A New Adaptive Array of Vibration Sensors

A New Adaptive Array of Vibration Sensors
Hartono Sumali
Dissertation submitted to the Faculty of the Virginia Polytechnic Institute and State University in
partial fulfillment of the requirements for the degree of
Doctor of Philosophy
in
Mechanical Engineering
Harley H. Cudney, Chair
Chris R. Fuller
Daniel J. Inman
Larry D. Mitchell
Alfred L. Wicks
July 1997
Blacksburg, Virginia
Keywords: Modal Analysis, Algorithm, Eigenvectors
Copyright 1997, Hartono Sumali
ii
A New Adaptive Array of Vibration Sensors
Hartono Sumali
(ABSTRACT)
The sensing technique described in this dissertation produces modal coordinates for monitoring
and active control of structural vibration. The sensor array is constructed from strain-sensing
segments. The segment outputs are transformed into modal coordinates by a sensor gain matrix.
An adaptive algorithm for computing the sensor gain matrix with minimal knowledge of the
structures modal properties is proposed. It is shown that the sensor gain matrix is the modal
matrix of the segment output correlation matrix. This modal matrix is computed using new
algorithms based on Jacobi rotations. The procedure is relatively simple and can be performed
gradually to keep computation requirements low.
The sensor system can also identify the mode shapes of the structure in real time using Lagrange
polynomial interpolation formula.
An experiment is done with an array of piezoelectric polyvinylidene fluoride (PVDF) film
segments on a beam to obtain the segment outputs. The results from the experiment are used to
verify a computer simulation routine. Then a series of simulations are done to test the adaptive
modal sensing algorithms. Simulation results verify that the sensor gain matrix obtained by the
adaptive algorithm transforms the segment outputs into modal coordinates.
ACKNOWLEDGEMENTS
I would like to express my gratitude to all who have contributed to this research endeavor,
especially the following individuals. I thank my major advisor Dr. Harley Cudney for his support
throughout my years as a graduate student, both financially and personally. I have learned so
much from him.
My special thanks are due to Dr. Larry Mitchell. I am grateful for the first-class education and
training I received from him, especially during the final years of my graduate studies. I thank Dr.
Dan Inman for his help, especially with my future career opportunities. I thank Dr. Al. Wicks for
his advise on signal processing and modal analysis. I am grateful that I had a chance to learn first-
hand information from Dr. Chris Fuller, a world-renowned expert in vibration and acoustics,
whose insight into the physical meaning of every mathematical expression never ceases to amaze
me.
I thank the agencies and companies which supported the various research projects I was involved
in: ONR, Westinghouse, DARPA, Cessna, and ARO. I thank many people who helped me with
experiments: Karsten Meissner, Rich Lomenzo, Ben Poe, and James Garcia. I thank those who
helped me with their knowledge, advice and discussion: Dr. Ricardo Burdisso, Chris Niezrecki,
Dr. Chul-Hue Park, Dr. Nesbitt Hagood, and many others. I thank my parents, friends, teachers,
and above all, I thank God.
iv
TABLE OF CONTENTS
1 Introduction .................................................................................................................... 1
1.1 Modal Analysis and Modal Coordinates .................................................................... 1
1.2 The Quest for Modal Coordinate Sensors .................................................................. 2
1.2.1 Modal Filtering in Time Domain ....................................................................... 3
1.2.2 Modal Filtering in Spatial Domain .................................................................... 3
1.2.3 Segmentation of Modal Filtering Sensors ......................................................... 7
1.2.4 Segmentation of Modal Filtering Sensors ......................................................... 7
1.2.5 Modal Filtering Using Adaptive Algorithms ...................................................... 8
1.3 Conventional System Characterization and Mode Shape Extraction ......................... 10
1.4 On-line Modal Analyzer: a Novel Concept .............................................................. 13
1.5 Overview of Dissertation ........................................................................................ 14
2 Simulation Method and Sensor Model ............................................................ 17
2.1 Model of Variable Host Structure ........................................................................... 17
2.1.1 Description of Structure ................................................................................. 17
2.1.2 Eigen-properties of Structure ......................................................................... 18
2.1.3 Equation of Motion and Its State-Space Form ................................................ 19
2.2 Simulation of Structure ........................................................................................... 20
2.2.1 Sampling and Discrete-Time Model ................................................................ 20
2.2.2 Discrete-Time Model of One Mode ................................................................ 21
2.2.3 Discrete-Time Model of Multi-Mode Structures ............................................. 22
2.2.4 Selecting Sensor Configuration ...................................................................... 23
2.3 Segmented Sensor Model ........................................................................................ 25
2.3.1 Voltage Generated by Segment ...................................................................... 26
2.3.2 Array of Piezoelectric Film Segments on Beam ............................................... 28
2.3.3 Gain Matrix for Segmented Modal Sensor ...................................................... 31
2.4 Chapter Summary ................................................................................................... 32
3 Numerical Simulation and a Proof-of-Concept Experiment .................. 34
3.1 Verification of Digital Filter Model and Gain Matrix Formula .................................. 34
3.1.1 Verification of Digital Filter Model ................................................................. 34
3.1.2 Verification of Gain Matrix Formula ............................................................... 41
3.1.3 Modal Truncation and Spatial Aliasing .......................................................... 43
3.2 Proof-of-Concept Experiment ................................................................................. 47
3.2.1 Properties of Experimental Structure .............................................................. 47
v
3.2.2 Experimental Setup ........................................................................................ 49
3.2.3 Experiment Results ........................................................................................ 51
3.2.4 Discussions on Experiment Results.................................................................. 55
3.3 Chapter Summary ................................................................................................... 58
4 Adaptive Computation of Gain Matrices ........................................................ 59
4.1 Effects of Inaccurate Mode Shapes ......................................................................... 60
4.2 Adaptive Design of Modal Sensors ......................................................................... 62
4.2.1 Correlation between Modal Coordinates ......................................................... 62
4.2.2 Adaptive Computation of Sensor Gain Matrix ................................................ 65
4.3 Numerical Example ................................................................................................. 67
4.4 Chapter Summary ................................................................................................... 75
5 Eigenvector Computing Algorithms ................................................................. 79
5.1 Jacobi Rotation Algorithm ...................................................................................... 79
5.2 Algorithm A, Convergence and Rotation Angle ....................................................... 85
5.3 Algorithm B............................................................................................................. 87
5.3.1 Development .................................................................................................. 87
5.3.2 Numerical Example......................................................................................... 91
5.4 Algorithm C............................................................................................................. 95
5.4.1 Development .................................................................................................. 95
5.4.2 Numerical Example ...................................................................................... 100
5.4.3 Frequency-Domain Analysis ......................................................................... 102
5.5 Limitations............................................................................................................. 112
5.6 Chapter Summary ................................................................................................. 113
6 A Mode Shape Sensing Technique ................................................................. 115
6.1 Sensor Gain Matrices and Mode Shapes ................................................................ 115
6.2 Lagrange Interpolation .......................................................................................... 117
6.3 A Numerical example ............................................................................................ 118
6.4 Chapter Summary ................................................................................................. 120
7 Conclusions and Future Direction ................................................................... 121
7.1 Conclusions .......................................................................................................... 121
7.2 Future Directions .................................................................................................. 122
vi
References ....................................................................................................................... 125
Appendix A LMS Computation of Gain Matrix ........................................... 130
Appendix B LMS Algorithm ................................................................................. 136
Appendix C A Control-Model Identification Procedure ........................... 138
Appendix D Stability of IIR Filters ..................................................................... 142
Vita ...................................................................................................................................... 146
vii
LIST OF FIGURES
1.1 Modal filtering in time domain ........................................................................................... 3
1.2 Modal filtering in spatial domain ........................................................................................ 4
1.3 Mobility magnitudes at 4 points ......................................................................................... 5
1.4 Modal mobility magnitudes for modally combined sensors .................................................. 6
1.5 Modal sensor in continuous spatial domain ........................................................................ 6
1.6 Using an adaptive algorithm to create a modal sensor ........................................................ 8
1.7 Error signal history of adaptive modal filter ....................................................................... 8
1.8 Typical control-model identification ................................................................................. 12
1.9 Typical experimental modal analysis and modal filtering with conventional methods ......... 14
1.10 Modal analysis and modal filtering with the new modal analyzer ...................................... 15
2.1 Beam with pin and pin-with-torsion-spring boundary conditions ...................................... 17
2.2 Continuous-time and sampled (discrete time) systems ...................................................... 21
2.3 Representing a high-order structure with a parallel bank of second-order digital filters ..... 24
2.4 An infinitesimal piezoelectric element under strain ........................................................... 26
2.5 Strain in the film as a function of deflection of the beam ................................................... 27
2.6 Piezoelectric film and zero-impedance signal conditioner ................................................. 28
2.7 Spatial filters as modal sensors ......................................................................................... 30
2.8 Segment positions on beam .............................................................................................. 31
3.1 Mode shapes of beam ...................................................................................................... 36
3.2 a)Z-plane poles, b)Denominator coefficients of the IIR-filter-equivalent of the beam ........ 37
3.3 Comparison between beams driving-point mobility and digital filters FRF ...................... 38
3.4 Time-domain comparison between second-order digital filters and ideal modal
coordinates: impulse response .......................................................................................... 40
3.5 Contribution of each mode to segment outputs ................................................................ 41
3.6 Gain matrix for modal sensors ......................................................................................... 42
3.7 Modal sensor output compared to ideal modal coordinates: impulse responses ................. 44
3.8 Responses of a 10-mode filter to a 12-mode impulse excitation ........................................ 45
3.9 Modal sensor output compared to ideal modal coordinates: impulse responses of modes 8
and 9 ............................................................................................................................... 46
3.10 Modal sensor output compared to ideal modal coordinates: FRF from force to sensor
output and modal coordinates .......................................................................................... 46
3.11 Experiment setup ............................................................................................................. 49
3.12 Schematic picture of experiment setup ............................................................................. 50
3.13 FRF from force to segment outputs ................................................................................. 51
3.14 Sensor gain matrix W for transforming 20 segment outputs into 8 modal coordinates ....... 54
3.15 FRFs from force to sensor outputs .................................................................................. 55
3.16 FRFs from force to sensor outputs, linear scale ............................................................... 57
viii
3.17 End connection to approximate simple support ................................................................ 58
4.1 Magnitudes of the responses of the sensor on the T* = 1 structure and on the T* = 10
structure........................................................................................................................... 60
4.2 Phases of the responses of the sensor on the T* = 1 structure and on the T* = 10
structure .......................................................................................................................... 61
4.3 Real parts of the responses of the sensor on the T* = 1 structure and on the T* = 10
structure .......................................................................................................................... 62
4.4 Modal responses to random excitation ............................................................................. 63
4.5 Mode-1 coordinate. Mode-3 coordinate, product of mode-1 and mode-3 coordinates,
average of product of mode-1 and mode-3 coordinates .................................................... 64
4.6 Adjusting sensor gain matrix to diagonalize correlation matrix ......................................... 66
4.7 Sensor gain matrix adjustment using eigenvector matrix of segment output correlations ... 68
4.8 Sensor gain matrix computed with 16 sets of time data (W16) ......................................... 68
4.9 Ideal sensor gain matrix ................................................................................................... 69
4.10 Sensor output correlation matrix using W(16) ................................................................. 70
4.11 Sensor gain matrix W(256) .............................................................................................. 71
4.12 Sensor output correlation matrix resulting from W(256) .................................................. 72
4.13 Sensor gain matrix calculated using 32768 time data points .............................................. 72
4.14 Sensor output correlation matrix resulting from W(32768) .............................................. 73
4.15 Sensor gain matrix calculated using 49152 data points ..................................................... 74
4.16 Sensor output correlation matrix resulting from W(49152) .............................................. 74
4.17 Magnitudes of adaptive sensor outputs and ideal modal coordinates ................................ 76
5.1 Jacobi rotation example ................................................................................................... 83
5.2 Result of first sweep ........................................................................................................ 84
5.3 Result of second sweep .................................................................................................... 84
5.4 Algorithm A .................................................................................................................... 86
5.5 Typical rotation angle history of Algorithm A .................................................................. 87
5.6 Sensor output correlation matrix, Algorithm B, 256 time steps ........................................ 92
5.7 Sensor output correlation matrix, Algorithm B, 32768 time steps ..................................... 92
5.8 Rotation angle history, Algorithm B ................................................................................. 93
5.9 Sensor gain matrix (Algorithm B), 32768 time steps ........................................................ 93
5.10 Normalized magnitude responses of modal filter (Algorithm B) after 32768 time steps...... 93
5.11 Input connections to Algorithm C .................................................................................... 99
5.12 Rotation angle history, Algorithm C................................................................................ 100
5.13 Sensor gain matrix (Algorithm C), after 32768 time steps ............................................... 101
5.14 Sensor output correlation matrix (Algorithm C) after 32768 time steps........................... 101
5.15 Normalized magnitude response of modal filter with performance feedback: Mode 1 ..... 102
ix
5.24 Normalized magnitude response of modal filter with performance feedback: Mode 10 ... 107
5.34 Normalized magnitude response of modal filter with performance feedback: Mode 10 ... 112
6.1 Sensor gain matrix W .................................................................................................... 116
6.2 Mode shapes of beam, (x) ............................................................................................ 116
6.3 Beam with strain sensor segments .................................................................................. 117
6.4 Third row of sensor gain matrix ..................................................................................... 119
6.5 Reconstructed mode shape ............................................................................................ 120
x
LIST OF TABLES
3.1 Physical properties of beam ............................................................................................. 34
3.2 Eigenvalues of beam ........................................................................................................ 35
3.3 Natural frequencies of beam ............................................................................................. 35
3.4 Gain matrix for modal sensor ........................................................................................... 42
3.5 Physical properties of beam ............................................................................................. 47
3.6 Analytical natural frequencies of beam ............................................................................. 48
CHAPTER 1
INTRODUCTION
Vibration is a very important phenomenon in machinery and structures. In some cases vibration
causes breakdown, malfunction or discomfort. In other cases vibration is the principal means of
operation. In many systems, from ships to musical instruments, quality and performance are
closely related to vibration. In those cases it is very important to understand and control vibration.
To understand and control vibration of a structure, first it is necessary to characterize the
vibrational properties of the structure, i.e. to have certain knowledge of how the parts of the
structure vibrate. Vibration characterization is essential in anticipating the vibration levels or
determining what actions to be taken if the vibration is to be controlled. However, the quest for
the ultimate control of vibration has advanced to the extent that active forces are now used to
counteract the vibration. This relatively new vibration control method is called active control of
vibration. This method requires sensing of the vibration in real time. Regardless of the active
control strategy, either feedback or feedforward, sensing is necessary.
This dissertation was conceived of an aspiration to contrive a system that performs both the
characterization and the sensing of vibration of machinery or structural components. This system
operates on the basis of adaptive processing of signals from distributed sensor arrays. This chapter
will give the reader an idea of the basic concepts, purpose, and expected results of the research
endeavor to develop the system.
1.1 Modal Analysis and Modal Coordinates
Vibration of a multi-degree-of-freedom system can be expressed in terms of the motion of the
systems along several coordinates. The equations governing the motion of the system can be
relatively simple or relatively complicated depending on the choice of the coordinate system.
Some coordinate systems result in coupled equations of motion. Coupling means that one cannot
solve any of the individual equations without involving the others.
The choice of coordinate system determines the degree of coupling among the equations. As a
rule, the more coupling exists among the equations, the more complicated the solutions are. In
controlling the vibration of a multi-degree-of-freedom system, a coordinate system that leads to
no coupling among the equations also allows simple control schemes. In many cases, it is possible
to choose a coordinate system that results in no coupling among the equations of motion. The
coordinates in such a coordinate system are called the principal coordinates. These coordinates
are also called the natural coordinates.
2
The natural coordinates provide a basis on which to express mathematically the vibration of a
structure. On this basis, the vibration of a structure can be viewed as a summation of products of
a spatial function and a temporal function.
w x t x t
m m
m
M
( , ) ( ) ( ) =
=

1
, (1.1)
where x is position and t is time. The spatial function is the mode shape of the structure, which
is a characteristic of the structure. The temporal function is called modal coordinate.
Real-time monitoring of modal coordinates is very important in active vibration control of
continuous structures. The use of modal coordinates in feedback control can prevent control
spillover, a phenomenon that results in degradation of performance or in instability (Balas, 1978).
Feedback control problems of continuous structures using modal coordinates can be viewed as a
problem of controlling single-degree-of freedom (SDOF) systems in parallel, with no interaction
among the systems (Meirovitch and Baruh, 1982). Decades ago, Porter and Crossley

(1972)
published a book dedicated to this control method, which is called modal control. Modal control
has been developed for several control applications such as vibration control of large space
structures (Davidson, 1990).
Several control theories have been developed using the modal control concept for various control
problems including LQG optimal control (Bai and Shieh, 1995). Positive Position Feedback (Baz
and Poh, 1996), and neural-network-based control (Chen et al., 1994). Modal control theory has
also advanced beyond linear structures (Slater and Inman, 1995). Modal coordinates are not just
useful in feedback control. Clark (1995) developed a feedforward control strategy that relies
heavily on the availability of modal coordinates. Modal control experiments have been done on
various structures such as plates

(Clark ,1991, Zhou, 1992, Gu et al., 1994, Miller et al., 1996),
cylinders (Sumali and Cudney, 1991, Finefield et al., 1992, Clark and Fuller, 1993), and highway
bridges (Shelley et al., 1991).
All of the above modal control methods require monitoring of modal coordinates. Sensing modal
coordinates in real time is so important that many researchers have developed a special area
within structural control dedicated to obtaining modal coordinates in real time (Meirovitch and
Baruh, 1985, Ouyang, 1987, Shelley, 1991). This area is called modal sensing. A short description
of some previous work in modal sensing is presented below.
1.2 The Quest for Modal Coordinate Sensors
Many researchers have developed techniques to create sensors that can produce modal
coordinates in real time. The proposed research work in this dissertation adopts the concepts of
3
spatial filtering, segmentation, and adaptive signal processing. This section will mention a selected
sampling of previous work especially related to those concepts.
1.2.1 Modal Filtering in Time Domain
Monitoring modal coordinates of a vibrating structure can be done by processing signals from
sensors in time domain. Several researchers claimed that this processing can be done by filtering
sensor outputs with a bank of filters, each of which admits only certain frequency and filters out
other components of the signal. Balas (1978) introduced this modal filtering concept. Ouyang
(1987) developed a realization of this concept with a bank of special filters where each filter only
passes a single frequency that coincides with a natural frequency of the structure (See Fig. 1.1).
Davidson (1990) conceptually designed an electronic circuit that implements Ouyangs filters with
a set of phase-locked loops (PLLs) built with voltage-controlled oscillators (VCOs) that
generates pure sinusoidal signals. Davidson performed numerical simulation of a scenario where
his VCO-based modal filters are used to control many hundred modes of a large space structure.
Figure 1.1 Modal filtering in time domain.
1.2.2 Modal Filtering in Spatial Domain
Another method to obtain modal coordinates in real time from sensor outputs is by filtering the
sensor outputs in space. Basically this means assigning different weights to different sensor
outputs depending on which mode to be sensed. Sumali and Cudney (1991) performed some
experiments using this method. One set of weights produce one modal coordinate. This method
can be illustrated with the beam in Fig. 1.2. Assume that the exciting force is such that the
response is limited to combinations of the first four modes. In practice, this rather simplistic
assumption might be realized by several methods such as low-pass filtering the excitation. We
know that for the simple boundary conditions the mode shapes of the Euler-Bernoulli beam are
Point
sensor
Beam
v
f
( )
( )
v(t)
VCO- based
filter
&
$
( )
1
t
&
$
( )
2
t
&
$
( )
3
t
&
$
( )
4
t
4
sinusoidal. We use four point sensors to sense displacements at strategically assigned positions on
the beam, based on our knowledge of the mode shapes.
Figure 1.2 Modal filtering in spatial domain.
= Mode shape 1
= Mode shape 2
= Mode shape 3
= Mode shape 4
.414
1
1
.414
1
1
-1
-1
1
-.41
-.41
1
1
-1
1
-1
f
j
Point sensor
v
f
j
1
( )
( )
v
f
j
2
( )
( )
v
f
j
3
( )
( )
v
f
j
4
( )
( )
&
$
( )
( )

1
f
j
v
4
(t) v
3
(t) v
2
(t) v
1
(t)
&
$
( )
1
t
&
$
( )
( )

2
f
j
&
$
( )
( )

3
f
j
&
$
( )
( )

4
f
j
&
$
( )
2
t
&
$
( )
3
t
&
$
( )
4
t
5
This method can be described well with simulation results. The frequency response functions
(FRFs) from the force to the velocities at the four point sensor locations are shown in Fig. 1.3.
By combining the outputs of the four sensors with the right mixture as shown in Fig. 1.2, we can
obtain sensor outputs that are proportional to the individual modal coordinates, as shown in Fig.
1.4. The gains in Fig. 1.2 constitute a matrix that transforms the sensor coordinate system to the
modal coordinate system. It will be shown later that these gains are closely related to the modal
matrix (or eigenvector matrix) of the structure.
If we increase the number of sensors, we get a higher spatial resolution. In the limit, for an infinite
number of sensors, the sensor gain matrix becomes continuous functions, each row representing a
mode. This modal sensing concept was invented by Lee (1987) for strain sensors such as
piezoelectric film. The width of the film is varied along the beam as a function of the modal sensor
weight. For a mode 3 sensor, such a sensor is shown in Fig. 1.5. Theoretically, this sensor is an
ideal mode 3 sensor, insensitive to any other mode. Lee and Moon (1990) successfully applied
this type of spatially distributed modal filter to vibration control. Many other researchers have
implemented Lees modal filters in various forms, for example: Structure-borne acoustic sensors
(Clark, 1992), one-dimensional modal sensors on plates (Zhou, 1992), one-dimensional modal
sensors on cylinders (Sumali, 1992; Clark and Fuller, 1993) sensors for feedforward modal
control of structures (Clark, 1995), acoustic sensing by volume-velocity (Guigou et al., 1995).
Burke and Hubbard

(1990) explained the concept of spatial filtering sensors and applied the
techniques to control distributed parameter systems in general.
0 100 200 300 400 500 600 700 800
10
-4
10
0
0 100 200 300 400 500 600 700 800
10
-4
10
0
0 100 200 300 400 500 600 700 800
10
-4
10
0
0 100 200 300 400 500 600 700 800
10
-4
10
0
Frequency (Hz)
Y
i 1
,
m
Ns
Y
i 2
,
m
Ns
Y
i 3
,
m
Ns
Y
i 4
,
m
Ns
Figure 1.3 Mobility magnitude, Y
v
f
ij
i
j
=
( )
( )
, at four points.
6
0 100 200 300 400 500 600 700 800
10
-4
10
0
0 100 200 300 400 500 600 700 800
10
-5
10
0
0 100 200 300 400 500 600 700 800
10
-5
10
0
0 100 200 300 400 500 600 700 800
10
-5
10
0
Frequency (Hz)
&
$
( )
1
f
j
&
$
( )
2
f
j
&
$
( )
3
f
j
&
$
( )
4
f
j
Figure 1.4 Modal mobility magnitudes,
&
$
( )
( )

i
j
f
, for modally combined sensors.
Figure 1.5 Modal sensor in continuous spatial domain.
+
-
PVDF film
segment V
1
Beam
Point
sensor
Mode 3
sensor
Frequency
V
o
f
j
V
f
o
j
( )
( )
V
f
j
1
( )
( )
7
1.2.3 Polyvinylidene Fluoride (PVDF) Film Sensors
Throughout the research work, the material used for vibration sensor is thin film of a piezoelectric
polymer polyvinylidene fluoride (PVDF). The use of this material in the structural dynamics
community was promoted mainly by Lee (1987). This material is particularly suited for the
structural modal sensing applications because it offers many advantages, including the following.
PVDF is lightweight. Compared to accelerometers (even very small ones), PVDF film has
negligible mass-loading effects, even on light structures. The polymer is also very compliant. This
feature enables us to neglect the changes in structural parameters due to stiffening by the film.
Another attractive feature of the PVDF film sensor is its strain-integrating nature. Unlike point
sensors, segments of piezoelectric film can be cut into shapes that convolve the strain on the
structures surface with a specified function (Lee and Moon, 1990, Collins et al., 1992). The
function can be selected such that the convolution results in a low-pass filtering effect in the wave
number domain. This effect results in the reduction of spatial aliasing, which is an important
property that we will address later in this dissertation.
PVDF film segments are much cheaper than accelerometers or other vibration transducers such as
strain gages, laser, or fiber optic transducers. The signal conditioning circuit for a PVDF sensor is
also much cheaper than the signal conditioning circuits for the other sensors. The simplicity to
attach large numbers of PVDF segments on structures is also very important in creating highly
distributed sensor arrays.
PVDF has a strong piezoelectric effect compared to most other piezoelectric materials (Lee,
1987). The material is also robust, both physically and chemically. It has been used also as
protective coatings against harsh environment, for instance, as vat liners for chemicals (Collins et
al., 1990). It endures time, temperature (up to about 120
o
C), and mechanical shock (up to several
hundred gs). The (mechanical) bandwidth of the film is very high (up to about 10
7
Hz). With
appropriate signal conditioning circuit, no dynamics is introduced by the sensor.
1.2.4 Segmentation of Modal Filtering Sensors
The idea of dividing the piezoelectric film layer into segments emerged primarily because a
piezoelectric sensor layer covering the whole host structure fails to detect anti-symmetric modes
(Cudney, 1992). Several researchers have investigated the use of segmentation in piezoelectric
film sensors. Tzou and Fu (1992, 1992b) developed a theory for applying the segmented sensors
and actuators to vibration control of plates. Clark (1992) developed an adaptively-computed
sensor array gains to apply to segmented PVDF film sensors so that the sensor array emulates
vibration sensors and structural acoustic sensors. Sumali and Cudney (1993) developed a
technique to create modal sensor from an array of segmented piezoelectric film sensors. Callahan
and Baruh (1994) developed another method to create modal sensors from the same segment
8
array configuration. Sullivan (1993) developed a special kind of distribution calculus to calculate
the responses of piezoelectric sensor patches of general shapes and to calculate the actuation of
arbitrarily shaped strain-induced actuators on multidimensional structures. Sullivan also developed
a sensor array that is weighted spatially according to a linearly varying function developed by
Burke and Hubbard (1990).
1.2.5 Modal Filtering Using Adaptive Algorithms
At least two methods have been developed to create modal sensors using adaptive signal
processing. The first method

(Shelley et al., 1992) uses the sensor configuration shown in Fig. 1.6.
The outputs of the sensors are input to a linear combiner with some gain vector. The gain vector
is computed on-line using the LMS algorithm. The structure is excited with a random force
excitation. The output of the sensors are used to adjust the gain matrix iteratively until the output
of the linear combiner matches the output of a pre-programmed second-order digital filter. The
natural frequency and damping of the digital filter are pre-programmed based on the knowledge of
the natural frequency, modal damping, and residue of the desired mode. These parameters must
be known in advance. The advantage of this method is that knowledge of the mode shapes of the
structure is not required in computing the gain matrix. An example of the application of this
method is shown below. The details of a simulation of the adaptive modal filter are given in
Appendix A.
Figure 1.6 Using an adaptive algorithm to create a modal filter.
...
Error
signal
(k)
Sensor output
&
$
3
(k)
-
+
W
3,1
(k)
Adaptive
algorithm
V
1
(k)
Beam
PVDF segments
f(k)
T
*
Torsion spring
W
3,2
(k) W
3,10
(k)
Desired signal
d(k)
V
2
(k)
V
10
(k)
...
9
This adaptive method seems to result in an impressive modal filtering effect. Figure 1.7 shows the
difference between the sensor output and the ideal modal coordinate. This error converging to
zero means that the sensor output converges to the desired modal coordinate.
0 0.01 0.02 0.03 0.04 0.05
-0.25
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
0.25
Time (sec)
Error
(Volt)
Mode 3 sensor, T* = 1
Figure 1.7 Error signal history of adaptive modal filter.
Despite the impressive modal filtering effect, a more critical examination of Appendix A reveals
the fact that this technique is not likely to be of practical utility. Figures A.1 and A.2 in the
appendix show that the key element in this modal filter is the filter that generates the desired
signal d(k). This filter must be programmed with the natural frequency, modal damping, and
modal residue. Therefore, this method still requires complete knowledge of the natural
frequencies, modal damping, and residues of the modes of the structure. If one knows these
parameters and hence the filter coefficients, then there is no need for the adaptive linear combiner.
The filter alone can be used to generate the desired signal d(k), which is precisely the modal
coordinate. The real utility of this method would be in obtaining the mode shape.
The second method uses a parallel bank of second-order recursive digital. This method is based
on the modal decomposition of the multi-degree-of-freedom structure into a parallel bank of
single-degree-of-freedom systems as in Eq. (1) (Horvath, 1976). Wimmel and Melcher

(1992)
used adaptive recursive filter algorithms developed by White (1975) and Hsia (1981). Part of this
method, namely modeling a multi-degree-of-freedom (MDOF) system as a parallel bank of
second-order filters, will be adopted in this research work for the purpose of simulating a
structure with already-known modal properties with digital filters.
10
1.3 Conventional System Characterization and Mode Shape Extraction
From the above discussion, we know that modal filtering requires knowledge of the characteristic
of the structure: either mode shapes or natural frequencies and damping ratios. In fact, the most
important step in virtually all vibration analyses is to obtain the natural frequency or eigenvalues
and mode shapes or eigenvectors of the system. The importance of eigenvalues and eigenvectors
can not be overemphasized. In all but the simplest structures, the modal properties of the system
must be obtained numerically or experimentally. The most popular methods to obtain the modal
properties of a structure by experiments can be easily classified into two broad categories. Each
category is different from the other in many ways: tradition and historical development,
mathematical foundation, and experimental knowledge base. Even the purposes of the different
categories are different. The first category is commonly called System Identification, the
second, Experimental Modal Analysis (EMA). In this section we will discuss the concepts and
efforts involved in computing mode shapes from experimental data using the two categories of
computation methods.
System identification, or more specifically Control-Model Identification (Juang, 1994) has
developed since the mid-sixties out of the necessity to control sophisticated systems such as
guidance and controls of aerospace structures. The main objective of this type of analysis is to
obtain the block diagram of the system to be controlled so that the control engineer can design the
controller, observer, sensors, and actuators. Modal properties of the structure can be obtained
easily from the resulting dynamic model of the structure. However, obtaining modal properties is
only a small part of system identification and often not an essential part. Much theoretical work
has been generated by many researchers on system identification (see for example the bibliography
of Juang, 1987).
EMA originated partly from such testing practices as Resonance Testing and Mechanical
Impedance Methods in the 1940s (Ewins, 1986). Development in electronics in the 1960s and
Cooley and Tukeys Fast Fourier Transform (FFT) algorithm created a revolution in signal
processing (Mitchell, 1986). Unlike Control-Model Identification, Modal Testing is especially
geared towards obtaining the modal properties of structures. The resulting estimates of modal
properties are used not mainly in active control of aerospace structures, but in various other tasks,
such as design modification and passive vibration control.
EMA is very commonly used in obtaining vibration modes to verify finite element or theoretical
models. Once the model is validated, it can be used to predict the responses to complex
excitations such as shock, or to proceed to more further stages of analysis. A validated model can
be used as a basis for further modeling and analysis. Modal Testing is often used to produce a
mathematical model of a component which may then be used in a structural assembly. Mode
shapes obtained from EMA can be used to modify the design of a component to improve its
vibrational characteristics, such as lower dynamic stresses, less acoustic radiation, relocation of
points of large vibration, etc.
11
In terms of knowledge requirements, Modal Testing requires a thorough integration of 1)
Vibration theory 2) Accurate measurement, and 3) Signal processing. The area of Modal Testing
is very rich in experimental knowledge and intuitive rules that complement, sometimes even
circumvent, the mathematics. On the other hand, Control-Model Identification requires the
mathematics of modern control systems theory. In particular, most Control-Model Identification
algorithms rely on Singular-Value Decomposition (SVD). The state-space is the standard domain
of Control-Model identification.
Keeping the backgrounds of Control-Model Identification and Modal Testing in mind, we can
now make a reasonable judgment of the two categories of methods in terms of an important goal
in this dissertation: Computing mode shapes. Then we propose a new method of computing mode
shapes, which also results in a means of obtaining modal coordinates in real time. The advantages
of this new method over Control-Model Identification and the classical Modal Testing will be
discussed later.
For an explanation of the Control-Model Identification procedure, Juang (1994b) is an excellent
reference, on which the following paragraphs are based. The purpose of describing the Control-
Model identification procedure here is only to give an idea of the computational process. Figure
1.8 shows a typical sequence of variables to calculate in Control-Model identification of a
structure. Knowledge of SVD-based procedures such as one described in Appendix C is desirable
for further understanding of the rest of this section.
From the identification procedure in Appendix C, we learn that obtaining mode shapes from
experimental data using Control-Model Identification requires at least the following operations:
1. Fast Fourier Transform (FFT) to transform force and response signals into the frequency
domain. This process does not only require the application of Fourier transform algorithms,
but also other procedures to ensure good results, e.g. windowing, coherence computation and
checking, averaging, sometimes zooming, and so on.
2. FRF Computation from all input forces to all sensor outputs. This complex operation requires
multiplication and division. FRF computation must be done to all time data. In a highly
distributed sensor system with a high number of input channels, FRF computation requires
significant computing power.
3. Inverse FFT to transform the FRF back into time domain to obtain Markov parameters.
4. Singular value decomposition (SVD). Equations (C.4) through (C.7) in Appendix C show that
SVD is a key step in the identification process. This process is a lengthy sequence of matrix
operations (see, for example Golub and Van Loan, 1989).
5. Eigenvalue and eigenvector computation. This is another computationally extensive operation.
6. Transformation of eigenvalues from the discrete z-plane to the continuous s-plane (Eq.
(C.19)). This transformation requires evaluation of logarithms, a computation process that is
much less elementary than addition or multiplication. In the implementation of real-time
algorithms, this nonlinear operation may take many computation steps.
12
7. Raising a matrix to the -1/2 power (Equation (C.9)). This process obviously needs extensive
computation.
8. Matrix inversion (Eq. (C.21)) and several multiplications. This is another computationally
extensive process.
From the above discussion, we can conclude that a typical Control-Model Identification process is
computationally extensive. If mode shapes are to be obtained and modal filtering are to be
performed in real time, this type of analysis must be done with tremendous computing power.
Less computing power may do the job if the computation is done with recursive methods.
Recursive algorithms are available mainly for single-input-single-output (SISO) systems. (Ljung
(1989) is an excellent reference.) The algorithms are mostly based on Recursive Least Squares
(RLS) or Kalman filter theories. These methods are not suitable for obtaining mode shapes
because they do not address the distributed nature of sensors required in mode shape
computation.
EMA is perhaps the best category of methods to obtain mode shapes. However, the current
procedures to compute mode shapes are not designed for on-line computation. Figure 1.9 shows
typical steps of modal testing and the computation steps to obtain mode shapes. The first steps in
EMA are identical to their counterparts in Control-Model Identification. The procedure shown in
the picture requires the first three operations performed in Control-Model Identification. EMA
programs use circle-fit, least-squares curve-fitting, and other averaging and error-minimizing
techniques. Like the processes in Control-Model Identification, those processes also require too
much computing power to apply in real time.
The foregoing discussion was not meant to argue that the current state of the art is full of
disadvantages and inefficiency. The argument is that the current practices of system identification
are not geared towards real-time application. Recent advances in adaptive algorithm have created
a trend towards streamlining the computational procedure by using recursive adaptive algorithms.
(See, for example, Ljung (1989b)). However, most of the current advances can be traced back to
SVD, least-squares curve-fitting including recursive least-squares error minimization and Kalman-
filtering-type algorithms, mainly because those are standard procedures. A typical system
identification expert is well-trained in, and feels comfortable with, those standard procedures.
Building on the current techniques can only improve the speed within the bounds inherent to the
nature of the underlying concepts.
The LMS algorithm (Widrow, 1985) is a notable exception to the above statements although the
underlying concept is least-squares error minimization. This algorithm is simple, versatile, and
requires very little computation, hence a lot of practical applications. The drawbacks of this
algorithm are, among others: Slow convergence, inability to take advantage of knowledge about
the system, lack of guaranteed stability. In terms of adaptive modal filtering, the application of this
algorithm is not very promising, mainly because it requires some kind of desired signal, as
discussed earlier.
13
Figure 1.8 Typical Control-Model Identification. (Adapted from Juang, 1994b)
1.4 On-line Modal Analyzer: a Novel Concept
The proposed research work is to develop a structural vibration sensor system that obtains the
mode shapes of the structure in real time and performs modal filtering. This sensor system will be
called the modal analyzer. This novel system does not require discrete Fourier transform (DFT or
FFT), FRF computation, SVD, or matrix inversion. The modal analyzer works in time domain
using adaptive algorithms. This method is based on correlation and computation of eigenvectors.
Markov parameters
Hankel matrix, H(0)
Singular values
Time-shifted Hankel
matrix, H(1)
SVD
State Matrix
Eigenvalues =
Natural frequencies
and modal damping
Left singular vectors Right singular vectors
Output matrix Input matrix
Eigensolution
Mode shapes Modal amplitudes
14
Therefore, it still needs eigenvector computation (step 5 in the Control-Model Identification
procedure described above.) This step is computationally extensive. However, the proposed
system performs this step recursively, improving the estimates of the eigenvectors gradually with
each iteration. Adaptive algorithms will be developed to perform this operation.
Comparison between Fig. 1.9 and Fig. 1.10 shows that the new modal analyzer bypasses many
computational steps between data acquisition and the computation of mode shapes and modal
coordinates. The modal analyzer does not compute natural frequencies and modal damping.
However, it performs modal filtering concurrently with mode shape computation. The
conventional methods require that the mode shapes be obtained first, and then the modal filter
constructed accordingly. If the structures parameters change, for example, due to temperature
change or drift in boundary conditions, then the fixed-parameter modal filter will no longer be
accurate. The modal analyzer, on the other hand, will track changes in structural parameters, and
adaptively adjust the modal filter so that the outputs are the correct modal coordinates.
1.5 Overview of Dissertation
Although a proof-of concept experiment will be conducted, most of the theory developed in this
dissertation will be tested on a simulated structure. This structure and its discrete-time simulation
will be described in chapter 2. Throughout the dissertation it is assumed that that the structure is
linear and self-adjoint, that the modes are real, and that the mode shapes are orthogonal. Most of
the theory developed here will apply only under those assumptions.
The sensor system is physically constructed from an array of piezoelectric polyvinylidene fluoride
(PVDF) film segments connected to electronic signal conditioning circuits. The design and
modeling of the sensor array are also described in chapter 2.
Numerical simulation of the structure-sensor system is presented in chapter 3. In this chapter we
also present an experiment to verify that the simulation procedure, indeed, represents the physical
system and to prove that the concepts used in the development of the adaptive sensor array are
physically realizable.
In chapter 4, we reveal the fundamental principle that we will use to develop a new algorithm to
perform adaptive mode shape calculation and modal filtering simultaneously. Numerical
simulation is presented to demonstrate the effectiveness of using this fundamental principle. The
development of the adaptive algorithms is presented in chapter 5. In chapter 6 we utilize a simple
formula to extract mode shapes from sensor gain matrices. Conclusions are presented in chapter
7. Also in this last chapter we discuss the possibility of constructing a hardware prototype of the
modal analyzer.
15
Figure 1.9 Typical experimental modal analysis and modal filtering with conventional methods.
Force
j
Time
Velocity
1
FFT
FFT
FFT
( )
( )
( )( )
FRF
Velocty FFT Force
Force FFT Force FFT
ij
ir jr
r
jr jr
r
=
FFT
*
*
FFT
Velocity
2
FRF
1
FRF
2
FRF
i
FRF
N
.
.
.
.
.
.
Modal
analysis
program
Natural
frequency
Damping
Mode shape 1
Mode shape 2
.
.
.
Modal filter
Mode
shapes
Velocity
Modal coordinates
16
Figure 1.10 Modal analysis and modal filtering with the new modal analyzer.
Force
Time
Velocity
N
Mode shape 1
Mode shape 2
.
.
.
Modal analyzer
Velocity
1
Velocity
i
Modal
coordinates
.
.
.
CHAPTER 2
SIMULATION METHOD AND SENSOR MODEL
Throughout this dissertation we will use examples to help explain the concepts and theories that
we will develop. In the examples we will use numerical simulation of a structure. In the first half
of this chapter we will discuss the structure, modeling, and simulation, and give some insight into
the choice of the adaptive identification algorithm. In the second half of the chapter, we will
discuss the development of the theory of segmented sensors, the configuration of the sensor, and
its principle of operation.
2.1 Model of Variable Host Structure
2.1.1 Description of Structure
The structure we choose should be simple enough to model mathematically, but also allows for
variations in mechanical properties. For this purpose, we choose a simply-supported beam with
some torsional stiffness at one of its simple supports (Fig. 2.1).
Figure 2.1 Beam with pin and pin-with-torsion-spring boundary conditions.
The modal properties of the structure depends partly on the torsional spring stiffness T*. The
normalized torsion spring constant is
T TL EI
*
( ) , (2.1)
where T* is the torsion spring constant in m
-1
, L is the length of the beam in m, E is the modulus
of elasticity in Pa, and I is the cross-section area moment of inertia in m
4
.
The following paragraphs explain the modeling and simulation of the structure. We assume a
linear Euler-Bernoulli beam with internal viscous damping. As in most vibrational analyses of
Force f(t)
T
*
Torsion
spring
18
structures, first of all we need to obtain the eigen-properties of the structure. Then we present the
equations of motion in the modal coordinates.
2.1.2 Eigen-Properties of Structure
The m
th
eigenvalue
m
of the beam is the m
th
non-trivial solution to the characteristic equation for
the undamped case is (Gorman, 1975).
( ) cos sinh sin cosh sin sinh
*
T 2 0. (2.2)
The m
th
natural frequency in rad/s,
m
, is

m m
m
L
f
L
EI
2
2
2
, (2.3)
where f
m
is the natural frequency in Hz,
m
is the m
th
root of Eq. (2.2), L is the length of the beam,
EI is the product of the modulus of elasticity and the area moment of inertia, and
L
is the mass
per unit length. The mass-normalized mode shape is

m
mass norm
m
m
m
m
x
c
x L x L ( ) sin( / )
sin
sinh
sinh( / )
_
,
1
, (2.4)
where the mass normalization constant c
mass norm
for each mode can be calculated by setting

L m
L
x x
2
0
1 ( ) d
(2.5)
for all m. This equation results in a mass normalization constant for each mode shape.
( )
c L
massnorm L
m
m
m
m
m
m
m
m
m m
m m m m
+
_
,
'

05
2
4 4 2
2
2
2
.
sin sin
sinh
sinh
sin
sinh
sin cosh cos sinh
(2.6)
19
2.1.3 Equation of Motion and Its State Space Form
The velocity response of the structure to a point force excitation can be expressed in terms of
modal expansion as
& ( , ) ( )
&
( ) w x t x t
m m
m
M

1
(2.7)
where
m
(x) is the m
th
mass-normalized mode shape of the structure, and
m
(t) is the modal
coordinate, which is the solution to the equation of motion
&&
( )
&
( ) ( ) ( ) ( )
m m m m m m m f
t t t x f t + + 2
2
(2.8)
where
m
(t) denotes modal coordinates. The dot above the variables denote time derivation. The
summation limit M in Eq. (2.7) is theoretically infinity, but in practice must be limited to the
number of modes included in calculations;
m
2
is the eigenvalue;
m
denotes the modal damping;
m
(x
f
) is the value of the mass-normalized mode shape at the forcing point position.
In later chapters, we will need a state-space representation of the structure. The state-space model
is
&
( )
&&
( )
( )
&
( )
( )

t
t
t
t
f t
'

'
+ A B , (2.9)
where
A
Z
n
1
]
1
0 I
2

, (2.10)
is the matrix containing the natural frequencies and damping ratios, i.e.,

n
2
1
2
2
1
]
1
1
1
0
0
O
M
, (2.11)
and
20

1
]
1
1
1
2
2
1 1

0
0
O
M M
. (2.12)
0 and I are M-by-M zero and identity matrices, respectively. In the second term, the input matrix
is
B
'
0
0
1
M
M
( )
( )
x
x
f
M f
. (2.13)
2.2 Simulation of Structure
2.2.1 Sampling and Discrete-Time Model
Because data are normally acquired by sampling and processed by digital signal processing
equipment, it is necessary in our analysis to model the input signal, the structure, and the output
signal as discrete-time processes.
Figure 2.2 shows a structure excited with a force f(t). At this point we consider the point sensor.
The output signal is a velocity v(t) at some point. Because of sampling, the measured force and
velocity will be discrete functions of time step or sample number k instead of continuous functions
of time. We call the sampled input f(k) and the sampled output v(k). We must represent the
frequency response function (FRF) of the structure by a discrete time FRF. This FRF must be
such that if the input f(k) is a sampled version of the force f(t), the output v(k) is a sampled
version of the velocity v(t).
21
Figure 2.2 Continuous-time and sampled (discrete time) systems.
2.2.2 Discrete-Time Model of One Mode
To give a simple example of the process of representing a structure with a digital filter, first we
consider only mode-m of the structure. The frequency response of a single mode is like that of a
second-order system. According to Eq. (2.8), the continuous-time transfer function from force to
a modal coordinate rate
&
in the Laplace (s) domain is
&
( )
( )
( )
( )
( )

m m
m f
m m m
s
f s
s s
f s
x s
s s

+ +
2 2
2
(2.14)
(Henceforth, what we denote by modal coordinate actually means the time rate of change of the
modal coordinate. This time rate is related to vibration velocity and is often more important than
the modal coordinate.) If we know the modal properties of a structure, i.e., the natural
frequencies, modal damping, mode shapes, and residues, we can synthesize the digital filter
representation using several methods. The digital filter is connected to the analog parts of the
system by A/D and D/A converters, usually with a zero-order hold. For this kind of signal
conversion, the discrete equivalent of the transfer function in Eq. (2.14) can be computed by
(Franklin et al., 1990)
H z z
s
s s
f s
( ) ( )
( )
( )

'
1
1
1
Z

(2.15)
Equivalent discrete-time system
Sampler
Continuous-time system (ideal structure)
f(t)
v(t)
f(k)
v(k)
22
where Z { } denotes the z-transform operator. It can be shown that substituting Eq. (2.14) into Eq.
(2.15) results in the discrete transfer function
&
( )
( )
( )
m m
m m
z
f z
b z z
a z a z
+
+ +

1 2
1
1
2
2
1
(2.16)
where the filter coefficients are
b x T
m m f
m m
m m
m m

( )
sin( )
exp( )
1
1
2
2
, (2.17)
a T
m m m m m 1
2
2 1 exp( ) cos( ) (2.18)
a T
m m m 2
2 exp( ) (2.19)
In discrete-time domain, the second-order filter representation of the single-mode structure is
&
( )
&
( )
&
( ) ( ) ( )
m m m m m m m
k a k a k b f k b f k + +
1 2
1 2 1 2 (2.20)
What we have done so far is:
1. Modeled the structure with its modal properties, i.e.,
m
,
m
, and
m
(x
0
)
2. Transformed the equations of motion in time domain into transfer functions in the s-domain
3. Transformed the continuous transfer functions in the s-domain into discrete-time transfer
functions in the z-domain
4. Transformed the discrete-time transfer functions into digital filter coefficients
2.2.3 Discrete-Time Model of Multi-Mode Structure
Continuous structures, such as the example structure we use in this research work, have an
infinite number of modes. In practical calculation and numerical analysis we can only include a
limited number of modes. Each mode increases the order of the transfer function by two. In
general, an M-mode structure has a 2M
th
order transfer function of the form
& ( )
( )
...
...
w s
f s
s s s
s s s
M
M
M
M
+ + + +
+ + + +
b b b b
a a a
0 1 2
2
2
2
1 2
2
2
2
1
(2.21)
23
The above transfer function can be represented by a 2M
th
order digital filter of the form
H z
v z
f z
z z z
z z z
M
M
M
M
( )
( )
( )
...
...

+ + + +
+ + + +

0 1
1
2
2
2
2
1
1
2
2
2
2
1
(2.22)
To decompose the multi-mode transfer function into several single-mode transfer functions, we
can decompose Eq. (2.22) by partial fractioning into
H z
b b z b z
a z a z
m m m
m m m
M
( )
+ +
+ +

0 1
1
2
2
1
1
2
2
1
1
(2.23)
The last equation can be realized as a bank of second-order infinite-impulse-response (IIR) digital
filters connected in parallel with a common input. The outputs of the filters are summed up to
produce the total output of the system. This digital-filter representation of the structure is shown
in Fig. 2.3.
The representation of the structure as a parallel bank of second order digital filters lends to modal
analysis because the transfer function of the parallel filters is exactly the same as the modal
expansion of the structures transfer function. Each second-order filter has three coefficients that
are the results of transformations from the natural frequency, modal damping, and modal residues.
Even modal analysis could be done adaptively if we find a way to compute the filter coefficients
recursively.
2.2.4 Selecting Sensor Configuration
The above conclusion seems very promising because it shows the possibility of doing on-line,
adaptive modal analysis. The most difficult part of realizing this concept is the adaptive algorithm
that would be necessary to adjust the filter coefficients without requiring a high computational
load that precludes the on-line signal processing. There is however another important problem
with the parallel IIR filter configuration, namely, stability. Basically, stability requires that
t < a a a
m m m 1 1
2
2
4 1, (2.24)
and
a
a
m
m
1
2
2
1
<
<
(2.25)
24
Figure 2.3 Representing a high-order structure with a parallel bank of second-order digital filters.
This issue is discussed separately in appendix D. The conclusions we can draw from the
discussion is given below.
If the system to be identified is a high order system such as continuous structures, the order of the
acceptable auto-regressive-moving-average (ARMA) model in Eq. (2.22) is very high. With high
order models, a slight error or variation from the true values can very easily drift the digital filter
poles out of the stability disc. Moreover, the algorithm used to compute the coefficients may
f
Beam
Force
transducer
v
Velocity
sensor
b z z
a z a z
11
1 2
11
1
2 1
2
1
,
, ,
( )

+ +
b z z
a z a z
m
m m
1
1 2
1
1
2
2
1
,
, ,
( )

+ +
b z z
a z a z
M
M M
1
1 2
1
1
2
2
1
,
, ,
( )

+ +
M
M
y
f
1
( )
( )
y
1
y
m
y
M
v
^
T
*
Torsion
spring
$( )
( )
v
f
v
f
( )
( )
25
easily turn to be unstable itself. The near-instability of the structure couples with the instability
of the algorithm to make stability problem even more difficult to deal with.
With the second-order filters, it may be easy to add a stability check routine into the adaptive
signal processing algorithm to makes sure that the two denominator coefficients for each mode
stay within the stable underdamped parabola of Fig. D.1.c in appendix D. However, this brute-
force approach may create more problems in the development and analysis of the adaptive
algorithms. One such problem is discontinuities in the evaluation of the gradients if the gradient-
search methods are used, or in the Kalman gains if the recursive-least-squares methods are used.
Because of the stability problem, we will not use the parallel IIR filter configuration in our system
identification algorithm. However, we will use this model in simulating mechanical structures with
known modal properties, and also as an important tool in the development of the on-line modal
coordinate sensor. For these two purposes, we know the exact values of the filter coefficients.
2.3 Segmented Sensor Model
Originally, segmentation of strain sensors was invented to avoid the problem of observability. If a
sensor is made of one continuous piece of piezoelectric film covering the whole structure, the
symmetry of the structure causes the charge resulting from some positive strain to be cancelled
out by the charge resulting from some negative strain. The result is that the sensor cannot sense
(or observe) the symmetric mode (Cudney, 1992, Tzou and Fu, 1992). Segmentation may solve
this observability problem.
This section discusses a sensor system which can sense modal coordinates of vibrating structures
in real time. If we used IIR filters, we would need three filter coefficients for each mode. Finite-
impulse-response (FIR) filters would require many more coefficients because this type of filters do
not reuse results from previous time steps. The lack of computational feedback is the reason for
the inherent stability of the FIR filters. In many cases, especially in adaptive sensing, this
guaranteed stability is worth the extra number of coefficients. One purpose of this research work
is to come up with sensor/filter configurations that require minimum number of filter coefficients
without tremendous increases in computational loads. A logical approach is to get as much spatial
information as possible from the structure. It is for this purpose that we devise the segments of
piezoelectric film as distributed sensors on the structures.
For the array of segments to be modal filters, the outputs must be multiplied by correct sets of
gains. If we know the mode shape of the structure, it is easy to construct a sensor that can
selectively sense one mode and filter out other modes. Lee (1987) is an excellent reference on
how to create on-line modal coordinate analyzers. Lees design method requires one spatially
continuous piezoelectric sensor film profile for each mode. The segmented design of modal
sensors does not require spatially continuous shaping of the film and does not require a separate
26
layer of sensor film for each mode. The most important advantage of our sensor is that it can be
made adaptive.
2.3.1 Voltage Generated by Segment
Within its linear region, piezoelectric film generates electric charge proportional to strain and to
the area of the film. Figure 2.4 shows an infinitesimal element of original length dx and width b.
The three orthogonal directions for an electrically anisotropic piezoelectric material are shown by
the arrows. The electrodes are on the 1-surfaces (top and bottom) of the element. The element is
stretched in the 3-direction of the piezoelectric material. The charge generated by this element
under strain is
dq e dx
31
, (2.26)
where e
31
is a piezoelectric charge-density-per-strain constant in the 3-direction.
Figure 2.4 An infinitesimal piezoelectric element under strain.
Figure 2.5 shows a segment of a beam with a rectangular piezoelectric film patch attached on the
top surface. The left end of the film is at x = x
1
, and the right end is at x = x
2
. The beams
deflection as a function of position x is denoted by w(x). The piezoelectric film is so compliant
compared to the beam that we can assume that the film does not affect the bending of the beam
with any loading. Assuming Euler-Bernoulli beam theory, we can calculate the strain in the film,
based on how the film is stretched or compressed as a function of the flexure of the beam.
dx
b
dx
1
2
3
Directions
27
Figure 2.5 Strain in the film as a function of deflection of the beam.
Neglecting the small thickness of the piezoelectric film, the strain on the piezoelectric film can be
considered the same as the strain on the surface of the beam, which is
( , )
( , )
x t
h w x t
x

2
2
2
, (2.27)
where h is the thickness of the beam. We obtain the total charge generated in the piezoelectric
film by substituting Eq. (2.26) into Eq. (2.27) and integrating over the length of the film to yield
q t e b
h w x t
x
dx e b
h w
x
w
x
x
x
x x
( )
( , )

_
,
31
2
2
31
2 2
1
2
2 1
. (2.28)
Thus, the charge generated by the film is proportional to the difference between the slope at the
right end and the slope at the left end of the piezoelectric film.
The electrical circuit model of the piezoelectric film is a charge generator in parallel with the
capacitance of the film (Kynar Piezofilm Technical Manual, 1987). The voltage generated by the
segment depends also on the signal conditioning circuit connected to the piezoelectric film. The
signal conditioning circuit that we will use is an Op-Amp circuit with theoretically zero input
impedance. The circuit representation of the film and the signal conditioner is shown in Fig. 2.6.
Deflection
w(x)
x
x
1
x
2
h
Piezoelectric
film
Neutral
axis
w(
n-1
)
w(
n
)
28
Figure 2.6 Piezoelectric film and zero-impedance signal conditioner.
Because we connect the film to a zero-impedance signal conditioner, the output terminals of the
film are effectively shorted together. The capacitor C does not draw any current and therefore can
be ignored. In this case, the output current is simply the time derivative of the charge generated by
the film. For the signal conditioning circuit, the output voltage is the product of the current i(t)
and the feedback resistance , thus,
V t
dq t
dt
( )
( )
(2.29)
Substituting Eq. (2.28) into Eq. (2.29), we obtain the output voltage of the piezoelectric film
segment as a function of the deflection of the beam.
V t e b
h w x t
x
w x t
x
x x
( )
& ( , ) & ( , )

_
,
31
2
2 1
(2.30)
2.3.2 Array of Piezoelectric Film Segments on Beam
As shown in Fig. 2.7, N piezoelectric sensor segments are mounted on the beam. We will
construct modal filters, i.e., linear combiners whose output emulates modal coordinates of the
vibrating beam. Each modal filter consists of an array of weights (gains) that multiply the outputs
of the segments. The mode-m filter is an array of gains W
nm
, where m denotes the mode and n
denotes the segment number to which the gain is connected. The output of each segment is
denoted by V
n
. The output of the mode-m sensor, denoted by &
$
( )
m
t , is the weighted sum of the
segment outputs,
q(t)
C
Charge
generator
Piezoelectric film
Op-Amp
+
-
Output
voltage
i(t)
V(t)
Signal conditioner
29
&
$
( ) ( )
m
t t
mn n
n
N
W V
1
. (2.31)
If we have M modal sensors, we can write the outputs of all of them as
&
$
= WV (2.32)
where
&
$
is an M-vector of sensor output, W is an (M x N) matrix of gains, and V is an N-vector
of segment outputs [V
1
, V
2
, ..., V
N
]
T
. Our task now is to determine the gain matrix W so that the
modal filter outputs
&
$
emulates modal coordinates.
As shown in Fig. 2.8, the left and right end positions of the n
th
segment are denoted by
n-1
and
n
,
respectively. We connect each segment to signal conditioning circuit with zero input impedance.
For the n
th
segment, the circuit translates the current into voltage V
n
, which can be obtained by
Eq. (2.30).
( ) V
n n n
t w t w t ( ) & ( , ) & ( , )

1
, (2.33)
where

h
e b
2
31
, (2.34)
the dot above the variable w denotes time derivative and prime denotes spatial derivative.
30
Figure 2.7 Spatial filters as modal sensors
V
n V
N
V
1
Beam
PVDF
segments
Force
transducer
f
Mode-1
sensor
&
$
( )
( )

1
f
&
$
( )
( )

M
f
.
.
.
.
.
.
Mode-m
coordinate
W
1,1
W
m,N
W
m,1
W
M,N
W
1,n
W
m,n
W
M,n
W
M,1
W
1,N
. . . . . .
31
Figure 2.8 Segment positions on beam.
2.3.3 Gain Matrix for Segmented Modal Sensor
The time rate of change of a deflection slope can be expanded into its modal components
according to Eq. (1.1),
& ( , ) ( )
&
( )
w x t x t
m
m
M
m

1
, (2.35)
where '
m
(x) is the slope of the m
th
mode shape, and
m
(t) is the m
th
modal coordinate.
Substituting Eq. (2.35) into Eq. (2.33), we obtain the segment outputs V
n
, n = 1, ..., N, as
V C
S
&
. (2.36)
where the (n,m) component of the segment output matrix C
S
is defined as
( ) C
S
( , ) ( ) ( ) ; ,..., ; ,..., n m n N m M
m n m n

1
1 1 . (2.37)
and
&
is the modal coordinate vector
&
&
( )
&
( )

'
1
t
t
M
M . (2.38)
The segment output voltages in Eq. (2.36) are input to the linear combiners (i.e., multiplied by the
gain matrix W) to produce the sensor outputs in Eq. (2.32). The sensor output vector
&
$
and the
gain matrix W contain as many rows as there are linear combiners. One linear combiner emulates
one modal filter.
PVDF segments
f(t)
Simple
Support
x
n-1

n
V
n
x
f
Beam
32
Substituting Eq. (2.36) into Eq.(2.32), we can express the outputs of the linear combiners as
functions of the gain matrix, the segment output matrix, and the modal coordinates,
&
$
&
WC
s
. (2.39)
For the segment array to behave as modal coordinate sensors, the segment outputs V must be
multiplied by correct gains. The gain matrix W is the matrix that ideally will transform the
segment outputs V into the modal coordinates
&
. Equation (2.39) shows that the estimated modal
coordinates
&
$
will be equal to the ideal modal coordinates
&
only if the gain matrix is
W C
S
1

(2.40)
We will use the above equation to obtain the ideal gain matrices of linear combiners.
If we substitute the above weight matrix into Eq. (2.39), at the first glance it seems that the
estimated modal coordinates are equal to the ideal modal coordinates. However, this is not
exactly the case. Note that the gain matrix W and the segment output matrix C
S
must be square.
From Eq. (2.40), it is clear that to create a sensor to sense M modes we need at least N = M
segments. Theoretically, the structure has an infinite number of modes, therefore, the size of the
segment output matrix C
S
should be infinity. In practice, we can only use a finite number of
segments. Therefore, the segment output matrix C
S
is actually a truncated version of the ideal,
and the weight matrix obtained by Eq. (2.40) is only an approximation of the ideal weights for
modal filters. This approximation will result in a non-ideal estimate for the modal coordinates, as
will be shown in a later chapter.
2.4 Chapter Summary
In this chapter we reviewed and developed some concepts that are required to create a sensor-
and-signal-processor system that functions as a modal coordinate sensor. This system will receive
its input from the mechanical vibration of a structure and produce outputs that are proportional to
the modal coordinates of the structural vibration. As a transducer to convert mechanical vibration
into electrical signal, we choose an array of segments of piezoelectric film. The signal-processing
part of the system consists of linear combiners that act as modal filters. Each modal filter shall be
sensitive to only one mode, and filter out any other modes. This chapter reviewed and developed
some concepts necessary to design the modal filters.
To facilitate understanding of the concepts, we developed a structural model to use as examples
in analytical development and numerical simulation. The model we have chosen is a simply
33
supported beam with an additional torsion spring at one end. The excitation is a point force. We
derived the natural frequencies, mode shape, and equation of motion for the structure.
We developed a discrete-time model for the structure. Mechanical structures can be modeled by
equivalent electrical circuits. When the signals are sampled, both can in turn be modeled with
digital filters. The modal-expansion form of the response of the mechanical structure is a
summation of second-order-type responses. Therefore, we derived a method to model a
mechanical structure with a filter composed of parallel second-order digital filters. This method
transforms the modal properties of the structure (natural frequency, damping, and residue) into
filter coefficients. The resulting discrete-time model is not only useful in simulating the structure
with a digital computer, but also useful in further analysis of the characteristics of the structure.
The relationship between the filter coefficients and the modal properties is known precisely.
Therefore, if we can adaptively adjust the filter to match the structure, we shall be able to extract
the modal properties of the structure from the filter coefficients. Thus, the parallel-second-order-
filter model of the structure could be used as a basis for developing an adaptive algorithm that can
be used in on-line modal parameter estimation. A problem with this approach is stability. Many
metal structures have very low damping that cause the z-plane poles of the structures to be very
close to the edge of the unit circle. Adaptive algorithms for identifying these poles can easily come
up with an unstable pole in the adaptation process, throwing the whole adaptive identification
process into instability. Stability problems force us to consider other filter configuration for
system identification. We chose a simple linear combiner configuration to do the filtering in space
rather than time.
CHAPTER 3
NUMERICAL SIMULATION AND PROOF-OF-
CONCEPT EXPERIMENT
The work discussed in the chapters subsequent to this will rely heavily on numerical simulations.
In this chapter we describe an experiment to ensure that our simulation closely represents real
structures which it is designed to simulate. Towards the end of this chapter, we will also
understand the need for adaptive modal filters.
3.1 Verification of Digital Filter Model and Gain Matrix Formula
3.1.1 Verification of Digital Filter Model
To verify the method we use to represent the structure with digital filters, we shall present a
numerical example. As an example, we simulate a beam with boundary conditions as described in
Fig. 2.1, which are a pin at x = 0 and a pin and a torsion spring at x = L. The properties of the
rectangular beam are given in Table 3.1.
Table 3.1 Physical properties of beam
With the above physical properties, the parameter
1
2
L
EI
L
23.9596 s
-1
. The eigenvalues can be
computed by solving Eq. (2.2). For several nondimensional torsion spring constant T* (defined in
Eq. (2.1)), these eigenvalues are as shown in Table 3.2. The T* = 0 case is the case of simply
supported beam without a torsion spring and is included in the table for comparison.
Property Value Unit Property Value Unit
Length (L) 0.4382 m Density () 2700 kg/m
3
Width (b) 0.0381 m Youngs modulus (E) 68(10)
9
Pa
Thickness 3.175 mm Poissons ratio 0.3
35
Table 3.2 Eigenvalues of beam
The natural frequencies can be computed by Eq. (2.3). For several nondimensional torsion spring
constant T*, the natural frequencies of the beam are as shown in Table 3.3.
Table 3.3 Natural frequencies of beam (Hz)
For the numerical example, we choose T* = 1 as the torsion spring stiffness. The mode shapes of
the beam as computed by Eq. (2.4) are shown in Fig. 3.1.
T* 0 0.032 0.1 0.32 1 3.2 10 32
1
3.1416 3.1466 3.1572 3.1888 3.2733 3.4461 3.6646 3.8193
2
6.2832 6.2857 6.2911 6.3076 6.3560 6.4769 6.6874 6.8915
3
9.4248 9.4265 9.4301 9.4412 9.4749 9.5657 9.7516 9.9748
4
12.566 12.567 12.570 12.578 12.604 12.676 12.839 13.067
5
15.708 15.709 15.711 15.717 15.738 15.798 15.942 16.167
6
18.849 18.850 18.852 18.857 18.875 18.926 19.054 19.217
7
21.991 21.991 21.993 21.998 22.013 22.057 22.173 22.383
8
25.132 25.133 25.134 25.139 25.152 25.191 25.296 25.498
9
28.274 28.274 28.276 28.279 28.291 28.327 28.422 28.616
10
31.415 31.416 31.417 31.420 31.431 31.463 31.551 31.737
T* 0 0.032 0.1 0.32 1 3.2 10 32
f
1
37.6 37.8 38.0 38.8 40.9 45.3 51.2 55.6
f
2
150.5 150.7 150.9 151.7 154.1 160.0 170.5 181.1
f
3
338.7 338.8 339.1 339.9 342.3 348.9 362.6 379.4
f
4
602.2 602.3 602.6 603.4 605.8 612.8 628.6 651.1
f
5
940.9 941.0 941.3 942.1 944.6 951.8 969.2 996.7
f
6
1354.9 1355.0 1355.3 1356.1 1358.6 1366.0 1384.5 1408.1
f
7
1844.1 1844.3 1844.5 1845.3 1847.9 1855.4 1874.8 1910.6
f
8
2408.7 2408.8 2409.1 2409.9 2412.4 2420.0 2440.1 2479.3
f
9
3048.5 3048.6 3048.9 3049.7 3052.2 3059.9 3080.6 3122.7
f
10
3763.6 3763.7 3764.0 3764.8 3767.3 3775.0 3796.2 3840.9
36
0 0.2 0.4 0.6 0.8 1.0
10
9
8
7
6
5
4
3
2
1
x / L
Mode
Figure 3.1 Flexural mode shapes of beam.
The beam is excited with a point force at position x
f
= 0.1304 times the length of the beam. A
typical loss factor value for a metal beam in bending is about 0.001 (Lazan, 1968). However, the
beam that we will use in an experiment described later in this dissertation is covered on one side
with a thin polymer double-sided adhesive tape and with segments of polyvinylidene fluoride
(PVDF) film. An experiment described later in a subsection 3.2.3 was used to estimate the
damping ratio of the beam with this covering. The damping ratios for all modes was found to be
around
m
= 0.01. With the above properties, we can compute the driving-point mobility of the
beam using Eq. (2.7) and Eq. (2.8). We limit our calculation to include only the first 10 modes of
the beam.
To simulate the beam, we can design a digital filter that consists of a bank of IIR filters with
coefficients computed by Eqs. (2.17) through (2.19). We chose a sampling rate 4 times the tenth
natural frequency f
10
. As shown in the z-plane in Fig. 3.2.a, the filter poles are barely inside the
unit circle because the damping ratios are quite low. Figure 3.2.b shows the filter denominator
coefficients are very close to the unstable region above the a
2
= 1 line.
37
-1 -0.5 0 0.5 1
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Real part
Imag-
inary
part
a)
-2 -1.5 -1 -0.5 0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
a
1
a
2
Coefficient plane
b)
Figure 3.2 a)Z-plane poles b)Denominator coefficients of the IIR-filter-equivalent of the beam.
38
To verify that the parallel bank of digital filters, indeed, simulates the driving-point mobility of the
beam, we compare the continuous-time driving-point mobility of the beam with the FRF of the
IIR filter. The driving point mobility of the structure is computed by substituting s = j into Eq.
(2.14) and the Laplace transform of Eq. (2.7). The FRF of the filter can be computed by
substituting z = exp(jT) into Eq. (2.23). The comparison between the beams driving-point
mobility and the IIR filters FRF is shown in Fig. 3.3. The magnitudes of the two FRFs agree
very well. However, the digital filter loses about 60
o
of phase at the high end of the frequency
range. Such a phase lag might be of some importance in a feedback control system design. A
higher sampling rate would bring the phase back up.
10
1
10
2
10
3
-100
0
Frequency (Hz)
10
1
10
2
10
3
10
-4
10
-2
10
0
Frequency (Hz)
100
__ Digital filter FRF
.... Beam mobility
Figure 3.3 Comparison between beams driving-point mobility and digital filters FRF.
Further verification of the digital filter characteristics can be done by comparing the impulse
response of the filter to the impulse response of the beam. Other excitation than an impulse force
could be used as well. We choose the impulse force because it is mathematically the simplest
excitation in the integration of the modal equation of motion.
The modal response of the beam to a unit impulse force (t) can be obtained as follows.
Integrating the modal equation of motion between t = 0 and t = , and taking the limit as 0 ,
we obtain
39
&
( ) ( )
m m f
x 0
+
(3.1)
The magnitude of the initial modal coordinate happens to be the same as the mass-normalized
mode shape value at the excitation point. (However, it can be shown that the unit is not kg
-1/2
but
kg
1/2
ms
-1
, due to the integral of (t) over time. The corrected unit is consistent with the modal
equation of motion.) Using the above initial modal coordinate as an initial condition and
m
( ) 0 0 (3.2)
as another initial condition, we can show that the modal coordinate as a response to the impulse
force is
&
( ) ( ) exp( ) cos( ) sin( )

m m f m m dm dm m
m
dm
dm
t x t t
'
t (3.3)
where

dm m m
1
2
(3.4)
is the damped natural frequency of mode-m.
The impulse response of each IIR filter can be evaluated in the discrete-time domain using Eq.
(2.20). The unit impulse is approximated with an input of 1/(sampling period) at k = 0 and zero
inputs thereafter. We can compare the responses of the digital filters with the modal responses of
the beam. The comparison is shown in Fig. 3.4. The agreement between the digital filter outputs
and ideal modal coordinates is very good. This agreement confirms that the parallel bank of
second-order IIR filters in Fig. 2.3, indeed, represents the dynamics of the beam with specified
modal properties.
3.1.2 Verification of Gain Matrix Formula
From the above comparison in frequency domain and in time domain, we have established
confidence in simulating the beam with a parallel bank of second-order digital filters. Since we can
simulate the responses of the beam, we can also simulate the outputs of the piezoelectric segments
on the beam. Now we can design a modal sensor for the beam by computing a gain weight matrix
W that will transform the outputs of the piezoelectric segments into modal coordinates. We will
take into account the first ten modes of the structure, use ten segments, and obtain a sensor that
outputs the first ten modal coordinates of the structure. Equation (2.37) gives the segment output
matrix C
S
. The m
th
column of this matrix represents the contribution of the m
th
mode to the
outputs of the sensor segments. These columns are shown graphically as strips in Fig. 3.5. We can
see that low-numbered modes contribute less to the outputs of the segments than high-numbered
40
modes. This difference is understandable since the output of a segment is related to the deflection
slope rather than the deflection itself. Higher modes generally give more slopes than lower modes.
kg
m
s
_
,
-1
0
1
-2
0
2
-4
0
4
-4
0
4
-2
0
2
-2
0
2
-1
0
1
-0.5
0
0.5
-2
0
2
0 0.01 0.02 0.03 0.04 0.05
0 0.005 0.01 0.015 0.02 0.025
0 0.004 0.008 0.012 0.016
0 0.002 0.004 0.006 0.008 0.01 0.012
0 0.002 0.004 0.006 0.008 0.01
0 0.002 0.004 0.008
0 0.002 0.004 0.006
0 0.002 0.004 0.006
0 0.002 0.004
0 0.002 0.004
-2
0
2
0.006
Time (s)
&
1
&
2
&
3
&
4
&
7
&
8
&
9
&
10
&
5
&
6
filter output =modal coordinate
&
$
;
&
L
Figure 3.4 Time-domain comparison between second-order digital filters and ideal modal
coordinates: impulse response.
41
2 4 6 8 10
10
9
8
7
6
5
4
1
2
3
1 3 5 7 9
Segment number
Mode
Figure 3.5 Contribution of each mode to segment outputs.
Finally, we compute the sensor gain matrix W according to Eq. (2.40). This gain matrix is shown
in Table 3.4. Each row of this matrix contains the gains for each mode, and is shown as a strip in
Fig. 3.6. If we multiply the outputs of the segments by this gain matrix according to Eq. (2.39),
the product will approximate the modal coordinates of the beam.
The mode-1 sensor has big gains compared to other modal sensor. These gains are large mainly
because the contribution of mode 1 to the outputs of the segments is small, as shown in Fig. 3.5.
Table 3.4 Gain matrix for modal sensor.
2.008 5.8254 9.0594 11.389 12.589 12.547 11.278 8.922 5.7183 1.9872
0.929 2.4178 2.9356 2.2640 0.6490 -1.293 -2.818 -3.339 -2.655 -1.014
0.6441 1.4021 1.0110 -0.191 -1.192 -1.146 -0.080 1.1269 1.4764 0.6726
0.4748 0.7601 -0.024 -0.798 -0.495 0.4539 0.7222 -0.072 -0.831 -0.502
0.3845 0.3860 -0.375 -0.364 0.4041 0.3962 -0.365 -0.345 0.4222 0.4007
0.3178 0.1166 -0.399 0.1167 0.3053 -0.335 -0.130 0.3757 -0.150 -0.334
0.2743 -0.046 -0.211 0.3094 -0.134 -0.126 0.3132 -0.207 -0.028 0.2860
0.2388 -0.150 -0.002 0.1417 -0.244 0.2326 -0.159 -0.007 0.1341 -0.251
0.2129 -0.190 0.1552 -0.095 0.0397 0.0383 -0.089 0.1596 -0.183 0.2244
0.0948 -0.096 0.0939 -0.097 0.0932 -0.098 0.0925 -0.100 0.0923 -0.101
42
10
0
5
10
15
9 8 7 6 5 4 3 2 1
1
2
3
6
7
8
9
10
Gain
Mode
Segment number
5
4
Figure 3.6 Gain matrix for modal sensors.
To verify that the gain matrix W really produces modal coordinates when multiplied by the
segment outputs, we shall do the following test.
1. Apply an impulse force to the beam with zero initial conditions.
2. Compute the ideal modal response for each mode by solving the modal equation of motion
(Eq. (2.8)) in time domain.
3. Compute the (time-domain) velocity response along the beam.
4. Compute the segment outputs using Eq. (2.36).
5. Multiply the segment outputs by the gain matrix to obtain the modal sensor output.
6. Compare the results of step 5 with the result of step 2.
Steps 1 and 2 were done earlier with Eq. (3.3) (see Fig. 3.4). The outputs of the segments (step 4)
can be computed using Eq. (2.35) through (2.37). The sensor gain matrix is obtained using Eq.
(2.40). In this step we use 10 modes to obtain the gain matrix for 10 segments. The impulse
responses of the modal filter
&
$
is compared with the ideal modal coordinates
&
in Fig. 3.7.
43
kg
m
s
_
,
0 0.01 0.02 0.03 0.04 0.05

-1
0
1
0 0.005 0.01 0.015 0.02 0.025
-2
0
2
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016
-4
-2
0
2
4
0 0.002 0.004 0.006 0.008 0.01 0.012
-4
-2
0
2
4
0 0.002 0.004 0.006 0.008 0.01
-2
0
2
0
-2
0
2
0
-1
0
1
0
-0.5
0
0.5
0
-2
0
2
0
-2
0
2
Time (s)
&
1
&
2
&
3
&
4
&
5
&
6
&
7
&
8
&
9
&
10
&
$
;
&
L
0.002 0.004
0.008
0.002 0.004 0.006
0.002 0.004 0.006
0.002 0.004
0.002 0.004
0.006
Figure 3.7 Modal sensor outputs compared to ideal modal coordinates: impulse responses.
44
3.1.3 Modal Truncation and Spatial Aliasing
As mentioned in subsection 2.3.3, the gain matrix calculation in Eq. 2.40 is approximate in nature.
Ideal modal coordinates can only be sensed with an infinite number of segments (i.e., Lees modal
sensor as in Fig. 1.5, with perfect geometry for the mode shapes). Using a finite number of
segments means using a finite number of modes in Eq. (2.37) and (2.40). This modal truncation
results in non-ideal modal coordinates. To investigate the effect of modal truncation, this time we
include the first 12 modes of the beam in the response calculation. The natural frequencies of
these modes are f
11
= 4558 Hz and f
12
= 5423 Hz. The modal damping ratios are assumed to be
0.01. We use the gain matrix in table 3.4 to obtain the outputs of the modal sensors (step 5). This
gain matrix was calculated using only 10 modes. The outputs of the modal sensors are shown in
Fig. 3.8. These outputs are generally in excellent agreement with the ideal modal coordinates,
except for modes 8 and 9, which are not shown in Fig. 3.8.
The outputs of mode-8 and mode-9 sensors are shown Fig. 3.9. These outputs do not agree with
the ideal modal coordinates. The disagreement is shown further in the frequency domain in Fig.
3.10.
Figure 3.9 shows that modes 8 and 9 sensors are very sensitive to modes 11 and 12. It can be
shown that if mode 13 were included in the beam response, this mode will be sensed by mode 7
sensor. This is a clear indication of spatial aliasing. This problem could be solved in spatial domain
by altering the shapes of the segments, or in frequency domain by using analog low-pass filters.
The frequency domain solution simply means filtering out the signal contribution from modes
with higher indices than the modes used in the gain matrix calculation. Spatial aliasing problem
needs special treatment outside the scope of this dissertation. For now, we just conclude that we
should low-pass filter the sensor output before the samplers, or that we should use band-limited
excitation forces so that aliasing is not a problem.
The above numerical simulation shows that the linear combiner is very effective in separating the
response of the beam into its modal coordinates. Experiments are required at this point of the
research work to verify the concept and numerical simulation. After experimental verification, we
shall be ready to develop an algorithm that would adjust the gain matrix automatically so that we
would not need to know the modal properties of the structure to create the modal sensor.
Development of this algorithm will be the main thrust of this dissertation.
45
kg
m
s
_
,
-1
0
1
-2
0
2
-4
0
4
-4
0
4
-2
0
2
-2
0
2
-1
0
1
-0.5
0
0.5
-2
0
2
0 0.01 0.02 0.03 0.04 0.05
0 0.005 0.01 0.015 0.02 0.025
0 0.004 0.008 0.012 0.016
0 0.002 0.004 0.006 0.008 0.01 0.012
0
0.002 0.004 0.006 0.008 0.01
0
0.002 0.004
0.008
0
0.002 0.004 0.006
0 0.002 0.004 0.006
0 0.002 0.004
0 0.002 0.004
-2
0
2
0.006
Time (s)
&
1
&
2
&
3
&
4
&
5
&
6
&
7
&
8
&
9
&
10
&
$
;
&
L
Figure 3.8 Responses of a 10-mode filter to a 12-mode impulse excitation.
46
&
,
&
$

9 9
kg
m
s
_
,
&
,
&
$

8 8
kg
m
s
_
,
0 1 2 3 4 5 6
x 10
-3
-5
0
5
0 1 2 3 4 5 6
x 10
-3
-5
0
5
Time (s)
&
$
;
&
L
Figure 3.9 Modal sensor output compared to ideal modal coordinates: impulse responses of
modes 8 and 9
1000 2000 3000 4000 5000
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
kgm
Ns
_
,

&
$
;
&
L
1
2
3
4
5
6
7
8
9
11
10
8
8
9
8
9
0
Frequency (Hz)
12
8 =
&
$
8
9 =
&
$
9
Figure 3.10 Modal sensor output compared to ideal modal coordinates: FRFs from force to
sensor outputs and to modal coordinates.
47
3.2 Proof-of-Concept Experiment
In the above section we verified the method we use to represent the structure with digital filters.
We also and also demonstrated numerically that the gain matrix calculated by Eq. (2.40), indeed,
transforms the segment outputs into modal coordinates. In this section, we will further verify the
simulation technique and the gain matrix calculation method by performing an experiment with a
simply supported beam equipped with piezoelectric film segments.
3.2.1 Properties of Experimental Structure
For the experiment, we use a beam with approximated pin supports at x = 0 and at x = L. The
properties of the beam are as tabulated in Table 3.5.
Table 3.5 Physical properties of beam
Property Value Unit Property Value Unit
Length (L) 0.4500 m Density () 2700 kg/m
3
Width (b) 0.0381 m Youngs modulus 68(10)
9
Pa
Thickness 3.175 mm Poissons ratio 0.3
The shaker point of action is at x
f
= 0.1742 m.
This structure can be considered a special form of the structure in Fig. 2.1 with T* = 0. For this
case, the modal properties of the structure can be found in many texts such as Blevins (1984). The
natural frequencies of the beam are

m
L
m
L
EI
1
2
2
(3.5)
where
L
is the mass per unit length of the beam. The m
th
eigenvalue is

m
m . (3.6)
With the above physical properties, the parameter
1
2
L
EI
L
21.0455 s
-1
, and the beams
analytical natural frequencies are as shown in Table 3.6.
48
Table 3.6 Analytical natural frequencies of beam (Hz)
The mass-normalized mode shapes of the beam are
m
L
x
L
m x
L
( ) sin

_
,
2
. (3.7)
The m
th
modal coordinate
m
(t) is the solution to the equation of motion, which is
&&
( )
&
( ) ( ) ( ) ( )
m m m m m f
t t t x f t + + 2
2
(2.8)
where
m
2
is the eigenvalue,
m
denotes the modal damping,
m
(x
f
) is the value of the mass-
normalized mode shape at the forcing point position. The summation limit M in Eq. (2.7) is set to
12.
The ideal responses from force to the sensor outputs are computed as follows. To simulate the
modal coordinates in frequency domain, we use MATLABs bode command with a state space
representation of the structure. The state space model is given in Eqs. (2.9) through (2.13).
To compute the FRFs from force to segment output voltages, we use Eq. (2.36). Then we
compute the gain matrix W using Eq. (2.40). Finally, we compute the FRF from force to sensor
output using Eq. (2.39). Obviously, these ideal sensor outputs are identical to the modal
coordinates. In the next sections, we will compare the experiment results with these ideal sensor
outputs.
3.2.2 Experimental Setup
Figure 3.11 shows the setup for the experiment, where we can see the beam covered with PVDF
segments attached to the steel frame. In the background we can see the shaker and the signal
conditioning circuit.
Figure 3.12 explains the setup in detail. Thin metal shim is used to connect each end of the
aluminum beam to the frame. This shim gives almost no resistance to bending, but is very stiff in
the beams transverse direction. Therefore, the shim approximates a simple boundary condition.
The shaker excites the beam at position x = x
f
. A force transducer is placed between the shakers
stinger and the beam.
f
1
f
2
f
3
f
4
f
5
f
6
f
7
f
8
f
9
f
10
f
11
f
12
37.6 150.5 338.7 602.2 940.9 1355 1844 2409 3049 3764 4554 5420
49
Figure 3.11 Experimental setup.
Figure 3.12 Schematic picture of experiment setup.
OP-07
+
-
100 K
Shaker
Stinger
Force
Transducer
Simple
Support
Beam
PVDF Film
Segment
Segment
Output
50
The beam is covered with N = 20 segments of piezoelectric polyvinylidene fluoride (PVDF) film.
Each piezoelectric film segment is connected to a current amplifying circuit shown in the figure.
This circuit is designed so that the output signal is proportional to the strain rate of the segment
(Lee, 1991). The input impedance of this circuit is theoretically zero. Current flows freely from
and to the PVDF film segment. The operational amplifier (Op Amp) used is type OP-07. The 100
k feedback resistor multiplies the current to produce the output voltage. A Tektronix 2630
Fourier Analyzer receives the force transducer signal and the segment outputs. A frequency range
of 0-3 kHz is chosen. Above this range, the experiment rig did not give very clean resonance
peaks in the FRFs.
The experiment in the 2-3kHz range was done with swept-sine excitation for two reasons: 1) to
ensure clean FRFs and 2) to prevent aliasing. The spatial resolution of the segmented sensor is
finite, which means that higher mode vibrations can be sensed falsely as lower mode vibrations. Of
course, limiting the excitation frequency below the upper limit of the calculation frequency range
still excites the modes above the frequency range. However, it is expected that the modal residues
of the higher modes excited this way will be lower than the modal residues of the modes within
the frequency range. From Table 3.6, we see that eight modes can be studied in this range.
3.2.3 Experiment Results
The magnitudes of the FRFs from force to the outputs of the segments are shown in Fig. 3.13.
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
(V/N)
Segment 1
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
(V/N)
Segment 2
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
(V/N)
Segment 3
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
(V/N)
Segment 4
Frequency (Hz) Frequency (Hz)
Figure 3.13 Magnitude responses from force to segment outputs
= Analytical = Experimental
51
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 5
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 6
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 7
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 8
(V/N) (V/N)
(V/N) (V/N)
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 9
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 10
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 11
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 12
(V/N) (V/N)
(V/N) (V/N)
Frequency (Hz) Frequency (Hz)
52
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 13
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 14
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 15
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 16
(V/N) (V/N)
(V/N) (V/N)
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 17
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 18
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 19
Frequency (Hz)
0 1000 2000 3000
10
-2
10
-1
10
0
10
1
10
2
Segment 20
Frequency (Hz)
(V/N) (V/N)
(V/N) (V/N)
53
As we can see in Fig. 3.13, the experimentally obtained outputs of the individual segments are
very close to the predicted outputs. The modal damping ratios in the theoretical calculation (0.01
for all modes) were actually obtained by trial and error so that the anti-resonances of the FRFs
predicted by theoretical calculation agreed with those obtained by this experiment.
The discrepancies between the experiment results and the analytical prediction in the 2k-3kHz
range is caused partly by the difference in the actual beam natural frequencies from the predicted
natural frequencies. The prediction gave lower frequencies than the experimental beam. The thin
metal shims at the ends of the beam may have added rotational stiffness to the boundary
condition, violating the zero-moment assumption. The ninth resonance, which was captured by
the experiment, was not even predicted to be in the frequency range. Spatial aliasing, i.e.,
contribution from higher modes which are not accounted for by the theoretical prediction, may
have caused the extra resonance peak in segment 13.
A small area around the shaker point of action caused a near field that was not predicted by the
theory. This local effect added to the deviation of the segment FRFs from the ideal FRFs. The
force transducer was attached to the beam by bolting to a whole through the beam. The force
transducer and its attachment loaded the beam and added local stiffness, thereby altering the mode
shapes. The mode shapes with small wavelength were affected most. This was likely to be why the
experiment FRFs differ from the theoretical FRFs in the frequency range above 2 kHz.
Despite the discrepancies, each experimentally-obtained FRF from force to segment output
generally agrees well with the predicted FRF. The next step is to calculate the sensor gain matrix
W by Eq. (2.40). Because we use 20 segments, we use M = 20 in Eqs. (2.37) and (2.40) so that
we obtain a 20-by-20 gain matrix. However, we only managed to capture the first eight
resonances in the experiment. The higher mode resonances did not show a good 180
o
phase
rolloff. For this reason we only include the lowest eight modes in further calculation. In other
words, we only use the first eight rows of the gain matrix W and convert the outputs from 20
segments into eight modal coordinates. The first eight rows of W are shown graphically in Fig.
3.14.
3.2.4 Discussion on Experiment Results
The sensor outputs are calculated from the segment outputs using Eq. (2.32). The magnitudes of
the FRFs from force to sensor outputs are shown in Fig. 3.15.
Figure 3.15 shows that the sensor outputs resulting from transforming the segment outputs with
the theoretically calculated gain matrix W exhibit some modal filtering effects, especially for mode
4 and mode 8. The highest response peaks always occur at the resonance frequency
corresponding to the intended modes. The responses tend to be lower at frequencies away from
the intended resonance frequencies. However, for each mode there are several resonance peaks at
natural frequencies of other modes than the intended one.
54
-0.1
0
0.1
2
-0.05
0
0.05
3
-0.04
0
0.04
4
0.2
0.4
0.6
Mode
1
-0.02
0
0.02
5
-0.01
0
0.01
6
-0.01
0
0.01
7
5 10 15 20
-0.01
0
0.01
Segment number
8
Figure 3.14 Sensor gain matrix W for transforming 20 segment outputs into 8 modal coordinates.
An unconventional method to show modal filtering effect, used in what is likely to be the most
famous treatise of modal filtering theory is to plot the sensor responses in linear scale (Lee, 1987).
Replotting Fig. 3.15 with this method visually accents the modal filtering effects. In Fig. 3.16, it
appears more clearly that the sensor outputs do tend to emulate modal coordinates. For most of
the modes, the outputs of the linear combiners are close to the desired single-mode outputs. Most
of the mode sensors show responses in some agreement with the predicted responses. They are
very sensitive to the mode that they are designed to sense.
55
0 200 400
10
-2
10
-1
10
0
10
1
(V/N)
Mode 1
10
-2
10
-1
10
0
10
1
(V/N)
Mode 2
10
-2
10
-1
10
0
10
1
(V/N)
Mode 3
10
-2
10
-1
10
0
10
1
(V/N)
Mode 4
1500 1000 500 0
0 200 400 600
0 500 1000
10
-3
10
-2
10
-1
10
0
(V/N)
Mode 5
10
-3
10
-2
10
-1
10
0
(V/N)
Mode 7
10
-3
10
-2
10
-1
10
0
(V/N)
Mode 8
10
-3
10
-2
10
-1
10
0
(V/N)
Mode 6
0 1000 2000
0 1000 2000 3000
Frequency (Hz)
0 1000 2000 3000
Frequency (Hz)
0 1000 2000
Figure 3.15 Magnitude responses from force to sensor outputs.
56
0 200 400
0
5
10
15
(V/N)
Mode 1
0 200 400 600
0
5
10
15
20
(V/N)
Mode 2
0 500 1000 1500
0
5
10
(V/N)
Mode 4
0 500 1000
0
5
10
15
(V/N)
Mode 3
0 1000 2000
0
1
2
3
(V/N)
Mode 5
0 1000 2000
0
0.5
1
1.5
(V/N)
Mode 6
0 1000 2000 3000
0
2
4
6
8
(V/N)
Frequency (Hz)
Mode 7
0 1000 2000 3000
0
1
2
3
(V/N)
Frequency (Hz)
Mode 8
Figure 3.16 Magnitude responses from force to sensor outputs: linear scale.
57
Figure 3.17 End connection to approximate simple support.
The modal filters are also sensitive to other modes than the intended ones. The worst case is the
mode-1 sensor, which seems to fail to filter out other modes than mode-1. This imperfect modal
filtering effect can be attributed to the following inaccuracies:
The segments were cut and attached by hand. They did not have perfect geometry and were not
attached in the perfect positions. Lee (1990) showed that even microscopic imperfection in a
shaped film sensor resulted in rather significant imperfection in modal filtering effects.
The boundary conditions of the beam were not ideal simple supports. The beam was attached to a
frame with thin sheet metal pieces attached to ends of the beam with small screws. This
connection, shown in Fig. 3.17, does not give zero resistance to bending. The mode shapes of the
beam were not exactly as predicted by the theory.
The point force causes evanescent (near-field) modes around the point of force application. These
modes were never accounted for in the theory, but may actually contribute more than negligibly to
the vibration of the beam. Additionally, the force transducer was attached to the beam by drilling
a hole through the beam and bolting the force transducer to the hole. This attachment creates
58
extra stiffness around the hole and some mass loading on the beam. The actual modes of the
experimental beam are different from the theoretically predicted modes.
Many other factors may have contributed to the imperfect modal filtering effects. However, in
general we can conclude that the theory of modal filtering with segments of piezoelectric film has
been somewhat supported by the simple experiment. We can also conclude that any imperfection
in the knowledge of the actual mode shapes and in the construction of the sensor may result in
severe leaking of the modal filters.
A method to improve this modal filtering technique in is to incorporate adaptive digital signal
processing algorithm to adjust the segment weights automatically based on the sensed
imperfections in the modal filtering effects. This method of improvement is the main thrust of the
remainder of this dissertation.
3.3 Chapter Summary
This chapter established confidence in the models and simulation procedures developed in
chapter 2. Numerical and experimental verification of the models and procedures showed the
following:
1. Simulations both in time and frequency domains showed that the parallel second-order digital
filter model agrees with the dynamic model of the structure.
2. Frequency-domain responses obtained by the experiment have verified:1) The digital filter
model of the beam. 2)The model for the responses of all sensor segments.
3. Transformation of experimental segment outputs by theoretical gain matrix showed that the
ideal gain matrix transforms the segment outputs into modal coordinates, albeit imperfectly.
Unintended modes are not completely filtered out by the modal filters.
4. Spatial aliasing is a problem that must be avoided in using the segmented sensor.
CHAPTER 4
ADAPTIVE COMPUTATION OF GAIN MATRICES
In the last chapter we showed a method to design a modal filter that separated the vibration signal
of a structure into its modal coordinates. Figure 2.7 shows the configuration of the sensor. In
particular, it shows that the outputs of the sensor segments are multiplied by a gain matrix to
produce the modal coordinates in real time. Equation (2.40) shows that the gain matrix is the
inverse of the segment output matrix, which in turn is calculated from the mode shapes of the
structure.
The analytical methods we have presented so far for computing segment weights for modal
sensors rely on precise knowledge of the mode shapes of the structures. The dependence of
precise knowledge of mode shapes has some drawbacks. First, if the actual mode shapes of the
structure are not exactly the same as the mode shapes used in the calculation of the gain matrix,
the modal filters will fail to cleanly filter out the modes that they should filter out. Second, it is
impossible to build a modal sensor without doing a thorough testing and analysis of the modal
properties of the structures. Mode shapes of real structures are usually obtained after much
computation, e.g., finite element methods or modal analysis. Calculation of gain matrices of the
modal sensors must be done after the calculation of the mode shapes.
Sometimes the following situations preclude the use of currently available techniques to design
modal sensors:
1. The modal properties (natural frequencies, modal damping ratios, and modal residues) are not
known precisely.
2. The global modal properties (natural frequencies, modal damping ratios) are known, but local
properties (modal participation factors) are not known. This means that we do not know the
mode shapes of the structures. As mentioned above, mode shapes are necessary in the
computation of the gain matrix for the modal filters.
3. Very limited modal information of the structure is available. For example, we only know the
number of modes in the frequency range of interest, but we dont know the natural
frequencies, modal damping ratios, and mode shapes of the structure.
In this chapter we shall discuss
1. How the design of modal filters is affected by incompleteness of knowledge of the modal
properties.
2. Techniques to compensate for the inaccuracies of the modal sensors
3. Methods to extract modal properties from the structure using the segmented sensor
60
4.1 Effects of Inaccurate Mode Shapes
In this section we shall use only a mode-3 sensor as an example. This sensor is designed to be
sensitive only to the 3
rd
vibration mode, and filter out signals from all other modes.
We showed in the previous chapter that the output of the mode-3 sensor consists only of the third
modal coordinate of the structural vibration. When the structure parameters change from their
original design values, the FRF from force to the output of the segmented sensor fails to exhibit
a Single-Degree-of-Freedom (SDOF) response as the sensor is designed to perform. In this
section we discuss how deviations from the design modal properties affect the performance of the
modal sensors.
Consider the response of a segmented sensor designed to sense the third modal coordinate of the
simply supported beam with torsion spring. Figure 4.1 shows the magnitudes of the FRFs from
force to sensor output. The sensor is designed for a the beam in Section 3.1 with torsion spring
stiffness T* = 1. However, if the structures actual torsion spring stiffness is T* = 10, the sensor
response deviates from the ideal SDOF response. Specifically, the differences between the two
responses are:
1. The resonance frequency of the T* = 10 structure is higher than that of the T* = 1 structure.
2. Resonance peaks appear in the T* = 10 response at the resonance frequencies of the modes
other than mode 3. These modes should be filtered out by the mode-3 sensor.
0 200 400 600 800 1000 1200 1400 1600 1800
-110
-100
-90
-80
-70
-60
-50
-40
-30
-20
Frequency, Hz
Mag.
0 dB =
1 V/N
FRF from force to sensor output: magnitude
---- T* = 1
___ T* = 10
Figure 4.1 Magnitudes of the responses of the sensor on the T*= 1 structure and on the T* =10
structure
61
Figure 4.2 shows the phases of the FRFs. The T* = 1 response phase rolls off from 90
0
to -90
0
in
the region around the third resonance frequency. At frequencies below this resonance frequency
region, the phase is 90
0
; and at frequencies higher than the resonance frequency region the phase
is -90
0
. The phase of the T* = 10 response rolls off at a higher frequency than the third resonance
frequency of the T* = 1 structure. At the resonance frequencies of other modes, the phase is
neither 90
0
nor -90
0
.
0 200 400 600 800 1000 1200 1400 1600 1800
-200
-150
-100
-50
0
50
100
150
200
Frequency, Hz
Phase,
deg
FRF from force to sensor output: phase
---- T* = 1
___ T* = 10
Figure 4.2 Phases of the responses of the sensor on the T* = 1 structure and on the T* = 10
structure
Figure 4.3 shows the real part of the FRFs. The real part of the T* = 1 FRF has a peak in the
region around the third resonance frequency. At frequencies below and above this resonance
frequency region, the real part is zero. The real part of the T* = 10 response, on the other hand,
fails to stay at zero outside the resonance frequency region.
62
0 200 400 600 800 1000 1200 1400 1600 1800
-0.01
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
Frequency, Hz
V/N
FRF from force to sensor output: real part
---- T*
___ 10 T*
Figure 4.3 Real parts of the responses of the sensor on the T* = 1 structure and on the T* = 10
structure.
4.2 Adaptive Design of Modal Sensors
The main thrust of this research is the development of a method to design a modal sensor for a
continuous structure without much prior knowledge of the modal properties of the structure.
Designing a modal sensor for a continuous structure means calculating the sensor gain matrix W.
In this section, we shall develop a such a method using adaptive signal processing techniques.
4.2.1 Correlation Between Modal Coordinates
The method we will develop is based on the correlations between modal coordinates. The first
step in developing the design method is to investigate these correlations. Figure 4.4 shows two
second order filters with different natural frequencies. A random input is common to both filters.
We shall investigate the correlations between the two outputs.
Let us call the outputs of the modal filters
&
( )
i
k and
&
( )
j
k , and assume that these are modal
coordinates. Let us define the cross correlation function between the two modal coordinates at
zero lag as
[ ]
R
ij i j
k k E
&
( )
&
( ) , (4.1)
63
Figure 4.4 Modal responses to random excitation.
where E[.] denotes expectation. Strictly speaking, the expectation should be obtained by ensemble
averaging. However, we assume that the modal coordinates are stationary and ergodic. Therefore,
we can replace the ensemble averaging with time averaging, and we can rewrite the last equation
as
R
ij
k
i j
h
k
k
h h

1
1
lim
&
( )
&
( ) . (4.2)
In the implementation, we can only approximate the expected value by choosing a large but finite
number for k and possibly checking for convergence. To avoid storage of all past records, we use
a recursive formula for the above approximation as
{ }
R R
ij ij i j
k
k
k k k k ( ) ( ) ( )
&
( )
&
( ) +
1
1 1 , (4.3)
where the index k denotes the time step in which R
ij
is computed. This index notation is different
from the usual notation for correlation functions, where the index k would normally show the
delay between
&
( )
i
k and
&
( )
j
k .
If
&
( )
i
k and
&
( )
j
k are modal coordinates, then they must be orthogonal to each other, therefore,
uncorrelated. Thus, the correlation between the two modal coordinates must be zero.
b z z
a z a z
i
i i
1
1 2
1
1
2
2
1
,
, ,
( )

+ +
&
( )
( )

i
f
&
( )
( )

j
f
b z z
a z a z
j
j j
1
1 2
1
1
2
2
1
,
, ,
( )

+ +
&
( )
i
k
&
( )
j
k
f(k)
k
k
k
64
lim ( ) ;
k
ij
k i j
R 0 . (4.4)
This principle will be used to develop a method to separate the structural response into its modal
coordinates.
Simulation of the filters in Fig. 4.4 with a Gaussian zero-mean random input results in signals in
Fig. 4.5 for modes 1 and 3. The product of the two outputs
&
( )
1
k and
&
( )
3
k is averaged over
time to estimate the zero-lag cross correlation function, R
13
( ) k . As expected, the computed
zero-lag cross correlation between the two modal coordinates approaches zero as more and more
samples are included in the computation.
0
5x 10
-3
0
2x 10
-5
0 0.5 1 1.5
0
1 x 10
-7
Time (s)
-0.01
0
0.01
Avg
[ & & ]
1 3
& &

1 3
&
1
&
3
-1 x 10
-7
-2x 10
-5
-5x 10
-3
Figure 4.5 Mode-1 coordinate, mode-3 coordinate, product of mode-1 and mode-3 coordinates,
average of product of mode-1 and mode-3 coordinates.
65
4.2.2 Adaptive Computation of Sensor Gain Matrix
We can compute the zero-lag correlations between all modes of a structure, and define a matrix of
correlations among sensor outputs at zero lag as
R
R R R
R
R R
1
]
1
1
1
1
11 12 1
12
1
L
M
M
L
M
M MM
, (4.5)
where the components R
ij
, i = 1, , M, j = 1, , M, are defined as in Eq. (4.2), and computed
as in Eq. (4.3). M is the number of modes included in calculations. Since the correlations between
any different modes is zero, R must be diagonal.
Now consider the system in Fig. 4.6. The correlations between the sensor outputs
&
$
m
(m = 1, ,
M) will form a diagonal matrix only if the sensor outputs are proportional to modal coordinates.
This condition will be achieved only if the gain matrix W has the correct components. If the gain
matrix does not have the correct component, the correlations between the sensor outputs will not
form a diagonal matrix. The foregoing notion can serve as a basis for developing algorithms to
adjust the sensor gain matrix W iteratively until the correlation matrix is diagonal. The
development of such algorithms will be explained as follows.
Define the correlation between sensor outputs as
]
R
'
1
]
1
1
1
E
&
$
&
$
&
$
&
$

1
1
M L
M
M
(4.6)
where M is the number of modes in the frequency range of interest. Substituting Eq. (2.32) into
Eq. (4.6), we can express the correlation matrix as
[ ]
R WVV W E
T T
. (4.7)
66
Figure 4.6 Adjusting sensor gain matrix to diagonalize correlation matrix.
In the adaptive system, the sensor gain matrix will be adjusted from time to time. However, the
sensor gain matrix W that transforms E[VV
T
] into modal coordinates is a constant matrix.
Therefore, for the purpose of obtaining W we can simplify Eq. (4.7) into
[ ]
R W VV W E
T T
. (4.8)
V
n
V
N
V
1
Beam
PVDF
segments
f(k)
Adaptive gain
adjuster
&
$
1
&
$
m
&
$
M
W
1,1
W
m,N
W
m,1
W
M,N
W
1,n
W
m,n
W
M,n
W
M,1
W
1,N
67
Next, we define the expected value term in the above equation as
[ ]
S VV E
T
. (4.9)
If S is diagonalizable, with an eigenvector matrix
[ ]
= =
1
L
M
, (4.10)
where
m
is the m
th
eigenvector of S, then the matrix
1 1
S is diagonal. Moreover, is an
eigenvector matrix, so
= =
1 T
. (4.11)
Thus, to diagonalize the sensor output correlation matrix R in Eq. (4.8) we can simply set the
gain matrix to
W
T
. (4.12)
4.3 Numerical Example
The above equation is a very important conclusion in this research. The equation states that: if we
set the sensor gain matrix to the transpose of the eigenvector matrix of the expected values of the
products between segment outputs, then the outputs of the sensor will be orthogonal to each
other.
To verify the above theory, we simulate the system in Fig. 4.6 with a computation program shown
in Fig. 4.7. The physical properties of the system are as in Subsection 3.1.1. The segment output
correlation matrix S in Eq. (4.9) is approximated with a recursive formula similar to Eq. (4.3).
The sensor gain matrix W is the transpose of the eigenvector matrix of S. As the estimate for S
becomes closer to the true expected values of the correlation between the segment outputs, the
sensor gain matrix W in Eq. (4.12) becomes closer to a gain matrix that results in orthogonal
sensor outputs. This means that the sensor output correlation matrix in Eq. (4.6) becomes closer
to a diagonal matrix. The orthogonal sensor outputs will be proportional to the modal
coordinates.
We run the above simulation with the beam example with a Gaussian random input force and
obtain the following results. At the outset, k = 16, the sensor gain matrix W (let us call this gain
matrix W(16)) is as shown in Fig. 4.8. Each plot shows the gains for one mode. This early guess
68
of sensor gain matrix W does not really resemble the ideal sensor gain matrix in Fig. 4.9 (which is
a rescaled version of Fig. 3.6.)
Figure 4.7 Sensor gain matrix adjustment using eigenvector matrix of segment output
correlations.
0.2
0.4
-0.6
0
0.2
-0.4
0
0.4
-0.6
0
0.2
-0.5
0
0.5
-0.4
0
0.2
-0.4
0
0.4
-0.4
0
0.4
-0.2
0
0.4
-0.4
0
0.4
Mode
1
2
3
4
5
6
7
8
2 4 6 8 10
9
10
Segment 1 3 5 7 9 2 4 6 8 10 3 5 7 9 1
Figure 4.8 Sensor gain matrix computed with 16 sets of time data, W(16).
Segment
output
matrix
C
S
Beam
dynamics
f(k)
Modal
coordinates
(no access)
Segment
outputs V
{ }
S S VV ( ) ( ) ( ) k
k
k k +
1
1 1
T
W ( ) Eigenvector matrixof
T
S
Adjustable
gain W
&
$
( ) k
Sensor
outputs
69
10
Mode 1
-2
0
2
2
-1
0
1
3
-0.5
0
0.5
4
-0.2
0
0.4
5
-0.2
0
0.2
6
-0.2
0
0.2
7
-0.2
0
0.2
8
2 4 6 8 10
-0.1
0
0.2
9
-0.1
0
0.05
10
Segment 1 3 5 7 9 2 4 6 8 10 3 5 7 9 1
0
Figure 4.9 Ideal sensor gain matrix.
With a sensor gain matrix that is so different from the ideal, we cannot expect the sensor outputs
to be uncorrelated. To verify this statement, we run another simulation with 8192 points using the
gain matrix in Fig. 4.8. Then we compute the correlation matrix of the resulting sensor outputs.
The correlation matrix is shown in Fig. 4.10.a. in a picture format suitable for showing symmetric
matrices with large diagonal elements. The height of each bar in this figure shows the value of the
corresponding component of the correlation matrix. Since the (10,10) component of the
correlation matrix is much greater than the other components, the scaling of the heights obscure
the depiction of the other components. Rescaling the picture as in Fig. 4.10.b gives a better idea
of the relative values of the matrix components. Notice that the sensor output correlation matrix is
far from being diagonal. This means that the sensor gain matrix results in sensor outputs that are
not uncorrelated.
At time step k = 256, the sensor gain matrix is as shown in Fig. 4.11. It appears that the segment
gains are starting to evolve towards the ideal gain matrix (Fig. 4.9). However, with this segment
gain matrix, the output correlation matrix is still not diagonal (See Fig. 4.12.)
A question that naturally arises from the above examples is: How many time steps should be used
to estimate the expected value of segment outputs (Eq. (4.9)) to obtain a sensor gain matrix (Eq.
(4.12)) that will uncorrelate the sensor outputs? To answer this question, we need to go back to
Eq. (4.1) and determine the number of time data points that will make Eq. (4.2) a reasonable
70
approximation to Eq. (4.1). To this end, we can check a few estimates for the correlation between
different modal coordinates. We know that the correlation should be close to zero. If we examine
Fig. 4.5 again, we can observe that the estimate for the correlation between
&
( )
1
k and
&
( )
3
k
settles around zero after about 30 000 time data points. For k > 30000, this estimate is almost
unaffected by large fluctuations in
&
( )
1
k or
&
( )
3
k . We also know that
&
( )
1
k is the most slowly
varying modal coordinate of the system. Therefore, if the product of
&
( )
1
k and another modal
coordinate settles around zero, evidently other products of modal coordinates also settle around
their expected values. Based on the above observation, we compute the sensor gain matrix W
using 32768 time-data points. The resulting sensor gain matrix is shown in Fig. 4.13. Notice that
each row of this sensor gain matrix is very nearly proportional to the corresponding row in the
ideal sensor gain matrix shown in Fig. 4.9. Figure 4.14 shows that correlation between the sensor
outputs for this sensor gain matrix is almost diagonal.
0
10 0
10
-0.5
0.5
1
2
2.5
x 10
-6
M
o
d
e
S
e
n
s
o
r

R

(
V
^
2
)
Using W(16)
a)
Figure 4.10.a Sensor output correlation matrix using W(16).
(Drawn to scale)
71
0
5
10 0
5
10
-5
0
5
x 10
-7
Mode
S
e
n
s
o
r

R

(
V
^
2
)
Using W(16)
scale)
(Out of
3e-006
b)
Figure 4.10.b Sensor output correlation matrix using W(16).
(Not to scale)
0.2
0.4
-0.4
0
0.4
-0.4
0
0.2
-0.5
0
0.5
-0.4
0
0.4
-0.4
0
0.4
-0.5
0
0.5
-0.6
0
0.4
-0.4
0
0.4
-0.4
0
0.4
Mode
1
2
3
4
5
6
7
8
2 4 6 8 10
9
10
Segment 1 3 5 7 9 2 4 6 8 10 3 5 7 9 1
Figure 4.11 Sensor gain matrix W(256).
0
5
10 0
5
10
0
2
4
6
8
x 10
-7
Mode
S
e
n
s
o
r

R

(
V
^
2
)
k = 256
scale)
(Out of
1.11e-006
Figure 4.12 Sensor output correlation matrix resulting from W(256).
0.2
0.4
-0.4
0
0.4
-0.4
0
0.4
-0.4
0
0.4
-0.2
0
0.2
-0.4
0
0.4
-0.2
0
0.4
-0.4
0
0.4
-0.2
0
0.4
-0.2
0
0.2
Mode
1
2
3
4
5 6
7 8
2 4 6 8 10
9
10
Segment 1 3 5 7 9 2 4 6 8 10 3 5 7 9 1
Figure 4.13 Sensor gain matrix calculated using 32768 time data points.
73
0
5
10 0
5
10
0
2
4
6
8
10
x 10
-7
Mode
S
e
n
s
o
r

R

(
V
^
2
)
Using W(32768)
scale)
(Out of
3.29e-006
If we continue using more and more time points, we may obtain better estimate for the segment
output correlation matrix and a better sensor gain matrix. However, the improvements may be
marginal because the estimates for the correlations have settled around their theoretical values.
For example, the sensor gain matrix obtained using 49152 time data points, shown in Fig. 4.15, is
very close to the sensor gain matrix obtained using 32768 points, shown in Fig. 4.13. The sensor
output correlation matrix obtained with 49152 time data points (Fig. 4.16) is also very similar to
the corresponding matrix obtained with 32768 points (Fig. 4.14). Evidently, the correlations have
settled around their theoretical values.
-0.4
-0.2
-0.4
0
0.4
-0.4
0
0.4
-0.5
0
0.5
-0.4
0
0.4
-0.4
0
0.4
-0.2
0
0.4
-0.4
0
0.4
-0.4
0
0.4
-0.2
0
0.2
Mode
1
2
3
4
5 6
7 8
2 4 6 8 10
9
10
Segment 1 3 5 7 9 2 4 6 8 10 3 5 7 9 1
Figure 4.15 Sensor gain matrix calculated using 49152 time data points.
0
5
10 0
5
10
0
2
4
6
8
10
x 10
-7
Mode
S
e
n
s
o
r

R

(
V
^
2
)
Using W(49152)
scale)
(Out of
3.29e-006
75
Each row of the gain matrix represents a sensor that is supposed to sense only one particular
modal coordinate and filter out other modal coordinates. The rows of the gain matrix obtained
with the adaptive technique using 32768 time data points are not exactly proportional to the
corresponding rows in the ideal gain matrix that we would obtain analytically if we knew the
mode shapes of the structure. Therefore, the sensor outputs obtained with the adaptive sensor in
Fig. 4.7 may not be exactly proportional to the ideal modal coordinates. To compare the sensor
outputs to the ideal modal coordinates, we obtain the frequency response functions
(
&
$
( ) ( )
m
f , m = 1, , 10) of the system in Fig. 4.7, and compare them with the frequency
response functions (
&
( ) ( )
m
f , m = 1, , 10) of the same system with the ideal gain matrix.
The comparison is shown in Fig. 4.17.
Around the natural frequencies, the adaptive sensor outputs are nearly proportional to the modal
coordinates. Away from the natural frequencies some of the adaptive sensor outputs are
corrupted by signals from modes that the sensor should not be sensitive to. The imperfect filtering
of other modes means that the sensor gain matrix still needs some refinement. A method to refine
the sensor gain matrix will be presented in the next chapter. The important statement at this point
is that the above numerical simulation and frequency-domain verification show that the sensor
gain matrix computation scheme described in this chapter converges closely to the ideal sensor
gain matrix. Thus, we have established the basis for a new method to compute the sensor gain
matrix of a modal sensor without knowledge of the mode shapes of the structure.
Being in its conception, this method still needs much more development before it can be applied
as a standard modal filtering technique. In the next chapter we will discuss the problems that must
be solved to implement this method.
4.4 Chapter Summary
As mentioned in Chapter 1, most methods for designing modal sensors for structures rely on
precise knowledge of the modal properties of the structures, especially the mode shapes. Modal
sensors designed using inaccurate mode shape functions may not filter out undesired modes which
the sensors are not supposed to sense. Some of the currently available methods may enables us to
design a modal sensor without knowledge of the mode shapes. However, these methods require
that we know the natural frequencies and modal damping of the structure. In this dissertation we
attempt to develop another method to design a modal sensor without even knowing the above
modal properties of the structure.
In this chapter we established the basis for such a method. This method is based on the fact that
modal coordinates of a vibrating structure are uncorrelated to each other. In our case study, the
sensor is composed of strain sensing segments. We produce the uncorrelated outputs by
transforming the segment outputs with a correct sensor gain matrix. We obtain this gain matrix
from the estimated expected values of the correlation between the segment outputs.
76
0 200 400 600
0
1
Mode 1
0 200 400 600
0
1
Mode 2
0 500 1000
0
1
Mode 3
0 500 1000
0
1
Mode 4
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
0 500 1000 1500
0
1
Mode 5
0 1000 2000
1
Mode 6
Frequency (Hz)
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
Frequency (Hz)
Figure 4.17 Magnitudes of adaptive sensor outputs (solid lines) and ideal modal coordinates
(dotted lines).
77
0 1000 2000 3000
0.2
0.4
0.6
0.8
1
Mode 7
0 1000 2000 3000
0.2
0.4
0.6
0.8
1
Mode 8
0 2000 4000
0.2
0.4
0.6
0.8
1
Mode 9
Frequency (Hz)
0 2000 4000
0.2
0.4
0.6
0.8
1
Mode 10
Frequency (Hz)
Figure 4.17 Magnitudes of adaptive sensor outputs (solid lines) and ideal modal coordinates
(dotted lines).
We showed that the sensor outputs produced by the transformation with this sensor gain matrix
are, indeed, uncorrelated to each other. We also showed that each row of the sensor gain matrix is
proportional to the corresponding row in the ideal sensor gain matrix that would result from
analytical calculation using known mode shape functions.
The sensor gain matrix that will produce modal coordinates can be shown to be equal to the
eigenvector matrix of the segment output correlation matrix. The segment output correlation
matrix is defined as a matrix that contains the correlations between the segment outputs at zero
time lag. This segment output correlation matrix can be obtained on-line by processing the
outputs of the sensor segments when the structure is excited with a persistent excitation such as
Gaussian random force. These segment outputs are multiplied by each other. The products are
averaged over time to estimate the expected values of the products. The estimation improves with
78
time as more data points are used to estimate the correlations. The resulting sensor gain matrix
converges over time to the ideal sensor gain matrix.
The computation of the sensor gain matrix involves the computation of the eigenvector matrix of
the segment output correlation matrix. In the application, the eigenvector computation must be
done iteratively. However, the eigenvectors do not have to be updated frequently, since changes
in the eigen-properties of the structure are generally very slow.
The FRFs from force to sensor outputs are generally proportional to the FRFs from force to
modal coordinates. Away from natural frequencies, the sensor outputs may contain a little
contribution from unwanted modes. To eliminate this sensitivity to unwanted modes, the adaptive
design procedure described in this chapter still needs refinement.
CHAPTER 5
EIGENVECTOR COMPUTING ALGORITHMS
In Sec. 4.2, we learned that the real time calculation of the gain matrix that transforms sensor
outputs into modal coordinates involves cross multiplying the segment outputs, averaging the
resulting matrix, and computing the eigenvector matrix of the averaged matrix. The eigenvector
matrix computation in section 4.3 was done in batches. In practice, computation of eigenvector
matrices may be so extensive that it holds up other operations. A more efficient eigenvector
computation strategy would involve recursive computation concurrent with other operations.
In this chapter we will discuss a method to compute eigenvector matrices, also called modal
matrices, recursively using a version of Jacobi rotation technique. We will present a brief review
of the Jacobi rotation algorithm including numerical examples, and develop special algorithms for
solving our eigenvector problem.
5.1 Jacobi Rotation Algorithm
Jacobi rotation, also known as Givens rotation, is a very popular eigenvector computing algorithm
in digital signal processing because it is simple and stable in general. It is even simpler and more
robust in processing symmetric real matrices (Haykin, 1991). It has also been used extensively for
computing eigenvectors in structural dynamic finite element problems (for example, Bradbary and
Fletcher, 1968).
In Sec. 4.3, we showed that the matrix that transforms segment outputs into modal coordinates is
the matrix that diagonalizes the segment output correlation matrix. Most square matrices can be
diagonalized by similarity transformation. If S is a diagonalizable matrix, then its eigenvector
matrix is a transformation matrix such that the result of similarity transformation
Q S
T
(5.1)
is diagonal. Jacobi rotation is a method to compute the transformation matrix, . Basically, this
method works by transforming the matrix to another matrix that has a zero component in selected
positions. The derivation of this method can be found in many linear algebra or numerical analysis
textbooks such as Press et al. (1986). Without derivation, the procedure can be described as
follows.
First, initialize Q(1) = S, the matrix to be diagonalized. Also initialize the matrix ,
80
( ) 1 I . (5.2)
1. At time step h, choose an off-diagonal element of matrix Q(h) to be set to zero. Call this
element Q
pq
(h), where p is the row position of the element and q is the column position. In the
cyclic version of Jacobi rotation algorithms, all the off-diagonal elements in the upper
triangular half of the matrix are chosen systematically row by row. For example, the (1,2)
element is chosen for the first loop. Then the (1,3) element is chosen, and so on. The last loop
chooses the (N-1,N) element of the N-by-N matrix. This systematic selection process is called
sweeping.
2. Compute the rotation angle
( ) tan
( )
( ) ( )
( ) ( ) h
h
h h
h h
pq
qq pp
qq qq
_
,
1
2
2
1
Q
Q Q
Q Q if , (5.3)
Or (h) = 90
0
if Q
qq
(h) = Q
pp
(h).
3. Assemble the Jacobi rotation matrix, which is like an identity matrix except for the elements
indicated below:
P( )
cos( ( )) sin( ( ))
sin( ( )) cos( ( ))
h
h h
h h
p
q
p q
1
]
1
1
1
1
1
1
1
1
1

1
1
1
O
O

th
th
th th
row
row
column column
(5.4)
4. Apply similarity transformation to Q(h) according to
Q P Q P ( ) ( ) ( ) ( ) h h h h + 1
T
(5.5)
5. Update the estimate for the modal matrix, (h), by post-multiplying it by the Jacobi rotation
matrix P(h).
( ) ( ) ( ) h h h + 1 P . (5.6)
81
6. Repeat steps 2 through 5 until a suitable diagonality criterion is satisfied by Q.
Notice that step 5 means that the eigenvector matrix is the product of all the Jacobi rotation
matrices obtained at each iteration, that is,
P P P ( ) ( ) ( ) 1 2 3 L (5.7)
Step 4 in each iteration sets the (p,q) element of matrix Q to zero. In the next iteration, where the
indices p and q are changed, the element that was set to zero in the previous iteration will again be
nonzero. Nevertheless, the matrix Q will become progressively more diagonal with each iteration.
From the analytic geometry point of view, this transformation is a rotational coordinate
transformation in a plane, hence the name Jacobi rotation.
To present a numerical example, we would like to obtain the gain matrix that will transform the
following matrix into a diagonal matrix.
S Q
1
]
1
1
1
1
( )
. . .
. . .
. . . .
. .
1
38 21 0 45 2
21 4 5 054 1
045 054 2 3 0 3
2 1 0 3 2 6
The first off-diagonal element to annihilate is the (1,2) element of Q, that is, 2.1. For this element,
Eq. (5.3) gives (1) = -0.703; and the Jacobi rotation matrix according to Eq. (5.4) is
P( )
. .
. .
1
0 763 0 646 0 0
0 646 0763 0 0
0 0 1 0
0 0 0 1
1
]
1
1
1
1
The first iteration transforms the matrix into
Q P Q P ( ) ( ) ( ) ( )
. .
. .
. . . .
. . . .
2 1 1 1
2 0 0 006 088
0 6 0 7 2 06
0006 0 7 2 3 0 30
088 2 06 030 2 6

1
]
1
1
1
1
T
This transformation annihilated the (1,2) element of Q. The next transformation annihilates the
(1,3) element, resulting in
82
Q P Q P ( ) ( ) ( ) ( )
. .
. . . .
. . .
. . . .
3 2 2 2
2 0 01 0 089
001 6 3 0 7 2 06
0 0 7 2 3 0 30
089 2 06 0 30 2 6

1
]
1
1
1
1
T
This transformation annihilates the (1,3) element, but undid the annihilation of the (1,2) element
of the previous transformation. However, the resulting matrix is relatively more diagonal than
the previous matrix. The original matrix and the results of the first five transformations are
illustrated in Fig. 5.1. We can see that the matrix becomes progressively more diagonal with each
transformation.
The above 4-by-4 matrix has 6 off-diagonal elements to annihilate ((p,q) = (1,2), (1,3), (1,4),
(2,3), (2,4), and (3,4)). After one sweep (which means six rotations to annihilate the six upper off-
diagonal elements), the matrix is almost diagonal, as shown in Fig. 5.2.
The above sweep can be repeated indefinitely. However, Fig. 5.3 shows that the matrix is
practically diagonal. The second sweep result is
Q
( )
. . .
. .
. .
. .
second sweep

1
]
1
1
1
1
102 0028 0 014 0
0028 7 34 0 0
0 014 0 218 0
0 0 0 2 66
Further sweeps will diagonalize the matrix even further, and eventually only maintain the
diagonality. The eigenvector matrix is obtained by cumulatively multiplying the Jacobi rotation
matrices according to Eq. (5.7).
Notice that as the matrix resulting from rotations becomes more and more diagonal, the rotation
angle in Eq. (5.3) becomes closer and closer to zero, and the Jacobi rotation matrix in Eq. (5.4)
becomes closer and closer to identity. Eventually, multiplying the estimate of the eigenvector
matrix with more Jacobi rotation matrices does not change the estimate significantly. At this stage
the estimate has converged to the true eigenvector matrix.
83
0
1
2
3
4 0
1
2
3
4
0
1
2
3
4
Original

0
1
2
3
4 0
1
2
3
4
0
1
2
3
4
5
6
Sweep 1; Rotation #1
0
1
2
3
4 0
1
2
3
4
0
1
2
3
4
5
6

0
1
2
3
4 0
1
2
3
4
0
2
4
6
0
1
2
3
4 0
1
2
3
4
0
2
4
6

0
1
2
3
4 0
1
2
3
4
0
2
4
6
Figure 5.1 Jacobi rotation example.
84
0
1
2
3
4 0
1
2
3
4
0
2
4
6
Figure 5.2 Result of first sweep.
0
1
2
3
4 0
1
2
3
4
0
2
4
6
Figure 5.3 Result of second sweep.
85
The above example showed the effectiveness and simplicity of the Jacobi rotation algorithm.
Equation (5.7) shows that cumulative multiplication of the Jacobi rotations improves the estimate
of the modal matrix progressively. Equation (5.4) shows that multiplication with a Jacobi rotation
matrix effectively changes only two rows and two columns of the previous estimate of the modal
matrix. The transformed matrix in Eq. (5.4) becomes more and more diagonal with each rotation.
The rotation angle becomes closer and closer to zero with each rotation. Therefore, the Jacobi
rotation matrix P in Eq. (5.4) becomes closer and closer to identity. Eventually after sufficient
rotations the rotation matrix will just oscillate around identity, effectively maintaining the
diagonality of the rotated matrix Q.
5.2 Algorithm A, Convergence and Rotation Angle
The gain matrix calculation procedure shown in Fig. 4.7 can be done using any eigenvector
computation algorithm. The modal matrix computation in the simulation examples in section 4.3
was done with the Jacobi rotation algorithm, sweeping the upper off-diagonal elements of the
segment output correlation matrix S four times. For a ten-mode filter, each sweep annihilates 45
off-diagonal elements, with one Jacobi rotation for each element. For the purpose of discussion
and comparison, we refer to this algorithm as Algorithm A.
The segment output correlation matrix is estimated by recursive averaging. The average S(k) is
improved with each time step k. Each time a new average is obtained, its modal matrix is
computed with Eqs. (5.2) through (5.6). Figure 5.4 shows the connection between the segments
and the signal processing.
The rotation angle in Eq. (5.3) is a good indication of how much the modal matrix estimate will
change with each rotation. A small rotation angle means that the Jacobi rotation matrix in Eq.
(5.4) is close to the identity matrix, therefore, a small rotation angle does not change the modal
matrix estimate significantly. Figure 5.5 shows the first 8 time steps of the algorithm used in Sec.
4.3. Each time step includes four sweeps, which mean 4*45 = 180 Jacobi rotations. There is a
strong tendency that within each time step the rotation angle becomes smaller. However, each
new time step finds a new correlation matrix, discards the modal matrix, and starts forming a new
modal matrix by cumulatively multiplying the Jacobi rotation matrix all over again. The initial
guess is the same for each time step. Improvement of the modal matrix (and hence the sensor gain
matrix) is achieved only because the estimate of the segment output correlation matrix S improves
with time steps. In other words, W
k
is better than W
k-1
only if S
k
is better than S
k-1
.
86
Figure 5.4 Algorithm A.
.
.
.
.
.
.
W
1,1
W
m,N
W
m,1
W
M,N
W
1,n
W
m,n
W
M,n
W
M,1
W
1,N
V
n
V
N
V
1
Beam
PVDF
segments
f
Mode-m
filter
&
$
1
Mode-1
coordinate
( )
{ }
S S V V ( ) ( ) ( ) ( ) k
k
k k k k +
1
1 1
T
Jacobi rotations to obtain
(k) = eigenvector matrix of S(k)
W(k) =
T
(k)
Mode-m
coordinate
&
$
m
Mode-M
coordinate
&
$
M
. . .
. . .
87
Jacobi rotations in Algorithm A are merely a tool to solve for the eigenvectors of the segment
output correlation matrix S in each time step. For each S(k), the eigenvector estimate does not
improve beyond the four sweeps of Jacobi rotations. Thus, the rotation angle does not necessarily
become smaller with time step k. Because of the large number of operations per time step and the
little improvement in the eigenvectors with time steps, Algorithm A is of little practical utility. We
discussed it only to facilitate the understanding of the development of the two new algorithms that
we will discuss in the next sections.
0 200 400 600 800 1000 1200 1400
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Rotation angle
Rotation number
deg
Figure 5.5 Typical rotation angle history of Algorithm A.
5.3 Algorithm B
5.3.1 Development
A much more economical method for computing the modal matrix will be devised below. The
objective of the new algorithm is to retain the recent update for the modal matrix of S (the
estimate for the segment output correlation matrix) and to track the slow changes in S. If the
excitation force is stationary and if the structure does not change significantly with time, then after
sufficient time steps S can be assumed to be constant. This assumption introduces error that will
decrease with time. With this assumption, we can perform just one Jacobi rotation at each time
88
step instead of sweeping the entire upper off-diagonal elements. Instead of being discarded, the
modal matrix is stored and improved gradually with each time step.
The selection of the element on which to perform the Jacobi rotation is done based on a variable
which we will call the coupling factor. We define the coupling factor of the ij
th
off-diagonal
element of a real symmetric matrix T as
ij
ij
ii jj
T
T T
2
. (5.8)
The coupling factor can be used to determine which off-diagonal elements are to be annihilated.
This helps prevent a Jacobi rotation from operating on an off-diagonal element that is already very
close to zero. This strategy saves computational effort. In the Threshold Jacobi method (Bathe
and Wilson, 1976), the element is annihilated only if the coupling factor is larger than a prescribed
threshold. This method saves computational time if the eigenvector computation is implemented
with a computer or a programmable digital signal processing board. Another method is to
annihilate the element that has the largest coupling coefficient. In this section we adopt the latter
method. With the above ideas, we develop the following algorithm, which we refer to as
Algorithm B.
Before starting the algorithm, initialize a variable matrix
T V V ( ) ( ) ( ) 0 0 0
T
(5.9)
This matrix will later be the result of rotation of S.
Also initialize the sensor gain matrix W(0) to an initial guess. In the absence of such a guess,
initialize W(0) to identity.
1. Compute the coupling factors of the upper off-diagonal elements of T(k), using
ij
ij
ii jj
T
T T
2
(5.10)
2. Find the indices (p,q) of the element with the largest coupling factor.
3. Calculate the rotation angle (k) using
89
( )
( ) ( )
( ) arctan k
k
k k
pq
pp qq
_
,
1
2
2T
T T
if ( ) ( ) T T
pp qq
k k 0, (5.11a)
or

( ) k
4
if ( ) ( ) T T
pp qq
k k 0. (5.11b)
4. Assemble a Jacobi rotation matrix
( ) P k
p
q
p q
1
]
1
1
1
1
1
1
1
1
1

1
1
1
O
O
cos( ) sin( )
sin( ) cos( )

th
th
th th
row
row
column column
(5.12)
5. Update the modal matrix according to Eq. (5.7).
( ) ( ) ( ) k k k 1 P (5.13)
6. Estimate the segment output correlation matrix S by the recursive averaging process
( )
{ }
S S V V ( ) ( ) ( ) ( ) k
k
k k k k +
1
1 1
T
(5.14)
7. Perform similarity transformation on S(k) with (k), store the result as T(k)
( ) ( ) ( ) ( ) T S k k k k
T
(5.15)
8. Obtain the sensor gain matrix W by transposing the modal matrix
( ) ( ) W k k
T
(5.16)
90
9. Repeat steps 1 through 8.
The repetition of this algorithm can be terminated after a chosen convergence criterion is satisfied.
However, we can also run this algorithm indefinitely, especially if we implement it with a
dedicated digital electronic circuit, as we intend to do in the future. In this case, if the structures
modal parameters change, the computation of the sensor gain matrix can track the changes.
Therefore, the sensor gain matrix will always decorrelate the sensor outputs.
The intermediate matrix T in step 7 is a unique feature of the new algorithm. To understand the
logic behind this algorithm, keep in mind that it is just a Jacobi rotation algorithm to solve for the
modal matrix of the segment output correlation matrix S, as done in Eq. (4.10). Equation (5.15)
shows that T is the estimate of the diagonal eigenvalue matrix of S. Substitution of Eq. (5.13) into
Eq. (5.15) results in
( ) ( ) ( ) ( ) ( ) ( ) T P S P k k k k k k
T T
1 1 . (5.17)
If S is quasi-constant, as usually the case after enough time steps, then we can assume that
( ) ( ) S S k k 1 . (5.18)
Under this assumption, Eq. (5.17) can be written as
( ) ( ) ( ) ( ) ( ) ( ) T P S P k k k k k k
T T
1 1 1 . (5.19)
The three middle terms on the right-hand side of the last equation can be simplified by applying
Eq. (5.15)
( ) ( ) ( ) ( ) T S k k k k 1 1 1 1
T
(5.20)
Thus, Eq. (5.19) becomes
( ) ( ) ( ) ( ) T P T P k k k k
T
1 . (5.21)
According to Eq. (5.5), the last equation is simply a Jacobi rotation to bring the matrix T closer to
a diagonal matrix.
Equation (5.18) is a key assumption in the derivation of the algorithm because it effectively
enables us to obtain the modal matrix of a slowly varying matrix. Consequently, this algorithm can
91
track changes in the modal parameters of the structure, since this algorithm uses the most recent
updates of the segment output correlation matrix S.
Algorithm B is much faster than Algorithm A. In section 4.3, Algorithm A needs to sweep the
upper off-diagonal elements of S four times at each time step to achieve a modal matrix that
will transform S into a reasonably diagonal matrix. For a ten-mode sensor, there are 45 upper
non-diagonal elements to annihilate with a Jacobi rotation. Four sweeps per time step involve
4*45 = 180 Jacobi rotations per time step. On the other hand, Algorithm B only does one rotation
per time step.
Algorithm B also results in better sensor gain matrices than algorithm A. Equation (5.15) shows
that Algorithm B keeps a matrix T that is stored and rotated towards a diagonal form at each time
step. The estimate for (the modal matrix of S) is the cumulative product of the rotation
matrices. This means that the initial guess for the modal matrix at each time step is the modal
matrix at the previous time step. Therefore, the modal matrix improves with each time step. In
Algorithm A, on the other hand, the initial guess for the modal matrix at each time step is the
identity matrix as in Eq. (5.2).
5.3.2 Numerical Example
Algorithm B results in fast convergence with the number of rotations. After only 256 time steps,
one rotation at each time step, the correlation matrix of the sensor output is already close to a
diagonal matrix (See Fig. 5.6). After 32768 time steps, the sensor output correlation matrix,
shown in Fig. 5.7, is practically diagonal.
The time history of the rotation angle, which is an indication of the convergence of the algorithm,
is shown in Figure 5.8. This figure has been decimated by a factor of 8 for clarity. After the
rotation angle has converged around zero, some spikes appear occasionally. The largest
rotations (close to + 45
o
or - 45
o
) are done because the random excitation causes the denominator
of Eqs. (5.11) to be zero. However, these rotations are usually followed by rotations in the
opposite direction immediately. Therefore, the modal matrix is quickly restored to its steady-state
value.
92
0
5
10 0
5
10
0
1
2
3
x 10
-5
Using W(256)
scale)
(Out of
0.000156
Figure 5.6 Sensor output correlation matrix (Algorithm B) after 256 time steps.
0
5
10 0
5
10
0
1
2
3
4
x 10
-6
Using W(32768)
Figure 5.7 Sensor output correlation matrix (Algorithm B) after 32768 time steps.
93
0 0.5 1 1.5 2 2.5 3 3.5
x 10
4
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Rotation angle
Time step
Rad
Figure 5.8 Rotation angle history, Algorithm B.
The sensor gain matrix resulting from Algorithm B with 32768 time steps is shown in Fig. 5.9.
The magnitudes of the FRFs from force to sensor outputs are shown in Fig. 5.10.
0.2
0.4
-0.4
0
0.4
2
-0.4
0
0.4
3
-0.4
0
0.4
4
-0.2
0
0.2
5
-0.4
0
0.2
6
-0.4
0
0.2
7
-0.4
0
0.4
8
0
0.4
9
-0.2
0
0.2
10
Mode
2 4 6 8 10 Segment 1 3 5 7 9 2 4 6 8 10 3 5 7 9 1
1
Figure 5.9 Sensor gain matrix (Algorithm B), 32768 time steps.
94
0 200 400 600
0
1
Mode 1
0 200 400 600
0
1
Mode 2
0 500 1000
0
1
Mode 3
0 500 1000
0
1
Mode 4
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
0 500 1000 1500
0
1
Mode 5
0 1000 2000
0.2
0.4
0.6
0.8
1
Mode 6
Frequency (Hz)
0.2
0.4
0.6
0.8
Frequency (Hz)
Figure 5.10 Normalized magnitude responses of modal filter (Algorithm B) after 32768 time
steps.
95
Figure 5.10 Normalized magnitude responses of modal filter (Algorithm B) after 32768 time
steps.
5.4 Algorithm C
5.4.1 Development
The above reasoning and numerical example show that Algorithm B is very efficient in computing
the gain matrix W of the adaptive modal sensor. There is, however, another fact to be considered
in the adaptive computation of the gain matrix. For the outputs of the adaptive modal filter to be
proportional to modal coordinates it is necessary that these outputs be orthogonal to each other.
The discussion in section 4.2 verifies that this requirement is satisfied by the gain matrix obtained
0 1000 2000 3000
0.2
0.4
0.6
0.8
1
Mode 7
0 1000 2000 3000
0.2
0.4
0.6
0.8
1
Mode 8
0 2000 4000
0.2
0.4
0.6
0.8
1
Mode 9
Frequency (Hz)
0 2000 4000
0.2
0.4
0.6
0.8
1
Mode 10
Frequency (Hz)
96
by Algorithm B, since the gain matrix is the transpose of the modal matrix of S. However, we can
exploit this orthogonality requirement to refine the modal matrix computing scheme as follows.
As long as any off-diagonal element of the sensor output correlation matrix R is not zero,
coupling still exists between the outputs of the sensor. Therefore, the sensor gain matrix W
should be modified to eliminate this coupling. Thus, checking the diagonality of the output
correlation matrix R provides a sort of performance feedback to the gain matrix computation
scheme. We will develop a new algorithm to incorporate performance feedback, which was called
for in our discussion in Subsection 4.2.3.
To understand the intuitive derivation of the algorithm, imagine that we could rotate the output
correlation matrix R with a Jacobi rotation matrix P to annihilate an off-diagonal nonzero p,q
th
element R
pq
. The matrix P that would do this task can be obtained using Eqs. (5.11) and (5.12). In
reality, we cannot rotate the output correlation matrix R directly. However, we can create the
same effect by modifying the sensor gain matrix W. Recall that when W is quasi-constant and
close to the ideal gain matrix, Eqs. (4.8) and (4.9) can be combined into
R WSW
T
(5.22)
and, since W
T
,
R S
T
(5.23)
The last equation points us to Eq. (5.15) of Algorithm B, which has been proven successful in
computing the correct modal matrix. In particular, the similarity of Eq. (5.23) to Eq. (5.15) shows
that we can now use the matrix T in place or R to create a Jacobi rotation matrix. However, this
time we take advantage of the performance feedback by aiming to annihilate the p,q
th
element of
R instead of T as in Algorithm B. The resulting algorithm, referred to as Algorithm C, is listed
below.
Before starting the algorithm, initialize a variable matrix
T V V ( ) ( ) ( ) 0 0 0
T
(5.24)
This matrix will later be the result of rotation of S.
Also initialize the sensor gain matrix W(0) to an initial guess. In the absence of such a guess,
initialize W(0) to identity.
1. Estimate the sensor output correlation matrix R recursively by
97
( )
{ }
R R ( ) ( )
&
$
( )
&
$
( ) k
k
k k k k +
1
1 1
T
(5.25)
2. Compute the coupling factors of the upper off-diagonal elements of R(k), using
ij
ij
ii jj
R
R R
2
(5.26)
3. Find the indices (p,q) of the element with the largest coupling factor.
4. Calculate the rotation angle (k) using
( )
( ) ( )
( ) arctan k
k
k k
pq
pp qq
_
,
1
2
2T
T T
if ( ) ( ) T T
pp qq
k k 0, (5.27a)
or

( ) k
4
if ( ) ( ) T T
pp qq
k k 0. (5.27b)
5. Assemble the Jacobi rotation matrix
( ) P k
p
q
p q
1
]
1
1
1
1
1
1
1
1
1

1
1
1
O
O
cos( ) sin( )
sin( ) cos( )

th
th
th th
row
row
column column
(5.28)
6. Update the modal matrix according to Eq. (5.7).
( ) ( ) ( ) k k k 1 P (5.29)
98
7. Estimate the segment output correlation matrix S by averaging recursively
( )
{ }
S S V V ( ) ( ) ( ) ( ) k
k
k k k k +
1
1 1
T
(5.30)
8. Perform similarity transformation on S(k) with (k), store the result as T(k)
( ) ( ) ( ) ( ) T S k k k k
T
(5.31)
9. The sensor gain matrix W is the transpose of the modal matrix , as mentioned in Eq. (4.12).
The sensor output is the segment output multiplied by the sensor gain matrix. Therefore,
( ) ( ) ( ) ( )
&
$
( ) k k k k k W V V
T
(5.31)
10. Repeat steps 1 through 9.
The repetition of this algorithm can be terminated after a chosen convergence criterion is satisfied.
However, we can also run this algorithm indefinitely, especially if we implement it with a
dedicated digital electronic circuit. In this case, if the structures modal parameters change, the
computation of the sensor gain matrix can track the changes. Therefore, the sensor gain matrix
will always decorrelate the sensor outputs.
Figure 5.11 shows the connections between the segments and Algorithm C. This algorithm is
more sophisticated than Algorithm B, particularly because of the performance feedback. Equation
(5.25) constitute extra computation over algorithm B. This is the cost of performance feedback
which does not exist in Algorithm B.
99
Figure 5. 11 Input connections to Algorithm C.
V
n
V
N V
1
Beam
PVDF
segments
f
Mode-m
filter
&
$
1
( )
{ }
S S V V ( ) ( ) ( ) ( ) k
k
k k k k +
1
1 1
T
&
$
m
&
$
M
. . .
. . .
.
.
.
.
.
.
W
1,1
W
m,N
W
m,1
W
M,N
W
1,n
W
m,n
W
M,n
W
M,1
W
1,N
( )
{ }
R R ( ) ( )
&
$
( )
&
$
( ) k
k
k k k k +
1
1 1
T
Performance
feedback
Find off-diagonal element of R that has
greatest coupling factor.
Form Jacobi rotation matrix P(k).
( ) ( ) ( ) k k k 1 P ; W =
T
.
( ) ( ) ( ) ( ) T S k k k k
T
100
5.4.2 Numerical Example
Algorithm C results in a faster convergence and more robust gain matrix than Algorithm B. The
eight-time decimated history of the rotation angle in Figure 5.12 shows that the rotation angle
converges to zero in fewer time steps than in Algorithm B. Moreover, after the rotation angle
converges around zero, it does not spike to a big value and flip to the opposite direction, as
occurred with Algorithm B.
0 0.5 1 1.5 2 2.5 3 3.5
x 10
4
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
With performance feedback
Time step
Rad
Figure 5.12 Rotation angle history, Algorithm C.
The sensor gain matrix W obtained after 32768 time steps with Algorithm C is shown in Fig.
5.13. We can see that this gain matrix is very close to the ideal gain matrix shown in Fig. 4.9. The
sensor output correlation matrix is shown in Figure 5.14. This correlation matrix is practically
diagonal, meaning that the sensor outputs are uncorrelated. From the last two figures, we can
conclude that the modal filter with Algorithm C effectively decouples the segment output voltages
V into modal coordinates.
101
0.2
0.4
1
-0.4
0
0.4
2
-0.4
0
0.2
3
-0.4
0
0.4
4
-0.2
0
0.2
5
-0.4
0
0.4
6
-0.4
0
0.4
8
-0.4
0
0.2
9
-0.2
0
0.2
10
-0.4
0
0.2
7
Mode
2 4 6 8 10 Segment 1 3 5 7 9 2 4 6 8 10 3 5 7 9 1
Figure 5.13 Sensor gain matrix (Algorithm C), after 32768 time steps.
0
5
10 0
5
10
0
0.5
1
1.5
2
2.5
3
x 10
-6
Figure 5.14 Sensor output correlation matrix (Algorithm C) after 32768 time steps.
102
5.4.3 Frequency-Domain Analysis
The sensor output correlation matrix resulting from the adaptive modal sensor configured with
Algorithm C indicates that the correlation between different sensor outputs is very small.
However, a more critical analysis of the modal sensor outputs can be done in frequency domain.
Figures 5.15 through 5.24 show the magnitudes of the FRFs from force to each sensor outputs.
(This time we plot the magnitudes in log scale, so that we can see the small modal leaks of
undesired modes through the modal filters.) These figures show very little coupling among the
outputs. Each modal sensor effectively passes only one modal coordinate and filter out other
modal coordinates. Signals from undesired modes is generally only around 1% of the intended
modes. This leaking is worst with modes 8 and 9.
Now, to further illustrate the effectiveness of the adaptive modal filters, we redraw Figs. 5.15
through 5.24 on a linear scale. The resulting plots are shown in Figs. 5.25 through 5.34. This
analysis shows that Algorithm C is indeed a powerful algorithm to separate the segment outputs
into uncorrelated modal coordinates.
0 100 200 300 400 500
10
-3
10
-2
10
-1
10
0
Mode 1
Frequency (Hz)
= Adaptive
= Ideal
Figure 5.15 Normalized magnitude response of modal filter with performance feedback:
Mode 1.
103
0 100 200 300 400 500
10
-3
10
-2
10
-1
10
0
Mode 2
Frequency (Hz)
= Adaptive
= Ideal
Mode 2.
0 200 400 600 800 1000
10
-4
10
-3
10
-2
10
-1
10
0
Mode 3
Frequency Hz
= Adaptive
= Ideal
Mode 3.
104
0 200 400 600 800 1000
10
-4
10
-3
10
-2
10
-1
10
0
Mode 4
Frequency (Hz)
= Adaptive
= Ideal
Mode 4.
0 500 1000 1500
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Mode 5
Frequency (Hz)
= Adaptive
= Ideal
Mode 5.
105
0 500 1000 1500 2000
10
-4
10
-3
10
-2
10
-1
10
0
Mode 6
Frequency (Hz)
= Adaptive
= Ideal
Mode 6.
0 500 1000 1500 2000 2500 3000
10
-4
10
-3
10
-2
10
-1
10
0
Mode 7
Frequency (Hz)
= Adaptive
= Ideal
Mode 7.
106
0 500 1000 1500 2000 2500 3000
10
-4
10
-3
10
-2
10
-1
10
0
Mode 8
Frequency (Hz)
= Adaptive
= Ideal
Mode 8.
1000 2000 3000 4000 5000
10
-4
10
-3
10
-2
10
-1
10
0
Mode 9
Frequency (Hz)
0
= Adaptive
= Ideal
Mode 9.
107
1000 2000 3000 4000 5000
10
-4
10
-3
10
-2
10
-1
10
0
Mode 10
Frequency (Hz)
0
= Adaptive
= Ideal
Mode 10.
0 100 200 300 400 500
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mode 1
Frequency, Hz
= Adaptive
= Ideal
Mode 1.
108
0 100 200 300 400 500
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mode 2
Frequency, Hz
= Adaptive
= Ideal
Mode 2.
0 200 400 600 800 1000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mode 3
Frequency, Hz
= Adaptive
= Ideal
Mode 3.
109
0 200 400 600 800 1000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mode 4
Frequency, Hz
= Adaptive
= Ideal
Mode 4.
0 500 1000 1500
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mode 5
Frequency, Hz
= Adaptive
= Ideal
Mode 5.
110
0 500 1000 1500 2000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mode 6
Frequency, Hz
= Adaptive
= Ideal
Mode 6.
0 500 1000 1500 2000 2500 3000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mode 7
Frequency, Hz
= Adaptive
= Ideal
Mode 7.
111
0 500 1000 1500 2000 2500 3000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mode 8
Frequency, Hz
= Adaptive
= Ideal
Mode 8.
0 1000 2000 3000 4000 5000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mode 9
Frequency, Hz
= Adaptive
= Ideal
Mode 9.
112
0 1000 2000 3000 4000 5000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mode 10
Frequency, Hz
= Adaptive
= Ideal
Mode 10.
5.5 Limitations
All three algorithms developed in this chapter are effective in computing the modal filter gain
matrix W without much knowledge of the modal properties of the structure. These algorithms
require only the number of modes that contribute to the segment signals in the anti-aliased
frequency range. These algorithms are much more powerful than, for example, the LMS-based
algorithm discussed in appendix A, which requires exact knowledge of the natural frequency and
damping of the structure. Likewise, these algorithms can track slow changes in the modal
properties of the structure by virtue of keeping track of changes in the segment output correlation
matrix. Despite the above interesting features, some of the limitations to the algorithms are as
follows.
1. Spatial aliasing must not be present in the segment outputs. Spatially-aliased higher-mode
coordinates will appear as lower-mode coordinates.
2. The Jacobi rotation algorithm results in a sensor gain matrix that contains eigenvectors that are
arbitrarily scaled. Consequently, the outputs of the modal filter are scaled modal coordinates.
The scaling factor for each mode is not likely to be the same as the scaling factors for other
modes. The scaling factors may even be negative. This limitation is a result of the fact that
arbitrarily scaled eigenvectors are orthogonal.
113
3. The ordering of the eigenvectors in the sensor gain matrix is also arbitrary. Therefore, the
sensor outputs are not necessarily ordered from the lowest mode coordinate to the highest.
4. In their present forms, the new algorithms cannot solve for repeated eigenvectors. The
algorithms will not work if the sensor output correlation matrix is singular.
The frequency-domain analysis also revealed that some unintended modal coordinates still leak
through the modal filters. This leaking indicates that perfect modal filtering is still not achieved by
the adaptive modal filtering algorithms. However, considering that the adaptive modal filters were
obtained without knowing the modal properties of the structure, the adaptive algorithm is
successful. Moreover, independence from exact knowledge of mode shapes gives the adaptive
modal filters potential for better modal filtering effects than those shown in Fig. 3.14. The modal
filters whose responses were shown in Fig. 3.14 were obtained with good knowledge of the
structures modal properties. However, discrepancies between the predicted and the actual
segment outputs result in imperfect modal filtering.
Additionally, recall that virtually all the work discussed in this dissertation relies on the
assumptions that the structure and transducers are linear and that the mode shapes are orthogonal.
The newly developed algorithms are still in their primordial stage. The above limitations may
determine the direction of further development. So may issues need to be addressed that it is
impossible to include all of them in the scope of this dissertation. The main contribution of this
research work is the conceptual creation of the novel method of modal sensing, which is a step
towards the future in adaptive modal analysis.
5.6 Chapter Summary
The Jacobi rotation algorithm solves for the eigenvector matrix (modal matrix) of a symmetric
real matrix Q by a series of norm-preserving similarity transformations. Each transformation is
done with a matrix called Jacobi rotation matrix. Each transformation makes Q progressively
more diagonal by annihilating one off-diagonal element of Q. Sweeping the upper off-diagonal
terms of Q a few times diagonalizes Q. The modal matrix of Q is the product of the Jacobi
rotation matrices.
As shown in chapter 4, the sensor gain matrix that transforms the segment outputs into
uncorrelated sensor outputs is the transpose of the modal matrix of the segment output
correlation matrix S. Algorithm A solves for the modal matrix of S(k) at each time step k by
sweeping the off-diagonal upper triangular part of S(k) with Jacobi rotations. This algorithm is
computationally extensive because the number of Jacobi rotations for each time step equals the
number of sweeps times the number of off-diagonal upper triangular elements of S.
114
Algorithm B performs only one Jacobi rotation for each time step. The Jacobi rotation is done on
an intermediate matrix T which is a rotated version of the segment output correlation matrix S.
This algorithm was developed with an assumption that the segment output correlation matrix S
does not vary quickly with time steps.
Algorithm C also performs one Jacobi rotation for each time step on the intermediate matrix T.
However, this algorithm checks for correlation between different sensor outputs and uses this
undesired correlation to correct the sensor gain matrix. This performance feedback scheme
accelerates convergence and produces more robust steady-state gain matrices.
All of the above algorithms are effective in computing the modal filter gain matrix W without any
knowledge of the modal properties of the structure other than the number of modes in the given
frequency range. Likewise, the algorithms can track slow changes in the modal properties of the
structure. A few limitations were discussed in this chapter. Much research is required to bring the
new algorithms into more practical implementation.
CHAPTER 6
A MODE SHAPE SENSING TECHNIQUE
In this chapter we discuss the development of a method to extract the mode shapes of structures
using the adaptive modal filter.
6.1 Sensor Gain Matrices and Mode Shapes
The sensor gain matrix in Fig. 1.2 in the Introduction give an idea that the sensor gain matrix of a
spatial modal filter is closely related to the mode shapes of the structure to which the sensors are
attached. Figures 3.1 and 3.13 Indicates that the rows of the sensor gain matrix are similar to the
mode shapes of the structure. These examples suggest that structural mode shapes can be
recovered from sensor gain matrices.
In this chapter we assume that the sensor gain matrix of the adaptive modal filter has converged
to the ideal gain matrix. Only then can we compute the mode shape function using the procedure
developed below. Notice that the m
th
row of the sensor gain matrix W contains the information on
the m
th
mode shape function,
m
. This information is in discrete form, i.e., if we use a sensor with
N segments, each row of the sensor gain matrix contains only N numbers whose relationship to
the mode shape function is not always obvious. In most cases where we need mode shape
functions, what we need is the values of the mode shape functions at given positions on the
structure.
The problem now is how to process the sensor gain matrix (for example, the one in Fig. 6.1) to
extract mode shapes as continuous functions of positions on the structure (for example, functions
shown in Fig. 6.2).
116
10
-2
2
-1
1
-0.5
0.5
-0.2
0.4
-0.2
0.2
-0.2
0.2
-0.2
0.2
2 6 10
0
9
0
10
3 7 2 6 10 5 9
Mode
Figure 6.1 Sensor gain matrix .
0 0.2 0.4 0.6 0.8 1.0
10
9
8
7
6
5
4
3
2
1
x / L
Mode
number
Figure 6.2 Mode shapes of beam, x).
117
6.2 Lagrange Interpolation
Based on the observation of the above sensor gain matrix and the mode shape functions, we can
make an approximating assumption that the mode shape functions are related to the rows of the
sensor gain matrix. Under this approximation, we assume that the m
th
row of the sensor gain
matrix are the values of the m
th
mode shape at points c
n
in the middle of the segments for n = 1,
N. These positions are shown in Fig. 6.3. Additionally, we know from the boundary conditions
in this example that the values of the mode shapes at the ends of the beam are equal to zero.
Figure 6.3 Beam with strain sensor segments.
We only have a limited amount of (discrete) data, i.e., the third row of the sensor gain matrix,
while the domain of the mode shape function (x) is continuous, i.e., it has an infinite number of
possible values. For structures in general (x) is not known to take the form of any specific
function. Therefore, we can only estimate the deflection. The quality of our estimation depends on
the function we assume for the deflection. In static cases, the deflection is often a polynomial
function of x. In many cases, polynomials provide close approximation to other functions
provided that the polynomials are of sufficiently high order. Therefore, we choose polynomials as
approximating functions. The estimated mode shapes will be inherently biased. However, if the
estimate is good, the bias can be made negligibly small for many purposes.
For simplicity and easier understanding, we limit our discussion to the extraction of the third
mode shape function (m = 3) of the beam. We will use the third row of the sensor gain matrix
accordingly. Of course, the method we will develop can be applied to other mode shapes as well.
For m = 3, we assume that the value of the mode shape function in the middle of the n
th
segment
is equal to the n
th
element of the third row of the sensor gain matrix, that is,

3 3
( )
, n n
W (6.1)
PVDF segments
x
c
0
= 0
c
1
W
m,n
c
N+1
= L
c
N
Segment gain
118
We can estimate the mode shape function between the above points by interpolating the known
N+1) order polynomial from N+
interpolation formula. (For explanation of this formula, see Burden and Faires, 1985).
( )
( )
( )
( )
( )
( )
( )
( )
( )
( )
( )
3
0
0
1
0
0
0
1
0
1
1
1
0
1
1
0
1
0
1
0
1
1
1
0
1
1
3
3 1
3
3
0
x
x c
c c
x c
c c
x c
c c
x c
c c
L
i
i
i
N
i
i
i
N
i
i
i
N
i
i
i
N
i
i
i N
N
N i
i
i N
N
i
i
i N
N
N i
i
i N
N
N
1
]
1
1
1
1
1
1
1
'
+
+
+
+
+
L M
W
W
,
,
(6.2)
In this example the mode shape functions at the ends of the beam are zero, so
( ) ( )
3 3
0 0 L , (6.3)
In the implementation of the segmented sensors to simulate a mode shape sensor at any given x,
we shall pre-compute the ratios in Eqs. (6.2) to form the appropriate shape sensor weight
coefficients. Therefore, as soon as the sensor gain matrices are available from the adaptive modal
filters, the only real-time computation required to obtain the slope and the shape is multiplication
of the sensor gain matrices by the pre-computed weight coefficients. The above formula results in
a single (scalar) number for each position x. To obtain the mode shape function at an array of
positions, we can enter a vector of values of desired positions.
6.3 Numerical Example
As an example, we will estimate the third mode shape of the beam,
3
(x), using the third row of
the sensor gain matrix, W
3,n
, where n = 1, , 10 denotes the segment number. This row is shown
in Fig. 6.4.
119
2 4 6 8 10
-1.5
-1
-0.5
0
0.5
1
1.5
Segment number
Gain
Row 3 of W
Figure 6.4 Third row of sensor gain matrix.
We apply Eq. (6.3) to the above W
3,n
with an array of 65 points equally spaced between x = 0 and
x = L as an example. (The equal spacing is not necessary, since the mode shape value at any point
can be evaluated.) The result of the reconstructed mode shape agrees well with the analytical
mode shape calculated using Eq. (2.4). This agreement is shown in Fig. 6.5. The analytical mode
shape has been normalized to a maximum value of 1 using Eq. (2.6). To make comparison easier,
the recovered mode shape is also normalized so that its maximum absolute value = 1.
120
reconstructed
true
0 0.1 0.2 0.3 0.4 0.5
-1
-0.5
0
0.5
1
1.5
Position (m)
Reconstructed mode shape
Mode
shape
Figure 6.5 Reconstructed mode shape.
6.4 Chapter Summary
The m
th
row of the sensor gain matrix W provides discrete spatial information that can be
interpolated to recover the m
th
mode shape of the structure. We used Lagrange interpolation
formula to create a matrix that premultiplies the sensor gain matrix W to recover a mode shape of
the structure. The recovered mode shape agrees well with the analytical mode shape.
CHAPTER 7
CONCLUSIONS AND FUTURE DIRECTIONS
7.1 Conclusions
Looking back to Eq. (1.1) in the Introduction, we recall that knowledge of mode shapes and real-
time monitoring of modal coordinates are very useful in vibration measurement and control. For
that reason, we have developed the concepts of adaptive modal filtering and real-time mode shape
extraction. Some of the accomplishments of the research work documented in this dissertation can
be summarized as follows.
1. A new distributed sensor configuration:
We devised a sensor system consisting of segments of strain-sensing film connected to
adjustable gains. This adaptive distributed sensor system can be programmed to perform as
various types of sensors. A proof-of-concept experiment showed very promising potential to
create adaptable distributed modal sensors from this configuration.
2. An adaptive modal filter:
The sensor system is programmed to produce modal coordinates of a vibrating structure in
real time. This system does not require knowledge of the modal properties of the structure. If
the modal properties of the structure changes slowly with time, the system can adapt to keep
acting as a modal filter.
3. Adaptive eigenvector computing algorithms:
We developed two new algorithms to compute the sensor gain matrix that transforms the
segment outputs into uncorrelated modal coordinates. The new algorithms use recursive
averaging, zero-order correlation matrices, and Jacobi rotation matrices in computing the
eigenvectors of real symmetric matrices. The algorithms compute the eigenvector matrices
recursively over many time steps, so as to lower the computational load. One of the
algorithms takes advantage of performance feedback in the form of non-zero correlation
between different sensor outputs.
122
4. A mode shape extraction formula:
We developed a simple method to extract the mode shapes of the structure from the
adaptively computed sensor gain matrix. This method is based on Lagrange interpolation
formula.
7.2 Future directions
The experiment with the segmented strain-sensing film and the signal conditioning circuits for the
sensor system showed that their outputs agree with the theoretical prediction. The transducers
and the associated signal conditioning circuit have then proven to be promising hardware for
future development. The intentions are to build a prototype of the adaptive modal filter, including
the signal processing hardware and the associated software. At this point, the signal processing
part has only been simulated on a computer. While a digital signal processing board appears to be
the most practical first try, the hardware implementation would utilize hard-wired dedicated
electronic circuits in the future if the concept of adaptive modal sensing and mode-shape analysis
becomes common. Much further development of the adaptive modal sensing theory is necessary
before a prototype can be constructed.
A natural extension to the completed work includes the following :
1. Study the sensitivity of the sensors to errors in sensor parameters, such as:
a. The piezoelectric constants of the sensor material.
For example, a deviation in the e
31
value (Eq. (2.30)) from the theoretical value will result
in a proportional deviation in the segment output voltages and the estimated modal
coordinates.
b. The positions of the segments on the beam.
For any set of segment dimensions and positions, the adaptive algorithms will converge to
the correct gain matrix to make sure that the outputs of the sensor are uncorrelated.
Therefore, the positions of the segments on the beam should not introduce any
unnecessary coupling between the modal sensor outputs. However, the mode shape
extraction technique in Chapter 7 assumes perfect knowledge of the segment dimensions
and positions. The accuracy of the mode shape estimates depends on the accuracy of the
segment positions and dimensions.
2. Study the effects of segment resolution on sensor performance.
123
As mentioned in subsection 2.3.3, even a hypothetical gain matrix derived from a perfect
modal matrix will not always result in exact modal coordinates. The reason for the error is
modal truncation. Although the adaptive modal filter described in this work requires almost no
information about the modal properties of the structure, it still requires information on how
many modes contribute significantly to the structural vibration within the specified frequency
range. It was shown in Subsection 3.1.2 that contribution from modes that are not accounted
for in the sensor design will falsely be sensed as lower mode (spatially aliased) coordinates.
The number of modes to be included in the design of a modal filter determines the minimum
number of segments. This number of modes can be obtained by several kinds of tests. One of
the simplest test is to sweep the structure with a point vibration transducer to locate nodes.
The number of modes is closely related to the upper limit of the frequency range, which is in
turn related to the sampling rate, the excitation frequency range, anti-aliasing filter range, etc.
The relationships among these parameters are simple but very important. The development of
a comprehensive set of simple rules is a necessary step towards the implementation of the
adaptive modal sensors.
The number of segments also determines the accuracy of the mode shape estimates. As
mentioned in Chapter 6, the mode shape estimates are polynomial functions that cannot be
identical to the actual mode shape functions. The difference between the estimates and the
actual mode shapes can be small only if the orders of the polynomial functions are sufficiently
high. The order of the polynomial can only be as high as the number of segment output
voltages plus the number of known boundary conditions. Thus, the accuracy of the mode
shape estimates depends on the number of segments. Further study is necessary to quantify the
relationship between the number of segments and the error in the mode shape estimates.
3. Study the effects of spatial aliasing and devise methods to reduce those effects.
As mentioned in Subsection 3.1.2, spatial aliasing is a serious problem in the discretized modal
filters. One of the advantages of a strain-integrating film segment over a point sensor is that
the film segment can acquire strain information over its entire area. This integrating feature
can be exploited to reduce spatial aliasing. Shaping a segment with a function that has a low-
pass characteristic in the wavenumber domain reduces the sensitivity of the segment to modes
with high wavenumbers, thereby reducing spatial aliasing. The relationship between spatial
shaping functions and their spatial aliasing reduction effects are important in the development
of the segmented modal filtering sensors.
4. Study the effects of noise and develop methods to reduce those effects.
Random, uncorrelated noise in the segment output signals is filtered out by the averaging
process in obtaining the segment output correlation matrix. However, other kinds of noise that
contaminates the segment output correlation matrix will introduce errors in the modal matrix
124
and the sensor gain matrix. Statistical analysis should be done to determine the effects of noise
on the modal filters. Furthermore, methods should be developed to reduce these effects.
5. Refine the adaptive algorithms to eliminate their limitations.
As mentioned in Section 5.5, naturally arbitrary scaling of eigenvectors and unpredictable
ordering of the modal sensor outputs are some of the problems of the adaptive modal sensors.
Further development of the algorithms may resolve these problems.
At its current stage, the adaptive modal filtering theory has only been applicable to self-adjoint
structures with orthogonal mode shapes, such as the beam used as the example structure
throughout the dissertation. To create modal sensors for plates, the theory has to be modified.
The mode shape functions of plates with most boundary conditions cannot be shown to be
orthogonal (Blevins, 1984b). Another problem with the segmented sensor on plates is that the
segment output voltage is caused by flexure in two directions. For each individual segment, it
is impossible to separate the segment output voltage into the part generated by curvature in
one direction from the part generated by curvature in the other direction. In the cases where
the adaptive modal filter theory can be applied to plates, point sensors are easier to use than
strain-sensing segments.
Future modifications to the algorithms we developed may reduce the computational requirements.
For example, Eqs. (5.4) and (5.5) show that Jacobi rotation of a matrix is a very efficient
operation because this matrix transformation only affect two rows (or columns) of the matrix to
be transformed. On the other hand, Eq. (5.15) shows that Algorithm B requires transformations
with a modal matrix, which is almost always less sparse than the Jacobi rotation matrix. The
transformation is therefore computationally more demanding than a simple Jacobi rotation. An
encouraging fact is that the symmetry of the matrices allows us to work with only the upper
triangular half of the matrices.
The writer hopes that this research work helps inspire efforts to devise other novel techniques to
perform modal filtering and mode shape identification adaptively in real time. Advances in digital
signal processing will then help make the concept of real-time adaptive modal analysis a reality in
the near future.
REFERENCES
Bai, M.R, Shieh, C., 1995, Active Noise Cancellation by Using the Linear Quadratic Gaussian
Independent Modal Space Control, J. Acoustical Society of America, Vol. 97, No. 5, pp. 2664-
2270.
Balas, M. J., 1978, Feedback Control of Flexible Systems, IEEE Trans. Automatic Control,
AC-23, pp. 673-679.
Bathe, K-J., and Wilson, E. L., 1976, Numerical Methods in Finite Element Analysis, Prentice-
Hall, Englewood Cliffs, NJ. P. 446.
Baz, A., Poh, S., 1996, Optimal Vibration Control with Modal Positive Position Feedback,
Optimal Control Applications & Methods, Vol. 17, No. 2, pp. 141-147.
Blevins, R. D., 1984, Formulas of Natural Frequency and Mode Shape, Krieger, Florida, pp. 261
- 264.
Blevins, R. D., 1984b, Formulas of Natural Frequency and Mode Shape, Krieger, Florida, p 236.
Bradbary, W. W. and R. Fletcher, 1968, New Iterative Methods for Solution of the
Eigenproblem Arising in Structural Dynamics, Proc. 2
nd
Conf. On Matrix Methods in Structural
Mechanics, Wright-Peterson A. F. B., OH.
Burden, R. L., and Faires, J. D., 1985, Numerical Analysis, 3rd ed., Prindle, Weber and Schmidt,
Boston, MA, p. 86.
Burke, S. E., and Hubbard, J. E., 1990, Spatial Filtering Concepts in Distributed Parameter
Control, J. Dynamic Systems, Measurement, and Control, No. 112, December 1990, pp. 565-
573.
Callahan, J., and H. Baruh, 1994, Modal Analysis Using Piezoelectric Sensors, Proc. AIAA
Structures, Structural Dynamics and Materials Conference, Hilton Head, S. C. p. 83-94.
Chen, C. I., Napolitano, M. R., Smith, J. E., 1994, Active Vibration Control Using the Modified
Independent Modal Space Control (M.I.M.S.C.) Algorithm and Neural Networks as State
Estimators, Journal of intelligent material systems and structures, Vol. 5, No. 4, pp. 550-557.
Clark, R. L., 1991, Modal Sensing of Efficient Acoustic Radiators with Polyvinylidene Fuoride
Distributed Sensors in Actrive Structural Acoustic Control Approaches, J. Acoust. Soc.
America, Vol. 91, No. 6, pp. 3321-3329.
126
Clark, R. L., 1992, Advanced Sensing Techniques for Active Structural Acoustic Control,
Ph.D. Dissertation, Virginia Polytechnic Institute and State University, Blacksburg, VA, pp. 124-
131.
Clark, R. L., 1995, Adaptive Feedforward Modal Space Control, J. Acoust. Soc. America, Vol.
98, Number 5, pp. 2639-2650.
Clark, R. L., and Fuller, C. R., 1993, Active Control of Structurally Radiated Sound from an
Enclosed Finite Cylinder, Proceedings of the Second Conference on Recent Advances in Active
Control of Sound and Vibration, Blacksburg, VA, pp. 381-402.
Cudney, H. H., 1989, Distributed Structural Control Using Multilayered Piezoelectric
Actuators, Ph.D. Dissertation, State University of New York, Buffalo, NY, p. 53.
Davidson, R. A., 1990, Compensation of Controller-Structure Interaction Using Adaptive
Residual Mode Filters, Ph.D. Thesis, University of Colorado, Boulder, CO.
Ewins, D. J., 1986, Modal Testing: Theory and Practice, Research Studies Press, Letchworth,
England, p. 2.
Finefield, J. K., H. Sumali, and H. H. Cudney, Combined Feedback and Adaptive Digital
Control of Cylinder Vibrations Using Smart Materials, Proceedings of IEEE 31st Conference on
Decision and Control, Tucson, AZ, December 1992, pp. 1809-1814.
Franklin, G. F., Powell, J. D., and Workman, M. L., 1990, Digital Control of Dynamic Systems,
Addison-Wesley, Reading, MA, p.42.
Golub, G. H., and Van Loan, C. F., 1989, Matrix Computations, John Hopkins University Press,
Baltimore, MD, pp.427-435.
Gorman, D. J., 1975, Free Vibration Analysis of Beams and Shafts, John Wiley & Sons, New
York, NY, p.16.
Gu, Y., Clark, R. L., Fuller, C. R., Zander, A. C, 1994, Experiments on Active Control of Plate
Vibration Using Piezoelectric Actuators and Polyvinylidene Fluoride (PVDF) Modal Sensors, J.
Vibration and Acoustics, Vol.116, No. 3, pp. 303-309.
Guigou, C. R. J., Berry, A., and Charette, F.,1995, Active Control of Plate Volume Velocity
Using Shaped PVDF Sensor, Proceedings of the Symposioum on Adaptive Structures and
Composite Materials, Chicago, IL, pp. 247-255.
Haykin, S., 1991, Adaptive Filter Theory, 2
nd
ed., Prentice Hall, Englewood Cliffs, NJ, p. 418.
127
Horvath, S. Jr., 1976, Adaptive IIR Digital Filters for On-Line Time Domain Equalization and
Linear Prediction, paper presented at the IEEE Arden House Workshop on Digital Signal
Processing, Harriman, NY.
Hsia, T. C., 1981, A Simplified Adaptive Recursive Digital Filter, Proceedings of the IEEE,
Vol. 69, No. 9, pp 1153-1155.
Juang, J-N, 1987, Mathematical correlation of Modal-Parameter-Identification Methods Via
System-Realization Theory, J. Modal Analysis, January 1987, pp. 1-18.
Juang, J-N, 1994, Applied System Identification, Prentice Hall, Englewood Cliffs, NJ, pp. 3-8.
Juang, J-N, 1994b, Applied System Identification, Prentice Hall, Englewood Cliffs, NJ, pp. 131-
148.
Kynar Piezo Film Technical Manual, 1987, (Pennwalt Corporation, Valley Forge, PA,) p. 27.
Lazan, B. L., 1968, Damping of Materials and Members in Structures, Pergamon Press, Oxford,
England, p. 205.
Lee, C. K., 1987, Piezoelectric Laminates for Torsional and Bending Modal Control: Theory and
Experiment, Ph. D. Thesis, Cornell University, Ithaca, NY.
Lee, C. K., and Moon, F. C., 1990, "Modal Sensors/Actuators", J. Applied Mechanics, Vol. 57,
pp. 434-441.
Lee, C. K., and OSullivan, T. C., 1991, Piezoelectric Strain Rate Gages, J. Acoustical Society
of America, Vol. 90, No. 2, pp. 945-953.
Ljung, L., 1989, System Identification: Theory for the User, Prentice Hall, Upper Saddle River,
New Jersey.
Ljung, L., 1989b, System Identification: Theory for the User, Prentice Hall, Upper Saddle River,
New Jersey, pp. 303-326.
Meirovitch, L. and Baruh, H., 1982, Control of Self-Adjoint Distributed Parameter Systems, J.
Guidance, Control, and Dynamics, Vol. 5, No. 1, pp. 60-66.
Meirovitch, L. and Baruh, H., 1985, The Implementation of Modal Filters for Control of
Structures, J. Guidance and Control, Vol. 8, pp. 707-716.
Miller, S. E, Oshman, Y., Abramovich, H., 1996, Modal Control of Piezolaminated Anisotropic
Rectangular Plates. Part 1: Modal Transducer Theory, AIAA Journal, Vol. 34, No 9, pp. 1868.
128
Mitchell, L. D., 1986, Signal Processing and the Fast-Fourier-Transform (FFT) Analyzer A
Survey, International J. of Modal Analysis, January 1986, pp. 24-36.
Ouyang, J. J., 1987, Adaptive Residual Mode Filter Control of Distributed Parameter Systms for
Large Space Structure Application, Ph.D. Dissertation, Rensselaer Polytechnic Institute, Troy,
NY.
Porter, B., and Crossley, R., 1972, Modal Control, Taylor & Francis, London, the United
Kingdom, p. 2.
Press, W. H., Flannery, B. P., Teukolsky, S. A., and Vetterling, W. T., Numerical Recipes,
Cambridge University Press, Cambridge, England, pp. 339-379.
Shelley, S. J., 1991, Investigation of Discrete Modal Filters for Structural Dynamic
Applications, Ph. D. Dissertation, University of Cincinnati.
Shelley, S. J., Lee, K. L., Aksel, T., Aktan, A. E., 1995, Active-Control and Forced-Vibration
Studies on Highway Bridge, J. Structural Engineering, September 1995, pp. 1306-1312.
Shelley, S. J., Schultze, J. F., Rost, R. W., Allemang, R. J., 1992, Active Vibration Control
Utilizing A Discrete Modal Filter Based Control Technique, Proc. 17
th
International Seminar on
Modal Analysis, pp. 1747-62.
Slater, J. C., Inman, D. J., 1995, Nonlinear Modal Control Method, J. Guidance, Control, and
Dynamics, Vol. 18, No. 3, pp. 433.
Sullivan, J. M., 1993, Distributed Transducer Design for the Active Control of Multi-
dimensional Elastic Structures, Ph.D. Thesis, Massachusetts Institute of Technology, Cambridge,
MA.
Sumali, H., 1992, Demonstration of Active Structural Acoustic Control of a Cylinder, Masters
Thesis, Virginia Polytechnic Institute and State University, Blacksburg, VA.
Sumali, H., and H. H. Cudney, Experimental Investigation of Piezoelectric Film Used as Modal
Sensors, Proceedings of the 9th International Modal Analysis Conference, Florence, Italy, April
15-18, 1991, pp. 1174-1180.
Sumali, H., and H. H. Cudney, Experimental Investigation of Piezoelectric Film Used as Modal
Sensors, Proceedings of the 9th International Modal Analysis Conference, Florence, Italy, April
15-18, 1991, pp. 1174-1180.
Sumali, H., and H. H. Cudney, Segmented Two-dimensional Modal-filtering Sensors, Vibration
and Control of Mechanical Systems, Proceedings of the 1993 ASME Design Technical
129
Conferences - 14th Biennial Conference on Mechanical Vibration and Noise, Albuquerque, NM,
September 19-22, 1993, pp. 59-66.
Sumali, H., H. Cudney, and J. Vipperman, Vibration Control of Cylinders Using Piezoelectric
Sensors and Actuators, Proceedings of ADPA/AIAA/ASME/SPIE Conference on Active
Materials and Adaptive Structures, Gareth Knowles, ed., Philadelphia, PA, Institute of Physics
Publishing Company, 1991, pp. 467-472.
Tzou, H. S., and Fu, H., 1992,A Study on Segmentation of Distributed Piezoelectric Sensors
and Actuators: Part 1 - Theoretical Analysis, Active Control of Noise and Vibration, p 239-246.
Tzou, H. S., and Fu, H., 1992b,A Study on Segmentation of Distributed Piezoelectric Sensors
and Actuators: Part 2 - Theoretical Analysis, Active Control of Noise and Vibration, p. 247-253.
White, S. A., 1975, An Adaptive Recursive Digital Filter, Proc. 9
th
Asilomar Conf. On Circuits,
Systems and Computers, Pacific Grove, CA, pp. 21-25.
Widrow, B., and Stearns, S. D., 1985, Adaptive Signal Processing, Prentice Hall, Englewood
Cliffs, NJ, pp. 99-113.
Wimmel, R. and Melcher, J., 1992, Application of Recursive Adaptive Algorithms for System
Identification and Vibration control, J. Intelligent Material Systems and Structures, Vol. 3, July
1992, pp. 519-535.
Zhou, N., 1992, Active Control of Sound Transmission Through Plates in a Reverberant
Environment, M.S. Thesis, Virginia Polytechnic Institute and State University, Blacksburg, VA.
130
APPENDIX A
LMS COMPUTATION OF GAIN MATRIX
In this appendix we shall show that there are methods to compute a sensor gain matrix without
knowledge of the mode shapes. These methods use adaptive digital signal processing. As a brief
example, we present below the design of a mode-3 sensor. This sensor is sensitive to mode-3 only
and filters out other modes.
We use as an example the beam that is simply supported at one end and supported with a pin and
a torsion spring at the other end. The surface of the beam is covered with N rectangular segments
made of piezoelectric film. The electric voltages V
n
from the segments, caused by the strain
associated with the deflection of the beam, are multiplied by a series of weights W
3,n
. The
products are added together to produce the sensor output
&
$
3
, which is supposed to approximate
mode-3 coordinate.
Figure A.1 Training a segmented sensor to approximate a mode-3 filter.
With this method, we use a known signal to train the segment weights W
3,n
to approach the
weights of a mode-3 sensor. By training the segment weights we mean adjusting the weights
adaptively so that the sensor output (that is, the weighted sum of the segment outputs)
approaches a desired output in the least-squares sense.
...
Error
signal
(k)
Sensor output
&
$
3
( ) k
-
+
W
3,1
(k)
Adaptive
algorithm
V
1
(k)
Beam
PVDF segments
f(k)
T
*
Torsion spring
W
3,2
(k) W
3,10
(k)
Desired signal
d(k)
V
2
(k)
V
10
(k)
...
131
We use a random signal to excite the beam. The desired sensor output is the output that would
result from a modal sensor on the beam under this excitation. To provide this desired signal, we
use a second-order filter with specified natural frequency and damping. Since natural frequency
and damping are global modal properties, they can be obtained from a single FRF of the structure
using, for example, peak-picking and the half-power method.
A representation of the system is the block diagram in Fig. A.2. Note that unlike the fixed gains in
the previous chapter, this time the gains are adjusted from time-step to time-step. The index k
denotes time step.
Figure A.2 Block diagram of the adaptive system.
We shall use the LMS adaptive algorithm

(Widrow and Stearns, 1985) to train the weights
towards the weights of a modal sensor. The algorithm is explained in Appendix C. To do the
weight adjustment, the algorithm needs the sensor output signals V and the error signal , which
is the difference between the desired signal d and the actual sensor output y. All these signals are
available at any time step k up to the present. A digital signal processing (DSP) board can be used
to accept these signal and provide and adjust the weights.
To study the adjustment process, we simulate the system with a discrete-time model. As before,
we model the dynamics of the beam and the second-order filter. Based on the model, we calculate
the responses of the above systems to the random excitation for one time step. Then we calculate
the outputs of the segments from the beam response. The output of a segment is proportional to
the difference between the slopes at the ends of the segment. These segment outputs are then
multiplied by their corresponding weights to give the sensor output. The sensor output is
Error
(k)
-
+
LMS
algorithm
d(k)
Second-order
filter
Segment
output
matrix
Beam
f(t)
Modal
coordinates
(no access)
A/D
A/D
Digital signal
processor
Sensor
output
&
$
( ) k
Adjustable
gain W
Segment
outputs
132
compared to the second-order filter output to produce the error signal. The error signal and the
segment outputs are used to adjust the weights. Then we continue with the next time step and
repeat the above calculations.
In this numerical example, the properties of the beam are as in Table 3.1. The normalized torsion
spring stiffness T* = 1; the excitation point is 0.05715 m from the end that has no spring. The
first ten natural frequencies range from 41 Hz to 3767 Hz. The sampling frequency is set at 15070
Hz. Ten sensor segments are used.
The segment weights were initially set arbitrarily. As training progresses, the weights evolve as
shown in Fig. A.3. At iteration number k = 100, the weights still change rapidly. But the weights
settle after about 500 iterations. At k = 10000, the weights are very similar to the weights at k =
500.
In real (hardware) implementation, all signals outside the DSP box in Fig. A.2 occur as a natural
response of the system to the excitation force. The DSP box adapts the weights towards the
analytical weights for a modal sensor based on the sensor outputs and the excitation force. Note
that the DSP box does not have any information about the mode shapes, sensor properties, etc.
The only information about the structure that is programmed into the DSP box is the natural
frequency and damping ratio of the desired mode. Therefore, only minimum knowledge of modal
properties is required to arrive at the correct segment weights.
The error signal is closely related to the difference between the ideal gains and the gains obtained
from the adaptation process. In the above example, the error signal settle close to zero at time =
0.035 s. At the sampling rate of 15070 Hz, this time corresponds to iteration step k of around
550. Figure A.4 shows the convergence of the error signals (hence the gain matrix).
133
2 10
0
0.15
Segment number
k = 1
-0.15
3 4 5 6 7 8 9 1 2 10
0
0.2
Segment number
k = 100
-0.2
3 4 5 6 7 8 9 1
2 10
0
0.5
Segment number
k = 150
-0.5
3 4 5 6 7 8 9 1 2 10
0
1.0
Segment number
k = 250
-1.0
3 4 5 6 7 8 9 1
2 10
0
1.5
Segment number
k = 500
-1.5
3 4 5 6 7 8 9 1 2 10
0
1.5
Segment number
k = 10000 (-); analytical (--)
-1.5
3 4 5 6 7 8 9 1
Figure A.3 Evolution of segment weights W with adaptation number k.
134
0
1.2
0
0.5
Mode 1
Time (s)
Error
(V)
0
1.2
-0.25
0
0.25
Mode 2
Time (s)
Error
(V)
0
1.2
-0.25
0
0.25
Mode 3
Time (s)
Error
(V)
0
1.2
-0.25
0
0.25
Mode 4
Time (s)
Error
(V)
1.2
-0.1
0
0.1
Mode 5
Time (s)
Error
(V)
1.2
-0.05
0
0.05
Mode 6
Time (s)
Error
(V)
1.2
-0.05
0
0.05
Mode 7
Time (s)
Error
(V)

1.2
-0.05
0
0.05
Mode 8
Time (s)
Error
(V)
Figure A.4 Convergence of error signals.
135
1.2
-0.05
0
0.05
Mode 9
Time (s)
Error
(V)
1.2
-0.05
0
0.05
Mode 10
Time (s)
Error
(V)
Figure A.4 Convergence of error signals.
For the structure in this example, the correct gain matrix is achieved in fewer than 10000
iterations. The above example showed that we can build modal filters without knowledge of the
mode shapes of the structure. In essence, the above method lets the structure compute the
segment gain matrix. Therefore, this modal sensor design method is an improvement over the
fixed-gain method.
Notice that this method uses second-order digital filters to train the gain matrix to converge to
such value that the outputs of the sensor emulate modal coordinates. The coefficients of these
second-order filters must be obtained from the natural frequencies and damping. Therefore, some
knowledge of the modal properties is still required in building modal sensors using this adaptive
method. Sometimes we still have to do a partial modal analysis to obtain the natural frequencies
and damping ratios. At least we need an FRF of the structure to obtain the global modal
properties. A more advanced method of computing the gain matrix for modal sensors would be a
method that does not require prior knowledge of any modal properties of the structure.
136
APPENDIX B
LMS ALGORITHM
Figure B.1 shows N piezoelectric film segments attached to a beam. The output of the n
th
segment
is weighted by its weighting factor W
nk
, where k denotes the time step. The sensor output is,
therefore,
y
k k k
W V . (B.1)
where W
k
is a row vector of the N segment weights, and V
k
is a column vector of the N segment
output voltages.
Figure B.1 Adaptive sensor on a beam.
We want to adjust the weights W
nk
(n = 1, ..., N) so that the sensor output y
k
approaches the
desired signal d
k
. Specifically, we want to minimize the expected value of the squared difference
between d
k
and
k
. The function to be minimized can be shown to be
[46]
[ ] [ ]
+ E E d
k k
T 2 2
2 WRW WP , (B.2)
where E[.] denotes expected value, R is the input correlation matrix
...
d
k
V
2k
V
Nk
k
Error
y
k
Sensor output
-
+
V
1k
W
1k
W
2k
W
Nk
137
[ ]
R V V E
k k
T
, (B.3)
and P is the cross correlation between the desired signal and the input,
[ ]
P V E d
k k
. (B.4)
Minimization of the mean squared error function in Eq. B.2 can be achieved step-by-step using
the gradient of the function, which is
( )

W
WR P
T
T
2 . (B.5)
The steepest-descent method adjusts the weights W
k
according to
( ) W W
k k k +
+
1
, (B.6)
where is a gain constant that determines the adaptation speed and stability.
Equation B.6 shows that the true steepest-descent method requires that the gradient be known at
any time step. Gradient computation is extensive since it involves the computation of the input
correlation matrix R and the cross-correlation vector P. To greatly reduce computation time for
real-time application, we can use an estimate of the gradient instead of the true gradient by
replacing = E[
k
2
] in Eq. B.2 with

k k
2
. (B.7)
This approximation results in an estimated gradient
$

k
k
k
k
k

W
V 2
T
. (B.8)
Substituting this estimated gradient for the true gradient in Eq. B.5 results in a new weight
adjustment scheme,
W W V
k k k k +
+
1
2 . (B.9)
This weight adjustment scheme is the LMS algorithm.
138
APPENDIX C
A CONTROL-MODEL IDENTIFICATION
PROCEDURE
The Markov parameters of the system to be identified are the pulse-response matrix. For an r-
input system the Markov parameters are the response of the system to an excitation
u i r
u k k
i
i
( ) ; , ,...,
( ) ; , ,...
0 1 1 2
0 1 2

(C.1)
Markov parameters can be computed from experimental data by transforming the force and
response signals into the frequency domain using Fast Fourier Transform (FFT), computing the
FRFs of the system, and transforming the FRF back into the time domain. For an m-input-r-
output system, the Markov parameters are m-by-r matrices denoted as
Y
k
th
k Markov parameter (C.2)
The Hankel matrix H(0) is composed of the Markov parameters ordered as
( ) H
Y Y Y
Y Y Y
Y Y Y
0
1 2
2 3 1
1
1
]
1
1
1
1
1
+
+ +
L
L
M M O M

(C.3)
The number of rows and the number of columns are chosen to utilize a sufficient number of
Markov parameters. The higher the order of the system, the more data points are required. The
above matrix can be constructed only if more than + Markov parameters are available.
Modified versions of the above matrix can be formed by deleting some rows and columns to
discard noisy or undesirable data. The Eigensystem Realization Algorithm (ERA)
[43]
uses such a
modified Hankel matrix.
The next step is to perform singular-value decomposition (SVD) of the Hankel matrix, i.e., to
factorize the Hankel matrix H(0) into
H R S ( ) 0
T
(C.4)
139
In the above equation, R and S are matrices with orthonormal columns, i.e. R
T
R = I, and S
T
S = I.
R and S are called the left singular vectors and the right singular vectors, respectively. The
matrix in the middle, , is a square matrix of singular values

1
]
1
n
0
0 0
, (C.5)
with
[ ]

n i i n
diag
+

1 2 1
, , , , , , K K (C.6)
where
i
is monotonically non-increasing,

1 2 1
0
+
L L
i i n
. (C.7)
Among the above values, examine the singular values and discard the last singular values that are
much smaller than the others. The number of remaining singular values is the order of the system.
Next, define R
n
and S
n
as the matrices formed by the first n columns of R and S, respectively. The
SVD of the Hankel matrix becomes
( ) H R S 0
n n n
T
(C.8)
where R
n
and S
n
are orthonormal, i.e. R
n
T
R
n
= I, and S
n
T
S
n
= I.
Now, form the shifted block Hankel matrix
( ) H
Y Y Y
Y Y Y
Y Y Y
1
2 3 1
3 4 2
1 2 1
1
]
1
1
1
1
1
+
+
+ + + +
L
L
M M O M

The next step is to compute the estimates of the state matrix A, input matrix B, and output matrix
C, using
( )
$
/ /
A H S

N N n N
R
1 2 1 2
1
T
(C.9)
$
/
B S E

N n
T
r
1 2
(C.10)
140
$
/
C E R

m
T
n n

1 2
(C.11)
In the last two equations, the column vectors E
m
and E
r
are
E
I
O
O
m
m
m
m
1
]
1
1
1
1
M
(C.12)
where m is the number of outputs, and
E
I
O
O
r
r
r
r
1
]
1
1
1
1
M
(C.13)
where I
r
denotes an r-by-r identity matrix and r is the number of inputs.
The parameters denoted with the caret sign in Eqs. (C.9), (C.10), and (C.11) are the estimates of
the state matrix A, input matrix B, and output matrix C, of the discrete-time system
x k x k u k ( ) ( ) ( ) + + 1 A B , (C.14)
y k x k u k ( ) ( ) ( ) + C D . (C.15)
The parameter D is already known as the zeroth-time-step Markov parameter. Recall that the
Markov parameters are the system response to a pulse excitation. Substituting Eq. (C.1) into Eqs.
(C.14) and (C.15) shows that the Markov parameters are
Y D, Y CB, Y CAB, Y CA B
0
1

1 2 k
k
. (C.16)
If this system in Eqs. (C.14) and (C.15) is the sampled version of a continuous system with
linearly independent eigenvectors, then the modal properties of the continuous system can be
obtained easily, provided that the state matrix A of order n has a complete set of linearly
independent eigenvectors. To obtain the modal properties of the continuous system, first obtain
the eigenvalues and the eigenvectors of A. Define a diagonal matrix of eigenvalues
diag( , , , )
1 2
L
n
(C.17)
141
and a matrix of eigenvectors
[ ]

1 2
, , , L
n
(C.18)
Transform the eigenvalue matrix into the continuous Laplace (s) domain using
( )

c
T
ln
(C.19)
where T is the sampling period.
The natural frequencies and modal damping of the continuous system are the imaginary and real
parts of the eigenvalues
c
. Of course, this transformation assumes that there is no aliasing in the
sampling process. The experimental data must be obtained with a sufficiently high sampling rate.
The mode shapes of the system at the sensor points are

11 2 1 1
1 2
1
, , ,
,
, ,
L
M M
L
n
r n r
1
]
1
1
1
1
C (C.20)
where
i,j
denotes mode shape i at sensor point j.
The modal residues are computed by
b
b
b
n
1
2 1
M
_
,

B (C.21)
142
APPENDIX D
STABILITY OF IIR FILTERS
The mobility of a single-degree-of-freedom (SDOF) system can be represented by a digital filter
&
( )
( )
( )
m m
m m
z
f z
b z z
a z a z
+
+ +

1 2
1
1
2
2
1
(D.1)
where the filter coefficients are
b x T
m m f
m m
m m
m m

( )
sin( )
exp( )
1
1
2
2
, (D.2)
a T
m m m m m 1
2
2 1 exp( ) cos( ) (D.3)
a T
m m m 2
2 exp( ) (D.4)
Equations (D.2) through (D.4) constitute a transformation from the (
n
,
n
, residue) plane into the
filter coefficient plane. While this transformation may seem complicated, the important boundaries
between areas in the s-plane transform into simple lines in the coefficient plane. Figure D.1 gives
some insight into the transformations among the domains by showing the pole locations and the
digital filter coefficients. The poles of the transfer function in Eq. (D.1) determine the stability and
oscillatory behavior of the filter. In the Laplace domain (Fig. D.1.a), the poles must lie in the left-
hand side of the graph, where the real part of the poles are negative. In the z-domain (Fig. D.1.b) ,
the stability region is where the z-plane poles lie inside the unit circle. From Eq. (D.1), this region
is
t < a a a
m m m 1 1
2
2
4 1 (D.5)
Stability also requires that
a
a
m
m
1
2
2
1
<
<
(D.6)
143
The above stability criteria limit the filter denominator coefficients inside a triangular area in Fig.
D.1.c. Any (a
1m
, a
2m
) pair outside this area result in an unstable second-order filter.
In this work, we are interested only in stable, oscillatory, underdamped structures. For these
structures, the discriminant in inequality D.5 must be negative. This condition limits the filter
denominator coefficients inside a parabola in Fig. D.1.c. Thus, we shall limit the filters
denominator coefficients to (a
1m
, a
2m
) pairs inside the parabola for underdamping, and below the
a
2m
= 1 line for stability. The area outside the parabola but inside the V-shaped boundary
represents overdamped filters and, therefore, corresponds to the negative real axis in the s-plane.
-3 -2 -1 0 1 2 3
-3
-2
-1
0
1
2
3
s-plane
Real
Imag-
inary
m
=0.2
m
T/=0.6
m
=0.2
m
=0.8
m
T/=0.2
a)
144
-1.5 -1 -0.5 0 0.5 1 1.5
-1.5
-1
-0.5
0
0.5
1
1.5
m
=0.2
m
T/=0.6
Imag-
inary
Real
m
T/=0.2
m
=0.8
b)
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
a1
a2
m
=0.2
m
T/=0.2
m
T/=0.8
m
=0.8
Overdamped
region
Unstable
region
Unstable
region
c)
Figure D.1 Lines of constant natural frequency and constant damping: a)in the s-plane, b)in the z-
plane, and c)in the coefficient (a
1
,a
2
) plane
145
While the structure to be identified is almost always a stable structure (a plant without feedback
or control), the digital filter that may emerge during the identification process may be unstable.
Many continuous structures have very low damping coefficients. For such structures, the
eigenvalues (poles) of the structures are likely to be very close to the edge of the stability disc in
Fig. D.1.b. For example, for = 1%, which is typical for the structure used in the examples in this
dissertation, the pole locations are as shown in Fig. D.2.
-1 -0.5 0 0.5 1
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Real part
Imaginary
part
Figure D.2 Typical pole locations for a lightly damped structure
146
VITA
Hartono Sumali
Hartono Sumali, also known as Anton, was born in Jakarta, Indonesia, on January 4, 1964. He
graduated from a Jesuit Catholic high school in the same city in 1982. That year he began his
study of mechanical engineering at Bandung Institute of Technology in West Java, where he
received various recognition awards and scholarships. He graduated in 1987, first in his class.
Upon graduating, he was hired as an instrumentation engineer by McDermott, Inc., an offshore oil
platform construction company.
After working for more than two years at that company, he enrolled at the Mechanical
Engineering Department at Virginia Tech, where he received his Masters in mechanical
engineering in 1992. Just before he received his Ph.D. from the same department in 1997, he
accepted an Assistant Professors position at Purdue University.

A New Adaptive Array of Vibration Sensors - Tesis

Enviado por

Dados do documento

Descrição original:

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

A New Adaptive Array of Vibration Sensors - Tesis

Enviado por

Direitos autorais:

Formatos disponíveis

0 0.01 0.02 0.03 0.04 0.05

filter output =modal coordinate

Você também pode gostar