q
m
(t) + 2
m
m
q
m
(t) +
2
m
q
m
(t) = Q
m
(t), m = 1, 2, ... (2.30)
From (2.30) it can be seen that the system is described by an innite set of
second-order dierential equations that can be solved independently.
In control design, due to the versatility of the frequency domain method, the
frequency domain model is often required. The solution to each second-order
dierential equation in the frequency domain can be found using Laplace trans-
forms. In many cases, the generalized force Q
m
can be written as a decomposition
of the spatial and temporal components [96]
Q
m
(t) = P
m
u(t), (2.31)
where u(t) is the input to the system and P
m
is the time-independent forcing
term. Taking the Laplace transform of the second-order equation (2.30), the
transfer function of the system can be shown to be
G(s, x) =
m=1
y
m
(x)P
m
s
2
+ 2
m
m
s +
2
m
. (2.32)
Equation (2.32) is an innite-dimensional transfer function due to the existence
of an innite number of modes. This is a general solution for PDE (2.12). The
solution for a particular structure is then solved by nding the eigenfunction (y
m
),
the natural frequency (
m
), and the structural damping (
m
). The solutions for
the experimental models are now discussed. The solution for the cantilever beam
without loads is solved rst, followed by the solution for the cantilever beam with
loads located at arbitrary positions along the beam. In nding the solutions the
following assumptions are made:
1. The beam is perfectly clamped to the concrete.
2. The eects of the cables connected to the magnetically clamped loads on
the dynamics of the structure can be ignored.
3. The masses of the piezoceramic patches are negligible compared to the
mass of the beam and the masses of the loads, thus the loading eects of
the piezoceramic patches on the dynamics of the structure can be ignored.
4. The placement of the loads are accurately positioned and the loads are
perfectly clamped to the beam.
Solution for a cantilevered beam
Consider a cantilevered uniform beam of length L where a point force u is acting
at point x = x
1
as shown in Fig. 2.6. For a linear system, such as the experimental
plant in this research, the solution for general excitation can be obtained from
the solution with a point force. The PDE that describes the dynamics of the
beam is (2.8), and the boundary conditions are described as in (2.9) and (2.10).
For a uniform beam the boundary conditions in (2.9) and (2.10) can be written
as
y(0) = 0,
dy(x)
dx
x=0
= 0,
d
2
y(x)
dx
2
x=L
= 0,
d
3
y(x)
dx
3
x=L
= 0. (2.33)
Comparing the general notations in (2.12) and (2.13) with the notations in (2.8)
and (2.33) one obtains
L = EI

4
x
4
,
M = A,
.
.
.
m,EI
u(t)
x
1
L
x
y
Figure 2.6. A cantilever beam.
B
1
= 1,
B
2
= EI

2
x
2
f(x, t) = u(t)(x x
1
) (2.34)
and D = [0, L]. C is dened as in (2.14) to include damping in the structure.
Using (2.34), the orthogonality conditions (2.21) and (2.22) can be written as
L
_
0
y
m
(x)y
n
(x)Adx =
mn
, (2.35)
L
_
0
y
m
(x)y
n
(x)EIdx =
mn
2
m
, (2.36)
where y
m
and
m
are the solutions to the eigenvalue problem (2.15). The general
solution to the eigenvalue problem (2.15) is [27, 120]
y
m
(x) = C
1m
cos
m
x +C
2m
sin
m
x +C
3m
cosh
m
x +C
4m
sinh
m
x, (2.37)
where
4
m
=
A
2
m
EI
(2.38)
and C
1m
, C
2m
, C
3m
and C
4m
are constants. The constants C
1m
, C
2m
, C
3m
and
C
4m
can be found from the boundary conditions.
To nd the frequency equation for the cantilevered beam, apply the frequency
determinant method to the eigenfunctions (2.37) and the boundary conditions
(2.33). Dropping the subscript m in (2.37) to simplify the notation, and substi-
tuting it into the rst and the second rows of (2.33) yields
y(0) = 0 = C
1
+C
3
,
dy(x)
dx
x=0
= 0 = (C
2
+C
4
). (2.39)
Using results from (2.39), (2.37) can be simplied to
y(x) = C
1
(cos x cosh x) +C
2
(sin x sinh x). (2.40)
Substituting (2.40) into the third and fourth rows of (2.33), results in the following
two homogeneous equations
d
2
y(x)
dx
2
x=L
= 0
=
2
[C
1
(cos L + cosh L) +C
2
(sin L + sinh L)] ,
(2.41)
and
d
3
y(x)
dx
3
x=L
= 0
=
3
[C
1
(sinh L sin L) +C
2
(cos L + cosh L)] .
(2.42)
For a nontrivial solution of C
1
and C
2
from (2.41) and (2.42), the determinant of
their coecients must be zero, i.e.,
(cos L + cosh L) (sin L + sinh L)

(sinh L sin L) (cos L + cosh L)
= 0. (2.43)
Solving the determinant gives the frequency equation
cos Lcosh L = 1. (2.44)
The roots
m
L are found by inserting the subscript m into in (2.44) and solving
(2.44). Applying
m
L into (2.38), gives the natural frequency of vibration
m
= (
m
L)
2
EI
AL
4
, m = 1, 2, ... (2.45)
Furthermore, solving (2.41) for C
2
in terms of C
1
and substituting it into (2.40),
the equation of modes shapes is obtained as
y
m
(x) = C
m
_
(sin
m
x sinh
m
x)
(sin
m
L + sinh
m
L)
(cos
m
L + cosh
m
L)
(cos
m
x cosh
m
x)
_
.
(2.46)
The coecient C
m
can be obtained by substituting (2.46) into the orthogonality
conditions (2.35) and (2.36). After nding the eigenfunction (y
m
) and the natu-
ral frequencies (
m
), introduce damping into the system, where
m
denotes the
damping ratio associated with mode m. Then following (2.32), the transfer func-
tion from the applied force U(s) to the transverse deection of the beam Y (s, x)
is found to be
Y (s, x)
U(s)
=
m=1
y
m
(x)y
m
(x
1
)
s
2
+ 2
m
m
s +
2
m
(2.47)
where y
m
(x) and y
m
(x
1
) are the m
th
eigenfunctions at the degrees of freedom x
and x
1
respectively.
Solution for a cantilevered beam with attached masses at arbitrary
positions
The analysis of a beam carrying multiple concentrated masses has been of interest
to researchers since the late sixties [10, 80, 83, 84, 85, 86, 139]. This is due to the
fact that the problem of a beam carrying concentrated masses is widely found
in engineering applications such as robotics, aerospace, railway bridges and long
highway bridges. Analytical methods and numerical methods can be used to
obtain the natural frequencies of beam structures. Stanisic and Hardin [139]
developed a method based on Fourier analysis to solve the natural frequencies
of a simply supported beam carrying a moving mass. The equation of motion
is given in terms of -dirac functions and is solved through the use of the rst
term Fourier nite sine and cosine series. Due to the use of the rst term series
approximation only, this method gives inaccurate results. Approximation using
higher terms can only be solved numerically because the solution is not in a
closed form. Other methods use the Laplace transform [36, 80] and a combined
analytical and numerical method [151]. In the Laplace transform method, the
eect of the attached masses is expressed by the use of -dirac function and
the equation of motion is derived by using the Timoshenko or Euler-Bernoulli
beam theories. The dierential equation is then solved by means of the Laplace
transform method. In the combined analytical-numerical method the eigenvalue
equation is derived analytically and then the eigenvalues and eigenfunctions are
calculated numerically. Low [83, 84, 85, 86] has done extensive research on the
vibration analysis of beams carrying concentrated masses at arbitrary locations.
He compared the eectiveness of the Laplace transform method and the frequency
determinant method to generate the eigenfrequency equation of a beam carrying
multiple concentrated masses at arbitrary positions. He also compared the results
from the analytical method and numerical method using Rayleighs method and
he concluded that the solution obtained using the frequency determinant method
can be obtained faster than by using the Laplace transform method [85], and that
the Rayleighs numerical method gives a good approximation to the analytical
method with the advantage of a saving in computation time.
From the above discussion, it can be concluded that the combination of ana-
lytical and numerical methods is the best way to nd the natural frequencies of
a beam carrying concentrated masses at arbitrary positions along the beam. In
this research, the frequency equation of the system is derived analytically using
the frequency determinant method, and the natural frequencies of the system are
determined numerically. Due to the complexity of the equations, the Matlab
TM
symbolic mathematics toolbox is used to derive the frequency equations of the
systems, and the natural frequencies are found numerically and/or graphically
using Matlab.
Consider a cantilever beam of mass m carrying n concentrated masses M
1
to M
n
at arbitrary positions as shown in Fig. 2.7. The beam is assumed to be
segmented into n + 1 segments. The general solutions of the eigenvalue problem
for the loaded beam system are as in (2.37). To simplify the notation drop the
subscript m and dene the solution for dierent segments as
m,EI
x
1
L
x
y
x
2
x
n
y
1
y
2
y
n
y
n+1
M
1 M
2
M
n
Figure 2.7. A cantilever beam with n attached masses.
_
_
y
1
(x) = C
1
cos x +C
2
sin x +C
3
cosh x +C
4
sinh x,
y
2
(x) = C
5
cos x +C
6
sin x +C
7
cosh x +C
8
sinh x,
.
.
.
y
n+1
(x) = C
4(n+1)3
cos x +C
4(n+1)2
sin x+
C
4(n+1)1
cosh x +C
4(n+1)
sinh x,
(2.48)
where C
1
to C
4(n+1)
are constants to be determined, and y
1
to y
n+1
are the
transverse displacements associated with the respective concentrated masses M
1
to M
n
. The compatibility conditions at the locations of n concentrated masses
are given as follows [86]
_
_
y
1
(x
1
) = y
2
(x
1
),
y
1
(x
1
) = y
2
(x
1
),
y
1
(x
1
) = y
2
(x
1
),
y
1
(x
1
) y
2
(x
1
) +
1
4
y
1
(x
1
) = 0,
_
_
y
2
(x
2
) = y
3
(x
2
),
y
2
(x
2
) = y
3
(x
2
),
y
2
(x
2
) = y
3
(x
2
),
y
2
(x
2
) y
3
(x
2
) +
2
4
y
2
(x
2
) = 0,
.
.
.
_
_
y
n
(x
n
) = y
n+1
(x
n
),
y
n
(x
n
) = y
n+1
(x
n
),
y
n
(x
n
) = y
n+1
(x
n
),
y
n
(x
n
) y
n+1
(x
n
) +
n
4
y
n
(x
n
) = 0,
(2.49)
where
i
=
M
i
m
, i = 1, 2, ..., n (2.50)
is the corresponding mass ratio. The boundary conditions for a cantilever beam
are as in (2.9) and (2.10) and can be written as
y
1
(0) = 0,
y
1
(0) = 0,
y
n+1
(L) = 0,
y
n+1
(L) = 0. (2.51)
There are 4(n+1) constants, C
1
to C
4(n+1)
in (2.48), 4n equations in (2.49) and 4
equations in (2.51). Substituting (2.48) into (2.49) and (2.51) results in 4(n +1)
equations with 4(n + 1) constants which can be written as
BC = 0, (2.52)
where C =
_
C
1
, C
2
, ..., C
4(n+1)
_
T
and B is the associated 4(n + 1) 4(n + 1)
matrix. For a non trivial solution of C
1
to C
4(n+1)
, the determinant of their
coecient must be zero i.e.,
|B| = 0. (2.53)
Solving the determinant (2.53) gives the frequency equation. The eigenvalues
which are associated with the natural frequencies can be found by solving the
generated frequency equation. The maximum number of loads attached to the
cantilever beam in the experimental plant is two. Therefore, only the frequency
equations for a cantilever beam with one and two loads at arbitrary positions are
derived in this section.
Solution a for cantilevered beam with one attached mass
For only one mass M attached to the cantilever beam, the beam is considered as
being split into two segments y
1
and y
2
with the general solution
y
1
(x) = C
1
cos x +C
2
sin x +C
3
cosh x +C
4
sinh x,
y
2
(x) = C
5
cos x +C
6
sin x +C
7
cosh x +C
8
sinh x. (2.54)
The compatibility conditions at the location of the concentrated mass are
y
1
(x
1
) = y
2
(x
1
), (2.55)
y
1
(x
1
) = y
2
(x
1
), (2.56)
y
1
(x
1
) = y
2
(x
1
), (2.57)
y
1
(x
1
) y
2
(x
1
) +
1
4
y
1
(x
1
) = 0. (2.58)
The boundary conditions are
y
1
(0) = 0, (2.59)
y
1
(0) = 0, (2.60)
y
2
(L) = 0, (2.61)
y
2
(L) = 0. (2.62)
Subtituting y
1
(x) from (2.54) to (2.59) gives
C
1
+C
3
= 0. (2.63)
Deriving y
1
(x) in (2.54) once against x and substituting the results into (2.60)
gives
C
2
+C
4
= 0. (2.64)
Substituting the second derivative of y
2
(x) against x in (2.54) into (2.61) gives
C
5
2
cos L C
6
2
sin L +C
7
2
cosh L +C
8
2
sinh L = 0. (2.65)
Substituting the third derivative of y
2
(x) against x in (2.54) into (2.62) gives
C
5
3
sin L C
6
3
cos L +C
7
3
sinh L +C
8
3
cosh L = 0. (2.66)
Substituting (2.54) into (2.55), (2.56), (2.57) and (2.58) gives the following equa-
tions
C
1
cos x
1
+C
2
sin x
1
+C
3
cosh x
1
+C
4
sinh x
1
C
5
cos x
1
C
6
sin x
1
C
7
cosh x
1
C
8
sinh x
1
= 0, (2.67)
C
1
sin x
1
+C
2
cos x
1
+C
3
sinh x
1
+C
4
cosh x
1
+C
5
sin x
1
C
6
cos x
1
C
7
sinh x
1
C
8
cosh x
1
= 0, (2.68)
C
1
2
cos x
1
C
2
2
sin x
1
+C
3
2
cosh x
1
+C
4
2
sinh x
1
+C
5
2
cos x
1
+C
6
2
sin x
1
C
7
2
cosh x
1
C
8
2
sinh x
1
= 0, (2.69)
C
1
_
3
sin x
1
+
4
cos x
1
_
+C
2
_
4
sin x
1
3
cos x
1
_
+C
3
_
3
sinh x
1
+
4
cosh x
1
_
+C
4
_
3
cosh x
1
+
4
sinh x
1
_
C
5
3
sin x
1
+C
6
3
cos x
1
C
7
3
sinh x
1
C
8
3
cosh x
1
= 0. (2.70)
Using (2.63) to (2.70), (2.52) can be formed with Bof dimension 88. Solving the
det(B) = 0 using Matlab
TM
symbolic mathematics toolbox, gives the frequency
equation as follows
2 cosh x
1
sin x
1
2 cos x
1
sinh x
1
+
2sin (L +x
1
) cosh (L +x
1
) 2cos (L +x
1
) sinh (L +x
1
)
sinh Lcos (L + 2x
1
) +sin Lcosh (L + 2x
1
) +
sin Lcosh L cos Lsinh L 4 cos Lcosh L 4 = 0.
(2.71)
Equation (2.71) is the frequency equation for a cantilever beam with a load at
arbitrary position (x
1
) along the beam. The validity of this equation needs to be
checked prior to the application of the equation to nd the natural frequencies
of the system. This is done by making comparisons against known valid models
using a comparison technique [128]. The two valid frequency equations available
from the literature are frequency equation for a cantilever beam without load
[92, 120] which is (2.44), and the frequency equation for a cantilever beam with
a load at the tip of the beam [71, 92]
sin Lcosh L cos Lsinh L cos Lcosh L 1 = 0. (2.72)
To compare frequency equation (2.71) against the two valid frequency equations
(2.44) and (2.72), the parameters and x
1
in (2.71) need to be set properly. To
compare frequency equation (2.71) against the frequency equations of a beam
without load (2.44), set = 0 in (2.71). By setting = 0 in (2.71) frequency
equation (2.44) is obtained. To compare the frequency equation (2.71) against
the frequency equations of a beam with load at the tip of the beam (2.72) set
x
1
= L in (2.71). By setting x
1
= L in (2.71) frequency equation (2.72) is
obtained. Therefore, from the comparisons with two dierent valid frequency
equations (2.44) and (2.72), it can be seen that the frequency equation (2.71) is
valid. The natural frequencies of a beam with a load at an arbitrary position can
then be found using frequency equation (2.71).
Solution for cantilevered a beam with two attached masses
For two masses M
1
and M
2
attached to the cantilever beam, there are eight com-
patibility conditions, four boundary conditions and twelve constants (C
1
to C
12
).
Using the same procedure as for a cantilever beam with one mass attached to the
cantilever beam, (2.52) can be formed with B of dimension 12 12. Solving the
det(B) = 0 using Matlab
TM
symbolic mathematics toolbox, gives the frequency
equation as follows
16 8
1
cosh x
1
sin x
1
+
2
sinh Lsin (L 2x
2
)
2
2
cos Lcosh (L 2x
1
) +
2
sin (L 2x
1
) sinh (L 2x
2
)
2
2
cos Lcosh (L 2x
2
) +
2
sin (L 2x
2
) sinh (L 2x
1
)
2
2
sin Lsinh (L 2x
1
)
2
cos Lcosh (2x
2
+L + 2x
1
) +
2
2
cosh Lcos (L 2x
2
)
2
sinh Lsin (L 2x
1
) +
2
2
cosh Lcos (2x
2
+L + 2x
1
) +
2
sin Lsinh (L 2x
2
) + 4
2
cos Lsinh L +
2
cosh Lcos (L 2x
1
) + 16 cos Lcosh L +
4
2
2
sin (x
1
+L) sinh (x
1
+L) + 2
2
1
sin x
2
2
sinh (x
2
+ 2x
1
)
4
2
sin Lcosh L + 2
2
2
sinh x
2
sin (x
2
+ 2x
1
)
2
2
2
cos (x
1
+L x
2
) cosh (x
1
+L x
2
) +
2
2
2
cosh x
2
cos (x
2
+ 2x
1
)
2
2
2
cos (2x
2
+x
1
+L) cosh (x
1
+L) +
2
2
2
cos (x
1
+L x
2
) cosh (x
1
+L x
2
)
2
2
2
sin (2x
2
+x
1
+L) sinh (x
1
+L) +
2
2
2
sin (x
1
+L x
2
) sinh (x
1
+L x
2
)
2
2
1
cos x
2
2
cosh (x
2
+ 2x
1
) +
8
2
2
sin (x
2
+x
1
) sinh (x
2
+x
1
) +
2
2
2
cos (x
1
+L) cosh (2x
2
+x
1
+L)
2
2
2
sin (x
1
+L) sinh (2x
2
+x
1
+L)
2
2
2
sin (x
1
+L x
2
) sinh (x
1
+L x
2
)
2
1
2
sin Lsinh L 4
1
2
sin x
2
2
sinh x
2
+
4
1
cos Lsinh L 4
1
sin Lcosh L 8 sin x
2
2
cosh x
2
+
8 cos x
1
1
sinh x
1
+ 8 cos x
2
2
sinh x
2
4
2
sin Lcosh (L 2x
2
) +
4
1
sinh Lcos (L 2x
1
) + 4
2
sinh Lcos (L 2x
2
) +
8
1
sin (x
1
+L) cosh (x
1
+L) 4
1
sin Lcosh (L 2x
1
) +
8
2
sin (L x
2
) cosh (L x
2
) 8
1
cos (x
1
+L) sinh (x
1
+L)
8
2
cos (L x
2
) sinh (L x
2
) = 0.
(2.73)
Equation (2.73) is the frequency equation for a cantilever beam with two masses
at arbitrary positions (x
1
) and (x
2
) along the beam. The validity of the frequency
equation (2.73) is tested by the same procedure as for the one mass case. The
frequency equation (2.71) is now a valid equation, therefore, the validity of (2.73)
can be tested by comparing it against (2.71). To compare (2.73) against (2.71),
set
2
= 0 in (2.73). Setting
2
= 0 in (2.73) makes all terms in (2.73) except the
1
st
, 2
nd
, 16
th
, 33
rd
, 34
th
, 36
th
, 39
th
, 41
st
, 42
nd
and 44
th
terms equal to zero. All the
ten non-zero terms form an equation which is the same as (2.71). Furthermore,
setting
1
=
2
= 0 in (2.73) results in (2.44) and setting x
2
= L,
1
= 0 in (2.73)
results in (2.72). Therefore, comparisons with all the valid frequency equations
(2.44), (2.72) and (2.71) show that the frequency equation (2.73) is valid. The
natural frequencies of a beam with two masses at arbitrary positions then can be
found using frequency equation (2.73).
To nd the natural frequencies of all the models in the experimental plant,
insert the corresponding parameters values x
1
, x
2
,
1
and
2
from Table 2.1
and Table 2.2 into frequency equation (2.73), nd the equation roots (
m
) then
insert into (2.45). The experiments with the experimental plant show that the
most dominant modes for all the dierent loading conditions are the rst three
modes. Therefore, only natural frequencies of the rst three modes are identied
here. The roots (
m
) for each model are found graphically using Matlab
TM
. It is
achieved by plotting the frequency equation (2.73) and nding the values where
the graph intercepts the frequency axis. The rst three roots (
m
), m= 1 to 3
are shown in Table 2.4. From the values of (
m
), the natural frequencies of the
Model
m
, m = 1-3
1 1.4477
3.9842
6.8439
2 1.5486
4.3478
7.5599
3 1.6482
4.5595
7.1418
4 1.8749
4.6941
7.8547
Table 2.4. The rst three frequency equations roots of the models.
models can be calculated. The rst three natural frequencies of all the models
compared to the rst three natural frequencies from the experimental results are
shown in Table 2.5.
From Table 2.5, it can be seen that the analytical results are very close to the
experimental results with the errors ranging from 0.32% for the second natural
frequency of Model 2 to 4.82% for the rst natural frequency of Model 4.
There are several factors that may cause the discrepancies between the an-
Model Experimental results (Hz) Analytical results from (Hz) error (%)
1 6 5.9726 0.46
44.8 45.2367 0.97
132.7 133.48 0.58
2 7.1 6.834 3.89
53.7 53.8702 0.32
155.1 162.87 4.77
3 8.1 7.7416 4.63
59.4 59.2439 0.26
147.5 145.3532 1.48
4 10.5 10.0176 4.82
64 62.7934 1.92
174 175.8200 1.04
Table 2.5: Comparison of natural frequencies obtained from experi-
mental and analytical results.
alytical and experimental results, such as imperfect clamping conditions at the
xed end, imperfect clamping of the loads to the beam, imperfect positioning of
the loads, the loading eect of cabling from the magnetically clamped loads, and
the loading eect of the piezoceramic patches. These factors are ignored or as-
sumed to be perfect in the four assumptions for nding the solution analytically.
The discrepancies, however, are reasonably small (less than 5%). Hence the mod-
els are considered to be reasonably accurate for the purpose of the experiments.
Furthermore, the proposed controllers are designed to cope with plants with nat-
ural frequency variations and unmodeled dynamics. Therefore, in the context
of the modeling for control design purposes as proposed in this thesis, the four
assumptions for nding an analytical solution do not detract from the validity
and usefulness of the proposed design methods. The analytical obtained models
are valid models, with the small variations between analytical and experimental
results explicable by the assumptions made.
2.4 Modal Analysis Using ANSYS (Numerical
Models)
The previous section shows that modal analysis can be used to nd the solution
of a cantilever beam with loads attached at arbitrary positions. However, it
can be seen from (2.52) that the determinant of a matrix needs to be solved
to nd the frequency equation of the structure. The dimension of the matrix
increases with an increase in the number of the attached loads, which increases
the complexity of the matrix. This complexity is associated with an increase in
the number of terms in the frequency equation as shown in (2.71) and (2.73).
Increasing the number of loads from one to two, causes the number of terms in
the frequency equation to increase from 10 to 45. For more than two attached
loads, the frequency equation becomes very long. Therefore, even though the
analytical method provides a solution for nding the natural frequencies of a
cantilever beam with n attached masses at arbitrary positions, it is impractical
and computationally intense.
In practice, especially for complex structures, a software tool such as ANSYS
can be used to nd the natural frequencies of the structures. ANSYS uses the
nite-element method (FEM) to solve the underlying governing equations and the
associated problem-specic boundary conditions. FEM tools are used widely in
industry to simulate the responses of a physical system to structural loading, and
thermal and electromagnetic eects. In this research, ANSYS software version
9.0 is used to nd the natural frequencies of the exible beam with varying load
conditions and the results are validated through comparison with the results from
the analytical method.
The natural frequencies of the models are obtained using modal analysis in
ANSYS. For comparison purposes, the rst three natural frequencies of the mod-
els are shown along the results from the analytical solution in Table 2.6. From
Model Results from ANSYS (Hz) Results from analytical (Hz) error(%)
1 5.9752 5.9726 0.04
45.2360 45.2367 0.0015
133.50 133.48 0.015
2 6.8345 6.834 0.0073
53.8710 53.8702 0.0015
162.86 162.87 0.0061
3 7.7432 7.7416 0.02
59.2430 59.2439 0.0015
145.3400 145.3532 0.009
4 10.0200 10.0176 0.02
62.7920 62.7934 0.002
175.8000 175.8200 0.01
Table 2.6: Comparison of natural frequency results from ANSYS with
results from analytical method.
the comparison, it can be seen that the dierence between the results from the
analytical method and the results from ANSYS are very small (maximum 0.02%).
From this comparison it can be concluded that the ANSYS results are accurate.
Based on the accuracy of the results, numerical models from ANSYS will be used
here to form simulation models for the experimental plant.
2.5 Simulation Models
In this research, computer simulation using Simulink is used to design and eval-
uate the proposed controllers. The simulation results are then used as a bench-
mark for real-time implementation of the proposed controllers in the experimental
plant. To use Simulink as a simulation platform, the numerical models from AN-
SYS need to be modied into simulation models in the form of transfer functions
or state space equations.
In this section simulation models to represent the experimental plant (Model
1 to Model 4) are designed and implemented. The simulation models are created
in transfer functions form. A description on how to obtain the transfer functions
using modal analysis in ANSYS together with the transfer functions for all the
models are given in Appendix A.
In the implementation, the order of the simulation models, which is related to
how many modes need to be included in the models, must be determined. From
a control design perspective, only those modes within a particular bandwidth
need to be included in the model. A seemingly natural method to determine how
many modes must be included in a model is to simply ignore the modes which
correspond to frequencies that lie outside of the bandwidth of interest [95]. Other
methods are to use a mode selection criterion based on modal information such
as damping ratios and natural frequencies [52], or to use dc gain sorting [46]. In
these methods, the inclusion of modes in the model is not simply based on the
frequencies of the modes, but also on how dominant the modes are in forming
the overall response of the system.
In this research, mode truncation is based on the dc gain sorting method as
presented in [46]. The dc gain for each mode can be obtained from (2.47) by
setting s = 0 as explained in Appendix A. All the numerical models from ANSYS
show that the rst three modes of vibration are the most dominant modes, and
the dc gains of the modes higher than the tenth mode are very small (< 10
2
)
of the dc gain of the rst three modes. Moreover, the inclusion of the rst 10
modes gives a bandwidth of more than 2 kHz, which is as large as one might
reasonably expect for a cantilever beam system to be controlled with an active
control. Therefore, the simulation models for all the plants are based on the
rst ten modes, and implemented in Simulink as twentieth-order systems. When
compared to the experimental results where the most dominant modes are the rst
three modes, the simulation models with 10 modes are adequate for representing
the responses of the real plant with an active an control system where most of
the disturbance frequencies are less than 2 kHz[45].
The frequency responses of the models are shown in Fig. 2.8. It can be
seen that in every interval between the consecutive resonance frequencies there
is a frequency where the amplitude of the frequency response function (FRF)
tends to zero. This is the anti-resonance frequency. This is because the transfer
functions of the models have alternating poles and zeros near the imaginary axis,
as shown in Fig. 2.9. This frequency response is typical of a collocated system
[115].
Figure 2.8. Frequency responses of simulation models.
Remark 2.1. A collocated system is obtained by the collocated placement of
sensor-actuator pairs. One advantage of a collocated system is that it is always
asymptotically stable for a wide class of single-input single-output (SISO) control
systems even if the system parameters are subject to large perturbations [115].
The experiments with the experimental models and the results from the sim-
ulation models show that the most dominant modes are the rst three modes.
Im(s)
O
Re(s)
Figure 2.9: Alternating pole-zero pattern of exible structures with
collocated sensor and actuator.
Based on this, the proposed controllers will be designed to attenuate the rst
three modes of vibration. Therefore, it is necessary to ensure that the contribu-
tions of modes higher than the rst three are negligible to the overall vibration.
To show that the modes higher than the rst three modes would has insigni-
cant eect on the overall responses of the system, simulations that compare the
responses of three-mode models with ten-mode models for dierent input signals
are conducted.
Two types of signal are chosen as input signals: white noise and pulses. White
noise is used as an input signal to give a uniform amplitude over a wide range of
frequencies and to make sure that all the modes are excited with approximately
the same amplitude. The pulse signal is used to observe the responses of the
systems when subjected to impulse or shock. This type of signal contains higher
harmonics, which can be expected to excite the higher modes and contribute to
the overall vibration. To approximate an impulse signal, a pulse signal with short
duration is used in the simulation. A pulse signal with duration 0.05 second is
chosen, because for shorter durations, the models would not respond due to dead-
time. Of all the models, Model 1 and Model 4 are the extreme loading condition
models. The responses of Model 2 and Model 3 are intermediate between those
of Model 1 and Model 4. Therefore, only a three-mode model and a ten-mode
model of Model 1 and Model 4 are chosen for the simulations.
Four simulations are conducted. In the rst simulation, white noise is used as
an input to the three-mode model and the ten-mode model of Model 1. In the
second simulation the white noise is replaced by the pulse signal. The third and
fourth simulations are similar to the rst and second simulations with Model 1
replaced by Model 4.
The responses of the simulations are shown in Figs. 2.10 - 2.13. From the
gures it can be seen that the dierences between the three-mode models and
the ten-mode models are very small. For a pulse signal input, the dierences
between the three-mode models and the ten-mode models only occurs in the rst
0.05 second, corresponding to the duration of the pulse signal and it vanishes
quickly after that period, as can be seen in Figs. 2.11 and 2.13. From the
simulations, it can be veried that the rst three modes of vibration are the most
dominant modes and the contribution of the modes higher than the rst three
to the overall vibration are negligible. Hence, attenuating the rst three modes
of vibration would in this case adequately control the vibration of the exible
beam with varying load conditions. This would also be generally true for many
structure encountered in practice.
To see how close the responses of the rst three modes of the simulation models
are to those of the experimental plant, comparisons in the frequency domain are
made. In the experiments each model is excited with a combined sinusoidal signal
containing signals with the frequencies corresponding to the rst three natural
frequencies of each model. The responses are then compared with the responses
of the rst three modes of the simulation models excited with the same signals.
Figure 2.10: Comparative responses of the three-mode model and the
ten-mode model of Model 1 with white noise input.
The comparison results are shown in Figs. 2.14 to 2.17. From the gures it can
be seen that the simulation models responses are close to the experimental plant
responses. In the gures, the natural frequencies are displayed in rad/second,
whereas the corresponding natural frequencies in Table 2.5 are shown in Hz. In
agreement with the nding of other researchers [26, 95, 96], the simulations show
that mode truncation has no eect on the natural frequency of the system. The
natural frequencies of the system are still the same after the truncation. The
truncation only changes the zero locations of the systems, indicated by a slightly
dierent shape between the frequency response functions (FRFs) of experimental
models and simulation models as shown in Figs. 2.14 to 2.17. Therefore, the
truncated model is adequate for use as a tool for designing controllers which only
need the natural frequency as a design parameter.
Figs. 2.14 to 2.17 also reveal that the vibration amplitude of the second mode
ten-mode model of Model 1 with pulse signal input.
vibration for all the models is greater than those of the rst and third modes.
This phenomenon occurs due to the placement of the sensor-actuator pair close
to the root of the cantilever beam as stated in Section 2.2. Close to the root of
the beam, the curvatures for the second mode vibration for all the models are
the largest, as can be seen from the mode shape of the simulation models shown
in Figs. 2.18 to 2.21.
2.6 Summary
In this chapter an experimental plant and simulation models of the plant are
implemented. To obtain the simulation models of the experimental plant several
steps are undertaken. In the rst step, the frequency equation of the system
is solved analytically using the frequency determinant method, and the natural
frequencies of the system are obtained graphically. In the next step, numerical
ten-mode model of Model 4 with white noise input.
models obtained from modal analysis using ANSYS are compared to the analyt-
ical results to verify the validity and accuracy of the ANSYS generated models.
In the last step, the simulation models are implemented in transfer function form
in Simulink.
Comparisons show that the responses for the simulation models are in agree-
ment with the responses for the experimental plant. The following conclusions
can be drawn from the results:
The rst three modes of the system are the dominant modes, and the con-
tributions of the higher modes to the overall vibration are negligible. There-
fore, it would be sucient to suppress the vibration of the rst three modes
to control the vibration of the exible beam considered in this research.
In the modeling process, the loading eects of the piezoceramic patches on
ten-mode model of Model 4 with pulse signal input.
Figure 2.14: Comparative responses of real plant and simulation model
for Model 1.
for Model 2.
for Model 3.
for Model 4.
Figure 2.18. Mode shapes for Model 1.
the accuracy of the natural frequencies of the models are determined to
be negligible. The objective of the modeling in this research is to obtain
reasonable and usable models that can be used for control design. All the
proposed controllers considered in this research only require the natural
frequency as the design parameter. Therefore, neglecting the patch loading
eect in the modeling is valid for the purposes of the resonant controller
design.
Mode truncation in the modeling process does not aect the accuracy of the
natural frequency estimates for the structure. Therefore, for the design of
controllers such as resonant controller that only use the natural frequency
as the design parameter, truncated models are sucient.
Chapter 3
Multiple Model Resonant
Control
In this chapter, a new control method, able to attenuate the multi-mode vibration
of a exible structure with varying natural frequencies, is presented. The method
is referred to as multi-model multi-mode resonant control (M
4
RC). Two M
4
RC
design cases are discussed. In the rst case, the M
4
RC design assumes that all of
the possible variations to the plant natural frequencies are a priori known. In the
second case, this assumption is relaxed. Only the upper and lower bounds of the
operating region are presumed to be a priori known. This chapter starts with a
brief introduction to resonant control. The resonant controller structure and its
characteristics are presented in Section 3.2. In Section 3.3, discrete-time resonant
control is derived using a bilinear transformation, and the stability of the discrete-
time resonant control system is proven using the passivity theorem. The multiple
model control approach that forms the basis of the proposed control method is
presented in Section 3.4. The proposed multiple model resonant control method
is then presented in Section 3.5. An evaluation of the proposed control method
through simulation studies is given in Section 3.6. The experimental results used
to verify the simulation results are reported in Section 3.7.
74
CHAPTER 3. MULTIPLE MODEL RESONANT CONTROL 75
3.1 Introduction
As shown in Chapter 2, one of the characteristics of exible structures is their
highly resonant nature. For the case of the cantilever beam used in this research
this characteristic is made evident by the relatively large vibration at or near
to the natural frequencies of the structure, as shown in Fig. 3.1. From the
Figure 3.1. Frequency response of a exible cantilever beam.
gure, it is clear that suppressing the vibration of a structure at or very close to
the natural frequencies of the structure is more important than suppressing the
vibration at other frequencies. However, suppressing the vibration at one or more
natural frequencies may excite or amplify other natural frequencies. Therefore,
one important design requirement for exible structure control is to achieve high
attenuation for modes of interest without driving the other modes into instability
or exciting and amplifying the vibration of other modes.
As stated in Chapter 1, a resonant controller has necessary characteristics that
satisfy the design requirements for multi-mode vibration attenuation of exible
structures investigated in this research. Exploiting the highly resonant charac-
teristic of the exible structure, the resonant controller only applies high gain at
or close to the natural frequencies of interest, and is therefore able to suppress
the vibration at those frequencies without causing adverse eects at other fre-
quencies. In the following section, a further analysis on the characteristics of a
resonant controller is presented.
3.2 Structure of a Resonant Controller
Consider a exible structure with a collocated piezoelectric sensor-actuator pair
attached to it as shown in Fig. 3.2 [114]. The piezoelectric patch at the top
of the beam is used as an actuator, while the patch on the bottom serves as a
sensor. The controller output is u(t), the sensor voltage is y(t), and the external
disturbance is f(t). The block diagram for the control system can be depicted
as in Fig. 3.3. The approximate transfer function of the exible structure is the
truncated version of (2.47):
G
M
(s, x) =
M
m=1
y
m
(x)y
m
(x
1
)
s
2
+ 2
m
m
s +
2
m
m = 0, 1, 2, ..., M (3.1)
where M is the highest resonant mode to be controlled.
The resonant controller K(s) as given in [113, 114] is described by
K(s) =
M
m=0
K
m
(s) (3.2)
with
K
m
(s) = k
cm
s
2
+ 2
cm
m
s
s
2
+ 2
cm
m
s +
2
m
(3.3)
where
m
is the m
th
controller centre frequency, which is the same as the m
th
natural frequency of the vibrating system,
cm
is the m
th
mode controller damping
factor, and k
cm
is the gain for the m
th
mode controller.
The controller K
m
(s) is a second-order lter with two zeros and two poles.
One zero is at the origin and the other is a real zero in the Left Half Plane
K(s)
f(t)
u(t)
r(t)
y(t)
Figure 3.2. Flexible structure control system.
G
M
(s,x)
U(s)
R(s)
-
F(s)
+
+ Y(s)
K(s)
Figure 3.3. Block diagram of resonant control.
(LHP) of the s-plane. The controller has a band-pass lter which has a high
Q characteristic, with a high gain at the natural frequency of the system and a
sharp drop in gain away from the natural frequency of the system. This highly
resonant characteristic of the controller is shown in Fig 3.4. The transfer function
Figure 3.4. Frequency response of a dual mode resonant controller.
of the closed loop system is
Y (s)
R(s)
=
G
M
(s)
1 +G
M
(s)K(s)
. (3.4)
To suppress vibration the term on the right hand side of (3.4) must be small.
Hence it can be seen that the feedback is eective only for the frequency range
close to the natural frequency of the system where |G
M
(j)K(j)| is large. This
characteristic makes the controller robust to spillover eects caused by mode
truncation, because the controller will not excite the unmodeled dynamics away
from the frequency of concern. The eect of the controller is localized to each
mode of a structure. Due to this localized eect of the controller, a single sensor-
actuator pair can be used to suppress multi-mode vibration in the structure.
Equation (3.2) shows that to control M modes of vibration, M controllers can
be arranged in parallel to form a summation of the dierent mode controllers.
Each of the mode controller parameters,
m
,
cm
, and k
cm
in (3.3) is designed
independently, and solely based on the plant parameters for the corresponding
mode. The parameter k
cm
in (3.3) can be set to zero for a particular m if that
mode is not to be controlled, giving designers the freedom to choose which mode
they wish to control.
Of all the three controller design parameters, only the controller centre fre-
quency is sensitive to variations of the plant parameters. To achieve optimum
attenuation, the controller centre frequency must be set to be the same as the
natural frequency of the plant. In contrast, the controller damping ratio,
cm
,
and the controller gain, k
cm
, are not sensitive to variations of the plant param-
eters, and there is no simple relationship between
cm
and k
cm
with the plant
parameters. Optimum values of
cm
and k
cm
can be found through a trial and
error procedure. The parameters
cm
and k
cm
have diering eects on control
performance. As the chosen value of
cm
is decreased, the attenuation at the
corresponding natural frequency becomes higher. If, however, the selected value
of
cm
is too small, the vibration at the other frequencies outside this natural
frequency will be amplied. The eect of k
cm
selection demonstrates a converse
result to that of the
cm
selection. The higher the value of k
cm
, the higher the
attenuation associated with that natural frequency. However, if the selected value
of k
cm
is too high, then the vibration amplitudes for the other frequencies out-
side this natural frequency will be amplied. The sensitivity of the controller
parameters to variations of the plant parameters, which aect the control system
performance, is demonstrated in the following example.
Example 3.2.1
A resonant controller K(s) with k
cm
= 10,
cm
= 0.01, and = 62.8 rad/sec is
used to control a second-order plant in ve dierent cases. In the rst case the
controller is used to control the nominal plant, G
1
(s). In the second and third
cases, the natural frequency of G
1
(s) is decreased and increased by 5% to form
G
2
(s) and G
3
(s), respectively. In the fourth and fth cases the damping factor
of the plant are two times larger and ten times lower, respectively, compared to
the damping factor of G
1
(s). This yields G
4
(s) and G
5
(s). The parameters and
the transfer function of the plant for all ve cases are shown in Table 3.1.
Case (rad/sec) Transfer function
1 0.001 62.8 G
1
(s) =
1
s
2
+0.126s+3947.8
2 0.001 59.7 G
2
(s) =
1
s
2
+0.1196s+3562.9
3 0.001 65.9 G
3
(s) =
1
s
2
+0.1326s+4352.5
4 0.002 62.8 G
4
(s) =
1
s
2
+0.251s+3947.8
5 0.0001 62.8 G
5
(s) =
1
s
2
+0.0126s+3947.8
Table 3.1. Models parameters used in the example.
The responses of the controlled system for all the cases are shown in Fig. 3.5
and Fig. 3.6. Fig. 3.5 shows that a signicant degradation of the attenuation
performance occurs for a small change in the natural frequency of the plant. The
attenuation performance decreases around 13 dB for only a 5% change in the
plant natural frequency. However, the attenuation performance is not sensitive
to the plant damping factor variation. The attenuation performance only varies
around 1 dB for 2 times (100%) and 10 times variations in the plant damping
factor as shown in Fig. 3.6.
In the implementation of the resonant controller, it is assumed that the natu-
ral frequencies of the structure are known through a modeling process or modal
61.5 62 62.5 63 63.5 64 64.5
-60
-55
-50
-45
-40
-35
-30
-25
-20
-15
-10
Frequency (rad/sec)
M
a
g
n
i
t
u
d
e

(
d
B
)
open loop
case 1
case 2
case 3
Figure 3.5: Closed-loop responses for variations in a plant natural
frequencies.
61.5 62 62.5 63 63.5 64 64.5
-60
-55
-50
-45
-40
-35
-30
-25
-20
-15
-10
Frequency (rad/sec)
M
a
g
n
i
t
u
d
e
(
d
B
)
open loop
case 1
case 4
case 5
Figure 3.6: Closed-loop responses for variations in a plant damping
factors.
testing measurement. Then, based only on the natural frequencies of the struc-
ture, the resonant controller can be designed. This promotes a simple design
procedure suitable for a real-time implementation. However, since a resonant
controller is designed for specic narrow frequencies and is very sensitive to the
structure natural frequencies variations, then it becomes ineective if the struc-
ture natural frequencies are altered due to changes in its conguration and or
loading.
3.3 Discrete-time Resonant Control
To embed the resonant control algorithm in a real-time micro controller platform,
the continuous resonant controller (3.2) must be transformed into its correspond-
ing discrete-time form. There are ve common methods that can be used to
discretize continuous systems into discrete-time systems: the impulse-invariant
method, the backward-dierence method, the forward-dierence method, the bi-
linear transformation method, and the step-invariance method [107].
In this research, the selection of the discretization method is based on preser-
vation of the passivity of the continuous systems through the discretization pro-
cess and practical minimisation of the computational load of the discretization
method. A method that preserves the passivity characteristic of continuous res-
onant control is chosen to ensure stability. A proof for this is given in Section
3.3.1 and 3.3.2.
Not all of the ve methods mentioned above preserve the passivity of the
continuous systems. Jiang [58] examined the passivity preservation properties of
the ve methods. The conclusion of the study is that only backward-dierence,
bilinear transformation and impulse-invariant methods are capable of preserving
the passivity of the continuous system. Furthermore, of these three methods,
the bilinear transformation method has the lowest computational demand [107].
Therefore, the bilinear transformation is chosen in this research as the discretiza-
tion method.
Applying the bilinear transformation
s = 2f
s
z 1
z + 1
(3.5)
to (3.3), where f
s
is the sampling frequency, one obtains the discrete-time resonant
controller for the m
th
mode as
K
m
(z) = k
dm
A
m
z
2
+B
m
z +C
m
z
2
+D
m
z +E
m
(3.6)
with
A
m
=
4f
2
s
+ 4f
s
cm
m
4f
2
s
+ 4f
s
cm
m
+
2
m
, (3.7)
B
m
=
8f
2
s
4f
2
s
+ 4f
s
cm
m
+
2
m
, (3.8)
C
m
=
4f
2
s
4f
s
cm
m
4f
2
s
+ 4f
s
cm
m
+
2
m
, (3.9)
D
m
=
2
2
m
8f
2
s
4f
2
s
+ 4f
s
cm
m
+
2
m
, (3.10)
E
m
=
4f
2
s
4f
s
cm
m
+
2
m
4f
2
s
+ 4f
s
cm
m
+
2
m
, (3.11)
and k
dm
is the gain for the m
th
controller. From (3.6), the discrete-time resonant
control law u(k) for sampling time k is given by
u(k) =
M
m=1
u
m
(k)k
dm
, (3.12)
with
u
m
(k) = A
m
y(k)+B
m
y(k1)+C
m
y(k2)D
m
u
m
(k1)E
m
u
m
(k2) (3.13)
where y(k) is the controlled output at sampling time k.
The continuous resonant control system is proven to be stable [114]. In the
next two sections there is a discussion on system stability and a proof that for
a discrete-time resonant control system preservation of a system passivity using
bilinear transformation also preserves the stability of the system.
3.3.1 Input-output Stability
One general approach that is widely used to address the stability problems that
arise in control systems is the input-output stability theory [56, 99]. This stability
theory considers the eect of external inputs on a systems stability. If a system
is described by an operator W that maps an input space U to an output space Y,
the concept of the input-output stability is based on the properties of U and Y.
If a property L
p
space of the input is invariant under the transformation W, the
system is said to be L
p
-stable [99]. For any xed p [1, ), variable u is said to
belong to L
p
if the L
p
norm of u exists [56] or
u
p
0
|u(t)|
p
dt
1/p
< , (3.14)
and for p =
u
= sup
t0
|u(t)| < . (3.15)
Then the system represented by the operator W is said to be L
p
-stable if u
L
p
is mapped into y L
p
. When p = , L
p
-stability is also referred to as
bounded-input bounded-output (BIBO) stability. The familiar BIBO stability
notion contends that if the input signal is nite, then the output signal also must
be nite.
Consider now a canonical feedback system as shown in Fig. 3.7. G
1
and G
2
are operators that act on input u
1
and u
2
, respectively, to produce output y
1
and
y
2
. r is an external input to the system. The equations describing the system are
given by
y
1
= G
1
u
1
y
2
= G
2
u
2
(3.16)
The stability problem then is to determine conditions for G
1
and G
2
, so that if
u
1
and u
2
are in the same class L
p
, then y
1
and y
2
are in the same class.
G
1
G
2
r
-
+
u
1
y
1
y
2
u
2
Figure 3.7. A canonical feedback system.
One concept, which is widely used in input-output stability problems, is that
of passivity [99]. In the passivity concept, a system is considered as a device
which interacts with its environment by transforming inputs into outputs. From
an energy viewpoint, a passive system can be dened as a system which cannot
store more energy than is supplied by some source, with the dierence between
stored energy and supplied energy, being the dissipated energy [82]. Hence for a
bounded input, a passive system will produce bounded output. Therefore it is
clear that a passive system is stable in the input-output stability sense.
A fundamental property of passive systems is that the negative feedback in-
terconnection of two passive systems yields a passive system. This property is
formalized in the passivity theorem which states that a negative feedback inter-
connection of two passive system is stable [29]. Therefore, the stability of the
feedback system in Fig 3.7 can be achieved by proving that G
1
and G
2
are pas-
sive. Another property of passive systems is that a system resulting from the
parallel interconnection of passive systems is passive [70].
3.3.2 Stability of a Discrete-time Resonant Control Sys-
tem
As shown in Fig. 3.3, a resonant control system is a feedback interconnection
between the truncated plant G
M
and the resonant controller K. The stability
of the discrete-time resonant control system will be proven using the passivity
theorem. To accomplish this, it is sucient to prove that both the discrete-time
controller K(z) and the discretization of truncated plant (3.1) are passive. Fur-
thermore, due to the passivity of the parallel interconnection of passive systems,
it is sucient to prove that the discrete-time controller (3.6) is passive.
Two steps are undertaken in the stability proof. Step 1 proves that the
discrete-time controller (3.6) is passive, and Step 2 proves that the discretize
truncated plant (3.1) is also passive.
Step 1
In Appendix B, it is stated that a discrete-time system is passive if all the poles
of the system lie inside the unit circle and the real value of the system for z = 1
is greater than or equal to zero. Therefore it can be proven that the discrete-time
resonant controller (3.6) is passive by conrming that all the zeros of (3.6) lie
inside the unit circle and that the real value of (3.6) for z = 1 is greater than or
equal to zero. Write (3.6) as
K
m
(z) =
P(z)
Q(z)
= k
dm
A
m
z
2
+B
m
z +C
m
z
2
+D
m
z +E
m
. (3.17)
According to the Jury criterion [107] all the poles of (3.17) are inside the unit
circle if
1. E
m
< 1,
2. Q(z)|
z=1
> 0, and
3. Q(z)|
z=1
> 0.
From (3.11) it can be seen that the rst condition is satised. Substituting z = 1
and z = 1, respectively, into Q(z) in (3.17) results in
Q(1) =
4
2
m
4f
2
s
+ 4f
s
cm
m
+
2
m
> 0 (3.18)
and
Q(1) =
16f
2
s
4f
2
s
+ 4f
s
cm
m
+
2
m
> 0, (3.19)
respectively, which shows that the second and third conditions are also satised.
Thus all the poles of K
m
(z) lie inside the unit circle. Inserting z = 1 in (3.17)
results in
Re[K
m
(z)] = 0. z = 1 (3.20)
Therefore, (3.17) satises the conditions for passive systems as in Denition B.2
of Appendix B, and hence the controller (3.6) is passive.
Step 2
Applying the bilinear transformation (3.5) to (3.1) one obtains the transfer func-
tion of the discrete-time plant as
G(z, x) =
R(z)
S(z)
=
M
m=1
y
m
(x)y
m
(x
1
)(z
2
+ 2z + 1)
a
m
z
2
+b
m
z +c
m
(3.21)
with
a
m
= 4f
2
s
+ 4f
s
m
+
2
m
, (3.22)
b
m
= 8f
2
s
+ 2
2
m
, (3.23)
and
c
m
= 4f
2
s
4f
s
m
+
2
m
. (3.24)
From (3.21) to (3.24), it can be seen that
a
m
> c
m
, (3.25)
S(z)|
z=1
= 4
2
m
> 0, (3.26)
and
S(z)|
z=1
= 16f
2
s
> 0. (3.27)
Hence all the poles of (3.21) lie inside the unit circle. Furthermore, by substituting
z = 1 in (3.21) and because eigenfunctions, y
m
(x) and y
m
(x
1
), have the same sign
for collocated systems, then
Re[G(z, x)] =
4y
2
m
(x)
4
2
m
> 0, z = 1. (3.28)
Following Denition B.3 of Appendix B, the discrete plant (3.21) is strictly pas-
sive.
Hence, the closed loop system comprising the discrete-time resonant controller
(3.6) and the plant (3.1) is stable according to passivity theorem.
The control system is stable with respect to incorrect resonant frequencies.
The incorrectly specied natural frequencies of the model will only make the
controller perform sub optimally without destabilizing the closed-loop system.
To obtain optimum performance for a system with varying natural frequencies, a
resonant controller based on the multiple model control approach, referred to as
M
4
RC, is proposed. Prior to the discussion of the M
4
RC design, a brief review
of the multiple model control method is presented in the next section.
3.4 Multiple Model Control
As mentioned in Chapter 1, the multiple model control (MMC) method is a
control method that is designed to cope with systems that have varying param-
eters. MMC is a model-based control method which integrates a bank of model-
controller pairs to handle all possible operating conditions. Two approaches to
MMC can be found in the literature: the weighting function scheme [5, 22, 40,
48, 119] and the supervisor scheme [6, 38, 100, 75].
Weighting Function Scheme
In the weighting function scheme approach, as shown in Fig. 3.8, a model bank
is designed based on a priori knowledge of the plant. Corresponding to each
model in the bank, a controller is designed. At each step, the output of the plant,
y, is compared with the output of the models y
i
, i = 1 to L. The dierences
generated from the comparisons are the errors e
i
, i = 1 to L. Using these errors,
a weighting function is used recursively to measure the probability of each model
in the model bank representing the current plant. Based on the measurement,
suitable weights are given to individual controllers in the controller bank such that
the most probable model carries the highest weight. The sum of the weighted
controllers outputs is then used as the control signal to the plant. In this way,
control signals from the controllers, u
i
, i = 1 to L, for models which closely
represent the current plant will exert greater inuence on the nal control signal
u.
The most common method used in the weighting function design employs a
probability estimate based on Bayes rule [22, 40, 48]. Using this method, the
likelihood of the i
th
model tting the plant can be measured, and the weighting
value for each corresponding controller can be determined. At the k
th
step, the
probability for the i
th
model being the true model of the plant is computed as
[22]
p
i,k
=
exp(
1
2
e
T
i,k
C
f
e
i,k
)p
i,k1
L
j=1
exp(
1
2
e
T
i,k
C
f
e
j,k
)p
j,k1
, (3.29)
where
e
i,k
= y
k
y
i,k
, i = 1 . . . L, (3.30)
is the error at the k
th
step. L denotes the total number of models in the model
bank and C
f
is the convergence factor used to tune the rate of convergence of
the probabilities. The recursion is initialized by assigning equal probability
1
L
to all models in the bank. At each iteration the new probability, p
i,k
, is calcu-
K
1
K
2
K
L
Plant
e
1
Weighting
function
u
1
r +
u y
M
L
M
2
M
1
Controller Bank
Model Bank
+
+
+
-
L
y
~
-
-
u
2
u
L
e
2
e
L
-
1
~
y
2
~
y
Figure 3.8: Multiple model control method using weighting function
scheme.
lated, improving upon the probability calculated at the previous iteration, p
i,k1
.
Equation (3.29) shows that once a probability reaches zero it will remain zero
thereafter. A threshold is added to prevent p
i,k
from becoming zero,
p
i,k
= p
i,k
p
i,k
> ,
p
i,k
= p
i,k
. (3.31)
At the k
th
step, a weight W
i,k
for the i
th
model is calculated as [130]
W
i,k
=
p
i,k
L
j=1
p
j,k
p
i,k
> ,
W
i,k
= 0 p
i,k
= . (3.32)
Then the nal control signal applied to the plant is
u
k
=
L
j=1
W
i,k
u
i,k
. (3.33)
From (3.33), it can be seen that the unique feature of this approach is the
capability of the weighting function to nd a proper combination of control signals
from the individual controllers which can be applied to the plant as the nal
control signal. Due to the highly resonant characteristic, the resonant controller
works eectively only for a narrow-band frequency. For a certain plant, only a
controller with a centre frequency that matches or is close to the plant natural
frequency will produce an optimum control signal. A weighted combination of
several controllers with dierent centre frequencies will not produce the optimum
control signal. For that reason, a multi-model control with a weighting function
scheme will not improve the performance of resonant control when the natural
frequencies of the system under control are varied. Hence, this multiple-model
control method is not suitable for use as a basis for a multi-model controller based
on resonant control.
K
1
u
2
u
L
K
2
K
L
Plant
u
1
r +
e
1
e
2
e
L
L
y
~
1
~
y
2
~
y
-
Controller
Switch
u y
-
M
1
M
2
M
L
+
Controller Bank
Model Bank
+
+
-
-
Supervisor
Figure 3.9. Multiple model control method using supervisor scheme.
Supervisor Scheme
In the supervisor scheme approach, as shown in Fig. 3.9, a bank of models and a
supervisor scheme are used to choose a model that gives the best approximation
to the current plant condition. A controller corresponding to the model is then
used to supply the control action to the plant. In contrast to the weighting
function scheme, in the supervisor scheme only a single controller is chosen for
the nal control action through a switching process in the controller bank. Since
only one controller is selected at any instant, then horizon-based error tracking
such as minimum mean-squares error (MMSE) is commonly used to design the
supervisor [6, 38, 75, 101]. This method identies which model produces the
lowest error over a xed period of time and selects the corresponding controller
as the most suitable for the given condition.
The algorithm for the supervisor scheme can be written as [101]
J
l
(t) = e
2
l
(t) +
t
0
e
(t)
e
2
l
()d, 0, , > 0 (3.34)
for a continuous-time system or
J
l
(k) = e
2
l
(k) +
k
j=0
e
(kj)
e
2
l
(j) (3.35)
for a discrete-time system. The error e
l
is the dierence between the output of the
plant, y, and the output of the model L, y
l
, as shown in Fig. 3.9. The constants
and are the weighting factors which are used to weight the importance of
current and past errors, respectively. For = 0 the index performance (3.34) can
be written as
J
l
(t)

= e
2
l
(t). (3.36)
The index performance (3.36) is based only on the current errors. It will quickly
detect transient peaks in e
2
l
(t) resulting in very rapid switching. However, very
rapid switching between the controllers may lead to poor system performance
[100]. If the current errors weighting factor, , is set to be zero, the index
performance (3.34) becomes
J
l
(t)

=
t
0
e
(t)
e
2
l
()d. (3.37)
In the index performance (3.37), the constant plays an important role in deter-
mining the response of the system. The selection of an appropriate value for for
a given problem is essentially heuristic and can be dierent for each model [6]. It
is aected by several factors such as the parameter variation period, disturbance
eect, and noise eect [65]. For a system with slow parameter variations, small
should be selected. For a small , the index performance (3.37) will approach
J
l
(t)

=
t
0
e
2
l
()d, (3.38)
and this index is a good indicator of steady-state identier accuracy which leads
the system to select the best model in the model bank. On the other hand,
for a system with frequent parameter variations a large value of might be
chosen. A large value of will give more weight to the latest errors resulting in
a quick response to abrupt parameter changes. However, disturbance and noise
aect the error signal e
l
, and a large value of makes the index performance
very sensitive to noise and disturbance, leading to unwanted switching resulting
in poor performance. Hence, for a noisy system and/or a system subjected to
disturbance, a small value of needs to be chosen.
From the discussion above it can be seen that the choice of for a noisy system
with frequent parameter variations that are subject to disturbance is conicting.
In such a situation choosing a dierent set of parameters (i.e., , , and ) can
be a dicult task, hence greater knowledge about the system is required. Besides
the diculty of selecting parameters, as can be seen from the second term of
(3.34) or (3.35) the switching scheme poses a high computational demand that
increases signicantly with the number of models employed. Index performance
is calculated for each model, and then a comparison is made to determine the
minimum index performance. For example, if 100 samples (i.e., k = 100 in (3.35))
are required to calculate the performance index for each model, then a further
100 samples are required for every additional model. The computation demand
can be reduced by setting = 0, as shown in (3.36). However, as discussed above
extremely rapid switching between controllers may occur, resulting in ineective
control. To reduce the computational demand and to avoid too rapid switching,
a simple supervisor scheme for use with the M
4
RC method is proposed in the
next section.
3.5 Multi-model Multi-mode Resonant Control
(M
4
RC)
As discussed in Section 3.2, due to its highly resonant characteristic, resonant
control will only give optimum performance when the controller centre frequency
coincides with the natural frequency of the system. For a system with varying
natural frequencies, even though the system is stable, the closed loop system will
not achieve optimum performance. In order to cope with the system natural
frequency variations, M
4
RC is proposed.
The M
4
RC design is based on the multiple model control with supervisor
scheme. Two design cases are presented. In the rst case, a priori information
about all the plant condition is assumed to be available. From that information,
a model bank and its associated controller bank can be designed. In the second
case, only the upper and lower bounds of the operating region are a priori known.
3.5.1 Case 1: All the Possible Loading Condition are a
priori Known
Two design problems need to be addressed in the design process for the multiple
model control method with supervisor scheme: (i) the design of the supervisor
scheme and (ii) the determination of how many models in the model bank are
required to span the operating region. Due to the assumption that all of the
possible loading conditions are a priori known, the second problem is not an
issue. The number of required models is determined from the a priori information
regarding how many dierent loading conditions exist in the operating region.
This leaves the design of the supervisor scheme as the main focus of the M
4
RC
design.
M
4
RC Design
In principle, the role of the supervisor in the multi-model control method is to
determine which model in the model bank most closely represents the current
plant condition. In the M
4
RC, the models are designed to represent the natural
frequencies of the dierent plant conditions. Since the eectiveness of the resonant
control depends on how close the controller centre frequency is to the excitation
frequency, a simple supervisor can be designed to replace the MMSE scheme
described in (3.35).
From the linearity principle, a linear structure will vibrate with the same
frequency as the frequency of excitation signal. Thus the excitation frequency
applied to a linear structure can be measured by measuring the output frequency.
Using this principle, the determination of the closest model to the current plant
can be determined by comparing the plants output vibration frequency with the
centre frequencies of the models. To achieve this for single-mode control, the
model bank Model 1 to Model L used in Fig. 3.9 is replaced by the lter bank
system shown in Fig. 3.10.
In this approach, a bank of band-pass lters, BPF
m1
to BPF
mL
, is used to
represent the m
th
natural frequency of Model 1 to Model L. The representation is
achieved by employing a narrow band-pass lter where the centre frequency is the
same as the natural frequency of the corresponding model. In this way if the plant
vibration signal is injected into the lter bank, the BPF with the closest centre
frequency to the vibration signal frequency will produce the maximum output.
To identify which model (BPF) gives the maximum output, a decision making
component, formed from an absolute (abs) block, a low-pass lter (LPF), and a
maximum (MAX) block, is used. The abs block obtains the absolute value of the
vibration signal. The absolute value of the signal is then passed through a LPF
to obtain the dc value of the signal. The MAX block is then used to select which
LPF produces the maximum dc value. In this way the most appropriate known
model and its corresponding controller for the current single-mode frequency can
be identied. An output of the lter bank system then sets a switch to select the
appropriate plant controller.
y
BPF
m2
abs LPF
BPF
m1
abs
Max
LPF
out
BPF
mL
abs
LPF
Figure 3.10. Filter bank system for the m
th
mode.
Theoretically, a bank of xed-parameter controllers needs to be implemented
with one controller for each model in the model bank. Since each model only
retains the plant natural frequency, and each xed-parameter resonant controller
only needs the natural frequency as a design parameter, then only one adjustable
controller is used to replace the controller bank. At each sampling time the pa-
rameter
m
of the closest model to the current plant is loaded into the adjustable
controller through a switching system, as shown in Fig. 3.11.
To enable the multi-mode control, the above principle can be simply extended
by replicating the lter bank system and the switching system repeatedly for all
From m
th
mode filter
bank systems output
S
m
1 m
2 m
mL
To m
th
mode
adjustable
controller
Figure 3.11. Switching system for the m
th
mode.
the modes of interest. In this way the most appropriate models for all the modes
of interest can be identied. A schematic diagram of the M
4
RC for M modes of
interest is shown in Fig. 3.12.
Band-pass lter design
The complexity of the proposed supervisor scheme is signicantly reduced relative
the complexity of the existing MMSE supervisor scheme. Only implementation of
the BPF in the lter bank system is required in the design process of the proposed
scheme. The centre frequency of the BPF is chosen to be the same as the natural
frequency of each corresponding model. The pass-band of the BPF is chosen to
be as narrow as possible so as to accurately represent the natural frequency of
the model. However, to implement a very narrow pass-band, a high order lter
is required. In the digital lter implementation, higher order lters require more
memory and more computational time. Therefore, it is desirable to design a
low order BPF with a narrow pass-band close to the lter centre frequency. To
obtain a narrow low order BPF, a BPF with resonant controller structure is used
in the lter bank of the M
4
RC. The BPF with resonant controller structure is
a second-order lter which produces a sharper shape than the commonly used
K
1
K
2
K
M
Plant
Adjustable
controller
Switch
system
Filter bank
system
r = 0
u
y
S
1
S
2
S
M
u
1
u
2
u
M
-
BF
1
BF
2
BF
M
Figure 3.12. Block diagram of M

4
RC for controlling M modes.
high order Butterworth lter. A comparison of the BPF with resonant controller
structure (k
c
= 10 and
c
= 0.01) with Butterworth BPFs of dierent order is
shown in Fig. 3.13. The gure shows that for frequencies close to the lter centre
frequency, the BPF with resonant controller structure produces a sharper shape
than a tenth-order Butterworth BPF.
Since the lter bank system selects the closest model that represents the cur-
rent vibration frequencies at every sample time, it requires lower computational
demand than the MMSE supervisor scheme. The new supervisor scheme only
needs to determine which BPF produces the maximum output. Moreover, rapid
switching does not occur in this scheme because the supervisor does not measure
the error signal between the plants output and the models output, which could
change rapidly, but instead uses the proximity of the plant natural frequencies
to the xed mode frequencies of the known models. Another advantage of the
proposed supervisor scheme is that the parallel lter bank system can be easily
embedded in a hardware system, which would reduce the computational time
even more thus making it more suitable for real-time implementation.
Figure 3.13: Comparative responses of BPF with resonant controller
structure with Butterworth BPF.
The M
4
RC Algorithm
The algorithm for each mode is as follows:
1. Using the lter bank system select the BPF that gives the maximum output.
2. Load the parameters of the corresponding model into the adjustable reso-
nant controller.
3. Calculate the control signal
u
m
(k) = A
m
y(k) +B
m
y(k 1) +C
m
y(k 2) D
m
u
m
(k 1) E
m
u
m
(k 2)
(3.39)
The nal control signal is the summation of control signal for each mode
u(k) =
M
m=1
u
m
(k)k
dm
, (3.40)
M
4
RC Stability
The multiple model control system using a supervisor scheme such as the
M
4
RC can be categorized as a linear switched system. A linear switched system
can be considered as a system that is composed of linear sub-systems. It is
well-known that when all the linear systems which make up a switched system
are stable, then the entire system is stable for any switching signal if the time
between consecutive switching, known as the dwell time, , is suciently large
[76, 156]. Furthermore, as stated in [100], the stability of multiple model control is
guaranteed under the assumption that at every time there is at least one controller
in the controller bank stabilize the system, and the interval between successive
switches have a nonzero lower bound, which can be arbitrarily chosen to be small.
For the proposed M
4
RC, stability is guaranteed because all the controllers in the
controller bank stabilize the system and no rapid switching occurs between the
controllers. To ensure stability an arbitrary small can be applied to the M
4
RC.
3.5.2 Case 2: Only the upper and Lower Bounds of Op-
erating Region are a priori Known
Determining the number of required models needed to span an operating region in
the multi-model control approach becomes a design problem if a priori knowledge
about all the possible loading conditions is not available. The number of required
models is determined by the bandwidth of the operating region and the choice of
controller design. For a given bandwidth and operating region, a large number
of models is required if narrow-band controllers are used, and fewer models are
required if wide-band controllers are used.
The number of required models in the M
4
RC is very large due to the very
narrow-band of the resonant controller, as demonstrated in the following illustra-
tive example.
Example 3.5.1
Consider a resonant controller that is designed with an operating point at 10 Hz
with a damping ratio
c
= 0.01 and a gain k
d
= 10. The frequency response of the
controller is shown in Fig. 3.14. From the gure, it can be seen that the controller
gain decreases by 15 dB for only a 0.05 Hz operating point variation. Assuming
that a 15 dB reduction in the controller gain gives an acceptable closed-loop
attenuation performance, then one model is required for every 0.1 Hz operating
point variation, as shown in Fig. 3.15. As can be seen from Table 2.3, the
operating regions for the physical plant under investigations are : 6 Hz - 10.5 Hz
; 44.8 Hz - 64 Hz ; 132.7 Hz - 174 Hz for the 1
st
, 2
nd
, and 3
rd
mode, respectively.
Therefore, to span the operating region of the experimental plant the M
4
RC would
require 45, 192, and 413 models for the 1
st
, 2
nd
, and 3
rd
mode, respectively or
650 models for the rst three modes of vibration. The number of required models
would increase for a tighter requirement (e.g., 5 dB reduction in the controller
gain) or a wider operating region which would lead to a system with a very large
number of models.
The example demonstrates that if only the lower and upper bounds of the
operating region are a priori known in the M
4
RC method then a very large
number of xed-parameter models are required. A fast computer with lots of
memory is required to implement such a control method. Therefore, the M
4
RC
method is only practical for systems where all the loading conditions are a priori
known. Simulation and experimental studies comparing the performance of the
M
4
RC method for known loading conditions and for unknown loading conditions
are reported in the following sections.
Figure 3.14: Frequency response of resonant controller with
c
= 0.01
and k
d
= 10.

Figure 3.15. Model array in the M
4
RC model bank.
3.6 Simulation Studies
In the following simulation studies, the resonant control method and the M
4
RC
method are applied to control the cantilever beam when loaded with the various
loads (Model 1 to Model 4), as described in Chapter 2.
The objectives of the rst simulation study (3.6.1) of the resonant control
method are to demonstrate that:
1. Resonant control is able to attenuate multi-mode vibration using only a
single sensor-actuator pair.
2. Resonant control has an independent characteristic, in the sense that the
controller is able to control a particular mode without destabilising the
other modes.
3. Resonant control fails to give optimum performance when the natural fre-
quencies of the system are altered due to load changes.
The objective of the second simulation study (3.6.2) is to demonstrate that:
1. The proposed M
4
RC supervisor scheme does avoid rapid switching.
2. The proposed M
4
RC supervisor scheme has a reduced computational de-
mand vis-a-vis a multiple model control system using the MMSE supervisor
scheme (3.35) [6, 38, 75].
3.6.1 Resonant Controller
The resonant controller (3.6) is used to control the rst three vibration modes of
simulation models. The parameters of the controller are chosen as k
d1
= k
d2
=
k
d3
= 10 and
c1
=
c2
=
c3
= 0.01 through a trial and error process. Six
simulation cases, referred to as RC.1 to RC.6, are conducted in this study. In
the RC.1 to RC.4 cases, controllers that are specically designed for Model 1 to
Model 4 are used to control each corresponding model. These cases are designed to
test the performance of the resonant controllers ability to attenuate multi-mode
vibration when the controller centre frequencies match the natural frequencies
of the plant. The RC.5 case is similar to the RC.1 case, however, only the
controller for the second mode is activated by setting k
d1
= k
d3
= 0. The objective
of the RC.5 case is to demonstrate the capability of the resonant controller to
independently attenuate a single mode for multi-mode excitation. In the RC.6
case, a controller that is specically designed for Model 3 is used to control the
system where the loading conditions change from Model 13 4. This case is
designed to test the performance of the controller when the natural frequencies
of the plant vary due to changes in the plants loading. The congurations of the
plant and the controller for all the cases are shown in Table 3.2. The excitation
signal for all cases is a summation of three 1 volt sinusoidal signals representing
the rst three natural frequencies of the particular model.
Case Loading condition Controller design
represented by Model based on Model
RC.1 1 1
RC.2 2 2
RC.3 3 3
RC.4 4 4
RC.5 1 1
RC.6 134 3
only the 2
nd
mode controller is activated
Table 3.2: Plant and controller congurations for resonant controller
simulation study.
RC.1 to RC.4 cases
The system responses and the control signals for cases RC.1 to RC.4 are shown
in Figs. 3.16 to 3.19. From the gures it can be seen that the resonant control is
able to attenuate multi-mode vibration using only a single sensor-actuator pair.
The attenuation level for the dierent models is shown in Table 3.3, and the
attenuation of each mode for all the dierent models are shown in Figs. 3.20 to
3.23. The gures show that for each model and for each mode the attenuation
level is dierent. These dierences in levels of attenuation are attributable to the
sensitivity of the sensor to the dierent vibration modes and the authority of the
actuator in attenuating the same modes, which are in-turn determined by the
position of the sensor-actuator pair along the beam. It is observed in Chapter 2
that positioning the sensor close to the cantilevered end makes the sensor most
sensitive to the second mode vibration. Similarly due to the collocated position
of the sensor and actuator, the control authority for the second mode is at its
highest.
Attenuation (dB)
mode 1 mode 2 mode 3 overall
Model 1 10.0 17.8 17.2 19.7
Model 2 15.2 22.3 20.3 23.9
Model 3 17.2 19.9 16.2 21.4
Model 4 18.9 31.5 23.6 30.3
Table 3.3. Attenuation level for the range of models.
RC.5 case
The response of the system in the frequency domain for the RC.5 case is shown in
Fig. 3.24. From the gure it can be seen that the resonant controller produces the
same attenuation as the RC.1 case for the second mode but gives no attenuation
or amplication for the other two modes. The result shows that the resonant
controller is capable of attenuating specic modes independently of the other
modes, which is an advantage for certain applications where only specic modes
need to be controlled.
Figure 3.16: Response of Model 1 and the corresponding control signal.
Figure 3.20. Frequency response of Model 1.
Figure 3.24: Frequency response of Model 1 with only the 2
nd
mode
controller active.
RC.6 case
In the RC.6 case, the loading condition changes from Model 1 to Model 3 at t
= 30 seconds, and to Model 4 at t = 60 seconds. The open-loop and closed-
loop time domain responses of the system are shown in Fig. 3.25 and Fig. 3.26,
respectively. From the comparison of the two gures, it can be seen that between
t = 30 and t = 60 when the system loading condition is represented by Model
3, the controller gives a good performance with 21.4 dB attenuation. However,
when the system loading conditions are represented by either Model 1 or Model 4,
although the system is stable, the controller gives poor attenuation performance
with only 0.7 dB and 8.7 dB attenuation, respectively. When the Model 1 loading
condition is used, the controller only produces a maximum amplitude control
signal of around 0.3 volts, as shown in Fig. 3.26(b). Since this control signal only
generates a small force at the actuator, only a small amount of attenuation is
achieved. When the Model 4 loading condition is used, the controller produces a
maximum amplitude control signal of around 3 volts. However, due to mismatches
between the controller centre frequencies and the natural frequencies of the system
under control, optimum attenuation cannot be achieved. Therefore, it can be
evident that the resonant controller is sensitive to unforeseen changes in the
natural frequencies of the system for which it was not designed and for which it
could not compensate. In the next subsection, it will be shown that the multiple
model resonant controller is able to give better performance than the single model
resonant controller for systems with varying natural frequencies.
Model 1
Model 3
Model 4
Figure 3.25: Open-loop system response for the 1 3 4 model
sequence.
3.6.2 M
4
RC
To control the rst three modes of vibration, the proposed multiple model res-
onant control uses three sets of lter banks (BF
1
, BF
2
, and BF
3
), one for each
Model 1
Model 3
Model 4
Figure 3.26: Closed-loop system response for the 1 3 4 model
sequence.
mode of vibration. The band-pass lters (BPFs) in the lter bank are imple-
mented as BPFs with a resonant controller structure. The centre frequency of
each lter is set according to the frequencies of the corresponding rst 3 modes
of Model 1, Model 3 and Model 4, respectively. The performance of the proposed
M
4
RC supervisor scheme is compared with that of the MMSE supervisor scheme
(3.35). To give balance to both instantaneous and long-term measurement accu-
racy, the parameters and are chosen to have the same value. A moderate
value of is chosen to accomodate both disturbances and sudden changes. A
moderate value of should avoid unwanted switching due to disturbances in the
system and yet be fast enough to follow sudden changes in the system parameters.
Based on this reasoning, the supervisor scheme is implemented with = = 1
and = 0.5, and the performance index is computed for every 100 samples. The
model bank includes three models, Model 1, Model 3, and Model 4 where each
model is of order 20, as obtained in Chapter 2. The controllers parameters k
di
and
ci
are the same as those of the resonant controller described in the previous
subsection. The schematic diagrams of the implemented M
4
RC and of the multi-
model control with the MMSE supervisor scheme are shown in Fig. 3.27 and Fig.
3.28, respectively.
Two simulation cases referred to as M
4
RC.1 and M
4
RC.2 are tested. The
loading sequences and the controllers congurations for both cases are shown in
Table 3.4. The loading conditions in both cases are changed at t = 5 seconds and
t = 15 seconds.
S
1
S
2
S
3
BF
1
BF
2
BF
3
Plant
K
Adjustable
controller
Switch
system
Filter bank
system
r = 0 u y
-
Figure 3.27. Schematic diagram of the implemented M
4
RC.
Case Loading condition is Models in the model bank
represented by Model are based on Model
M
4
RC.1 134 1,3,4
M
4
RC.2 124 1,3,4
Table 3.4: Plant and controller congurations for M
4
RC simulation
study.
K
1
K
3
K
4
Plant
r = 0
Controller
Switch
u y
Model 1
Model 3
Model 4
-
+
Controller Bank
Model Bank
+
+
-
-
Supervisor
-
Figure 3.28: Schematic diagram of the multi-model control with
MMSE supervisor scheme.
M
4
RC.1 case
For the M
4
RC.1 case, all the loading conditions are represented in the model
bank. In this case, both supervisor schemes select the appropriate controller
corresponding to the current loading condition, as can be seen from the switching
behaviour shown in Figs. 3.29 (a) to (f). Fig. 3.30 shows that both supervisor
schemes give similarly good attenuation performance. This result conrms that
if the assumption that all the loading conditions are a priori known is satised,
then the proposed supervisor scheme will select the appropriate controller, and
the system performance will be optimum, similar to the more complex MMSE
supervisor scheme.
M
4
RC.2 case
For the M
4
RC.2 case, Model 2 is not included in the model bank. As the loading
condition changes to Model 2, rapid switching occurs between all the controllers
in the controller bank when using the MMSE supervisor scheme. The rapid
switching shown in Figs. 3.31(a) to (c) leads the controller to produce a satu-
Model
1 3 4 1 3 4
Model
Figure 3.29: M
4
RC switching behaviour for the 1 3 4 model
sequence. (a)-(c) using MMSE supervisor scheme for mode 1, 2 and 3
respectively. (d)-(f) using proposed supervisor scheme for mode 1, 2
and 3 respectively.
Model 1 Model 4
Model 3
Figure 3.30: Closed-loop multiple model resonant control responses for
the 1 3 4 model sequence. (a) using MMSE supervisor scheme.
(b) using proposed supervisor scheme.
rated control signal as shown Fig. 3.32(a), resulting in poor performance. The
controller amplies the vibration during the period 9 to 14 seconds as shown in
Fig. 3.33 (a). Conversely, rapid switching does not occur for the M
4
RC using
Model
1 2 4 1 2 4
Model
Figure 3.31: M
4
sequence. (a)-(c) using MMSE supervisor scheme for mode 1, 2 and 3
respectively. (d)-(f) using proposed supervisor scheme for mode 1, 2
and 3 respectively.
the proposed supervisor scheme. Instead, the supervisor chooses the model with
the closest natural frequencies to the excitation signal when the exact model is
not in the model bank. Due to the values of the models natural frequencies, as
shown in Table 2.6, when the loading condition changes to Model 2 the supervi-
sor chooses Controller 1, Controller 3 and Controller 4 for mode 1, mode 2, and
mode 3, respectively, as shown in Figs. 3.31 (d) to (f). In this way, although the
controller cannot give optimum performance for the unknown Model 2, it still
gives 3.64 dB attenuation as shown in Fig. 3.33(b).
Relative to the MMSE supervisor scheme, the proposed supervisor requires
less computational time. With three models in the model bank, the actual time
Model 1 Model 2 Model 4
Figure 3.32: Control signals generated by the multiple model resonant
control for the 1 2 4 model sequence. (a) using MMSE supervisor
scheme. (b) using proposed supervisor scheme.
Model 1
Model 2
Model 4
Figure 3.33: Closed-loop responses of multiple model resonant control
for the 1 2 4 model sequence. (a) using MMSE supervisor scheme.
(b) using proposed supervisor scheme.
required to compute 25 seconds of simulation time is 100 seconds using the MMSE
supervisor scheme compared to only 50 seconds using the proposed supervisor
scheme. In the MMSE supervisor scheme, the time needed for the computation
will signicantly increase as the number of models increase, as mentioned in
Section 3.4. In the case of the proposed supervisor scheme, increasing the number
of models will only increase the number of band-pass lters without signicantly
increasing the time needed to compute which band-pass lter gives the maximum
output. The dierence in the computation times between the two supervisor
schemes shows that for the real-time implementation, the proposed supervisor
scheme is more practical than the MMSE supervisor scheme.
3.7 Experimental Studies
Experimental studies are used to verify the results of the simulation studies.
The resonant controller and the proposed multiple model resonant controller are
implemented on a dSPACE
TM
DS1103 data acquisition and control board using
Matlab
TM
, Simulink
TM
and Real-Time Workshop
TM
software. The schematic
diagram of the experimental set-up is shown in Fig. 3.34. The power supply
drives the coils that hold the loads to the beam through a switch box. Since
the maximum output voltage from the digital to analogue converter (DAC) in
the dSPACE is 5 volts and the maximum allowable voltage to be applied to the
piezoactuator is 100 volts, a piezo power amplier with a gain of 20 is used to drive
the piezoactuator. The output signal from the piezosensor which has a maximum
amplitude in order of 500 volts is divided by 100 using a dierential probe before
feeding it back to the analogue to digital converter (ADC) in the dSPACE board.
As a precaution, and so as not to break the piezoactuator, the voltage applied
to the piezoactuator is limited to 80 volts. Therefore, a saturation block with
a maximum and a minimum output 4 is added to the controllers output to
limit the control signal. In the experiment, the control signal and the disturbance
signal are applied to the system through the same piezoactuator. All the lter
and controller parameters are the same as those used for the simulation studies in
the previous section. Similarly to the simulation studies two sets of experimental
studies are conducted, one using the resonant controller and the other using the
multiple model resonant controller.
dSpace
Piezo
amplifier
Differential
probe
Power
supply
A B
Switch box
Figure 3.34. The experimental set-up.
3.7.1 Resonant Controller
The same protocol is followed for the physical experiments as for the simulation
studies in Section 3.6.1. Once again, six experimental cases, RC.1 to RC.6, as
shown in Table 3.2 are conducted.
RC.1 to RC.4 cases
For the RC.1 to RC.4 cases, the system responses and the control signals in
the time domain are shown in Figs. 3.35 to 3.38, and the system responses in
the frequency domain are shown in Figs. 3.39 to 3.42. The gures show good
agreement with the simulation results. The attenuation levels for the dierent
models are shown in Table 3.5.
Attenuation (dB)
mode 1 mode 2 mode 3 overall
Model 1 14.0 23.3 12.0 21.0
Model 2 17.6 20.8 18.2 25.1
Model 3 16.7 24.8 20.2 24.4
Model 4 22.6 27.3 21.4 30.5
Table 3.5. Attenuation level for the range of models.
RC.5 case
The frequency response of the system for the RC.5 case, as shown in Fig. 3.43,
conrms that the resonant control is able to control a specic mode of vibration
without destabilizing uncontrolled modes.
Figure 3.43: Frequency response of Model 1 with only the 2
nd
mode
controller active.
RC.6 case
For the RC.6 case, the open loop response of the system, when the system changes
from one loading condition to the next, is shown in Fig. 3.44. The gure shows
that there is good agreement with the corresponding simulation result. However,
the experiment result shows that there is a peak in the downward direction when
the system changes from one loading condition to the next. The explanation for
this phenomenon is that releasing a load is equivalent to applying a pulse of force
which causes a large non-symmetric oscillation. The amplitude of the upward
oscillation is smaller because the force of gravity makes the upward resultant
force smaller than the downward resultant force. The closed-loop response and
the control signal for the system are shown in Fig. 3.45. This gure also shows
that there is good agreement with the corresponding simulation result.
Model 1
Model 3
Model 4
Figure 3.44: Open-loop system response for the 1 3 4 model
sequence.
Model 1
Model 3
Model 4
Figure 3.45: Closed-loop system response for the 1 3 4 model
sequence with the controller designed based on Model 3.
3.7.2 M
4
RC
For the M
4
RC experimental study, only the proposed supervisor scheme is tested.
Three cases are tested in this study. The rst two cases M
4
RC.1 and M
4
RC.2
are the same as those in the simulation studies in Section 3.6.2 which are shown
in Table 3.4. The third case referred to as M
4
RC.3 is similar to the M
4
RC.1
case. However, in the M
4
RC.3 case the loading is directly changed from Model 1
(full load) to Model 4 (unloaded), resulting in a larger percentage change of the
parameters.
M
4
RC.1 case
For the M
4
RC.1 case, the controller is turned on at t = 8.7 seconds, and the
loading condition is changed to Model 3 at t = 27.3 seconds. In order to observe
the vibration amplitude for Model 3 without control, the controller is turned o
at t = 63.8 seconds. The controller is then turned on again at t = 82.5 seconds,
before the loading condition is changed to Model 4 at t = 100 seconds. Again, the
amplitude of Model 4 without control is observed by turning o the controller at
t = 137 seconds. Finally the controller is turned back on at t = 155 seconds. The
system response and the control signal are shown in Fig. 3.46. From the gure,
it can be seen that if all the possible loading conditions are included in the model
bank, the M
4
RC will give optimum attenuation for a system with varying loading
condition. The supervisor scheme selects the corresponding controller associated
with the current loading condition as shown in Fig. 3.47. The gure shows good
agreement with the corresponding simulation result.
controller on
model
change
model change
controller on
controller on
controller off controller off
Figure 3.46: M
4
RC closed-loop response for the 1 3 4 model
sequence.
M
4
RC.2 case
For the M
4
RC.2 case, the times at which the controller is turned on and o,
and the loading condition changed are similar to the M
4
RC.1 case. The system
Model 1
Model 3 Model 4
Figure 3.47: M
4
sequence.
response and control signal are shown in Fig. 3.48. The gure shows that the
controller performs well for the Model 1 and Model 4 loading conditions. When
the Model 2 loading condition, which is not included in the model bank, is ap-
plied, the supervisor selects the closest mode controller as shown by the switching
behaviour given in Fig. 3.49. Since Controller 3 has the closest centre frequencies
to those of Model 2 for all modes, as shown in Table 2.3, the supervisor then
selects Controller 3 for all the modes. However, because Controller 3s centre
frequency are signicantly dierent from the natural frequencies of the current
loading condition the attenuation performance could be poor, as can be seen for
the Model 2 case. Overall, although there are times when the supervisor selects a
less than ideal controller, the rapid switching that can degrade the performance
of the system or lead to instability of the system does not occur in the proposed
M
4
RC.
controller on
model change
model change
controller on
controller on
controller off
controller off
Figure 3.48: M
4
RC closed-loop response for the 1 2 4 model
sequence.
M
4
RC.3 case
For the M
4
RC.3 case the controller is turned on at t = 8.7 seconds and the
loading condition is changed to Model 4 at t = 27.3 seconds. To obtain a clear
comparison between the amplitude of the vibration for Model 4 without and with
control, the controller is turned o at t = 63.8 seconds and turned back on at
t = 82.5 seconds. The system response and corresponding control signal, and
the supervisor scheme switching behaviour are shown in Fig. 3.50 and Fig. 3.51,
respectively. Fig. 3.50 shows that as long as all the possible loading conditions are
included in the model bank, the M
4
RC will give optimum attenuation, regardless
of the magnitude of the parameter variations. By comparing Fig. 3.50 with Fig.
3.46 in the M
4
RC.1 case, it can be seen that the only dierence in the system
response is the transition time when the loading condition is changed. The results
show that larger load changes produce larger transition times.
Figure 3.49: M
4
sequence.
controller on
model change controller on
controller off
Figure 3.50: M
4
RC closed-loop response for the 1 4 model sequence.
Model 1 Model 4
Figure 3.51: M
4
RC switching behaviour for the 1 4 model sequence.
From the M
4
RC.2 case, it can be seen that the controller fails to give optimum
performance when the range of possible loading conditions is not fully represented
in the model bank. The optimum performances of the M
4
RC can be achieved
by adding additional models to the model bank to cover all loading possibilities.
However, this approach could be impractical especially for a large number of
models. Therefore, it is important to nd an alternative method for improving
the performance of resonant control for systems with varying natural frequencies.
3.8 Summary
Due to its sensitivity to the variations of the natural frequencies of the system, the
resonant control cannot give optimum performance when controlling the vibration
of a exible beam under varying load conditions. To improve the performance of
the resonant control, the M
4
RC with a reduced complexity supervisor scheme is
proposed in this chapter. The supervisor scheme design uses a lter bank scheme
to determine how close the plants vibration frequencies are to the natural fre-
quencies of the a priori known models. At every sampling instant the supervisor
identies the closest model to the plants current condition and selects the corre-
sponding controller to produce the control signal. Because the basis for selection
is only dependant on how close the models are to the plants current conditions
in the frequency domain, the supervisor scheme is computationally simple and is
able to avoid rapid switching between controllers.
The simulation experimental studies are in agreement and show that the pro-
posed M
4
RC gives good performance for the control of vibration in a cantilever
beam with varying loads. However, when not all of the possible loading conditions
are represented in the model bank, the performance of the controller is degraded.
Although the performance can be improved by increasing the models in the model
bank to represent all the possible loading conditions, the large number of models
required make this method impractical. The large number of models necessitate
the use of fast computers with lots of memory. Therefore, an alternative method
such as adaptive control should be considered to obtain better performance for
systems with unknown possible loading conditions.
Chapter 4
Natural Frequency Estimator
A parameter estimator is required to form an adaptive control method. In this
chapter, a new method for estimating the natural frequency of a exible structure
is presented. The chapter starts with a review of the various methods for esti-
mating the natural frequency of a system, along with their shortcomings. This
is followed by a description of the proposed method. A brief discussion of the
general concept of on-line parameter estimation, and a review of the standard
RLS algorithm are given in Section 4.2. Stability analysis and an analysis of
the factors that inuence the accuracy of RLS are given in Sections 4.3 and 4.4,
respectively. The proposed natural frequency estimator method is then given in
Section 4.5. The results from the simulation studies and the experimental imple-
mentation to evaluate the performance of the proposed method are presented in
Sections 4.6 and 4.7, respectively.
4.1 Introduction
As can be seen from Chapter 3, the M
4
RC method is capable of eectively con-
trolling a vibrating system with varying loads if all the possible loading conditions
are a priori known. If, however, the loading conditions cannot be previously pre-
dicted (e.g., when the damage to an aircraft wing changes its loading or damping
properties, or when a robot arm is lifting samples of unknown mass and/or is
133
CHAPTER 4. NATURAL FREQUENCY ESTIMATOR 134
being subjected to unforeseen disturbances), then an adaptive resonant control
method is required to handle such variability, which goes beyond the control
capability of its xed-parameter multiple-model counterpart.
In an adaptive control method, the plant parameters are identied on-line at
every instant and the controller parameters are updated in response to variations
of the plant parameters. This adaptation of the parameters enables the controller
to achieve optimum attenuation during load variations.
Three parameters (i.e., controller gain, damping ratio and centre frequency)
are involved in the resonant control design, as can be seen in (3.3). To implement
an ecient and eective adaptive resonant controller, an analysis to determine
which parameter(s) need to be updated is required. The analysis takes into
account the amount of computation needed to update the controller parameters,
and the sensitivity of the controllers attenuation performance to variations of
the plant parameters.
To achieve optimum attenuation, the controller centre frequency can be readily
set to be the same as the natural frequency of the plant. However, there is
no simple relationship between the controller gain and damping ratio with the
plant parameters. Optimum values of the controller gain and damping ratio can
be found through a trial and error procedure. From the relation between the
controller parameters and the plant parameters, it can be seen that the amount
of computation needed to calculate the controller centre frequency is much less
than that needed to calculate the other controller parameters.
As can be seen from example 3.2.1 in Chapter 3, the attenuation of the reso-
nant control is sensitive to the plant natural frequency variation and not sensitive
to the plant damping ratio variation. A small discrepancy between the plant nat-
ural frequency and the controller centre frequency produces poor attenuation,
while a large variation to the plant damping ratio only produces a small variation
to the attenuation. Therefore the controller centre frequency plays a major role
in the determination of the controllers performance.
From this discussion, it can be concluded that eective adaptive resonant
control can be obtained by setting the damping ratio and gain of the controller
to xed values, and choosing the controller centre frequency as the adjustable
parameter.
Since in and adaptive resonant control method the controller centre frequency
must be set to the same value as the plant natural frequency, a natural frequency
estimator is required to update the controller centre frequency. In the context
of adaptive control, the overall performance of the control system is signicantly
aected by the performance of the estimator. Therefore to obtain a high perfor-
mance adaptive control, a high performance estimator is required.
To implement adaptive resonant control several requirements for the natural
frequency estimator are essential. The estimator must be: on-line, accurate,
multi-mode, and simple. To be able to up-date the controller centre frequency at
every instant, the estimator must be able to operate on-line. Due to the sensitivity
of the controller to the plant natural frequencies estimation accuracy is essential
for optimum performance. Since the controller is intended to control multi-mode
vibration, the estimator must be able to estimate multiple modes of natural
frequency simultaneously. Finally, a feasible real-time implementation requires
an estimator that uses the minimal amount of computation. An estimator with
a simple algorithm is favoured for adaptive resonant control.
Several methods which can be used to estimate the natural frequencies of
exible structures can be found in the literature: modal testing [31, 140], goal
programming optimisation [124], maximum likelihood [145, 146], transformation
[111, 127, 155], adaptive input shaping [28], modal lter [17, 77, 78] and the
combined RLS-Bairstow method [121, 122].
Modal testing is the most widely used method [31]. Modal testing is an exper-
imental procedure in which the natural frequencies of a structure are determined
by vibrating the structure with a known excitation. While it vibrates, the struc-
ture will behave in such a way that some of the frequencies will not respond
at all or be highly attenuated, and some frequencies will be amplied in such a
way that the only limiting factor is the energy available to sustain the vibration.
These frequencies, where the structure resonates, are the natural frequencies of
the structure. Stobener and Gaul [140] used modal testing to nd the modal
parameters of a car body, and using the parameters they obtained they designed
a modal controller to attenuate the vibration in the car body. Although the re-
sults show that modal testing is able to accurately nd the natural frequencies of
the structure, this method can only be applied to structures with xed parame-
ters. Computing the modal parameters for variable structures is computationally
demanding and not feasible for on-line adaptation.
The goal programming optimisation method [124] and the maximum likelihood
method [145, 146] are methods that can be used to accurately estimate modal pa-
rameters (i.e., natural frequency and damping ratio) in noisy frequency response
functions (FRFs). In both methods, an initial estimation of modal parameters
is used to form an estimate of FRFs. The accurate values of the modal parame-
ters are then estimated by minimising the dierence between the measured FRFs
and the estimated FRFs iteratively, using the goal programming algorithm or the
maximum likelihood algorithm. Both of the methods, however, can only be used
o-line since they require the full set of experimentally measured FRFs as the
basis for minimisation.
Transformation methods such as the Fast Fourier Transform (FFT) [155] and
the Wavelet Transform (WT) [111, 127] are also used to estimate the natural
frequency of a structure. In the transformation methods an excitation is applied
to vibrate a structure. The structures output signal is then transformed into the
frequency domain to obtain a frequency spectrum. The natural frequencies of the
system are identied in terms of the spectrums peak responses. One shortcom-
ing of the transformation methods is computational complexity. The number of
calculations for the Radix-2 FFT algorithm is
N
2
log
2
N complex multiplications
and Nlog
2
N complex additions, where N is the number of samples of the signal
[117]. The FFT also requires batch processing of N time samples of the signal
simultaneously to produce spectrum estimates at discrete frequency spacings of
N
1
times the sampling frequency. As an illustration, to obtain 0.5 Hz esti-
mation accuracy using a 1 kHz sampling frequency, 2048 samples are required.
For this number of samples, 11,264 complex multiplications and 22,528 complex
additions need to be executed in the FFT algorithm. Therefore to estimate the
natural frequencies of multiple modes with reasonable accuracy, a large compu-
tational eort is required. While the FFT computational complexity may be
acceptable for some signal processing methods, it is less likely to be satisfactory
for adaptive control methods, where estimates of the plant are used to tune the
controller [15]. Similarly to the FFT method, the WT method is computationally
complex.
The adaptive input shaping method (AIS) [28] is a method which can be used
to estimate and track the natural frequency of a system. In this method, a learn-
ing rule based on the measurement of the phase and the vibration percentage of
the residual vibration is used to estimate the natural frequency of the system.
The vibration percentage is calculated by comparing the magnitude of vibration
before and after the shaping process, and the phase is estimated using a least-
squares method based on trigonometric identity. This method oers more ecient
computation than the FFT. For N signal samples, AIS only needs 14N + 6 real
multiplications, 10N 7 real additions, N power calculations, 2N table look-ups
and one division. To obtain the same accuracy as with FFT only 0.05N signal
samples are required with AIS [28]. As an illustration, to obtain a 0.5 Hz estima-
tion accuracy using a 1 kHz sampling frequency, 103 samples are required. The
method then needs to calculate 1448 real multiplications, 1023 real additions,
103 power calculations, 206 table look-ups and one division, around 10% of the
number of computations required by the FFT. This method can be used to es-
timate the natural frequency of a multi-mode system by isolating each mode of
interest using lters, and then applying the algorithm to the ltered responses,
one algorithm for each mode. The computational load, however, will increase in
proportion to the number of modes. Another limitation of this method is that
the convergence of the estimation to the actual frequency is only guaranteed if
the initial estimate is within the region 50% of the actual frequency. Therefore
for systems with large natural frequency variations, the method cannot track the
variations and consequently accurate estimation cannot be obtained.
Lim et al. [77] used an adaptive lter to estimate the natural frequency of
a exible structure to detect structural damage. The adaptive lter is formed
as a modal lter to represent a single-degree-of-freedom transfer function, and
the coecients of the lter are updated using the least-mean-squares algorithm.
Due to the use of a modal lter, which represents a single mode system, this
method is only able to estimate the natural frequency for a single mode. To
enable multi-mode frequency estimation, Bosse et al. [17] combined the modal
lter with neural networks in their endeavour to estimate the natural frequencies
of a space truss structure. The neural networks updates the lters coecients
and works as a band-pass lter to isolate any one mode from the other modes. To
estimate multi-mode natural frequencies, a set of parallel combined modal lter
- neural networks is used. The use of a modal lter requires one sensor for each
mode of interest, which is a disadvantage. Furthermore, due to the large number
of computations required, this method converges slowly. This slow convergence
rate is acceptable for a space structure, but it is not suitable for structures that
have a relatively short time constant compared to that of a space structure.
Rew et al. [121, 122] surveyed the frequency estimation approaches and devel-
oped a real-time multi-mode natural frequency estimator based on the Recursive
Least-Squares (RLS) method combined with the Bairstow method. The Bairstow
method is an algorithm for nding the roots of a real polynomial of arbitrary or-
der. Using this method an even order polynomial can be decomposed into several
second-order polynomials. In the combined RLS-Bairstow method, estimation
is achieved in two steps. In the rst step the coecients of the characteristic
equation of the system are estimated using the RLS algorithm. The result from
the estimation is a polynomial of order 2M, where M is the number of modes of
interest. In the second step, the polynomial is decomposed into M second-order
polynomials. From each second-order polynomial, the natural frequency of each
mode can be extracted. The RLS method is selected for this application because
it gives fast parameter convergence allowing fast adaptation under changing con-
ditions and requires a relatively small computational eort, which is crucial for
real-time applications. The total number of calculations required in the RLS is
only 4n
2
where n is the order of the system [61], which is very low compared to
the FFT and AIS methods. Rew et al. [121, 122] then used the estimator in
an adaptive positive position feedback control to attenuate vibration in a wing
like composite structure. The experimental results from those papers show that
for multi-mode frequency estimation, the estimator is sensitive to the choice of
sampling rate and to unmodeled high frequency modes. The estimator will fail to
give reasonable estimations for the lowest mode if the sampling rate is too high
relative to the natural frequencies of those modes. Thus, for a system with reso-
nant modes that are spaced more than a decade apart in frequency, the estimator
cannot give reasonable results.
Of the above methods, only FFT, AIS, combined modal lter - neural networks
and combined RLS-Bairstow are able to work on-line and estimate the multi-mode
natural frequency. Comparing the four methods, RLS is the simplest to compute,
while the estimation accuracy of all the methods are similar. Hence, RLS is the
design method chosen for the natural frequency estimator used in this research.
A further analysis of the RLS method is undertaken with a view to improve the
estimation accuracy for wide-band systems.
One reason why the RLS method fails to give an accurate estimation of the
natural frequency for wide-band systems is that it has to deal with data that
are scattered across the frequency domain [21]. Frequency domain analysis by
Wahlberg and Ljung [148] reveals that the RLS method has a high-pass charac-
teristic which becomes more pronounced as the sampling rate is increased. This
high-pass characteristic gives a higher weighting to the higher frequencies and
a lower weighting to the lower frequencies. Hence, the estimation for the lower
frequency modes becomes inaccurate. This analysis leads designers to use pre-
ltering to compensate for the low weighting at low frequencies [20, 125, 126] by
giving more weighting to the lower frequencies and removing frequencies above the
highest mode of interest. However, although Rovner and Franklin [126] show that
preltering improves the estimation accuracy, particularly for the highest modes,
the estimators high-pass lter characteristic still produces inaccurate results for
the lowest mode. Another reason for the estimation inaccuracies is round-o er-
rors as a result of nite word-length in digital system realizations [9, 59]. For
limited bit representations, the estimation accuracy will decrease as the sampling
rate and/or the order of the estimator increase.
These reasons prompt further consideration of preltering design and of cri-
teria for selecting minimal-order and appropriate sampling rates for the imple-
mentation of the RLS estimator. A new approach to obtain an accurate natural
frequency estimator for exible structure based on the preltering design, sam-
pling rate selection, and structure selection of an RLS estimator is proposed in
this thesis.
In the following section, the on-line parameter estimation concept is briey
discussed and the implementation of an on-line estimator based on the RLS al-
gorithm is reviewed.
4.2 On-line Parameter Estimation Using RLS
On-line plant parameter estimation is a technique for solving system identication
problems by using parametric estimation methods to process sequential pairs of
input-output data measurements. Fig. 4.1 [68] illustrates a schematic diagram of
the on-line parameter estimation principle. An input u(t) is applied to both the
plant and the adjustable model. The adjustable models parameter vector is
then updated by the parameter adaptation algorithm. The parameter adaptation
algorithm updates the adjustable models parameters by reducing the error (t)
between the plant output, y(t), and the output predicted by the model, y(t), at
each sampling instant t.
In general the parameter adaptation algorithm has a recursive structure as
illustrated in Fig. 4.2 [70]. The new values of the estimated parameters are
equal to the previous values plus a correcting term that depends on the most
recent measurements. There are several approaches for deriving the parameter
adaptation algorithm such as gradient descent, least-mean-squares minimization,
least-squares minimization, and fuzzy logic or neural networks. While the gradi-
ent descent technique and the least-mean-squares method have a lower compu-
tational load than the least-squares method, they have a slower convergence rate
[18, 47].
The rst step in the estimator design is to determine a class of model. In
Plant to be
identified
Adjustable
mathematical
model
Parameter
adaptation
algorithm
+
-
) ( t y
) ( ) ( ) ( t y t y t =
) (t y
) (t u
Figure 4.1. Block diagram of an on-line parameter estimator.
Newly
estimated
Parameter
(vector)
=
Previously
estimated
Parameters
(vector)
+
Adaptation
Gain
(matrix)
Measurement
Function
(vector)
Prediction
Error
Function
(scalar)

Figure 4.2. Recursive structure of parameter adaptation algorithm.
general a linear time-invariant discrete-time model can be given by [81]
y(t) = G(q, )u(t) + H(q, )e(t), (4.1)
where u(t) and y(t) are the input and output of the system, respectively, e(t) is
the disturbance with zero-mean value and variance , and is the set of system
parameters. The model spectrum of the disturbance is [81]
e
(, ) =
H(e
i
, )
2
. (4.2)
The transfer function G(q, ) and H(q, ) are functions of the delay operator q
1
,
where q
k
x(t) = x(t k). If the disturbance lter H(q, ) is minimum phase, the
disturbance e(t) in (4.1) can be written as
e(t) = H
1
(q, )y(t) H
1
(q, )G(q, )u(t). (4.3)
Then the optimal predictor for y(t) is
y(t|) = y(t) e(t)
y(t|) = H
1
(q, )G(q, )u(t) +
_
1 H
1
(q, )
_
y(t). (4.4)
Then if H
1
(q, )G(q, ) and H
1
(q, ) are both stable, one can write:
H
1
(q, )G(q, ) =
k=1
b
k
q
k
, (4.5)
H
1
(q, ) 1 =
k=1
a
k
q
k
. (4.6)
Truncating these expansions at k = n
b
and k = n
a
, respectively where n
b
and n
a
are the order of the polynomial B(q) and A(q), respectively, (4.1) can be written
as
_
1 +
n
a
k=1
a
k
q
k
_
y(t) =
n
b
k=1
b
k
q
k
u(t) + e(t) (4.7)
or in the compact form
A(q)y(t) = B(q)u(t) + e(t) (4.8)
with
A(q) = 1 + a
1
q
1
+ a
2
q
2
+ . . . + a
n
a
q
n
a
(4.9)
and
B(q) = b
1
q
1
+ b
2
q
2
+ . . . + b
n
b
q
n
b
. (4.10)
It can be seen that (4.8) corresponds to (4.1) with
G(q, ) =
B(q)
A(q)
, H(q, ) =
1
A(q)
. (4.11)
Dening
= [a
1
, . . . , a
n
a
, b
1
, . . . , b
n
b
]
T
(4.12)
and
(t) = [y(t 1), . . . , y(t n
a
), u(t 1), . . . , u(t n
b
)]
T
, (4.13)
(4.8) can then be written as
y(t) =
T
(t) + e(t). (4.14)
Model (4.8) is known as the AutoRegressive with eXternal input (ARX) model
or the least-squares model. Since the disturbance term e(t) in (4.8) enters as a
direct error in the equation, the model (4.8) is often referred to as an equation
error model or structure [81, 135].
The limitation of the ARX model (4.8) is the lack of adequate freedom in
describing the properties of the disturbance term in the governing equations of
a system, resulting in a possible inaccurate estimation [20]. The estimation ac-
curacy can be improved by describing the equation error as a moving average of
white noise [20, 81]. This modies (4.8) into
A(q)y(t) = B(q)u(t) + C(q)e(t) (4.15)
with
C(q) = 1 + c
1
q
1
+ c
2
q
2
+ . . . + c
n
c
q
n
c
(4.16)
where n
c
is the order of polynomial C(q). Clearly (4.15) corresponds to (4.1)
with
G(q, ) =
B(q)
A(q)
, H(q, ) =
C(q)
A(q)
(4.17)
where now
= [a
1
, . . . , a
n
a
, b
1
, . . . , b
n
b
, c
1
, . . . , c
n
c
]
T
. (4.18)
The model given in (4.15) is referred as the AutoRegresive Moving Average with
eXternal input (ARMAX) model.
Although the ARMAX model is better than the ARX model in terms of the
estimation accuracy, computationally it is much more complex than the ARX
model. In the computation of the coecients , the ARMAX model involves
non-linear optimisation [32, 47], which is complicated, computationally intensive,
and also sensitive to the initial guess parameter values [32]. From this viewpoint,
the ARX model oers the advantages of simpler computation and simpler im-
plementation compared to the ARMAX model. Furthermore, since only the AR
parameters (i.e., a
1
, a
2
, , a
na
) are required to estimate the natural frequency of
the system, as will be explained in Section 4.5, the ARX model is chosen as a
basis for the on-line estimator.
In the computation of the unknown parameter in the ARX model, the RLS
estimator ignores the disturbance e(t), and only uses the input pair u(t) and y(t).
Therefore, in the derivation of the RLS estimator algorithm, the disturbance e(t)
can be omitted from (4.14), then (4.14) can be written as
y(t + 1) =
n
a
k=1
a
k
y(t + 1 k) +
n
b
k=1
b
k
u(t + 1 k) =
T
(t) (4.19)
where
= [a
1
, . . . , a
n
a
, b
1
, . . . , b
n
b
]
T
(4.20)
and
(t) = [y(t), . . . , y(t n
a
+ 1), u(t), . . . , u(t n
b
+ 1)]
T
. (4.21)
From (4.19), the adjustable prediction model can described by [70]
y
o
(t + 1) = y
_
(t + 1)|
(t)
_
=
n
a
k=1
a
k
(t)y(t + 1 k) +
n
b
k=1
b
k
(t)u(t + 1 k)
=

T
(t) (4.22)
where y
o
(t + 1) is the a priori predicted output and
(t) = [ a
1
(t), . . . , a
n
a
(t)
,
b
1
(t), . . . ,
b
n
b
(t)]
T
. (4.23)
is the estimated parameter vector at instant t. In a similar way the a posteriori
predicted output can be given by
y(t + 1) = y
_
(t + 1)|
(t + 1)
_
=
n
a
k=1
a
k
(t + 1)y(t + 1 k) +
n
b
k=1
b
k
(t + 1)u(t + 1 k)
=

T
(t + 1)(t). (4.24)
From (4.19) and (4.22) an a priori prediction error can be dened as
o
(t + 1) = y(t + 1) y
o
(t + 1). (4.25)
Similarly, from (4.19) and (4.24), an a posteriori prediction error can be dened
as
(t + 1) = y(t + 1) y(t + 1). (4.26)
Based on the prediction models (4.22) and (4.24), a recursive parameter adap-
tation algorithm with the structure
(t + 1) =

+
(t + 1) =

(t) + f
_
(t), (t),
o
(t + 1)
_
(4.27)
can be formed. To design an RLS estimator, the parameter adaptation algorithm
(4.27) is solved by minimizing the least-squares criterion
N
= arg min
M
J
N
(
, t),
J
N
(
, t) =
1
N
N
t=1
_
y(t)
T
(t)(t 1)
_
2
, (4.28)
where the arg min function is the value of

which minimizes the cost function
J
N
(
, t), and M is the domain of admissible parameters related to the ARX

model. By setting the derivative of the cost function J
N
(
, t) in (4.28) with
respect to

equal to zero, one can obtain the RLS algorithm as [70]:
(t + 1) =

+ F(t)(t)(t + 1), (4.29)
F(t + 1)
1
=
1
(t)F(t)
1
+
2
(t)(t)
T
(t), (4.30)
F(t + 1) =
1
1
(t)
_
_
F(t)
F(t)(t)
T
(t)F(t)
1
(t)
2
(t)
+
T
F(t)(t)
_
_
, (4.31)
(t + 1) =
y(t + 1)
T
(t)(t)
1 +
T
F(t)(t)
, (4.32)
with
0 <
1
(t) 1; 0
2
(t) < 2; F(0) > 0 (4.33)
where F(t) is the adaptation gain matrix and
1
(t) and
2
(t) are weighting se-
quences.
From (4.31), it can be seen that the gain matrix F(t) is varying with time,
where the type of variation is determined by the choice of
1
(t) and
2
(t). The
weighting sequences
1
(t) and
2
(t) have an opposite eect to the variation of
F(t). While
1
(t) < 1 tends to increase F(t),
2
(t) > 0 tends to decrease it. For
1
(t) =
1
= 1 and
2
(t) =
2
= 1, the algorithm gives progresively less weight
to the new prediction error and thus to the new measurements. Consequently,
this type of variation of the adaptation gain is not suitable for the estimation
of time-varying parameters. For
1
(t) =
1
; 0 <
1
< 1 and
2
(t) =
2
= 1
the algorithm gives less weighting to the old data (k < t) and maximum weight
is given to the most recent error. This RLS algorithm is known as RLS with
constant forgetting factor, where
1
is the forgetting factor. The value of the
forgetting factor,
1
, determines the response of the algorithm. A small value of
1
produces a fast response for tracking parameter variations, but large uctuations
in the steady-state. On the contrary, small steady-state uctuations with a slow
tracking response are obtained for a large value of
1
. The typical values for
1
which give a relatively fast but small uctuation response are between 0.95 and
0.99 [70]. This algorithm is suited to the estimation of systems with time-varying
parameters. Due to its suitability, the RLS with constant forgetting factor is used
as the estimator in this research.
In the context of adaptive control, the stability of the estimator is a necessary
condition for the stability of the control system. Therefore, a stable estimator
must be employed in the adaptive control method.
4.3 Stability of the RLS Algorithm
The stability of an estimator is closely related to the convergence of the estimator.
An estimator that converges can be said to be stable. The convergence of an
estimator is determined from the behaviour of the prediction error. An estimator
converges if for bounded input, u(t), and bounded output, y(t), the prediction
error (t) is bounded. For the RLS estimator the convergence is guaranteed if the
prediction error in (4.32) is bounded or
lim
t
(t + 1) = 0. (4.34)
The stability of the RLS algorithm can be explained by the passivity theorem.
Using the passivity theorem it can be shown that for a bounded input, u(t) and
bounded output, y(t), the prediction error (t) is bounded. To apply the passivity
theorem, the equivalent feedback representation of the RLS algorithm needs to
be derived.
Dene the parameter error as
(t) =

(t) . (4.35)
Subtracting both sides of (4.29) by , and taking into account (4.35), one obtains
(t + 1) =

(t) + F(t)(t)(t + 1). (4.36)
From the denition of the a posteriori prediction error (t + 1) given by (4.26)
and taking into account (4.19), (4.24) and (4.35), one gets
(t + 1) = y(t + 1) y(t + 1)
=
T
(t)
T
(t + 1)(t)
=
T
(t + 1)(t)
= (t)
T

(t + 1). (4.37)
and using (4.36) one can write
T
(t)
(t + 1) =
T
(t)
(t) +
T
(t)F(t)(t)(t + 1). (4.38)
Equation (4.36), (4.37) and (4.38) dene an equivalent feedback system which
can be represented as in Fig. 4.3. From the gure, it can be seen that the RLS
algorithm can be described as the interconnection of a linear time invariant block
and a non-linear time-varying block. It is therefore reasonable to use the passivity
theorem to analyse the stability of the RLS algorithm.
After the equivalent feedback representation is obtained, it can then be proven
that the RLS algorithm is stable by showing that the interconnection of these two
blocks is passive (see [70] for detailed proof).
In the next section, the analyses of the eect of the high-pass characteristics
of the RLS, the eect of the sampling period selection, and the eect of the
word-length on the estimation accuracy, are discussed. The analyses provide
the justication for the use of the proposed method to improve the estimation
accuracy.
1
(X)
T
(X) F
q
-1
+
+
-
) 1 (
~
+ t
) (
~
t
Nonlinear/Time Varying
linear
) (t
) 1 (
~
) ( + t t
T

) 1 (
~
) ( + t t
T

Figure 4.3. Equivalent feedback representation of RLS.
4.4 Factors Which Inuence the Accuracy of
RLS
As mentioned in Section 4.1 there are several factors which aect the ability of
the RLS estimator to accurately estimate the natural frequencies of a exible
structure. The detailed analyses of these factors are presented in this section.
The discussion is divided into three subsections. In the rst subsection, the high-
pass characteristic of the RLS is analysed using frequency domain analysis. In the
second subsection, the eects of sampling period and nite word-length on the
estimators parameter tracking resolution are analysed. In the third subsection,
an examination of the sensitivity of the poles to the parameter perturbation
error is used to explain the eects of sampling period and nite word-length on
the estimation accuracy.
4.4.1 RLS Characteristics
The parametric identication for a exible structure gives biased results since
the order of the estimator is much smaller than the theoretically innite order of
the plant [2, 147]. To examine the bias distribution of the RLS estimator over a
range of frequencies, frequency domain analysis of the estimator is discussed in
this subsection.
A recursive estimator can be considered as an algorithm that recursively min-
imizes a quadratic criterion in terms of the plant model prediction error or the
adaptation error. The least-squares criterion in (4.28) can be written in terms of
the prediction error, (t), as
N
= arg min
M
J
N
(
, t),
J
N
(
, t) =
1
N
N
t=1
2
(
, t). (4.39)
For large N (i.e., N approaches innity) the cost function J
N
(
, t) can be written
as
lim
N
J
N
(
, t) = J
N
(
, t) (4.40)
where
J
N
(
, t) = lim
t
1
N
N
t=1
2
(
, t)
= E{
2
(
, t)}, (4.41)
is the asymptotic least-squares criterion. Assuming that the RLS estimator con-
verges, the estimated parameter

N
will converge to the optimal parameter esti-
mation

, and can be written as [148]
, as N (4.42)
where
= arg min
M
J
N
(
, t), (4.43)
Applying the Parseval Theorem to (4.41), a frequency interpretation of the asymp-
totic least-squares criterion can be obtained as [148]
= arg min
M
/T
_
/T
(e
j
)d, (4.44)
where
(e
j
) is the spectrum of the prediction error and T is the sampling period.
Equation (4.44) shows how the cost function of the RLS criterion in terms of the
adaptation error, (t), is distributed in frequency. By expressing the prediction
error in terms of the bias between the model and the estimation model, (4.44)
can be used to assess how the bias is distributed in frequency. To express the
prediction error, (t), in the estimation models parameters

, write (t) as the
dierence between the model output, y(t), and the estimation output, y(t),
(t) = y(t) y(t). (4.45)
The model output, y(t), is given in (4.1) and the optimal estimation output, y(t),
is given by
y(t) =

H
1
(q,

)

G(q,

)u(t) +
_
1

H
1
(q,

)
_
y(t), (4.46)
which is obtained by replacing the parameters H(q, ) and G(q, ) in (4.4) with the
estimation parameters

H(q,

) and

G(q,

). Using (4.1) and (4.46), the prediction
error (4.45) can be written as
(t) = y(t) y(t)
= G(q)u(t) + H(q)e(t)

H
1
(q)

G(q)u(t)
_
1

H
1
(q)
_
y(t)
=

H
1
(q)
_
G(q)

G(q)
_
u(t) +

H
1
(q)H(q)e(t)
=

H
1
(q)
__
G(q)

G(q)
_
u(t) +
_
H(q)

H(q)
_
e(t)
_
+ e(t).
(4.47)
To obtain the spectrum of (4.47), write
E{
2
(t,

)} = |
H
1
(q)|
2
_
|G(q)

G(q)|
2
E{u
2
(t)} +|H(q)

H(q)|
2
E{e
2
(t)}
_
+E{e
2
(t)}. (4.48)
Taking the Fourier transform of (4.48) and neglecting the terms that do not
depend upon the estimated parameters, one obtains
(e
j
) = |
H
1
(e
j
)|
2
_
|G(e
j
)

G(e
j
)|
2
u
() +|H(e
j
)

H(e
j
)|
2
e
()
_
,
(4.49)
where
u
() and
e
() are the spectrums of input and disturbance, respectively.
Inserting the spectrum of the prediction error (4.49) into (4.44), one obtains
= arg min
M
/T
_
/T
|
H
1
(e
j
)|
2
_
|G(e
j
)

G(e
j
)|
2
u
() +|H(e
j
)

H(e
j
)|
2
e
()
_
d.
(4.50)
Equation (4.50) is an interpretation of the asymptotic least-squares criterion as
a function of the system model in the frequency domain. It can be seen that
the optimum estimation value

is inuenced by the weighted error between the

true system parameter G and the estimated system parameter

G. Optimum

is obtained by minimizing the bias between

G and G. The weighting function,
|
H
1
(e
j
)|
2
, determines how much weight the estimator will give to minimize the
bias contribution at each particular frequency. For a large weight in a particu-
lar frequency range, the algorithm will tend to give small bias, resulting in an
accurate estimation in that range [88].
For RLS using the ARX model the weighting function |
H
1
(e
j
)|
2
is equal to
|
A(e
j
)|
2
(see (4.11)). For any order and any random coecients |
A(e
j
)|
2
has a
high-pass lter (HPF) prole as illustrated in Fig. 4.4. Therefore, for a broadband
input
u
(), the RLS algorithm will give a higher weighting to higher frequency
components and a lower weighting to the lower frequency components. This will
result in higher accuracies for the estimation of higher modes and, conversely,
lower accuracies for the estimation of lower modes. As T decreases, the pass-
band of |
A(e
j
)|
2
will shift to the right, resulting in a corresponding reduction
of |
H
1
(e
j
)|
2
for the same frequency. Consequently, for smaller T, the RLS will
produce even poorer estimates of the lower modes.
8
th
order
4
th
order
3
rd
order
2
nd
order
Figure 4.4. Frequency responses of

A(e
j
) for dierent order.
Preltering can be used to compensate for the low weighting in the low fre-
quency ranges, thereby improving the accuracy of estimation in the low frequency
ranges. Let L(q) be a prelter that is applied to both u(t) and y(t). Then from
(4.47), the ltered prediction error can be written as
(t) = L(q) [y(t) y(t)]
= L(q)
_
G(q)u(t) + H(q)e(t)

H
1
(q)

G(q)u(t)
_
1

H
1
(q)
_
y(t)
_
= L(q)
_

H
1
(q)
_
G(q)

G(q)
_
u(t) +

H
1
(q)H(q)e(t)
_
= L(q)

H
1
(q)
__
G(q)

G(q)
_
u(t) +
_
H(q)

H(q)
_
e(t)
_
+ L(q)e(t).
(4.51)
The optimal value of the estimated parameter in (4.50) can be written as
= arg min
M
/T
_
/T
|L(e
j
)|
2
|
H
1
(e
j
)|
2
_
|G(e
j
)

G(e
j
)|
2
u
() +|H(e
j
)

H(e
j
)|
2
e
()
_
d.
(4.52)
From (4.52) it can be seen that preltering modies the weighting function from
|
H
1
(e
j
)|
2
to |L(e
j
)|
2
|
H
1
(e
j
)|
2
, thus increase the exibility to set the weight-
ing function. Therefore, the overall weighting at the low frequencies can be
increased by choosing a prelter L(q) with a high gain at low frequencies.
From the frequency domain analysis in this subsection, it can be seen that
the RLS method has a tendency to emphasize the weightings attached to high
frequencies. This characteristic becomes more pronounced as T decreases, and
this implies that, without preltering and for very small T, the RLS method
will give inaccurate results for low frequency estimations. The eect of sampling
period and nite word-length on the estimation accuracy from the parameter
tracking resolution viewpoint is investigated further in the next subsection.
4.4.2 Parameter Tracking Resolution
The role of the estimator in an adaptive control system is to identify and track
changes in the plant parameters, no matter how small the changes are, and to pro-
duce accurate estimates which reect these changes. This requirement is directly
inuenced by the selection of the sampling period for the estimator as discussed
below.
The inuence of the sampling period T on the estimation tracking resolution
can be illustrated by mapping the plant poles from the s-plane to the z-plane for
dierent values of T. Assume that, for a xed-value of T, as the plant changes
from one model to another, one of the plant poles changes accordingly in the
s-plane from s
1
=
1
+ j
1
to s
2
=
2
+ j
2
. The corresponding pole locations
in the z-plane are
z
1
= e
s
1
T
= e
1
T
e
j
1
T
= |e
1
T
|
1
T,
z
2
= e
s
2
T
= e
2
T
e
j
2
T
= |e
2
T
|
2
T. (4.53)
The dierence between the arguments of z
1
and z
2
is
= |(
1
2
)T| = 2T|f
1
f
2
|. (4.54)
If, however, T changes from T
1
to T
2
, where T
2
= nT
1
(n > 0), then the ratio of
the corresponding dierences between the arguments of z
1
and z
2
can be written
as
2
=
T
1
|f
1
f
2
|
T
2
|f
1
f
2
|
=
T
1
nT
1
=
1
n
. (4.55)
From (4.55) it can be seen that, as T becomes n-times larger, the dierences
between the arguments of z
1
and z
2
also becomes n-times larger, thus resulting
in more accurate estimates. The following example illustrates this eect.
Consider a plant that changes from Model A with natural frequency 9.77 Hz
to Model B with natural frequency 7.53 Hz. The transfer functions of Model A,
M
a
, and Model B, M
b
, are as follows:
M
a
=
0.02768
s
2
+ 1.229s + 3775
, M
b
=
0.3398
s
2
+ 0.9461s + 2238
.
The models are both sampled with dierent sampling periods, T
1
= 1/500 second
and T
2
= 1/50 second (i.e., n = 10). The locations of the plant poles in the
z-plane for diering T are shown in Fig. 4.5. From the gure it can be seen
that, when the plant changes from M
a
to M
b
,
2
= 16.2
o
associated with T
2
is
10-times larger than
1
= 1.62
o
associated with T
1
. This demonstrates that the
estimators capacity to resolve small changes in natural frequency between one
model and another can be compromised by the use of inappropriately selected
high sampling rates when the bit-precision of the computing platform is limited.
The higher the sampling rate, the smaller the argument and, consequently, the
higher the bit-precision required.
2
Figure 4.5. Eect of sampling period on pole location.

The determination of the sampling period of the estimator is dependent on
the natural frequency of interest in the estimated system. If the sampling rate
is too high relative to the natural frequency, the resolution of the estimator will
be low and hence the estimation results will be inaccurate. Empirical results
show that a sampling rate between 4 to 20 times the natural frequency of the
system gives accurate estimates. This sampling rate range is in agreement with
recommendations from the literature [33, 45, 70, 107, 123, 150].
4.4.3 Parameter Perturbation Error
The RLS estimator (4.29) to (4.32) is realized as an Innite Impulse Response
(IIR) lter. In a digital hardware realization, the lters coecients are quantized
and stored as nite-word length values in a register. Due to the limitation of
the word-length, quantization errors occur which perturb the parameters of the
estimator and hence aect the estimator accuracy. This in turn is eected by the
sampling rate and the order of the estimator, as shown in the following analysis.
The IIR lter has the following transfer function [117]
G(z) =
n
b
k=0
b
k
z
k
1 +
n
a
k=1
a
k
z
k
. (4.56)
When the IIR lter is realized with quantized coecients, its tranfer function is
modied as follows
G
q
(z) =
n
b
k=0
b
qk
z
k
1 +
n
a
k=1
a
qk
z
k
(4.57)
where the quantized coecients {b
qk
} and {a
qk
} are obtained from their unquan-
tized counterparts {b
k
} and {a
k
}
a
qk
= a
k
+ a
k
k = 1, 2, . . . , n
a
b
qk
= b
k
+ b
k
k = 0, 1, 2, . . . , n
b
(4.58)
and where {a
k
} and {b
k
} are the quantization errors. The denominator of
G(z) in (4.56) can be expressed as
A(z) = 1 +
n
a
k=1
a
k
z
k
=
n
a
k=1
(1 p
k
z
k
) (4.59)
where {p
k
} are the poles of A(z). Similarly, the denominator of G
q
(z) in (4.57)
can be expressed as
A
q
(z) = 1 +
n
a
k=1
a
q
z
k
=
n
a
k=1
(1 p
qk
z
k
) (4.60)
where
p
qk
= p
k
+ p
k
, k = 1, 2, . . . , n
a
(4.61)
and p
k
is the error or perturbation resulting from the quantization of the lter
coecients. The relationship between the total perturbation error for the i
th
pole,
p
i
, and the quantization error {a
k
} can be given by [117]
p
i
=
n
a
k=1
p
i
a
k
a
k
=
n
a
k=1
p
n
a
k
i
n
a
l=1,l=i
(p
i
p
l
)
a
k
. (4.62)
Eectively, the expression for p
i
in (4.62) gives a measure of the sensitivity of
the i
th
pole to changes in the coecients {a
k
} due to the quantization error for
a
k
. From (4.62) it can be seen that the perturbation error also depends on:
1. the order of the estimator n
a
. The higher the order n
a
, the larger the error
for a xed word-length;
2. the sampling rate, which determines the separation of the poles. As the
sampling rate increases, the poles are drawn closer together, i.e., the term
(p
i
p
l
) representing the distance between poles {p
l
} to p
i
(where l = i)
becomes smaller. This results in a larger perturbation error p
i
. This eect
is illustrated in Fig. 4.5, where the separation of the poles for a given model
is decreased when the sampling rate is increased.
4.5 Design of the Natural Frequency Estimator
for Flexible Structures
The analysis in Section 4.4 shows that there are several factors which inuence
the accuracy of an RLS estimator. The factors can be summarized as the:
1. High-pass characteristic of the RLS algorithm. The RLS algorithm gives
higher gain for signals at higher frequencies, and consequently produces less
accurate results for lower frequencies.
2. Selection of the sampling period T. If the sampling period is too short with
respect to the mode of interest, it will make the estimation inaccurate for
that particular mode.
3. Parameter word-length representation. Higher estimation accuracy will be
obtained as the word-length for the IIR lter implementation is increased.
4. Order of the estimator. The higher the order of the estimator, the larger
the quantization error, thus resulting in less accurate estimation.
The high-pass characteristics of the RLS algorithm can be compensated for
by adding a prelter to the estimator. The prelter increases the weighting at
the frequency ranges of concern. However, as shown in [126] preltering alone
is not adequate for obtaining accurate estimations for the lower modes of wide-
band systems due to the miss match between the chosen sampling rate and the
frequencies of the lower modes. Using longer word lengths to improve the accuracy
is not always possible due to implementation cost and availability constraints.
Therefore a combination of preltering design, sampling period selection and
estimator order selection is required if improvements in estimation accuracy are
to be achieved.
What follow is an exposition of a design methodology for a multi-mode nat-
ural frequency estimator with superior performance in terms of accuracy and
simplicity to known natural frequency estimators described in the literature.
Rule of the estimator
In the proposed design, a bank of band-pass lters (BPFs) is used in the prelter-
ing process to decompose a system into several bands in the frequency domain.
Each band can be described as a second-order system which represents a system
mode of interest. A second-order estimator with an appropriate sampling pe-
riod for a specic mode is then used to estimate the natural frequency for that
mode. To estimate the multi-mode natural frequency, a parallel combination of
preltered second-order estimators can be employed.
Estimator Design
A linear time-invariant system can be band-wise approximated by the summation
of lower order systems [142]. Given a plant with a transfer function G(q), consider
the second-order approximation G
m
(q) = G(q)L
m
(q) where L
m
(q) is a BPF with
centre frequency
m
G
m
(q) =
b
1m
q
1
+ b
2m
q
2
1 + a
1m
q
1
+ a
2m
q
2
. (4.63)
Then the transfer function of the system can be described as
G(q) =
M
m=1
G
m
(q) (4.64)
where M is the maximum number of modes of interest. From (4.64), it can be seen
that the plant can be approximated as a sum of M decoupled systems G
m
(q).
The identication of plant G(q) can therefore be achieved by using cascaded
banks of M-parallel band-pass prelters L
m
(q) and M-parallel second-order RLS
estimators operated in a chain fashion, with each chain targeting an individual
mode. Assuming that the structure natural frequencies are widely spaced and
independent of each other (which is valid for many exible beam structures [90]),
each mode can be identied independently through their respective chains. The
cut-o of the band-pass prelter within each chain is chosen to admit the full
range of frequencies associated with its respective mode for all possible loading
conditions.
The RLS-based estimator within the chain uses a specic sampling rate ap-
propriate for that chains mode of interest, thus allowing the RLS pass-band to
be positioned over the mode of interest. The sampling rate is chosen to be greater
than 4 times the upper-bound frequency and less than 20 times the lower-bound
frequency of that mode, thus satisfying the consideration discussed in Section
4.4. The upper and lower bound frequencies of each mode for dierent loadings
can be predetermined through modal testing of the structure. In this way, the
proposed multi-mode estimator bank can eectively and accurately identify the
multiple resonant modes over a wide bandwidth for dierent loading conditions.
Dierent sampling rates can be applied to the various mode estimators without
aecting the stability of the overall estimator, because the estimator for each
mode is independent of the estimators for all the other modes.
To derive the natural frequency estimation algorithm for the m
th
mode using
RLS, rewrite (4.63) as
G
m
(z) =
b
1m
z + b
2m
z
2
+ a
1m
z + a
2m
. (4.65)
Applying the bilinear transformation
z =
2f
s
+ s
2f
s
s
(4.66)
to (4.65), where f
s
is the specic sampling rate for the m
th
mode, gives
G
m
(s) =
c
0
s
2
+ c
1
s + c
2
s
2
+ d
1
s + d
2
(4.67)
where
c
0
=
b
1m
+ b
2m
1 a
1m
+ a
2m
,
c
1
=
4f
s
b
2m
1 a
1m
+ a
2m
,
c
2
=
4f
2
s
(b
1m
+ b
2m
)
1 a
1m
+ a
2m
,
d
1
=
4f
s
4f
s
a
2m
1 a
1m
+ a
2m
,
d
2
=
4f
2
s
(1 + a
1m
+ a
2m
)
1 a
1m
+ a
2m
.
Comparing the denominator of (4.67) with the denominator of (2.47), the natural
frequency associated with the m
th
mode can be identied as
m
= 2f
s
1 + a
1m
+ a
2m
1 a
1m
+ a
2m
. (4.68)
From (4.68), it can be seen that
m
depends only on the coecients of the de-
nominator of (4.65) which decide the poles of that mode. As shown in Section
2.5 in Chapter 2, the locations of the systems poles are not inuenced by mode
truncation. Hence, truncation of the higher modes in the modeling will only give
inaccurate locations for the systems zeros, but not for the systems poles. The
natural frequency estimation obtained from (4.68) is therefore not aected by
mode-truncation. This implies that the proposed natural frequency estimator is
robust to unmodeled dynamics caused by mode truncation.
Estimator Algorithm
The algorithm for each component estimator is therefore specied as follows:
1. Find the parameters of (4.63) using the RLS algorithm
(t + 1) =

+ F(t)(t)(t + 1), (4.69)
F(t + 1)
1
=
1
(t)F(t)
1
+
2
(t)(t)
T
(t), (4.70)
F(t + 1) =
1
1
(t)
_
_
F(t)
F(t)(t)
T
(t)F(t)
1
(t)
2
(t)
+
T
F(t)(t)
_
_
, (4.71)
(t + 1) =
y(t + 1)
T
(t)(t)
1 +
T
F(t)(t)
. (4.72)
2. Calculate the corresponding natural frequency using (4.68).
Hence, the M-modes of interest can be collectively identied by a bank of M
parallel estimators.
For a single-input single-output system the number of multiply-accumulate
operations for the RLS algorithm is of the order of 4(2M)
2
, where 2M is the
order of the system [61]. Therefore, in addition to increasing the estimation
accuracy, the proposed M-parallel second-order estimator bank also reduces the
number of mathematical operations from 4(2M)
2
to around 16M operations. For
example, for M=3, there is a reduction in the number of operations from 144
operations to 48 operations (i.e., a 66.7% reduction). The reduction becomes
more signicant for larger M (e.g., 90% for M = 10).
The eectiveness of the proposed estimator is evaluated through simulation
and experimental studies which are discussed in the following sections. A real-
time estimator for estimating the natural frequency estimator of the rst three
modes of the beam is implemented using the Simulink Real-Time Workshop C
S-function.
4.6 Simulation Studies of the Proposed Natural
Frequency Estimator
In this section, two sets of simulation studies are undertaken to evaluate the
eectiveness of the proposed method.
The results of the rst set of simulations demonstrate how the choice of pre-
ltering and sampling period aect the accuracy of the estimator. They further
show that (i) the decomposition of a plant into several bands according to its
modes using a bank of BPFs improves the accuracy of estimation, and (ii) fur-
ther improvement can be achieved by using an appropriate sampling period for
each specic band.
In the second set of simulations, the estimator is used to estimate the natural
frequencies of the cantilever beam models. In this set of simulations, the accuracy
and convergence rate of the proposed estimator in tracking the changes to the
natural frequencies of the system are evaluated.
4.6.1 Eects of Preltering and T Selection
For the rst set of simulations, the twentieth-order Model 4 (the unloaded model)
obtained from Chapter 2 is considered to be a true model. Model 4 is chosen in
this simulation because this model has the highest rst mode natural frequency.
Therefore if inaccurate estimations of lower mode frequencies are obtained for
this model, inaccurate estimations would also be obtained for the lower modes in
the other models (Model 1 to Model 3). From the ten modes of the model, only
the rst three modes are estimated.
Four simulation cases for the beam model are designed as follows: the rst
case uses no preltering; the second case uses a single sixth-order estimator with
low-pass lter (LPF); the third case uses three parallel second-order estimators,
each with a BPF and a common value of T; the fourth case uses three parallel
second-order estimators, each with a BPF and a specically selected T according
to the mode of concern. To minimize frequencies outside the modes of concern
from entering the estimator, the cut-o frequency for the LPF is set slightly
higher than the natural frequency of the third mode. Similarly, the upper cut-o
frequencies of the three BPFs: f
u1
, f
u2
, and f
u3
, are set slightly higher than the
rst, second, and third natural frequencies, respectively. To isolate each mode
separately, the lower cut-o frequencies of the three BPFs: f
l1
, f
l2
, and f
l3
, are
set in such a way that the side lobes of each lter do not overlap. In the rst
three simulation cases, the sampling periods are all equal to T = 0.001 sec. In
the last simulation case, the sampling periods are T
1
= 0.02 sec, T
2
= 0.002 sec,
and T
3
= 0.001 sec. The FRF of the estimation results for the four simulation
cases are shown in Figs. 4.6 to 4.9.
From Figs. 4.6 to 4.9, it can be seen that the choice of prelter L(q) and
sampling period T inuence the estimation accuracy. Fig. 4.6 shows that without
preltering the estimator produces a poor result. Fig. 4.7 shows that using a
single LPF with a value of T, which is based on the third mode frequency, gives an
accurate estimation for the third mode, a less accurate estimation for the second
mode, and a false estimation for the rst mode. This demonstrates that the LPF
alone cannot compensate for the HPF characteristic of the RLS algorithm, and
that the RLS cannot handle a high-order system where the natural frequencies
are separated by more than one decade. This problem can be improved, as shown
in Fig. 4.8, by using a parallel set of second-order estimators with a bank of BPFs
where both third and second modes are correctly estimated. However, the rst
mode estimation is still inaccurate. This is because the 1 kHz sampling rate
used is far higher than the rst-mode frequency. Further improvement in the
estimation accuracy for all modes can be obtained by using the proposed multi-
rate, multi-mode estimator in which independent sampling rates are assigned to
the three parallel second-order estimators, as conrmed by Fig. 4.9.
true model
estimation result
Figure 4.6. Estimation result without preltering.
true model
estimation result
Figure 4.7. Estimation result with preltering using single LPF.
true model
estimation result
Figure 4.8. Estimation result with preltering using BPFs.
true model
estimation result
Figure 4.9: Estimation result with preltering using BPFs and dierent
sampling periods.
4.6.2 Natural Frequency Estimator for Cantilever Beam
Models
In the second set of simulations, a bank of three parallel second-order estimators,
E
1
to E
3
, is implemented. The same forgetting factor
1
= 0.98 is chosen for all
the estimators. This value is chosen as a trade-o between fast response to track
parameter variations and uctuations in the steady-state as mentioned in Section
4.2. The estimators are used to estimate the rst three natural frequencies of
the simulation models, Model 1 to Model 4, with natural frequencies as shown in
Table 2.5. To have appropriate sampling periods for each mode in the models,
as discussed in Section 4.4, the sampling periods for E
1
, (T
1
), E
2
, (T
2
), and E
3
,
(T
3
), are chosen as 0.02 sec, 0.002 sec and 0.001 sec, respectively. The sampling
rate selected for the simulation is 1 kHz, therefore the sampling periods T
1
and
T
2
are obtained by down-sampling the simulation rate by 20 and 2, respectively.
Considering the natural frequencies of the models, the upper and lower cut-o
frequencies for band-pass lters L
1
, L
2
and L
3
are chosen as follows: f
u1
= 20 Hz,
f
l1
= 2 Hz, f
u2
= 90 Hz, f
l2
= 30 Hz, f
u3
= 200 Hz, f
l3
= 100 Hz. The lter order
chosen is based on a trade-o between attenuation and latency. Higher order
lters give better attenuation but they increase the delay and take longer to
compute. Experiments show that a sixth-order BPF gives adequate attenuation
for signals outside the ranges of interest, then all the lters are implemented as
sixth-order Butterworth digital lters. The schematic diagram of the natural
frequency estimator bank is shown in Fig. 4.10. Four simulation cases referred
to as Est.1 to Est.4 cases are considered. The Est.1 and Est.2 cases assess the
accuracy of the estimator bank, and the Est.3 and Est.4 cases investigate how
the estimator copes with dynamic loading conditions. A white noise disturbance
is used in all cases.
20
L
1
L
2
L
3
Plant
E
3
E
2
E
1
2
Input
signal
Bandpass
filters
Down-
samplers
estimators
f
1
f
2
f
3
u(t) y(t)
L
1
L
2
L
3
20
2
Bandpass
filters
Up-
samplers
20
2
Down-
samplers
Estimator bank
Figure 4.10. Schematic diagram for natural frequency estimator.
Est.1 and Est.2 cases
In the Est.1 and Est.2 cases, Model 1 and Model 4, respectively, are used. These
cases represent the two extreme loading models for the beam. The estimates for
Model 1 are shown in Fig. 4.11, and the magnied portion of the steady-state
condition, which illustrates the accuracy of the estimates, is shown in Fig. 4.12.
Similarly the estimates for Model 4 are shown in Fig. 4.13, and the corresponding
magnied capture is shown in Fig. 4.14. From Figs. 4.11 to 4.14, it can be seen
that the estimator converges to the true values of the models and gives very
accurate results for all modes (the rst three modes) of the models with bias
errors ranging from 0.06% for the rst mode of Model 4 to 0.2% for the rst
mode of Model 1. The results show that the largest bias error occurred for the
rst mode of Model 1, which has the lowest natural frequency of all the models.
Est.3 and Est.4 cases
In the Est.3 and Est.4 cases, two loading conguration sequences namely 1 3
4, and 1 2 4, respectively, are tested. The result for the Est.3 and Est.4
cases are shown in Figs. 4.15 and 4.16, respectively. Figs. 4.15 and 4.16 show
that the estimator is able to track the varying parameters of the models when the
loading is suddenly changed. The convergence time of the estimator depends on
the magnitude of the loading change between the initial condition and the nal
condition. The larger the magnitude the longer the convergence time. The two
gures also reveal that regarding the use of dierent sampling periods for each
mode, the convergence rate of the mode estimators is slower for the lower mode
than for the higher mode. From the gures, it can be seen that the convergence
time ranges from around 1 second (for the third mode when the model changes
from Model 1 to Model 2) to about 7 seconds (for the rst mode when the model
changes from Model 3 to Model 4), which is fast enough to be used in a practical
adaptive control system.
Figure 4.11. Estimation results for Model 1.
Figure 4.12. Magnied steady-state results for Model 1 estimation.
Figure 4.13. Estimation results for Model 4.
Figure 4.15. Estimation results for the 1 3 4 load sequence.
Figure 4.16. Estimation result for the 1 2 4 load sequence.
To test the eectiveness of the proposed estimator in a real-time implementation,
and to verify the simulation results given in Section 4.6, experimental studies of
the proposed estimator are conducted and described here. The on-line natural
frequency estimator is implemented on a dSPACE DS1103 data acquisition and
control board using Matlab, Simulink and Real-Time Workshop software. In the
implementation, a rst-order LPF with low frequency cut-o (2 Hz) is added to
the output of each mode estimator. The lters are placed to lter the coecients
a
1m
and a
2m
from (4.29) before feeding them to (4.68). The low frequency cut-o
is chosen to minimize the uctuation (i.e., non dc component) in the coecients
a
1m
and a
2m
, while retaining the tracking ability of the estimator.
The set-up for this experiment is the same as that employed for the controller
implementation described in Chapter 3, with the schematic diagram shown in
Fig. 3.34. All the estimators parameters and the lters parameters are the
same as those used in the simulation in Section 4.6.2. The sampling rate is set
at 1 kHz and the beam is excited with white noise. The simulation cases Est.3
and Est.4 are repeated in the experiments.
The experimental results for the estimated frequencies for each mode are
shown in Fig. 4.17 for the Est.3 case, and in Fig. 4.18 for the Est.4 case. From
the gures, it can be seen that the estimator tracks the changes in the natural
frequency of the system as the system changes from one model to the next. The
estimated natural frequencies converge to their respective true values between 1
second (for the third mode when the model changes from Model 1 to Model 2)
and 3.5 seconds (for the rst mode when the model changes from Model 3 to
Model 4).
The convergence rate of the estimator used in the experiments is faster than
the convergence rate of the same estimator used in the simulations. This is rea-
sonable since in the physical implementation, the estimator receives data input
from real plant, which is faster than the simulation model. Moreover, in the phys-
ical implementation, the algorithm is embedded in the DSpace board resulting in
very fast execution, while in the simulated version the algorithm is executed by a
PC processor which also processes other tasks such as calculating and displaying
model responses. The convergence rate of the estimator in the physical imple-
mentation conrms that the estimator is fast enough to be used in a practical
adaptive control system.
To examine the accuracy of the estimator, the magnied outputs of the esti-
mations during steady-state are shown in Figs. 4.19 to 4.22. From the gures,
it can be seen that the estimator gives very accurate results with a maximum
variance of 0.11% for mode 1 of Model 1.
The physical implementation shows that the proposed estimator is able to
give accurate natural frequency estimations for a wide range of frequencies. The
estimation accuracy of the lowest mode of interest is comparable to the estimation
accuracy of the highest mode of interest regardless of the high-pass characteristic
of the RLS algorithm employed. It can be concluded from the experimental results
that the preltering selection and sampling period selection can be designed so as
to increase the estimation accuracy of RLS-based estimator. Experimental results
reveal that the proposed estimator gives accurate estimations for the rst three
natural frequencies of the exible structure. This demonstrates that the proposed
natural frequency is able to give accurate results in the presence of unmodeled
dynamics due to mode truncation.
4.8 Summary
In this chapter, a real-time natural frequency estimator for exible structures is
proposed. The estimator is designed to perform with high estimation accuracy
for wide-band exible structures. An analysis of the aect of sampling period
and nite word-length on the estimation accuracy reveals that the use of a higher
order and a lower sampling period in estimator produces poor estimation results.
To achieve high accuracy, therefore, the proposed estimator is built as a set
of parallel second-order RLSs with band-pass prelters and with independent
sampling periods for each mode. The parallel structure is constructed using a
bank of BPFs, which decomposes the plant into several decoupled second-order
systems. This parallel structure enables dierent sampling periods to be applied
independently with dierent modes. In this way a lower order estimator with
a suitable sampling period for each particular mode is obtained. In addition to
increasing the estimation accuracy, the parallel structure of the estimator with
its own preltering for each mode also increases the robustness of the estimator
to unmodeled dynamics, and also reduces the computational complexity.
The simulation results and experimental results show that the proposed es-
timator gives accurate estimation, and the estimator is able to cope with the
varying plant parameters and unmodeled dynamics.
Chapter 5
Adaptive Resonant Control
In this chapter, the natural frequency estimator from Chapter 4 is combined
with the resonant controller from Chapter 3 to form a new adaptive resonant
control (ARC) method. The ARC method is proposed to control systems with
unknown loading conditions. The transient response of the ARC is then improved
by combining it with the M
4
RC method to form the proposed multi-model multi-
mode adaptive resonant control (M
4
ARC) method. This chapter begins with an
introduction to adaptive control, with an emphasis given to transient response
of adaptive control, and then it goes on to introduce a multiple model adaptive
control method. The design of the proposed ARC and M
4
ARC are presented in
Section 5.2 and Section 5.3, respectively. The results from simulation studies and
experimental studies are given in Sections 5.4 and 5.5, respectively.
5.1 Introduction
To achieve optimum attenuation in systems with varying parameters and un-
known loading conditions, the adaptive resonant control method, ARC, is pro-
posed. ARC is formed by combining the resonant controller discussed in Chapter
3 with the on-line natural frequency estimator presented in Chapter 4.
The on-line parameter estimator makes use of a recursive algorithm to identify
180
CHAPTER 5. ADAPTIVE RESONANT CONTROL 181
the plant parameters. The recursive algorithm uses previously estimated param-
eters as a starting point to predict the new parameters. Due to its recursive
characteristic, the estimator can easily track variations to the plant parameter
if they change gradually. In this situation, the adaptive control system will give
optimum performance with an acceptable transient response. However, for large
and sudden changes to the load in a mechanical system [64, 65] the recursive
characteristic of the estimator will produce a large transient response. This tran-
sient response in the estimator will in turn generate an unacceptable transient
response in the adaptive control system [30, 41, 50, 157].
To improve the transient response of adaptive control, the multiple model
method with supervisor scheme presented in Chapter 3 can be combined with the
adaptive control method to form the Multiple Model Adaptive Control (MMAC)
method as proposed by Narendra and Balakrishnan [100]. This method assumes
a set of L xed-parameter models (M
1
to M
L
) each representing an a priori
known plant condition, and an adaptive model (M
a
) to accommodate unmodeled
plant conditions. The block diagram of this method is shown in Fig. 5.1. For
each model M
i
(i = 1, 2, , L, a), a controller K
i
(i = 1, 2, , L, a) is designed to
satisfy the control objective for M
i
. At every sampling instant, each model, both
xed-parameter and adaptive, produces its own output y
i
(i = 1, 2, , L, a) which
is fed into a switching scheme (i.e., a supervisor). The supervisor selects between
these controllers based on a MMSE performance index described in (3.34) or
(3.35). Using this method, the supervisor will choose the most appropriate xed-
parameter model and its corresponding controller while the adaptive model is still
in the transient phase, and then switch to and stay with the adaptive controller
once its parameters have come out of their transient phase. The xed-parameter
models can provide speed whenever their parameters are close to those of the
plant, while the adaptive model can provide accuracy because its parameters can
be ne-tuned to match those of the plant.
K
1
K
L
K
a
Plant
-
e
L
L
y
~
a
y
~
Controller
Switch
u
1
r +
u y
-
1
~
y
+
M
1
M
L
M
a
Controller Bank
Model Bank
+
+
-
-
u
a
u
L
e
1
e
a
Supervisor
Figure 5.1. Block diagram of the MMAC.
The theory of MMAC is well developed. The stability proof of MMAC for a
continuous system is given in [75, 101]. The method is extended to discrete-time
linear systems in the presence of a disturbance in [105] and a proof of overall
system convergence is oered. The method is also extended to stochastic systems
[103, 159], non-linear systems [23, 62, 104] and non-minimum phase systems [149].
The MMAC concept has been applied to a broad range of applications that
includes: exible transmission systems [64], automobiles [62], robotics [102] and
chemical processes [38, 39]. All of these applications use the MMSE performance
index (3.34) or (3.35) for the supervisor scheme. As discussed in Chapter 3,
this supervisor scheme imposes a high computational demand that increases sig-
nicantly with the number of models. In this research, the MMAC concept is
modied to improve the transient response of the proposed ARC.
The transient response of the ARC is improved by integrating it with the
M
4
RC method discussed in Chapter 3. The resultant M
4
ARC uses a simple
supervisor scheme based on the utility of a lter bank system and the measure-
ment of the output of the natural frequency estimator. The performance of the
proposed controller is evaluated by comparing the M
4
RC, ARC, and M
4
ARC
methods through simulation and experimental studies.
5.2 Adaptive Resonant Control (ARC)
The ARC is implemented using the indirect self tuning regulator (STR) scheme.
The block diagram for controlling the rst three modes of vibration is shown
in Fig. 5.2. The gure shows that the ARC is formed by combining a natural
frequency estimator with a resonant controller. The adjustable controller param-
eters for each mode are updated independently.
20
L
1
L
2
L
3
Plant
E
3
E
2
E
1
2
L
1
L
2
L
3
20
2
2
20
K
1
K
2
K
3
u(k)
y(k)
Natural frequency estimator
r =0
-
Adjustable resonant
controller
Figure 5.2. Block diagram of the ARC for controlling three modes.
ARC Algorithm
The algorithm for each mode is described by the following recursive process:
1. Sample the plant output y(k).
2. Compute the control signal u
m
(k) using (3.13)
u
m
(k) = A
m
y(k) +B
m
y(k 1) +C
m
y(k 2) D
m
u
m
(k 1) E
m
u
m
(k 2).
(5.1)
3. Send the control signal u
m
(k) to the plant.
4. Compute
m
(k) using (4.68)

m
= 2f
s
1 +a
1m
+a
2m
1 a
1m
+a
2m
. (5.2)
5. Update the controller parameters A
m
, B
m
, C
m
, D
m
and E
m
using (3.7) -
(3.11).
A
m
=
4f
2
s
+ 4f
s
cm
m
4f
2
s
+ 4f
s
cm
m
+
2
m
, (5.3)
B
m
=
8f
2
s
4f
2
s
+ 4f
s
cm
m
+
2
m
, (5.4)
C
m
=
4f
2
s
4f
s
cm
m
4f
2
s
+ 4f
s
cm
m
+
2
m
, (5.5)
D
m
=
2
2
m
8f
2
s
4f
2
s
+ 4f
s
cm
m
+
2
m
, (5.6)
E
m
=
4f
2
s
4f
s
cm
m
+
2
m
4f
2
s
+ 4f
s
cm
m
+
2
m
, (5.7)
The nal control signal u(k) is the summation of the control signal for each mode
as in (3.12)
u(k) =
M
m=1
u
m
(k)k
dm
. (5.8)
Closed-loop stability of ARC
The ARC algorithm shows that the adaptation of the controller parameters for
an indirect adaptive control method is achieved in two steps: (i) estimation of
the plant parameters (step 4), and (ii) computation of the controller parameters
based on the estimated plant parameters (step 5). The resulting control method
should guarantee that the output of the plant remain bounded for bounded input
if separately the three following conditions are satised: (i) the plant parameter
estimates are bounded, (ii) the controller parameters are bounded for bounded
plant parameter estimates, and (iii) the estimated plant parameters are admissible
with respect to the control design strategy [70].
The rst and second conditions are satised by using a stable estimator and
controller, respectively. The third condition is necessary due to the fact that
even the estimated plant parameters are bounded at each time t. The current
estimated parameters may not be admissible in the sense that there is no solution
for the controller. This is known as a singularity problem. For example, estimated
parameters which cause a pole-zero cancellation cannot be used to compute the
controller in the pole-placement control method.
The stability of the resonant controller and the natural frequency estimator
was proven in Chapter 3 and Chapter 4, respectively. Thus only the admissibility
of the parameters from the estimator needs to be proven to guarantee the stability
of the ARC.
The stability proof of the resonant controller in Chapter 3 shows that the
controller will produce a solution for any positive natural frequency,
m
. Thus,
the admissibility of the estimated parameters can be guaranteed by proving that
the natural frequency estimator always produces a positive output.
Estimated parameters admissibility proof
Consider the estimated natural frequency (4.68) which is obtained from a
comparison of the transfer function of a exible structure (2.32) with the bilinear
transformation of a second-order discrete time system (4.65), as shown in Chapter
4. For convenience rewrite (4.68), (4.65), and (2.32) here
m
= 2f
s
1 +a
1m
+a
2m
1 a
1m
+a
2m
, (5.9)
G
m
(z) =
b
1m
z +b
2m
z
2
+a
1m
z +a
2m
, (5.10)
G(s, x) =
m=1
y
m
(x)P
m
s
2
+ 2
m
m
s +
2
m
. (5.11)
The admissibility of the estimated parameters is guaranteed if the term within
the square root operator of (5.9) is positive.
1 +a
1m
+a
2m
1 a
1m
+a
2m
> 0. (5.12)
The inequality (5.12) is valid if both its numerator and denominator are positive
or negative. To nd a formulation of coecients a
1m
and a
2m
in terms of the
structure parameters
m
and
m
, apply the bilinear transformation
s = 2f
s
z 1
z + 1
(5.13)
to (5.11) and compare the result with (5.10) to obtain
a
1m
=
2
2
m
8f
2
s
4f
2
s
+ 4f
s
m
+
2
m
, (5.14)
a
2m
=
4f
2
s
4f
s
m
+
2
m
4f
2
s
+ 4f
s
m
+
2
m
. (5.15)
From (5.14) and (5.15), the numerator and denominator of (5.12) can be written
as
num = 1 +a
1m
+a
2m
=
4f
2
s
+ 4f
s
m
+
2
m
4f
2
s
+ 4f
s
m
+
2
m
+
2
2
m
8f
2
s
4f
2
s
+ 4f
s
m
+
2
m
+
4f
2
s
4f
s
m
+
2
m
4f
2
s
+ 4f
s
m
+
2
m
=
4
2
m
4f
2
s
+ 4f
s
m
+
2
m
> 0 (5.16)
and
den = 1 a
1m
+a
2m
=
4f
2
s
+ 4f
s
m
+
2
m
4f
2
s
+ 4f
s
m
+
2
m
2
2
m
8f
2
s
4f
2
s
+ 4f
s
m
+
2
m
+
4f
2
s
4f
s
m
+
2
m
4f
2
s
+ 4f
s
m
+
2
m
=
16f
2
s
4f
2
s
+ 4f
s
m
+
2
m
> 0 (5.17)
respectively. From (5.16) and (5.17), it can be seen that both the numerator and
denominator of (5.12) are always positive which implies that the admissibility of
the parameter from the estimator is guaranteed. Therefore, by separately proving
the stability of the estimator and the controller, and proving the admissibility of
the parameter from the estimator, the ARC is proven to be stable.
5.3 Multi-model Multi-mode Adaptive Resonant
Control (M
4
ARC)
The M
4
ARC method is proposed to facilitate the handling of fast transients in
dynamic systems. M
4
ARC combines the adaptive capability of the ARC with the
fast response feature of M
4
RC from Chapter 3. M
4
ARC retains the characteristic
of the M
4
RC and ARC in that each mode is controlled independently. In the
M
4
ARC method, the system chooses a xed-parameter model from the M
4
RC
method to deal with the transient condition while the adaptive model from the
ARC method is still uctuating. The system then switches to the adaptive model
once the estimator has reached the vicinity of its steady state. To determine the
condition of the m
th
mode estimator, the following convergence criterion is used.
|
m
(k)
m
(k 1)|
m
(5.18)
where k is the time base and
m
is a positive small number. Using this convergence
criterion, the supervisor scheme (3.35) can be simplied to (5.18) in the M
4
ARC
method. The value of
m
is obtained empirically. If
m
is large then the supervisor
will switch to the estimated parameters faster. However, if
m
is too large the
supervisor may select the estimators outputs while they are still too far from the
steady-state condition. This results in poor transient performance.
The block diagram of M
4
ARC for controlling the rst three modes of vibration
is shown in Fig. 5.3. The gure shows that for each mode, a supervisor is used
to update the parameters of the adjustable controller. The parameter from the
natural frequency estimator, , is loaded into the adjustable resonant controller
if criterion (5.18) is satised, otherwise the parameter from the closest xed-
parameter model to the current loading condition, , is loaded into the controller.
20
L
1
Plant
E
1
u(k)
y(k)
L
1
20
2
20
K
2
Natural frequency estimator

Adjustable resonant
controller
-
r =0
S
1
BF
1
S
2
S
3
BF
2
BF
3
2
L
2
E
2
L
2
2
L
3
E
2
L
3 K
1
K
3
Switching and filter bank systems
Supervisor 1
Supervisor 2
Supervisor 3
) (k
)
) (k
) (k
Figure 5.3: Block diagram of the M
4
ARC for controlling three modes.
The lter bank system and switching system are the same as those in the
M
4
RC method shown in Fig. 3.10 and Fig. 3.11, respectively. In the M
4
RC
method, as stated in Section 3.5.2, a large number of xed-parameter models
is required to span an operating region if all the possible loads are not a priori
known. Because of the two following reasons, however, the number of xed-
parameter models in the M
4
ARC method can be signicantly reduced.
1. The inclusion of an adaptive model. Because an adaptive model can change
its parameters, it can span a very large operating region. Therefore, once
the steady-state condition is achieved, one adaptive model can accurately
represent all of the possible loading conditions in the operating region.
2. The use of wide band band-pass-lters in the lter bank to represent the
xed-parameter models. In the M
4
ARC method, the xed-parameter mod-
els are used only to handle the transient conditions while the accuracy of
the model to represent the loading conditions is handled by the adaptive
model. Therefore, a band-pass-lter with a wider pass-band can be used in
the lter bank. In this way a small number of xed-parameter models can
be used, as demonstrated in the following illustrative example.
Example 5.3.1
Similarly to Example 3.5.1, consider a resonant controller that is designed with
operating point 10 Hz and gain k
d
= 10. Change the damping ratio from
c
=
0.01 to
c
= 0.05 to obtain a wider pass-band. The frequency response of the
controller is shown in Fig. 5.4. The gure shows that a gain of more than 32
dB is still achieved for a 1 Hz variation of the operating point. Assuming that
a controller with a gain of 32 dB gives acceptable attenuation, then one model
for every 2 Hz variation of the operating point is required, as shown in Fig. 5.5.
The band-widths of the operating region from the unload condition to the fully
loaded condition in the experimental plant for the rst three modes are 4.5 Hz,
19.2 Hz, and 41.3 Hz, respectively. Therefore, to span the operating region of the
experimental plant the M
4
ARC requires 3, 10, and 21 xed-parameter models for
the 1
st
, 2
nd
, and 3
rd
mode, respectively, or 34 models for the rst three modes of
vibration. Compared to the number of models required for the M
4
RC method,
as illustrated in example 3.5.1, the number of models required in the M
4
ARC is
signicantly reduced from 650 to 34 (a 95% reduction).
Figure 5.4: Frequency response of resonant controller with
c
= 0.05
and k
d
= 10.
M
4
ARC algorithm
The M
4
ARC algorithm for each mode is described by the following iterative
procedure
1. Sample the plant output y(k).

Figure 5.5. Model array in the M
4
ARC model bank.
2. Compute the control signal u
m
(k) using (3.13)
u
m
(k) = A
m
y(k) +B
m
y(k 1) +C
m
y(k 2) D
m
u
m
(k 1) E
m
u
m
(k 2).
(5.19)
3. Send the control signal u
m
(k) to the plant.
4. Compute
m
(k) using (4.68)

m
= 2f
s
1 +a
1m
+a
2m
1 a
1m
+a
2m
. (5.20)
5. Compute |
m
(k)
m
(k 1)|. If the result
m
then send
m
(k) to the
controller else send
m
to the controller.
6. Update the controller parameters A
m
, B
m
, C
m
, D
m
and E
m
using (3.7) -
(3.11)
A
m
=
4f
2
s
+ 4f
s
cm
m
4f
2
s
+ 4f
s
cm
m
+
2
m
, (5.21)
B
m
=
8f
2
s
4f
2
s
+ 4f
s
cm
m
+
2
m
, (5.22)
C
m
=
4f
2
s
4f
s
cm
m
4f
2
s
+ 4f
s
cm
m
+
2
m
, (5.23)
D
m
=
2
2
m
8f
2
s
4f
2
s
+ 4f
s
cm
m
+
2
m
, (5.24)
E
m
=
4f
2
s
4f
s
cm
m
+
2
m
4f
2
s
+ 4f
s
cm
m
+
2
m
, (5.25)
The nal control signal u(k) is the summation of control signal for each mode as
in (3.12)
u(k) =
M
m=1
u
m
(k)k
dm
. (5.26)
Closed-loop stability of M
4
ARC
Similar to M
4
RC method, the closed-loop stability of the M
4
ARC is guaranteed if
it is assumed that the intervals between successive switches have a nonzero lower
bound T
min
> 0, which can be chosen to be arbitrarily small [100]. The condition
in M
4
RC that there must be at least one controller in the controller bank that
can stabilize the system is not necessary in M
4
ARC because this condition is
automatically satised by the adaptive model once it comes close to the steady-
state condition.
In the following sections, the eectiveness of the proposed ARC and M
4
ARC
are evaluated through simulation and experimental studies.
5.4 Simulation Studies of ARC and M
4
ARC
In these simulation studies, the performances of the M
4
RC, ARC and M
4
ARC
methods are evaluated using the dynamically loaded cantilevered beam. The
simulation models of the plant, Model 1 to Model 4, are obtained from Chapter
2. The objectives of the simulation studies are twofold. Firstly, to demonstrate
that in contrast to the M
4
RC, the ARC is able to optimally attenuate the plant
when the loading conditions are unknown. Secondly, to demonstrate that the
M
4
ARC can improve the transient performance of the ARC when there are large
and sudden changes to the plant parameters.
To achieve these two objectives, three simulation cases, namely M
4
ARC.1,
M
4
ARC.2, and M
4
ARC.3, are conducted. The M
4
ARC.1 case is designed to
achieve the rst objective while the M
4
ARC.2 and M
4
ARC.3 cases are designed
to achieve the second objective. For all the cases, the M
4
RC, ARC and M
4
ARC
are designed and implemented to control the rst three-modes. Each of the modes
in the M
4
RC and M
4
ARC is represented by a lter bank, and each lter bank
contains three band-pass lters representing the known loading models: Model 1,
Model 3, and Model 4, respectively. The loading condition is changed in accor-
dance with the model sequences as shown in Table 5.1. The parameters k
d1
, k
d2
,
and k
d3
and
1
,
2
, and
3
for all the controllers are the same as the parameters
of the resonant controller described in Chapter 3. The natural frequency esti-
mator of the ARC has the same parameters as the natural frequency estimator
described in Chapter 4. To make sure that the supervisor selects the adaptive
model when the model is close to the steady-state condition (that is, within 90%
of the steady-state value), the parameters
1
,
2
, and
3
in (5.18) for the M
4
ARC
are chosen to be 0.05 using a trial and error procedure. The sampling period T
is 1 kHz. The excitation signal is a summation of three sinusoids representing
the rst three natural frequencies of the current model. Each sinusoid has an
amplitude of 1 volt.
Case Loading condition is BPFs in the lter bank of M
4
RC
represented by Model and M
4
ARC is based on Model
M
4
ARC.1 124 1,3,4
M
4
ARC.2 134 1,3,4
M
4
ARC.3 141 1,3,4
Table 5.1: Plant and controller conguration for three dierent simu-
lation study cases.
M
4
ARC.1 case
In the M
4
ARC.1 case, model 2 is deliberately not represented in the three lter
banks, so as to allow the performances of the controllers with unmodeled dynamics
to be assessed. The responses of the three control methods in the time domain
are shown in Fig. 5.6. Plots (a), (b), and (c) represent the response of the M
4
RC,
ARC and M
4
ARC, respectively. The plots show that ARC and M
4
ARC produce
a comparable response, while the M
4
RC fails to give the optimum attenuation
when the Model 2 loading condition occurs. M
4
ARC cannot improve the transient
response of the ARC because all of the xed-parameter models (i.e., Model 1,
Model 3 and Model 4) in the lter bank are signicantly dierent from Model
2. When the loading condition changes to model 4, however, it can be seen that
M
4
ARC improves the transient response of the ARC. The overshoot percentage
and settling time are reduced from 685 % and 2.5 seconds, respectively in the
ARC to 187 % and 0.6 seconds, respectively in the M
4
ARC.
Figure 5.7 shows how the M
4
ARC switches between models in response to
parameter changes. For the sake of clarity, the behaviour of each mode controller
is shown separately. Labels, 1 to 4, on the Y-axes refer to the known xed-
parameter models, and label 5 refers to the adaptive model. From the gure, it
can be seen that for each mode the system switches to the controller with the
closest matching centre frequency while the adaptive model is still in the transient
phase. The system switches to the adaptive controller once the adaptive model
reaches the steady-state condition.
M
4
ARC.2 and M
4
ARC.3 cases
In the M
4
ARC.2 and M
4
ARC.3 cases, all the loading conditions are represented
in the lter bank. To demonstrate that transient response improvement is more
signicant for large changes, the loading condition is changed directly from Model
1 to Model 4 in the M
4
ARC.3 case. This loading change produces larger param-
Model 1
Model 2
Model 4
Figure 5.6: Simulation responses of the (a) M
4
RC, (b) ARC and (c)
M
4
ARC for the 1 2 4 model sequence.
Figure 5.7: M
4
ARC switching behaviour for the 1 2 4 model
sequence.
eter variations compared to when the loading condition changes gradually from
Model 1 to Model 3 and then to Model 4, as in the M
4
ARC.2 case.
Fig. 5.8 and Fig. 5.9 show the comparison of the three control methods
for the M
4
ARC.2 and M
4
ARC.3 cases, respectively. From the gures, it can be
seen that when all the loading conditions are included in the lter banks of the
M
4
RC and M
4
ARC, both controllers give better transient performance than the
ARC. The maximum overshoot percentage (M
o
) and the settling time (T
s
) of the
ARC and M
4
ARC for dierent loading model changes are shown in Table 5.2.
The table shows that the M
4
ARC improves the overshoot and settling time of
the ARC when a large and sudden change occurs in the loading condition. The
improvement is especially signicant for the largest sudden change in the plant
parameters when the loading condition changes from 1 4. This table also shows
that the overshoot and settling time responses of the ARC when the loading
conditions change from 1 4 and from 4 1 are signicantly dierent. The
overshoot response when the system releases the loads (i.e., change from Model 1
to Model 4) is larger than when the system picks up the loads (i.e., change from
Model 4 to Model 1). This is because the damping of the system is decreased
when the system releases the loads, causing a large overshoot. Conversely, the
damping of the system is increased when the system picks up the loads and thus
no large overshoot occurs even though the system experiences a large parameter
variation.
M
o
(%) T
s
(seconds)
Changes of Model ARC M
4
ARC ARC M
4
ARC
13 286 200 2.5 1
34 542 332 3 0.8
14 950 320 3.5 0.8
41 228 110 3.3 0.9
Table 5.2: Maximum overshoot percentage and settling time of ARC
and M
4
ARC for dierent loading changes.
Model 1
Model 3
Model 4
4
RC, (b) ARC and (c)
M
4
Model 1
Model 4
Model 1
4
RC, (b) ARC and (c)
M
4
The switching behaviour of the M
4
ARC for the M
4
ARC.2 and M
4
ARC.3 cases
are shown in Fig. 5.10 and Fig. 5.11, respectively. Fig. 5.11 shows that when
the loading is increased (i.e., 4 1), the controller continues to use the xed-
parameter model for a longer period than when the loading is decreased (i.e., 1
4). This behaviour shows that the estimator converges faster when the system
releases the load as opposed to when the system picks up the load. The switching
behaviour also reveals that as a result of the multi-rate sampling scheme, the
estimators for the lower modes, which use a longer sampling period, take a longer
time to settle than the estimators for the higher modes. Consequently with the
M
4
ARC, the transient times for the lower modes are longer than those for the
higher modes.
Figure 5.10: M
4
sequence.
From the simulation cases, it is observed that the M
4
RC gives the best per-
formance if all the possible loading conditions are represented in the lter banks,
but gives poor performance for unmodeled loading conditions. As expected, the
Figure 5.11: M
4
sequence.
M
4
ARC improves the transient response of the ARC when large and sudden
changes in the loading conditions occur provided that at least one model in the
lter bank is close enough to the new unknown loading condition. Since it is
impractical in real applications to construct a xed-parameter multi-model reso-
nant controller that can accommodate all possible loading conditions and allow
for drift of the model parameters, the M
4
ARC method oers the best compromise
in terms of transient performance and load matching.
To verify the results obtained from the simulations, the performances of the ARC,
M
4
RC, and M
4
ARC are evaluated on the physical beam system in the following
experiments. All the controllers used in the simulation studies are implemented on
a dSPACE DS1103 data acquisition board using Matlab, Simulink, and Real-Time
Workshop software to build a real-time experiment. The schematic diagram of the
experimental set-up is the same as that shown in Fig. 3.34 in Chapter 3. The three
simulation cases M
4
ARC.1 to M
4
ARC.3 are repeated in the experiments. For the
M
4
ARC.1 and M
4
ARC.2 cases in the experiments, however, the controllers are
turned on and o at several instants to compare the open-loop and closed-loop
responses of the three control methods. Furthermore, due to the limitation of the
experimental apparatus, the M
4
ARC.3 case in the experiments can only change
the loading condition by releasing the load (i.e., Model 1 4), and not by picking
up the load (i.e., Model 1 4).
M
4
ARC.1 case
The performances of the three control methods are shown in Fig. 5.12. The gure
shows that the performance of the M
4
RC is very poor compared to those of the
ARC and M
4
ARC when the Model 2 loading condition is encountered. This case
shows that the xed-parameter controller M
4
RC is unable to control unknown
models. However when the loading changes to Model 4, it can be seen that the
M
4
RC and M
4
ARC produce better transient performance than the ARC. The
maximum overshoot and settling time of the systems are reduced from 607 %
and 7 seconds, respectively, for the ARC to 362 % and 4 seconds, respectively,
for the M
4
ARC.
The switching behaviour of the M
4
ARC is shown in Fig. 5.13. By the onset of
each of the three load sequences, the natural frequency estimator has already de-
termined the three mode frequencies before the controller is turned on - hence the
adaptive model parameter set, 5, is initially selected for the adjustable controller.
Fig. 5.13 reveals that for each mode the system selects the xed-parameter model
closest to the current loading conditions when the adaptive model is in the tran-
sient condition and switches to the adaptive model once the adaptive model
reaches steady-state.
Figure 5.12: Responses of the (a) M
4
4
ARC for
the 1 2 4 model sequence.
Figure 5.13: M
4
sequence.
M
4
ARC.2 and M
4
ARC.3 cases
The performances of all the control methods for the M
4
ARC.2 and M
4
ARC.3
cases are shown in Figs. 5.14 and 5.15, respectively. Since all the loading condi-
tions are represented in the lter banks of M
4
RC and M
4
ARC, both controllers
give equally better performance than the ARC. The transient responses of the
system, as measured by the overshoot and settling time, are improved with the
M
4
ARC relative to ARC, as is shown in Table 5.3. In the M
4
ARC.2, when the
loading condition changes from 1 3, the dierences in performance between
the three control methods are not as clear as when the loading condition changes
from 3 4. This is because the changes in parameters from 1 3 are smaller
than those from 3 4. However, when the loading changes to Model 4, M
4
ARC
outperforms ARC in terms of transient performances. This observation is rein-
forced in the M
4
ARC.3 cases which produce the largest sudden changes in the
plant parameters. The corresponding responses shown in Fig. 5.15 demonstrate
that the M
4
ARC has an improved transient performance relative to that of ARC.
M
o
(%) T
s
(seconds)
Changes of Model ARC M
4
ARC ARC M
4
ARC
13 359 253 3 3
34 590 302 6 3
14 478 239 8 7
Table 5.3: Maximum overshoot percentage and settling time of ARC
and M
4
ARC for dierent loading changes.
Figs. 5.16 and 5.17 show the controller switching behaviour for the three
modes of the M
4
ARC method for the M
4
ARC.2 and M
4
ARC.3 cases, respec-
tively. Fig. 5.16 reveals that for the loading sequence 1 3 4, the system
momentarily switches to the xed-parameter Model 1 when the load changes from
1 3, then switches to the xed-parameter Model 3, before nally settling with
the adaptive model, 5, once the adaptive model reaches steady-state. For the load
4
4
ARC for
the 1 3 4 model sequence.
model change
4
4
ARC for
the 1 4 model sequence.
Figure 5.16: M
4
sequence.
Model 1 Model 4
Figure 5.17: M
4
ARC switching behaviour for the 1 4 model se-
quence.
change 3 4, the controllers for modes 2 and 3 rst switch to the xed-parameter
Model 3, then to the xed-parameter Model 4 before settling on the newly es-
timated adaptive model. Fig. 5.17 demonstrates the switching behaviour under
the extreme loading change 1 4. For the rst mode, the system switches from
Model 1 to Model 3 and onto Model 4, whereas for modes 2 and 3, the controllers
switch directly from Model 1 to Model 4. All the switching behaviours demon-
strate that while the estimator is in the transient phase, each mode controller will
independently switch to the model with centre frequencies that are closest to the
frequencies of the vibration and then eventually switch to the estimated model
parameters once the mode estimator attains steady-state. In agreement with the
simulation results, the switching diagrams in the experimental results show that
the estimator for the lower modes takes longer to settle than the estimator for the
higher modes. The experimental results also reveal that the proposed switching
scheme avoids undesirable rapid switching.
The experimental apparatus only allows the load(s) to be released. However,
in principle the proposed methods should also work for systems with incremental
loading, as shown in the M
4
ARC.3 cases in the simulations.
5.6 Summary
In this chapter, the ARC method is proposed to improve the attenuation per-
formance of resonant control by enabling it to control multi-mode vibration in
systems with unknown loading conditions. The ARC method is necessary because
the M
4
RC method fails to give optimum performance for unknown loading con-
ditions. The ARC method is developed by combining a resonant controller with
a natural frequency estimator, and is implemented as an STR indirect adaptive
control method. The M
4
ARC method is then proposed to improve the ARCs
transient response. The M
4
ARC method is designed by including an adaptive
model in the M
4
RC model bank. A simple supervisor, which makes use of a
lter bank to identify the closest matching model and the measurement natu-
ral frequency estimators output, is proposed to control the switching between
the xed-parameter models and the adaptive model. The proposed supervisor
signicantly reduces the computational load and avoids unwanted switching.
Simulations and experiments based on the dynamically loaded cantilever beam
with multiple-frequency excitation demonstrate that the proposed ARC has good
attenuation performance, while the proposed M
4
ARC has both good attenuation
performance and transient performance.
The proposed M
4
ARC method provides a basis for a controller that with
a minimum number of sensor-actuator pairs is robust to unmodeled dynamics,
able to respond quickly to large and sudden load changes and is simple enough
for real-time implementation. This method oers a solution to a broad range
of control problems where a high performance controller working to stringent
transient performance requirements is required to accommodate large variations
to a system parameters.
Chapter 6
Summary, Conclusion and Future
Work
6.1 Summary
Increasingly, engineers are having recourse to lightweight, exible structure in at-
tempts to improve dynamic performance and reduce energy demands, but these
structures are subject to detrimental vibration. Attention has focused on active
means of attenuating this vibration. The literature surveyed in the course of this
research reveals that most methods proposed are not fully capable of meeting
the demands placed on the controller by variable loading and/or unmodeled dy-
namic. The most promising methodology appears to be resonant control, which
possesses the advantages of robustness to spillover, independence between modes,
and simplicity. Even so, resonant control is sensitive to variations in the natural
frequencies of the loaded structure, which seriously aect the control performance.
In this thesis, three new control methods, namely: Multi-model multi-mode
resonant control (M
4
RC), adaptive multi-mode resonant control (ARC), and
multi-model multi-mode adaptive resonant control (M
4
ARC) are proposed. Also
proposed is a new natural frequency estimation method. The development of
these proposed methods is motivated by the need for eective real-time control of
multi-mode vibrations in exible structures. The performances of the proposed
207
CHAPTER 6. SUMMARY, CONCLUSION AND FUTURE WORK 208
controllers are evaluated against the following essential requirements: (i) the
controllers to be robust to the unmodeled dynamics that result from mode trun-
cation, (ii) the control system to be able to cope with large and sudden changes
to the plant parameters, (iii) the control methods to use a minimum number of
sensor-actuator pairs in controlling multi-mode vibration, (iv) the resulting closed
loop systems to have a fast transient response, and (v) the proposed controllers
structures to be simple enough to be implemented in a real-time system.
It is claimed that the methodologies proposed oer superior performance to
other known methods reported in the literature. In order to justify this claim,
the following summary of the work undertaken is presented.
Resonant Control
Of the principlal methods for multi-mode vibration control in exible structures,
resonant control oers three advantages.
1. Robustness to spillover. Resonant control applies a high gain at and very
close to the natural frequency of concern, but rolls-o quickly away from
the natural frequency. Consequently, it has a negligible eect on other
frequencies thus avoiding spillover.
2. Independence between modes. The controller for one mode can be designed
independently of the controllers for the other modes. In this case the con-
trol eect on one mode is isolated from the other modes. Because of this
isolation, a resonant control can be formed in a parallel structure to attenu-
ate multiple modes individually. The isolation also enables resonant control
to employ a single collocated sensor-actuator pair for multi-mode vibration
attenuation.
3. Simplicity. Resonant control designs are relatively simple, as they make use
of only the natural frequencies of the structure. This simplicity makes them
suitable for real-time implementations.
Due to these characteristics, resonant control is chosen as the basis for the pro-
posed control methods.
Resonant control, however, is very sensitive to variations of the system nat-
ural frequencies. The controller cannot give optimum attenuation if the natural
frequency of the system is altered from the frequency for which the controller is
specically designed. In the case of a dynamically loaded structure, where fre-
quency variations of the modes are most likely to occur, optimum attenuation
can be achieved only if the resonant controller is able to track the system param-
eter changes and to update its parameters accordingly. Hence for such cases the
resonant control method must be modied to allow for the tracking of changes
to the system parameters. Three control methods based on the resonant control
principle are proposed and evaluated in this thesis.
Experimental Plant and Simulation Models
In the design process, a physical plant and simulation models of the plant are
designed and implemented. The physical plant, which consists of a cantilever
beam with magnetically clamped loads placed along the length of the beam, is
used to evaluate the proposed controllers. Releasing the magnetically clamped
loads causes large and sudden changes to the plant parameters. Modal testing
of the structure reveals large vibrations at the natural frequencies and small
vibrations away from the natural frequencies. The rst three modes are the most
signicant and a 43.4%, 30.0% and 27.0% variation of the natural frequencies of
the respective rst three modes can be obtained by changing the loading from full
load to no load. The structure also is wide-band: the rst and the third modes for
all loading conditions are spaced more than a decade apart in frequency. These
observations reveal that the plant has all the essential characteristics of a exible
structure, and is thus a suitable research vehicle for testing and evaluating the
proposed controllers.
A simulation study of the plant is undertaken for evaluation purposes. Since
all the proposed control methods are based on modal control, which uses the nat-
ural frequency and the damping ratio as the design parameters, the mathematical
models of the experimental plant are derived using modal analysis. To check the
validity of the mathematical models, comparisons are made with models cited in
the literature. After validation, the valid mathematical models are then further
compared against experimental plant and against numerical models built from
nite element methods using ANSYS. A comparison of the results shows that
only small discrepancies occur between the models and the real plant. This re-
sult conrms that the ANSYS numerical models are an accurate representation
of the experimental plant for all the models of interest, and can be used to form
simulation models for control design purposes. This is important since the ana-
lytical technique for constructing a model of the plant with arbitrary loading is
very complex and time consuming, but much easier to construct using numerical
methods. The simulation models of the plant are then implemented in Simulink.
Four, twentieth-order, simulation models with dierent loading conditions are
implemented.
In agreement with the modal testing results, the simulation results demon-
strate that the rst three modes of vibration are the dominant modes for all the
loading conditions. Therefore, the proposed controllers were designed to attenu-
ate the rst three modes of vibration.
M
4
RC
The M
4
RC is an improvement of a resonant control method to accommodate
attenuating the vibration of a plant with varying natural frequencies. The design
of the M
4
RC is based on the assumption that all the possible loading conditions
in the plant are a priori known. The M
4
RC method incorporates a bank of
known loading models, each designed so that it gives optimum attenuation for a
particular loading condition. In contrast to other multi-model methods reported
in the literature, which require a controller for each model, the M
4
RC uses only
one adjustable controller, which is loaded with the xed-model parameters of the
corresponding mode.
In the rst step of the M
4
RC design, a discrete-time resonant controller is
formed by transforming the continuous-time resonant controller using a bilinear
transformation. The bilinear transformation is used to preserve the passivity of
the continuous time resonant control. Therefore, the stability in the continuous-
time resonant controller is preserved in the discrete-time version. In the second
step, a simple supervisor scheme is designed. The supervisor identies which
model for each mode has the closest frequency to the current vibration frequency
and loads the parameters from that model into the adjustable controller to at-
tenuate that mode. The supervisor uses a lter bank. For each mode, the lter
bank is composed of band-pass lters that represent the known xed-parameters
models and a decision making component that determines which model is the
closest model for the current vibration frequency. Narrow band-pass lters are
required to represent the natural frequencies of the xed-parameter models. To
obtain narrow band-pass lters of low order, band-pass lters with a resonant
control structure are incorporated.
The eectiveness of the M
4
RC approach is veried through simulation and
experiments. The results demonstrate that the M
4
RC is capable of attenuating
multi-mode vibrations using a single sensor-actuator pair. A comparison of the
proposed supervisor scheme with the customary minimum mean-squares error
supervisor scheme is also undertaken. The comparison shows that the proposed
supervisor scheme gives comparable performance to the commonly used MMSE
supervisor scheme with the advantages that it is able to avoid rapid switching
and it requires fewer computations. It is evident that the M
4
RC performs well in
dealing with plants with varying natural frequencies. This method oers a simple
design and a fast response for systems with varying parameters. In the event that
the loading conditions are not represented in the lter bank, the M
4
RC still gives
a stable performance albeit with a reduced attenuation performance.
Natural Frequency Estimator
To achieve optimum attenuation for unknown loading conditions, ARC is pro-
posed. Of all the design parameters associated with a resonant controller the
controller centre frequency is the most important. This parameter must be tuned
to the plant natural frequency. Small deviations between this parameter and the
plant natural frequency will result in ineective control and poor attenuation.
Hence, to implement an eective ARC, an on-line natural frequency estimator
is designed and implemented. To obtain a high estimation accuracy of the nat-
ural frequencies of a wide-band plant using an RLS-based estimator is dicult.
The diculty arises from two factors: the high-pass characteristic of RLS and
the bit-length limitation in the implementation. These two factors lead to an in-
crease of the estimation error as the sampling rate and estimator order increase.
To achieve accurate estimation of the natural frequencies of the plant, a parallel
second-order estimator is used. Each component in the estimator estimates the
frequency of a specic mode. A specic sampling period, appropriate for each
mode, is used for each component.
Formulation of the natural frequency estimator is obtained by rst approxi-
mating the plant as a summation of second-order systems. The decomposition of
the plant into several second-order systems according to the numbers of modes
is achieved by using a bank of band-pass lters. Each lter is designed to ac-
commodate the full range of frequencies associated with the respective mode for
all possible loading conditions. The relationship between the second-order sys-
tem coecients and the natural frequency of the system is then obtained using
a bilinear transformation. The relationship reveals that the natural frequency
of the system depends solely on the positions of the poles of the system. This
formulation gives an advantage in that the proposed natural frequency estimator
is robust to unmodeled dynamics. This robustness is a consequence of the fact
that the positions of the poles of the system are not aected by mode truncation,
thus the natural frequency estimation is not aected by unmodeled dynamics
resulting from mode truncation.
Simulations and experiments are conducted to evaluate the performance of
the natural frequency estimator. The natural frequency estimator is used to es-
timate the rst three natural frequencies for dierent loading conditions. The
results show that there is good agreement between the simulations and the ex-
periments. The results also demonstrate that the natural frequency estimator
produces accurate estimations for all the loading conditions of the plant.
ARC
The natural frequency estimator is next used to extend the resonant controller
into an adaptive resonant controller. The ARC is formed as an indirect self-
tuning regulator (STR) conguration by combining the resonant controller with
the natural frequency estimator. To control the rst three modes of vibration
of the plant, a parallel structure of three resonant controllers is used, one con-
troller component per mode. The controller is made adaptive by combining it
with the natural frequency estimator, which is also a parallel structure of three
second-order estimators. The robustness of the ARC to unmodeled dynamics
is guaranteed, because both the resonant controller and the natural frequency
estimator are robust to unmodeled dynamics. Taking advantage of the indirect
conguration, the stability proof of the ARC is given by proving separately: the
stability of the resonant control, the stability of the natural frequency estimator,
and the admissibility of the parameters obtained from the estimator. Although
the ARC is capable of adapting its parameters to suit any loading conditions,
transient delay increases while the estimator identies the new natural frequen-
cies.
M
4
ARC
To improve the transient performance of the ARC, the M
4
ARC is designed. The
M
4
ARC is an extension of the ARC. It is obtained by adding an adaptive model
to the model bank in the M
4
RC. The addition of the adaptive model reduces
the number of xed-parameter models and enables the system to cope with un-
known loading conditions. The improvement of the transient performance is due
to the modication of controller parameters adjustment mechanism. A simple
supervisor scheme similar to the supervisor scheme in the M
4
RC is used to up-
date the controller parameters. The supervisor decides to update the controller
parameters either using the parameters from the xed-parameter models or from
the adaptive models. The decision uses the natural frequency estimators out-
put. The parameters from the xed-parameter models will be used if the natural
frequency estimator is in its transient condition, and the parameters from the
adaptive model will be used once it is close to the steady-state condition. The
closeness of the natural frequency estimator to its steady-state condition is de-
termined using gradient measurement. The dierence between two consecutive
outputs from the estimator is measured. If the dierence is less than a certain
preset value, the estimator is considered as being in the steady-state condition.
The eectiveness of the ARC and the M
4
ARC are evaluated from simulations
and experiments. Comparisons are conducted to measure the performances of
the M
4
ARC compared to ARC during load variations. The results demonstrate
that when the loading condition changes, the M
4
ARC produces a better transient
response than the ARC. The improvement is signicant, especially if the change
is sudden and large such as when the loading condition changes from fully loaded
to unloaded. The experiments reveal that the real-time implementation of the
M
4
ARC is robust to ummodeled dynamics, fast, and in the case of SISO system
only requires a single sensor-actuator pair.
6.2 Conclusion
In conclusion, the proposed three control methods are eective for cancelling
multi-mode vibrations in exible structures. The eectiveness of the methods is
achieved by treating a system as a decomposition of individual modes so that a
parallel of second-order controller structure can be used to deal with each mode
independently. This approach reduces the complexity of the algorithm so that it
is more feasible to apply the proposed methods in real-time implementations. The
outcomes of this research provide a basis for further development of the theory
and the application of active control to exible structures with unforeseen con-
guration variations. The principle of the proposed multi-model adaptive control
can be applied to systems where fast adaptation of the controller is necessary in
order to accommodate structural changes and uctuating external disturbances.
This control method is applicable to any system that can be decomposed into
independent modes. The assumption that all the modes are well separated is a
limitation of the method. This method is not applicable to systems where there
are interactions between more than one mode such as nonlinear systems with
internal resonance [106].
6.3 Recommendations for Future Work
A variety of recommendations for future research can be made based on the results
of this research.
In the research presented in this thesis, the stability of the resonant controller
is guaranteed due to the collocation of the sensor-actuator pair. The collocation
of a sensor-actuator pair, however, may not always be easy to achieve due to the
physical complexity of some structures. To ensure the collocation of a sensor-
actuator pair, the use of the same piezoceramic patch as a sensor and actuator,
known as a piezo sensoriactuator [74] can be considered as an alternative solution.
For some systems, however, collocating the sensors and actuators will degrade the
system performance. For example, the collocation of the sensor and actuator at
the joints of a exible manipulator in a tracking system will not give a satisfactory
performance because the elastic modes of the exible manipulator are seriously
excited and not eectively suppressed [73]. Therefore, further investigation of
the application of the proposed methods to systems with non-collocated sensor-
actuator pairs may be fruitful.
A xed sampling period is used in the natural frequency estimator. Further-
more, information on the frequency bands for each mode is required in order to
determine the sampling period of the corresponding estimator. These frequency
bands are obtained from a priori knowledge about the natural frequencies of the
upper and lower boundaries of the system. If a priori knowledge about the sys-
tem natural frequencies is not available, the estimation of frequency bands will
be required. Moreover, once the estimation of the frequency bands is obtained, a
variable sampling period for the estimator can be used to replace the xed sam-
pling period. This will ensure that the sampling period is always appropriate for
each mode regardless of the variation in the natural frequencies of the system. As
an initial suggestion, a narrow band-pass lter array may be used to determine
in which band the natural frequency of each mode is located, then the sampling
period can be set accordingly.
The ARC and M
4
ARC are designed under the assumption that the order of
the plant is xed. In reality not all systems are of xed-order. For instance,
for a composite structure that is subjected to sudden delamination, the number
of dominant modes which reects the order of the system may change [98]. To
employ the ARC for a system with varying order, the utilization of a lattice lter
that has the capability to estimate both the natural frequency and the order of
system can be investigated as a step in this research direction.
Although the control methods proposed in this thesis are designed for multi-
mode vibration cancellation, the hardware conguration of the control system
is a single-input single-output (SISO) architecture. Extension of the proposed
method to a multiple-input multiple-output (MIMO) system (e.g., cancellation
of vibration in a plate) is worthy of investigation. In the case of a MIMO system
that can be divided into several independent SISO systems the proposed method
can be directly applied. However, for a MIMO system with light coupling between
the subsystems, the calculation of the optimum controller parameters needs to
be considered.
The decentralized characteristic of the proposed methods enables the con-
troller to control each mode separately using a second-order controller. A de-
centralized implementation, with a distributed processing architecture for the
real-time control of a large number of modes, could be a promising research di-
rection.
In the ARC and M
4
ARC designs, although the natural frequency estimator
uses a dierent sampling period for each mode of vibration, the controller itself
still uses a single-rate control method. Research into the design of a multi-rate
ARC may be useful in obtaining further reduction in the computational time or
energy consumption over the single-rate ARC. This is especially applicable for an
implementation in low power micro controllers.
Appendix A
Simulation Models
This appendix explains how to obtain simulation models for a cantilever beam
with dierent loading conditions. The models are referred to as Model 1 to Model
4 with the loading condition shown in Fig. 2.2. Using modal analysis, it can be
found that the transfer function from the applied force u(s) to the transverse
deection of the beam y(s, x) is
Y (s, x)
U(s)
=
m=1
y
m
(x)y
m
(x
1
)
s
2
+ 2
m
i
s +
2
m
(A.1)
where y
m
(x) is the eigenfunction of the m
th
mode at position x, y
m
(x
1
) is the
eigenfunction of the m
th
mode at position x
1
,
m
is the damping ratio of the
m
th
mode and
m
is the eigenvalue or the natural frequency of the m
th
mode.
Substituting s = j = j0 = 0 to (A.1), one obtains the m
th
mode frequency
response at dc (dc gain) as
Y (s, x)
U(s)
=
y
m
(x)y
m
(x
1
)
2
m
. (A.2)
Equation (A.1) and (A.2) show that in general every transfer function is a
superposition of single degree of freedom systems, with each system having its dc
gain determined by the proper eigenfunction entry product divided by the square
of the eigenvalue, and with resonant frequency dened by the proper eigenvalue.
218
APPENDIX A. SIMULATION MODELS 219
The dc gain can be used as a mode selection criterion to sort modes of vibration
so that only the most important modes are kept in the models.
To obtain the simulation models in the form of (A.1), the models are built
using the ANSYS program. Since the values of eigenfunctions (y(x)) and eigen-
values () are obtained as return parameters from modal analysis in ANSYS,
the only parameter that needs to be found is . The value of can be found
by measuring the deection at the resonance frequency. At resonance, the peak
amplitude of each mode is given by substituting s = j
i
, s
2
=
2
m
into (A.1)
Y
m
(x)
u
=
y
m
(x)y
m
(x
1
)
s
2
+ 2
m
i
s +
2
m
=
y
m
(x)y
m
(x
1
)
2
m
+ 2
m
2
m
j +
2
m
=
y
m
(x)y
m
(x
1
)
2
m
2
m
j
. (A.3)
From (A.3) it can be seen that
m
can be calculated if the maximum deection
y
i
(x) at each resonance frequency for a certain force amplitude u is known. To nd
y
i
(x) and u, harmonic analysis using ANSYS is conducted. The harmonic analysis
is conducted by sweeping the system with a certain u along the frequency band
of interest and measuring the deection at certain points x along the beam. The
output of the harmonic analysis using ANSYS is a frequency response function
(FRF). From the FRF the maximum deection y(x) at the resonance frequency
for each of the modes can be measured, then the value of can be found from
(A.3).
The modal analysis in ANSYS for the rst 25 modes of vibration for all models
investigated in this research shows that the dc gain for the rst three modes
are the most dominant and the dc gain of the modes higher than the tenth-
mode are very small (< 10
2
) compared to the rst three modes. Therefore,
the simulation modes for all the models will be based on the rst ten modes of
vibration. The natural frequencies, eigenfunctions and damping ratio, , for the
rst ten modes for Model 1 to Model 4 are tabulated in Table A.1 to Table A.4.
The eigenfunctions are obtained for position x
1
in the middle of the piezoceramic
patch (3.5 cm from the xed end).
Natural frequency (Hz) eigenfunction y(x)
5.9750 0.0125574 4.97E-02
45.236 -0.0772385 9.02E-04
133.50 -0.264567 4.42E-04
319.90 -0.594745 2.29E-05
542.84 0.697097 5.35E-05
733.80 -1.02103 5.99E-06
1113.5 -1.54031 3.58E-05
1576.1 -1.60791 4.59E-05
1844.0 1.41085 3.25E-05
2350.6 -2.12969 3.26E-05
Table A.1: The rst ten natural frequencies, eigenfunctions and damp-
ing ratios of Model 1.
Natural frequency (Hz) eigenfunction y(x)
6.8345 0.0140025 9.31E-03
53.871 -0.108671 8.41E-03
162.86 0.304482 8.31E-04
331.95 -0.565164 6.92E-04
560.58 0.858715 8.73E-06
846.88 -1.15324 2.89E-04
1187.5 1.41837 5.36E-04
1576.2 -1.62605 2.66E-04
2003.3 1.75746 1.52E-03
2460.5 -1.82730 7.27E-04
Since the eigenfunctions obtained from modal analysis using ANSYS are nor-
malized with respect to the mass of the beam without loads, the eigenfunctions
for Model 1 to Model 3 must be multiplied by a correction factor before they can
be substituted into (A.1). The correction factor can be dened as the ratio of
Natural frequency (Hz) eigenvalue y(x)
7.7432 0.0170449 6.17E-02
59.243 -0.105044 8.58E-04
145.34 0.268462 3.69E-03
326.86 -0.604026 6.59E-05
557.56 0.719626 4.47E-05
739.16 -0.997956 3.00E-05
1113.6 -1.53892 2.61E-05
1580.7 -1.64910 4.61E-05
1871.3 1.41052 4.08E-05
2370.7 -2.14833 3.73E-05
Natural frequency (Hz) eigenvalue y(x)
10.020 0.0217182 1.85E-03
62.790 -0.124932 2.69E-04
175.80 0.321682 2.36E-03
344.47 -0.574135 1.44E-04
569.36 0.857314 1.88E-05
850.39 -1.14589 -1.74E-05
1187.5 1.41674 1.02E-05
1580.7 -1.64939 -2.10E-05
2029.8 1.82671 -8.62E-06
2534.8 -1.93543 -8.05E-06
the mass of the beam plus the mass of the load to the mass of the beam without
a load. Based on the mass of the beam (m) and mass of the load (M
1
and M
2
),
the correction factors for Model 1, Model 2 and Model 3 are 1.8087, 1.3281 and
1.4806, respectively. By inserting the values of corrected eigenfunctions, damping
ratios and natural frequencies for each of the models, the transfer functions of
Model 1 to Model 4 for the rst 10 modes are obtained. Note that because the
sensor and actuator are collocated then y
m
(x) = y
m
(x
1
). The transfer functions
are then implemented in Simulink and formed in a twentieth-order system.
Y (s)
U(s)
=
b
19
s
19
+ b
18
s
18
+ ... + b
1
s + b
0
s
20
+ a
19
s
19
+ a
18
s
18
+ ... + a
1
s + a
0
. (A.4)
The coecients of the transfer function and the position of the poles of Model
1 to Model 4 are tabulated in Table A.5 - A.8. All the tables show that the poles
of all the models are located in the left side of the s-plane, therefore all the models
are stable.
i a
i
b
i
pole positions
0 1.125e067 8.814e060 -8.7382e+001 +1.4769e+004i
1 4.36e063 2.821e057 -8.7382e+001 -1.4769e+004i
2 8.143e063 5.272e057 -6.8550e+001 +1.1586e+004i
3 6.337e059 3.195e053 -6.8550e+001 -1.1586e+004i
4 1.138e059 5.443e052 -5.8591e+001 +9.9028e+003i
5 4.016e054 1.528e048 -5.8591e+001 -9.9028e+003i
6 1.876e053 6.761e046 -4.1394e+001 +6.9962e+003i
7 3.746e048 1.076e042 -4.1394e+001 -6.9962e+003i
8 6.542e046 1.78e040 -2.7279e+001 +4.6105e+003i
9 8.74e041 1.879e035 -2.7279e+001 -4.6105e+003i
10 7.872e039 1.625e033 -2.0180e+001 +3.4107e+003i
11 7.328e034 1.15e028 -2.0180e+001 -3.4107e+003i
12 4.028e032 6.071e025 -1.1892e+001 +2.0100e+003i
13 2.579e027 2.779e020 -1.1892e+001 -2.0100e+003i
14 9.447e024 9.767e017 -4.9628e+000 +8.3879e+002i
15 4.024e019 2.671e012 -4.9628e+000 -8.3879e+002i
16 1.052e017 6.734e009 -1.6816e+000 +2.8422e+002i
17 2.723e011 8481 -1.6816e+000 -2.8422e+002i
18 5.373e008 16.26 -2.2212e-001 +3.7541e+001i
19 644.3 -2.274e-013 -2.2212e-001 -3.7541e+001i
Table A.5: The transfer function coecients and pole positions of
Model 1.
i a
i
b
i
pole positions
0 6.989e067 5.506e061 -3.6238e+001 +1.5460e+004i
1 9.355e063 6.126e057 -3.6238e+001 -1.5460e+004i
2 3.861e064 2.537e058 -2.9505e+001 +1.2587e+004i
3 1.02e060 5.192e053 -2.9505e+001 -1.2587e+004i
4 3.824e059 1.832e053 -2.3214e+001 +9.9035e+003i
5 4.739e054 1.8e048 -2.3214e+001 -9.9035e+003i
6 4.59e053 1.642e047 -1.7489e+001 +7.4613e+003i
7 3.443e048 9.787e041 -1.7489e+001 -7.4613e+003i
8 1.375e047 3.706e040 -1.2473e+001 +5.3211e+003i
9 6.877e041 1.447e035 -1.2473e+001 -5.3211e+003i
10 1.443e040 2.888e033 -8.2562e+000 +3.5222e+003i
11 4.998e034 7.556e027 -8.2562e+000 -3.5222e+003i
12 6.382e032 9.201e025 -4.8890e+000 +2.0857e+003i
13 1.533e027 1.576e020 -4.8890e+000 -2.0857e+003i
14 1.314e025 1.293e018 -2.3986e+000 +1.0233e+003i
15 2.114e019 1.325e012 -2.3986e+000 -1.0233e+003i
16 1.305e017 7.851e009 -7.9341e-001 +3.3848e+002i
17 1.276e011 3688 -7.9341e-001 -3.3848e+002i
18 5.975e008 16.63 -1.0066e-001 +4.2942e+001i
19 270.7 4.547e-013 -1.0066e-001 -4.2942e+001i
Model 2.
i a
i
b
i
pole positions
0 4.523e067 3.498e061 -4.8345e+001 +1.4895e+004i
1 7.552e063 4.73e057 -4.8345e+001 -1.4895e+004i
2 1.951e064 1.217e058 -3.8161e+001 +1.1758e+004i
3 7.012e059 3.387e053 -3.8161e+001 -1.1758e+004i
4 1.692e059 7.593e052 -3.2235e+001 +9.9318e+003i
5 3.145e054 1.131e048 -3.2235e+001 -9.9318e+003i
6 2.32e053 7.896e046 -2.2709e+001 +6.9969e+003i
7 2.528e048 6.943e041 -2.2709e+001 -6.9969e+003i
8 7.549e046 1.984e040 -1.5073e+001 +4.6443e+003i
9 5.517e041 1.153e035 -1.5073e+001 -4.6443e+003i
10 8.705e039 1.758e033 -1.1370e+001 +3.5032e+003i
11 4.439e034 6.844e027 -1.1370e+001 -3.5032e+003i
12 4.33e032 6.447e025 -6.6655e+000 +2.0537e+003i
13 1.52e027 1.62e020 -6.6655e+000 -2.0537e+003i
14 9.956e024 1.022e018 -2.9639e+000 +9.1319e+002i
15 2.321e019 1.53e012 -2.9639e+000 -9.1319e+002i
16 1.089e017 6.942e009 -1.2081e+000 +3.7223e+002i
17 1.54e011 4777 -1.2081e+000 -3.7223e+002i
18 5.468e008 16.5 -1.5790e-001 +4.8652e+001i
19 357.8 -4.206e-012 -1.5790e-001 -4.8652e+001i
Model 3.
i a
i
b
i
pole positions
0 2.919e068 2.248e062 -1.5927e+001 +1.5927e+004i
1 1.204e064 7.642e057 -1.5927e+001 -1.5927e+004i
2 7.586e064 4.907e058 -1.2754e+001 +1.2754e+004i
3 7.804e059 3.534e053 -1.2754e+001 -1.2754e+004i
4 5.627e059 2.167e053 -9.9318e+000 +9.9318e+003i
5 2.871e054 8.744e047 -9.9318e+000 -9.9318e+003i
6 5.86e053 1.659e047 -7.4613e+000 +7.4613e+003i
7 1.85e048 4.16e041 -7.4613e+000 -7.4613e+003i
8 1.629e047 3.464e040 -5.3432e+000 +5.3432e+003i
9 3.459e041 5.74e034 -5.3432e+000 -5.3432e+003i
10 1.641e040 2.586e033 -3.5774e+000 +3.5774e+003i
11 2.418e034 2.876e027 -3.5774e+000 -3.5774e+003i
12 7.074e032 8.014e025 -2.1644e+000 +2.1644e+003i
13 7.218e026 5.829e019 -2.1644e+000 -2.1644e+003i
14 1.427e025 1.102e018 -1.1046e+000 +1.1046e+003i
15 9.718e018 4.791e011 -1.1046e+000 -1.1046e+003i
16 1.387e017 6.57e009 -3.9452e-001 +3.9452e+002i
17 5.714e010 1305 -3.9452e-001 -3.9452e+002i
18 6.18e008 13.63 -6.2958e-002 +6.2957e+001i
19 117.4 -2.842e-013 -6.2958e-002 -6.2957e+001i
Model 4.
Appendix B
Passivity
The following denitions of a positive real or passive system are found in the
literature [99, 56]
Denition B.1 A rational function H(s) of the complex variables s = +j is
positive real if
1. H(s) is real for real s.
2. Re[H(s)] 0 for all Re[s]> 0.
For discrete time systems the passivity or positive realness of a system is dened
in the following denitions:
Denition B.2 [70] An mm discrete matrix H(z) of real rational functions is
positive real if and only if:
1. All the elements H(z) are analytic in |z| > 1 (i.e., they do not have poles
in |z| > 1).
2. The eventual poles of any element of H(z) on the unit circle |z| = 1 are
simple and the associated residue matrix is a positive semidenite Hermi-
tian.
227
APPENDIX B. PASSIVITY 228
3. The matrix H(e
j
) + H
T
(e
j
) is a positive semidenite Hermitian for all
real values of which are not poles of any element of H(e
j
).
In the case of a scalar transfer function H(z) condition (3) is replaced by :
ReH(z) 0; |z| = 1. (B.1)
Denition B.3 [70] An mm discrete matrix H(z) of real rational functions is
strictly positive real if and only if:
1. All the elements H(z) are analytic in |z| 1.
2. The matrix H(e
j
) + H
T
(e
j
) is a positive denite Hermitian for all real
.
Appendix C
Simulink
TM
Models
This appendix presents the Simulink models that are used for simulation and
experimental implementations. The models are referred to as Model 1 to Model
4 with the loading condition shown in Fig. 2.2. The parameters of the models
are shown in Table 2.2, and the natural frequencies of the models for simulation
(obtained from analytical results) and experimental implementations are shown
in Table 2.5.
A set of C S-functions is implemented for the simulation and experimental
implementations. The functions are stored as les with .c extension. The les
are listed in Table C.1
C.1 Simulations
In the simulations all the models are formed as transfer functions where the
denominators and numerators for all the models are stored in the mdl TF.mat
le. A bank of Band-Pass Filters (BPFs) with resonant controller structure for
all the models is stored in the rescon.mdl le. Six Simulink model (with .mdl
extension) are built for the simulations. The models and the explanation about
the models are shown in Table C.2. The xed-parameter models used in the lter
bank in all the simulation models are based on Model 1, Model 3 and Model 4.
229
APPENDIX C. SIMULINK
TM
MODELS 230
File Note
rlsid2 RLS estimator with constant forgetting factor.
fcal1 All the algorithms are identical. These algorithms are used to
fcal2 convert the coecients from estimation results to natural
fcal3 frequencies.
adacon Resonant controller algorithm. To make the controller
adjustable, natural frequency is used as an input to this le.
super1 All the algorithms are identical. Supervisor algorithms that
super2 are used in the M
4
ARC method to determine whether the
super3 estimator is still in the transient phase or close to the steady
state condition, and to supply the adjustable controller with
appropriate parameters.
superx1 All the algorithms are identical. Algorithms that are used to
superx2 observe the switching behaviour in the M
4
ARC method.
superx3
maxi Supervisor algorithm that is used in the M
4
RC method to
determine which model is the closest model to the current
loading condition and to supply the adjustable controller with
the appropriate parameters.
supsor Existing supervisor algorithm (MMSE) that is used in the
M
4
RC method to determine which model is the closest model
to the current loading condition and to supply the adjustable
controller with the appropriate parameters.
saklar Algorithm that is used to observe the switching behaviour in the
M
4
RC method with the MMSE supervisor scheme.
Table C.1: C S-functions that are used for simulation and experimental
implementations.
TM
MODELS 231
Model Can be used for
adapt sim Simulate the adaptive resonant controller compared to the
xed-parameter resonant controller.
Test the natural frequency estimator.
m4rc sim Simulate the M
4
RC using Butterworth BPF in the lter
bank compared to the xed-parameter resonant controller or
to the adaptive resonant controller.
rc m4rc sim Simulate the M
4
RC using BPF with resonant controller
structure in the lter bank compared to the xed-parameter
resonant controller.
narendra sim Simulate the M
4
RC using the existing MMSE supervisor
scheme.
m4arc sim Simulate the M
4
ARC using Butterworth BPF in the lter
bank compared to adaptive resonant controller.
rc m4arc sim Simulate the M
4
ARC using BPF with resonant controller
structure in the lter bank compared to the adaptive resonant
controller.
Table C.2. Simulink model for simulations.
adapt sim
The schematic diagram of the adapt sim Simulink model is shown Fig C.1. In the
gure, the blocks in mdl1 to in mdl4 are the disturbances that are formed as a
combination of three sinusoidal signals with frequencies are the same as the rst
three natural frequencies of Model 1 to Model 4, respectively. The block plant
is the plant under control that can be changed to Model 1, Model 2, Model 3,
or Model 4. The blocks 1st mode natural frequency estimator, 2nd mode
natural frequency estimator and 3rd mode natural frequency estimator
are three parallel second-order estimators. The detail of the blocks are shown in
Fig. C.2 (a) to (c). The BPF and BPF1 blocks in all the natural frequency
estimator blocks are band-pass lter that are used to preltering the input to the
RLS Estimator block. The sampling time for the 1
st
, 2
nd
and 3
rd
estimator
are 0.02 sec, 0.002 sec, and 0.001 sec, respectively. The controller block is an
adjustable controller that is built from three parallel resonant controllers. The
detail of the controller block is shown in Fig. C.3.
TM
MODELS 232
Figure C.1. Schematic diagram of adap sim model.
TM
MODELS 233
(a)
(b)
(c)
Figure C.2: Detail schematic diagram of (a) 1st mode natural fre-
quency estimator block (b) 2nd mode natural frequency estimator
block (c) 3rd mode natural frequency estimator block.
TM
MODELS 234
Figure C.3. Detail schematic diagram of the controller block.
To start a simulation, set solver type under the simulation parameters menu as
variable-step (ode45). Set the constant value in the Constant3 block according
to which model is want to be controlled. Set the numerator and denominator
in the plant block accordingly (e.g. set the numerator as num mdl1 and the
denominator as den mdl1 to make Model 1 as the plant under control). The
model and the disturbance can be changed while the simulation is running by
using the set param command in the Matlab editor. To observe the output and
the control signal, the variable y and u can be plotted by using the plot(t,y) or
plot(t,u) command in the Matlab editor. To observe the natural frequency of the
estimation results, plot the variable f1 to f3 using the plot command from the
Matlab editor.
m4rc sim
The schematic diagram of the m4rc sim model is shown Fig. C.4. From the
gure, it can be seen that the m4rc sim model is similar to adapt sim model
TM
MODELS 235
with addition of the 1st mode lter bank and switching systems to 3rd
mode lter bank and switching systems blocks. These blocks are bank
lter systems and switching systems for M
4
RC method as shown in Fig. C.5.
The BPFs in the lter bank system for each mode are designed as sharp band-
pass lters with frequency cut-o the same as natural frequencies of Model 1,
Model 3 and Model 4 for that particular mode.
A similar procedure to adapt sim simulation is used to start a simulation. The
controller switch block is used to select whether the M
4
RC method or adaptive
method want to be used. The variable s, s1, and s2 can be used to observe the
switching behaviour of the M
4
RC method.
rc m4rc sim
The schematic diagram of the rc m4rc sim model is similar to the schematic
diagram of the m4rc sim model. The dierence is that for the rc m4rc sim model
the band-pass lters in the 1st mode lter bank to 3rd mode lter bank
blocks are formed from the band-pass lter with resonant controller structure.
narendra sim
The schematic diagram of the narendra sim model is shown Fig. C.6. In this
model, the plant block consist of Model 1 to Model 4, therefore, to change the
model under control while simulation is running can be achieved by changing the
constant value in the Constant3 block.
m4arc sim
The schematic diagram of the m4arc sim model is shown Fig. C.7. From the
gure it can be seen that the m4arc sim model is similar to m4rc sim model as
shown in Fig. C.4. The dierence is that in the m4arc sim model, the output of
the natural frequency estimator is fed into the lter bank and switching system.
The lter bank and switching system for m4arc sim model is shown Fig. C.8.
TM
MODELS 236
Figure C.4. Schematic diagram of m4rc sim model.
TM
MODELS 237
(a)
(b)
Figure C.5: Schematic diagrams of (a) Filter bank and switching sys-
tem block (b) Detail of bank lter system block.
TM
MODELS 238
Figure C.6. Schematic diagram of narendra sim model.
TM
MODELS 239
Figure C.7. Schematic diagram of m4arc sim model.
TM
MODELS 240
Figure C.8: Detail of lter bank and switching systems block of
m4arc sim model.
rc m4arc sim
The schematic diagram of the rc m4arc sim model is similar to the schematic di-
agram of the m4arc sim model. The dierence is that for the rc m4arc sim model
the band-pass lters in the 1st mode lter bank to 3rd mode lter bank
blocks are formed from the band-pass lter with resonant controller structure.
C.2 Experiments
The schematic of the experimental set-up is shown in Fig. 3.34. For the ex-
periments, the controllers are implemented in the C-S-function in the Real-Time
Workshop Simulink, and embedded into dSpace board to obtain a real-time im-
plementation. All the les that need for compilation process are stored in the le
with .dll extension. A bank of Band-Pass Filters (BPFs) with resonant controller
structure for all the models is stored in the rescon exp.mdl le. Five Simulink
TM
MODELS 241
model les are implemented for the experiments. The les and the explanation
about the les are shown in Table C.3.
Model Can be used for
adapt exp Simulate the adaptive resonant controller compared to
xed-parameter resonant controller.
Test the natural frequency estimator.
m4rc exp Simulate the M
4
RC using Butterworth BPF in the lter bank.
rc m4rc exp Simulate the M
4
RC using BPF with resonant controller
structure in the lter bank.
m4arc exp Simulate the M
4
ARC using Butterworth BPF in the lter
bank compared to xed-parameter resonant controller, and
compared to adaptive resonant controller.
rc m4arc exp Simulate the M
4
ARC using BPF with resonant controller
structure in the lter bank compared to xed-parameter
resonant controller, and compared to adaptive resonant
controller.
Table C.3. Simulink model for experimental implementation.
The schematic of all the Simulink models for the experiments are similar
to the corresponding model for the simulations. The dierence is that in the
experimental models, the plant block in the simulation models is replaced by a
pair of ADC-DAC block. For example, the schematic of adapt exp model is shown
in Fig. C.9. From the gure it can be seen that the control signal and disturbance
signal are fed into channel 1 DAC of the dSpace board (DS1103DAC C1 block),
and the output from piezosensor is fed into channel 1 ADC of the dSpace board
(DS1103MUX ADC CON1 block).
In the experiments, the loading condition is changed by changing the constant
value in the Constant12 block. With the cable from loads M
1
and M
2
are
connected to the switch box as in Fig. 3.34, Model 1, Model 2, and Model 4 is
chosen by setting the constant value as 1, 2 and 3, respectively. To choose Model
3, rst swap the cable from loads M
1
and M
2
so that M
1
connected to point B
in the switch box and M
2
connected to point A. Then set the constant value in
TM
MODELS 242
Figure C.9. Schematic diagram of adapt exp model.
TM
MODELS 243
the Constant12 block to be 2.
For all the Simulink models for the experimental implementation set the solver
type in the simulation parameters menu to xed-step, and set the xed step size
to 1/1000.
Execution of an experiment and acquisition of an experimental data can be
obtained by the following steps:
1. Open Matlab program.
2. Open a Simulink model le that want to be executed.
3. Open dSpace ControlDesk program.
4. With the Simulink as the active screen, press Ctrl B to start the compilation
process. For success compilation Matlab editor will display :
### Successful compilation of Real-time Workshop build procedure for
model: le name. *** Finish RTI build procedure.
5. To start the experiment, choose Simulation in the Simulink window menu
and select connect to target.
6. To be able to observe the variables and collect the experimental data, choose
sdf le on the bottom of the screen on ControlDesk program.
7. On the le menu choose New and select new layout.
8. On data acquisition choose plotter for data observation and collection.
9. Drag the variable that want to be observed (collected) from the labels to
the y-axis of the plotter.
10. Repeat steps 7 to 9 to observe (collect) more variables (one plotter for one
variable).
TM
MODELS 244
11. Change the length to the desired observation time (in seconds).
12. Press animation button to start the observation (collection).
Note: Current Matlab directory must be set to be the same as the directory
where the Simulink model resides.
Bibliography
[1] K. J.

Astrom and B. Wittenmark. Adaptive Control. Addison-Wesley, USA,
1989.
[2] K.J.

Astrom and P. Eykho. System identication - a survey. Automatica,
7:123161, 1971.
[3] A.N. Ashour. Nonlinear Control of plate vibrations. PhD thesis, Virginia
Polytechnic Institute and State University, 2001.
[4] O.N. Ashour and A.H. Nayfeh. Adaptive control of exible structures using
a nonlinear vibration absorber. Nonlinear Dynamics, 28:309322, 2002.
[5] M. Athans, D. Castanou, K.P. Dunn, C.S. Greene, W.H. Lee, N.R. Sandell,
and A.S. Willsky. The stochastic control of the f-8 aircraft using multiple
model adaptive control (mmac) method - part 1: Equilibrium ight. IEEE
Trans. Automatic Control, 22(5):768780, 1977.
[6] T. Autenrieth and E. Rogers. Performance enhancement for a class of
multiple model adaptive control schemes. Int. J. Adapt. Control Signal
Process., 13:105127, 1999.
[7] M.R. Bai and K.Y. Ou. Experimental evaluation of adaptive predictive
control for rotor vibration suppression. IEEE Trans. Control Systems Tech-
nology, 10(6):895901, 2002.
245
BIBLIOGRAPHY 246
[8] M.J. Balas. Feedback control of exible systems. IEEE Transaction on
Automatic Control, AC-23:673679, 1978.
[9] B. Bamieh. Intersample and nite wordlength eects in sampled-data sys-
tem. IEEE Transaction on Automatic Control, 48(4):639643, 2003.
[10] C.N. Bapat and C. Bapat. Natural frequencies of a beam with nonclassi-
cal boundary conditions and concentrated masses. Journal of Sound and
Vibration, 112:177182, 1987.
[11] A. Baz and J.T. Hong. Adaptive control of exible structures using modal
positive position feedback. Int. J. Adapt. Control Signal Process., 11:231
253, 1997.
[12] A. Baz and S. Poh. Performance of an active control system with piezo-
electric actuators. Journal of Sound and Vibration, 126(2):327343, 1988.
[13] A. Baz, S. Poh, and J. Fedor. Independent modal space control with positive
position feedback. Journal of Dynamic Systems, Measurement and Control,
114:96103, 1992.
[14] F.D. Bianchi, R.J. Mantz, and C.F. Christiansen. Gain scheduling control
of variable-speed wind energy conversion systems using quasi-lpv models.
Control Engineering Practice, 13:247255, 2005.
[15] R.R. Bitmead and B.D.O Anderson. Adaptive frequency sampling lters.
IEEE Transaction on Acoustics, Speech, and Signal Processing, 29(3):684
694, 1981.
[16] M. Bodson. Rejection of periodic disturbances of unknown and time-varying
frequency. Int. J. Adapt. Control Signal Process., 19:6788, 2005.
BIBLIOGRAPHY 247
[17] A. Bosse, T.W. Lim, and S. Shelley. Modal lters and neural networks for
adaptive vibration control. Journal of Vibration and Control, 6:631648,
2000.
[18] M. Bouchard and S. Quednau. Multichannel recursive-least square algo-
rithms and fast-transversal-lter algorithm for active noise control and
sound reproduction systems. IEEE Transaction on Speech and Audio Pro-
cessing, 8(5):606618, 2000.
[19] M.J. Brennan, S.J. Elliott, and R.J. Pinnington. Strategies for the active
control of exural vibration on a beam. Journal of Sound and Vibration,
186(4):657688, 1995.
[20] X. Bu, L. Ye, Z. Su, and C. Wang. Active control of a exible smart beam
using a system identication technique based on armax. Smart Materials
and Structures, 12:845850, 2003.
[21] F. Chaplais and K. Alaoui. Two time scaled parameter identication by
coordination of local identiers. Automatica, 32(9):13031309, 1996.
[22] B. Chaudhuri, R. Majumder, and B.C. Pal. Application of multiple-model
adaptive control strategy for robust damping of interarea oscillations in
power system. IEEE Trans. Control Systems Technology, 12(5):727735,
2004.
[23] L. Chen and K.S. Narendra. Nonlinear adaptive control using neural net-
works and multiple models. Automatica, 37:12451255, 2001.
[24] M.H. Chen and M.J. Brennan. Active control of gear vibration using spe-
cially congured sensors and actuators. Smart Material Structure, 9:342
350, 2000.
BIBLIOGRAPHY 248
[25] P.C. Chen and J.S. Shamma. Gain-scheduled
1
-optimal control for boiler-
turbine dynamics with actuator saturation. Journal of Process Control,
14:263277, 2004.
[26] R.L. Clark. Accounting for out-of-bandwidth modes in the assumed modes
approach: Implications on collocated output feedback control. Transac-
tion of the ASME Journal of Dynamic Systems Measurement and Control,
119:390395, 1997.
[27] R.L. Clark, W.R. Saunders, and G.P. Gibbs. Adaptive Structures Dynamics
and Control. John Wiley & Sons, Canada, 1998.
[28] C.F Cutforth and L.Y Pao. Adaptive input shaping for maneuvering exible
structures. Automatica, 40:685693, 2004.
[29] C.A. Desoer and M. Vidyasagar. Feedback Systems: Input-Output Proper-
ties. Academic Pres, USA, 1975.
[30] W. Eversman and I.D. Roy. Active utter suppression using multi-
input/multi-output adaptive least mean square control. Journal of Aircraft,
34(2):224250, 1997.
[31] D.J. Ewins. Modal Testing: Theory, Practice and Application. Research
Studies Press Ltd, England, 2000.
[32] S.D. Fassois. Mimo lms-armax identication of vibrating structures-part
i: The method. Mechanical Systems and Signal Processing, 15(4):723735,
2001.
[33] G.F. Franklin, D.J. Powell, and M.L. Workman. Digital Control of Dynamic
Systems. Addison-Wesley, USA, 1990.
BIBLIOGRAPHY 249
[34] M.I. Friswell and D.J. Inman. The relationship between positive position
feedback and output feedback controllers. Smart Material Structure, 8:285
291, 1999.
[35] C.R. Fuller, S.J. Elliott, and P.A. Nelson. Active Control of Vibration.
Academic Press, London, 1996.
[36] R.P. Goel. Vibrations of a beam carrying a concentrated mass. Journal of
Applied Mechanics, 40:821822, 1973.
[37] C.J. Goh and T.K. Caughey. On the stability problem caused by nite
actuator dynamics in the collocated control of large space structures. Int.
J. Control, 41(3):787802, 1985.
[38] G. Gregorcic and G. Lightbody. A comparison of multiple model and pole-
placement self-tuning for the control of highly nonlinear processes. In Pro-
ceedings of the Irish Signals and Systems Conference, pages 472479, 2000.
[39] R. Gundala, K.A. Hoo, and M.J. Piovoso. Multiple model adaptive control
design for a mimo chemical reactor. Ind. & Eng. Chem. Res., 39:15541563,
2000.
[40] J.A. Gustafson and P.S. Maybeck. Flexible spacestructure control via
moving-bank multiple model algorithm. IEEE Trans. on Aerospace and
Electronics Systems, 30(3):750757, 1994.
[41] P. O. Gutman. Robust and adaptive control: delity or an open relation-
ship? System & Control Letters, 49:919, 2003.
[42] D. Halim. Vibration analysis and control of smart structures. PhD thesis,
The University of Newcastle, 2002.
BIBLIOGRAPHY 250
[43] D. Halim and S.O.R. Moheimani. Spatial resonant control of a exible
structures- application to a piezoelectric laminate beam. IEEE Trans. Con-
trol Systems Technology, 9(1):3753, 2001.
[44] C.R. Halkyard and B.R. Mace. Feedforward adaptive control of exural
vibration in a beam using wave amplitudes. Journal of Sound and Vibration,
254(1):117141, 2002.
[45] C.H. Hansen and S.D. Synder. Active Control of Noise and Vibration.
Chapman & Hall, London, 1997.
[46] M.R. Hatch. Vibration Simulation Using Matlab and ANSYS. Chapman &
Hall/CRC, USA, 2001.
[47] S. Haykin. Adaptive Filter Theory. Prentice Hall, Upper Saddle River, N.J.,
2002.
[48] W. He, H. Kaufman, and R. Roy. Multiple model adaptive control proce-
dure for blood pressure control. IEEE Trans. Biomed. Eng., 33(1):1019,
1986.
[49] G. Hillerstrom. Adaptive suppression of vibrations- a repetitive control
approach. IEEE Trans. Control Systems Technology, 4(1):7277, 1996.
[50] J.J. Hollkamp and T.F. Starchville Jr. A self-tuning piezoelectric vibration
absorber. Journal of Intelligent Material Systems and Structures, 5:559
566, 1994.
[51] S.J. Huang and B.T. Shu. Adaptive control for a mechanical system with
oscillation disturbance. Journal of Sound and Vibration, 199(1):107123,
1997.
[52] P.C. Hughes. Space structure vibration modes: How many exist? which
one are important? IEEE Control System Magazine, 7:2228, 1987.
BIBLIOGRAPHY 251
[53] D.J. Inman. Engineering Vibration. Prentice Hall, USA, 1996.
[54] D.J. Inman. Active modal control for smart structures. Phil. Trans. R.
Soc. Lond. A, 359:205219, 2001.
[55] A. Inoue, K. Yabuki, M. Deng, N. Ueki, and Y. Hirashima. Adaptive
vibration control of exible structure with unknown and unmeasurable dis-
turbances. In Proceedings of IFAC Workshop on Adaptation and Learning
in Control and Signal Processing, and IFAC Workshop on Periodic Control
Systems, pages 585590, Yokohama, 2004.
[56] P.A. Ioannou and J. Sun. Robust Adaptive Control. Prentice Hall, Engle-
wood Clis, N.J., 1996.
[57] Y. Iwaya, T. Ichinoseki, Y. Suzuki, M. Sakata, and T. Sone. Adaptive
control of vibration intensity in a beam in the frequency domain. IEICE
Trans. Fundamentals, E82A(4):605610, 1999.
[58] J. Jiang. Preservations of positive realness under discretizations. Journal
of the Franklin Institute, 330(4):721734, 1993.
[59] J. Jiang and Y. Zhang. A novel variable-length sliding window blockwise
least-squares algorithm for on-line estimation of time-varying parameters.
Int. J. Adapt. Control Signal Process., 18:505521, 2004.
[60] T.A. Johansen, I. Petersen, J. Kalkkuhl, and J. L udemann. Gain-scheduled
wheel slip control in automotive brake systems. IEEE Trans. Control Sys-
tems Technology, 11(6):799811, 2003.
[61] J.N. Juang. Applied System Identication. Prentice Hall, Upper Saddle
River, N.J., USA, 1994.
BIBLIOGRAPHY 252
[62] J. Kalkkuhl, T.A. Johansen, and J. L udemann. Improved transient perfor-
mance of nonlinear adaptive backstepping using estimator resetting based
on multiple models. IEEE Trans. Automatic Control, 47(1):136140, 2002.
[63] Y.K. Kang, H.C. Park, W. Hwang, and K.S. Han. Optimum placement
of piezoelectric sensor/actuator for vibration control of laminated beams.
AIAA Journal, 24(9):19211926, 1996.
[64] A. Karimi and I.D. Landau. Robust adaptive control of a exible transmis-
sion system using multiple models. IEEE Trans. Control Systems Technol-
ogy, 8(2):321331, 2000.
[65] A. Karimi, I.D. Landau, and N. Motee. Eects of the design parameters of
multimodel adaptive control on the performance of a exible transmission
system. Int. J. Adapt. Control Signal Process, 15:335352, 2001.
[66] S.G. Kelly. Fundamentals of Mechanical Vibrations. McGraw Hill, USA,
2000.
[67] H. Khalil. Nonlinear Systems. Prentice Hall, Upper Saddle River, N.J.,
USA, 1996.
[68] I.D. Landau. System Identication and Control Design. Prentice Hall,
Englewood Clis, N.J., 1990.
[69] I.D. Landau, A. Constantinescu, and D. Rey. Adaptive narrow band distur-
bance rejection applied to an active suspension - an internal model principle
approach. Automatica, 41:563574, 2005.
[70] I.D. Landau, R. Lozano, and M. MSaad. Adaptive Control. Springer-
Verlag, 1998.
BIBLIOGRAPHY 253
[71] P.A.A. Laura, J.L. Pombo, and E.A. Susemihl. A note on the vibration of
a clamped-free beam with a mass at the free end. Journal of Sound and
Vibration, 37(2):161168, 1974.
[72] C.H. Lee, M.H. Shin, and M.J. Chung. A design of gain-scheduled control
for a linear parameter varying system: an application to ight control.
Control Engineering Practice, 9:1121, 2001.
[73] T.H. Lee, S.S. Ge, and Z.P. Wang. Adaptive robust controller design for
multi-link exible robots. Mechatronics, 11:951967, 2001.
[74] D. Li, S.W. Or, H.L.W. Chan, P.K. Choy, and P.C.K. Liu. Tuneable vibra-
tion absorber incorporating piezoceramic sensoriactuator. Japanese Journal
of Applied Physics, 45(5B):47874792, 2006.
[75] X. Li, S. Wang, and W. Wang. Improved multiple model adaptive con-
trol and its stability analysis. In Proceedings of the 15
th
Triennial World
Congress, Bercelona, Spain, 2002.
[76] D. Liberzon and A.S. Morse. Basic problems in stability and design of
switched systems. IEEE Control System Magazine, 19(5):5970, 1999.
[77] T.W. Lim, A. Bosse, and S. Fisher. Structural damage detection us-
ing real-time modal parameter identication algorithm. AIAA Journal,
34(11):23702376, 1996.
[78] T.W. Lim, A. Bosse, and S. Fisher. Adaptive lters for real-time system
identication and control. Journal of Guidance, Conrol, and Dynamics,
20(1):6166, 1997.
[79] T.X. Liu, H.X. Hua, and Z. Zhang. Robust control of plate vibration via ac-
tive constrained layer damping. Thin-Walled Structures, 42:427448, 2004.
BIBLIOGRAPHY 254
[80] W.H. Liu, J.R. Wu, and C.C. Huang. Free vibration of beams with elas-
tically restrained edges and intermediate concentrated masses. Journal of
Sound and Vibration, 122:193207, 1988.
[81] L. Ljung. System Identication: Theory for the User. Prentice-Hall, En-
glewood Clis, New Jersey, 1987.
[82] A. Loria and H. Nijmeijer. Nonlinear control systems:Output feedback design
methods. In John G. Webster,editor,Wiley Encycloaedia of Electrical and
Engineering. Article #1025 in Section Automatic Control. John Wiley
& Sons, 1998.
[83] K.H. Low. Closed-form formulas for fundamental vibration frequency of
beams under o-centre load. Journal of Sound and Vibration, 201(4):528
533, 1997.
[84] K.H. Low. Comparison of experimental and numerical frequencies for clas-
sical beams carrying a mass in-span. International Journal of Mechanical
Sciences, 41:15151531, 1999.
[85] K.H. Low. On the methods to derive frequency equations of beams carrying
multiple masses. International Journal of Mechanical Sciences, 43:871881,
2001.
[86] K.H. Low. Frequencies of beams carrying multiple masses: Rayleigh es-
timation versus eigenanalysis solutions. Journal of Sound and Vibration,
268:843853, 2003.
[87] K. Ma. Vibration control of smart structures with bonded PZT patches:
novel adaptive ltering algorithm and hybrid control scheme. Smart Mate-
rial Structure, 12:473482, 2003.
BIBLIOGRAPHY 255
[88] G.A. McGraw. Robust adaptive identication and control design tech-
niques. Control and Dynamics Systems, 36:275331, 1990.
[89] C. Mei, B.R. Mace, and R.W. Jones. Hybrid wave/mode active vibration
control. Journal of Sound and Vibration, 247(5):765784, 2001.
[90] L. Meirovitch. Dynamics and Control of Structures. John Wiley & Sons,
USA, 1990.
[91] L. Meirovitch. Principles and Techniques of Vibrations. Prentice-Hall, USA,
1997.
[92] L. Meirovith. Fundamentals of Vibrations. McGraw-Hill, 2001.
[93] T Meurers, S.M. Veres, and A.C.H. Tan. Model-free frequency domain
iterative active sound and vibration control. Control Engineering Practice,
11:10491059, 2003.
[94] P Micheau and P. Coirault. Adaptive controller using lter banks to reject
multi-sinusoidal disturbance. Automatica, 36:16591664, 2000.
[95] S.O.R. Moheimani. Experimental verication of the corrected transfer func-
tion of a piezoelectric laminate beam. IEEE Trans. Control Systems Tech-
nology, 8(4):660666, 2000.
[96] S.O.R. Moheimani. Minimizing the eect of out-of bandwidth dynamics in
the models of reverberant systems that arise in modal analysis: implications
on spatial H
control. Automatica, 36:10231031, 2000.

[97] S.O.R. Moheimani, I.R. Petersen, and H.R. Pota. Broadband distur-
bance attenuation over an entire beam. Journal of Sound and Vibration,
227(4):807832, 1999.
BIBLIOGRAPHY 256
[98] K. Nakatani, S. Kubo, T. Sukagami, D. Shiozawa, and M. Takagi. An
experimental study on identication of delamination in composite mate-
rialby the passive electric potential ct method. In Proceedings of the 5
th
International Conference on Inverse Problems in Engineering: Theory and
Practice, pages No2 110, Cambridge, U.K., 2005.
[99] K.S. Narendra and A.M. Annaswamy. Stable Adaptive Systems. Prentice
Hall, Englewood Clis, N.J., 1989.
[100] K.S. Narendra and J. Balakrishnan. Improving transient response of adap-
tive control systems using multiple models and switching. IEEE Trans.
Automatic Control, 39(9):18611866, 1994.
[101] K.S. Narendra and J. Balakrishnan. Adaptive control using multiple mod-
els. IEEE Trans. Automatic Control, 42(2):171187, 1997.
[102] K.S. Narendra, J. Balakrishnan, and K.M. Ciliz. Adaptation and learn-
ing using multiple models, switching and tuning. IEEE Control Systems
Magazine, June:3750, 1995.
[103] K.S. Narendra and O.A. Driollet. Stochastic adaptive control using multiple
models for improved performance in the presence of random disturbances.
Int. J. Adapt. Control Signal Process., 15:1287317, 2001.
[104] K.S. Narendra, O.A. Driollet, M. Feiler, and K. George. Adaptive control
using multiple models, switching and tuning. Int. J. Adapt. Control Signal
Process., 17:87102, 2003.
[105] K.S. Narendra and C. Xiang. Adaptive control of discrete-time systems
using multiple models. IEEE Trans. Automatic Control, 45(9):16691685,
2000.
BIBLIOGRAPHY 257
[106] A.H. Nayfeh and D.T. Mook. Nonlinear Oscillations. Wiley-Interscience,
New York, 1979.
[107] K. Ogata. Discrete-Time Control Systems. Prentice Hall, Englewood Clis,
N.J., 1987.
[108] R. Ortega and Y. Tang. Robustness of adaptive controllers - a survey.
Automatica, 25(5):651677, 1989.
[109] H. Petroski. To Engineer is Human - The Role of Failure in Successful
Design. St. Martins Press, New York, 1985.
[110] R. Pickhardt. Adaptive control of solar power plant using a multi-model.
IEE Proc. Control Theory Appl., 147(5):493500, 2000.
[111] B.A.D. Piombo, A. Fasana, S. Marchesiello, and Ruzzene. M. Modelling and
identication of the dynamic response of a supported bridge. Mechanical
Systems and Signal Processing, 14(1):7589, 2000.
[112] S. Poh and A. Baz. A demonstration of adaptive least-mean-square con-
trol of small amplitude vortex-induced vibrations. Journal of Fluids and
Structures, 10:615632, 1996.
[113] H.R. Pota, S.O.R. Moheimani, and M. Smith. Resonant controllers for
exible structures. In Proceedings of the 38
th
Conference on Decision and
Control, pages 631636, Phoenix, 1999.
[114] H.R. Pota, S.O.R. Moheimani, and M. Smith. Resonant controllers for
smart structures. Smart Material Structures, 11:18, 2002.
[115] A. Preumont. Vibration Control of Active Structures : An Introduction.
Kluwer Academic Publisher, Netherlands, 2002.
BIBLIOGRAPHY 258
[116] F. Previdi and E. Carpanzano. Design of a gain scheduling controller for
knee-joint angle control by using functional electrical stimulation. IEEE
Trans. Control Systems Technology, 11(3):310323, 2003.
[117] J.G. Proakis and D.G. Manolakis. Digital Signal Processing: Principles,
Algorithm, and Applications. Prentice Hall, Upper Saddle River, N.J., 1996.
[118] Z. Qu. Adaptive and robust controls of uncertain systems with nonlinear
parameterization. IEEE Transaction on Automatic Control, 48(10):1817
1823, 2003.
[119] R. Rao, B. Aufderheide, and B. Bequette. Experimental studies on multiple-
model predictive control for automated regulation of hemo-dynamic vari-
ables. IEEE Trans. Biomed. Eng., 50:277288, 2003.
[120] S.S. Rao. Mechanical Vibrations. Addison-Wesley, USA, 1986.
[121] K.H. Rew, J.H. Han, and I. Lee. Adaptive multi-modal vibration control
of wing-like composite structure using adaptive positive position feedback.
AIAA, 1422:664673, 2000.
[122] K.H. Rew, S. Kim, I. Lee, and Y. Park. Real-time estimations of multi-
modal frequencies for smart structures. Smart Material and Structure,
11:3647, 2002.
[123] C.E. Rohrs, M. Athans, L. Valvani, and G. Stein. Some design guidelines
for discrete-time adaptive controllers. Automatica, 20(5):653660, 1984.
[124] L.F.L. Rosa, C. Magluta, and N. Roitman. Estimation of modal parame-
ters through a non-linear optimisation technique. Mechanical Systems and
Signal Processing, 13(4):593607, 1999.
BIBLIOGRAPHY 259
[125] D.M. Rovner and R. Cannon. Experiments toward on-line identication and
control of a very exible one-link manipulator. Int. J. Robotics Research,
6:319, 1987.
[126] D.M. Rovner and G.F. Franklin. Experiments in load- adaptive control of
a very exible one-link manipulator. Automatica, 24:541548, 1988.
[127] M. Ruzzene, A. Fasana, L. Garibaldi, and B. Piombo. Natural frequencies
and dampings identication using wavelet transform: Application to real
data. Mechanical Systems and Signal Processing, 11(2):207218, 1997.
[128] R.G. Sargent. Verication and Validation of Simulation Models, Progress
in Modelling & Simulation,Cellier, F.E., (Ed). Academic Pres, USA, 1982.
[129] S. Sastry and M. Bodson. Adaptive Control- Stability, Convergence and
Robustness. Prentice Hall, Englewood Clis, N.J., 1989.
[130] K.D. Schott and B.W Bequette. Multiple model adaptive control. In R
Murray-Smith and T.A. Johanson,editors,Multiple Model Approaches to
Modelling and Control. Chapter 11. Taylor & Francis, London, UK, 1997.
[131] A.E. Schwenk, S.D. Sommerfeldt, and S.I. Hayek. Adaptive control of struc-
tural intensity associated with bending waves in a beam. J. Acoust. Soc.
Am., 96(5):28262835, 1994.
[132] J. Shan, H.T. Liu, and D. Sun. Slewing and vibration control of a single-
link exible manipulator by positive position feedback (ppf). Mechatronics,
15:487503, 2005.
[133] J Shaw. Adaptive control for sound and vibration attenuation: a compar-
ative study. Journal of Sound and Vibration, 235(4):671684, 2000.
[134] J Shaw. Active vibration isolation by adaptive control. Journal of Vibration
and Control, 7:1931, 2001.
BIBLIOGRAPHY 260
[135] T. Soderstrom and P. Stoica. System Identication. Prentice Hall Interna-
tional, UK, 1989.
[136] G. Song, S.P. Schmidt, and B.N. Agrawal. Active vibration suppression
of a exible structure using smart material and a modular control patch.
Proc. Instn. Mech. Engrs., 214:217229, 2000.
[137] O. Song, L. Librescu, and H.D. Kwon. Vibration and stability control of
robotic manipulator systems consisting of a thin-walled beam and a spin-
ning tip rotor. Journal of Robotic Systems, 19(10):469482, 2002.
[138] O. Song, L. Librescu, and C.A. Rogers. Adaptive response control of a can-
tilevered thin-walled beams carrying heavy concentrated masses. Journal
of Intelligent Material Systems and Structures, 5:4248, 1994.
[139] M.M. Stanisic and J.C. Hardin. On the response of beams to an arbitrary
number of concentrated moving masses. Journal of the Franklin Institute,
287(2):115123, 1969.
[140] U. Stobner and L. Gaul. Active vibration control of a car body based
on experimentally evaluated modal parameters. Mechanical Systems and
Signal Processing, 15(1):173188, 2001.
[141] L. Sun, J.M. Krodkiewski, and Y. Cen. Self-tuning adaptive control of
forced vibration in rotor systems using an active journal bearing. Journal
of Sound and Vibration, 213(1):114, 1998.
[142] Y. Tang, E.F. Camacho, and J.J. Flores. Frequency domain adaptive con-
trol: Band-wise compensation. Automatica, 31:735740, 1995.
[143] Y. Tipsuwan and M.Y. Chow. On the gain scheduling for networked pi
controller over ip network. IEEE/ASME Transactions on Mechatronics,
9(3):491498, 2004.
BIBLIOGRAPHY 261
[144] M.O. Tokhi and M.A. Hossain. Self-tuning active vibration control in ex-
ible beam structures. Proc. Instn. Mech. Engrs., 208:263278, 1994.
[145] S. Vanlanduit, P. Verboven, P. Guillaume, and J. Schoukens. An automatic
frequency domain modal parameter estimation algorithm. Journal of Sound
and Vibration, 265:647661, 2003.
[146] P. Verboven, P. Guillaume, B. Cauberghe, S. Vanlanduit, and E. Par-
loo. Modal parameter estimation from input-output fourier data using
frequency-domain maximum likelihood identication. Journal of Sound
and Vibration, 276:957979, 2004.
[147] S. de Waele and P.M.T. Broersen. Multirate autoregressive modeling. Se-
lected Topics in Signal, Systems and Control, 12:7580, 2001.
[148] B. Wahlberg and L. Ljung. Design variables for bias distribution in transfer
function estimation. IEEE Transaction on Automatic Control, 31:134144,
1986.
[149] X. Wang, S. Li, Z. Wang, and Y. Zhang. Multiple models adaptive decou-
pling controller for a nonminimum phase system. In Proceedings of the 5
th
Asian Control Conference, pages 166171, Melbourne, 2004.
[150] B. Wittenmark and K.J.

Astrom. Practical issues in the implementation of
self-tuning control. Automatica, 20(5):595605, 1984.
[151] J.S. Wu and T.L. Lin. Free vibration analysis of a uniform cantilever beam
with point masses by an analytical-and-numerical-combined method. Jour-
nal of Sound and Vibration, 136(2):201213, 1990.
[152] L.L. Xie and L. Guo. How much uncertainty can be dealt with by feedback?
IEEE Transaction on Automatic Control, 45(12):22032217, 2000.
BIBLIOGRAPHY 262
[153] S.M. Yang and Y.J. Lee. Vibration suppression with optimal sen-
sor/actuator location and feedback gain. Smart Material Structures, 2:232
239, 1993.
[154] H. Yu and S. Lloyd. Variable structure adaptive control of robot manipu-
lators. IEE Proc.-Control Theory Appl., 144(2):167176, 1997.
[155] S. Yurkovich, F.E. Pacheco, and A.P. Tzes. On-line frequency domain
information for control of a exible-link robot with varying payload. IEEE
Transaction on Automatic Control, 34(12):13001304, 1989.
[156] G. Zhai, B. Hu, K. Yasuda, and A.N. Michel. Stability analysis of switched
systems with stable and unstable subsystems: An average dwell time ap-
proach. In Proceedings of the American Control, pages 355362, Hongkong,
2001.
[157] Y. Zhang, A.G. Alleyne, and D. Zheng. A hybrid control strategy for active
vibration isolation with electrohydraulic actuators. Control Engineering
Practice, 13:279289, 2005.
[158] Y. Zhang and L. Guo. A limit to the capability of feedback. IEEE Trans-
action on Automatic Control, 47(4):687692, 2002.
[159] Y. Zhang and L. Guo. Stochastic adaptive switching control based on multi-
ple models. In Proceedings of the 15
th
Triennial World Congress, Barcelona,
2002.

Adaptive Multi Mode Vibration Control of Dynamically Loaded Flexible Structures PDF

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Adaptive Multi Mode Vibration Control of Dynamically Loaded Flexible Structures PDF

Diunggah oleh

Hak Cipta:

Format Tersedia

Adaptive Multi Mode Vibration Control

of Dynamically Loaded Flexible

controller, robustness is achieved

(cos L + cosh L) (sin L + sinh L)

Figure 3.12. Block diagram of M

, t), and M is the domain of admissible parameters related to the ARX

, and can be written as [148]

is inuenced by the weighted error between the

is obtained by minimizing the bias between

Figure 4.5. Eect of sampling period on pole location.

Natural frequency estimator

control. Automatica, 36:10231031, 2000.

Anda mungkin juga menyukai