Identification of Civil Engineering Structures Using Vector ARMA Models
Identification of Civil Engineering Structures Using Vector ARMA Models
Identification of Civil Engineering Structures Using Vector ARMA Models
Palle Andersen
Aalborg University
Department of Building Technology and Structural Engineering
Sohngaardsholmsvej 57, DK-9000 Aalborg, Denmark
(http://www.civil.auc.dk/~i6pa/thesis.htm)
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 System Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Non-Parametric Model Structures . . . . . . . . . . . . . . . . . . 3
1.1.2 Parametric Model Structures . . . . . . . . . . . . . . . . . . . . . . 5
1.2 System Identification of Civil Engineering Structures . . . . . . . . . . . 6
1.2.1 Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.2 Vibration Based Inspection . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.3 Excitation of Civil Engineering Structures . . . . . . . . . . . 9
1.3 Scope of the Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Two Experimental Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4.1 Simulation of a Five-Storey Building . . . . . . . . . . . . . . 12
1.4.2 Vibration Based Inspection of a Lattice Steel Mast . . . . 13
1.5 Reader’s Guide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Contents
3.3 Combined Continuous-time Systems . . . . . . . . . . . . . . . . . . . . . . . 52
3.3.1 Generalization to an Arbitrary System Output . . . . . . . . 57
3.3.2 Example 3.1: A White Noise Excited System . . . . . . . . 58
3.3.3 Example 3.2: A Nonwhite Excited System . . . . . . . . . . 58
3.4 Modal Decomposition of Combined Continuous-Time Systems . . 59
3.4.1 Modal Decomposition of the Combined System . . . . . . 60
3.4.2 Effects of Arbitrarily Chosen Output . . . . . . . . . . . . . . 62
3.4.3 Effects of the Shaping Filter Convolution . . . . . . . . . . . 63
3.4.4 Example 3.3: Modal Decomposition . . . . . . . . . . . . . . . 65
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Contents
6 Modal Analysis of Civil Engineering Structures . . . . . . . . . . . . . . . . . . . . . 111
6.1 Modal Analysis of Civil Engineering Structures . . . . . . . . . . . . . 111
6.1.1 Modal Decomposition of Discrete-Time Systems . . . 112
6.1.2 Modal Parameters of Discrete-Time Systems . . . . . . 115
6.2 Estimation of Standard Deviations of Modal Parameters . . . . . . . 116
6.2.1 Natural Eigenfrequencies and Damping Ratios . . . . . . 118
6.2.2 Scaled Mode Shapes . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.3 Spectrum Analysis of Civil Engineering Structures . . . . . . . . . . . 119
6.3.1 Spectrum Analysis using Discrete-Time Systems . . . 120
6.4 Identifying Structural Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.4.1 Mode Identification using Eigenvalues . . . . . . . . . . . . 122
6.4.2 Mode Identification using Mode Shapes . . . . . . . . . . . 123
6.4.3 Other Methods for Mode Identification . . . . . . . . . . . . 125
6.5 A Hybrid Modal Analysis Approach . . . . . . . . . . . . . . . . . . . . . . 126
6.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
Contents
9 Experimental Case No. 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
9.1 Data Acquisition and Signal Processing . . . . . . . . . . . . . . . . . . . . 191
9.2 Virgin State Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
9.2.1 Adequate Modelling of the Dynamic Properties . . . . . 192
9.3 Damaged States Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
9.3.1 Accuracy Check of Estimated Eigenfrequencies . . . . . 199
9.3.2 Determination of Sensitive Eigenfrequencies . . . . . . . 199
9.4 Damage Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
9.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
10.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
10.2 General Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
10.3 Future Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Contents
Preface
In the mid 1980s, researchers at the Department of Building Technology and
Structural Engineering at Aalborg University, Denmark, started using time domain
models for system identification of civil engineering structures. In the past decade the
results of this work have been reported in several papers and in three Ph.D. theses.
A common feature of the work has been the use of the so-called auto-regressive
moving average (ARMA) models for time series modelling. The reason is the ability
of these models to provide an accurate estimate of the modal parameters of a
structural system on the basis of discretely sampled response.
However, the link between the ARMA model and the mathematical description of
civil engineering structures has not been addressed to the same extend as mathemati-
cal description of dynamic systems in fields such as electrical engineering and
econometrics. Therefore, the model has been applied as a grey-box model in the
above references. The relation between the auto-regressive part of the model and the
modal parameters has been well understood, whereas the understanding of the
moving average part has been limited.
In order to obtain a deeper understanding of the ARMA models, and how they are
related to the modelling of civil engineering structures, a Ph.D. project, with this as
its primary objective, was granted as a part of the Danish Research Council frame
programme “Dynamics of Structures”. Another objective of the Ph.D. project was to
implement the obtained knowledge as, especially designed, time domain software for
system identification of civil engineering structures.
The present thesis System Identification of Civil Engineering Structures using Vector
ARMA Models has been developed as a part of this Ph.D. project from September
1993 to May 1997 at the Department of Building Technology and Structural
Engineering at Aalborg University.
I would like to thank my supervisors Rune Brincker, Associate Professor Ph.D., Poul
Henning Kirkegaard, Associate Professor Ph.D., and Lars Pilegaard Hansen,
Associate Professor Ph.D., from the Department of Building Technology and
Structural Engineering at Aalborg University for their guidance, patience and not
least technical and moral support during the years.
I would also like to thank George C. Manos, Professor Ph.D., Milton Demosthenous,
Associate Professor Ph.D., and the staff at the Department of Civil Engineering,
Aristotle University, Thessaloniki, Greece, for their kind support during my four
month stay.
Preface i
Thanks are also given to the staff at the Centre for Dynamic Measurements, Aalborg
University, led by Henning Andersen, engineering assistant, for help during the
performance of the experimental tests.
The figures in the thesis have been prepared by draughtsman Mrs. Norma Hornung
from Aalborg University, and the proof-reading has been performed by senior
secretary Mrs. Kirsten Aakjær from Alborg University. Their carefully performed
work is greatly appreciated.
The economic support given by the Danish Technical Research Council is gratefully
appreciated.
Finally, I would like to thank my daughter Lisa and my wife Ane for their great
patience and moral support during my Ph.D. study.
ii Preface
Nomenclature
General Rules
y( , ÿ( t )
dt dt 2
- Continuous-time stochastic processes are denoted y(t).
- Discrete-time stochastic processes are denoted y(tk).
Scalars
Nomenclature iii
C(q) Moving-average matrix polynomial (Innovation formulation).
Ci ith moving average coefficient matrix.
D Direct term matrix
F Continuous-time state matrix
h() Impulse response function
H() Transfer function
I Identity matrix
K Stiffness matrix of a second-order continuous-time system
K Kalman gain matrix
M Diagonal mass matrix of a second-order continuous-time system
Qr State space Reachability matrix
Qo State space Observability matrix
Modal Parameters
System Identification
Statistical Parameters
iv Nomenclature
$ Steady state covariance matrix of a discrete-time state vector process x(tk).
Steady state covariance matrix of the prediction of the state vector process
x(tk).
P Steady state covariance matrix of the state prediction errors.
W Continuous-time Gaussian white noise intensity matrix /
Discrete-time Gaussian white noise covariance matrix.
Operators
AR: Auto-Regressive
ARMA: Auto-Regressive Moving Average
ARV: Auto-Regressive Vector
ARMAV: Auto-Regressive Moving Average Vector
MDOF: Multi Degree Of Freedom
RMS: Root Mean Square
SDOF: Single Degree Of Freedom
Nomenclature v
vi
1 Introduction
The past 30 years have witnessed major developments in the theory and application
of linear systems, work that was heavily influenced by results of Kalman in the early
1960s, see Kalman [49] and Kalman et al. [50]. Certain essential advances have been
made in understanding the algebraic and topological structure of linear dynamic
systems, with much of this work originating from systems and control engineering.
At the same time, these systems, especially for the univariate (scalar) output case,
have been widely used to model and statistically treat data arising in signal
processing. In particular, electrical engineers have developed algorithms for on-line
and real-time model estimations, see Ljung et al. [73] and Young [113]. Finally,
statistical time-series analysts, motivated by applications coming from a wide variety
of fields, have developed theory and algorithms primarily for off-line model
estimation, see e.g. Akaike [1] and Hannan [32]. The work of Box and Jenkins, see
Box et al. [16], played a major part in this development. Statisticians have also
developed the required asymptotic theory associated with estimation procedures, see
e.g. Whittle [112].
In this chapter the basic concepts of system identification are introduced. Also
introduced, are some of the applications of system identification in civil engineering.
The introduced applications are modal analysis and vibration based inspection.
Finally, in the end of this chapter the scope of the work of this thesis is stated.
/ Physical modelling.
/ System identification.
Introduction 1
Thus, system identification should not be thought of as a substitute of physical
modelling, since identification can be based on model structures that have physical
origin. Basically there are two categories of model structures:
In any case, physical modelling will always be linked with parametric model
structures. Common to both categories of model structures is that they depend on the
applied excitation, which may be one of the following
/ Instantaneous excitation.
/ Periodic excitation.
/ Pseudo-random periodic excitation.
/ Stochastic excitation.
In the case of instantaneous excitation, the system is either given an impulse or step
excitation and the system is the left vibrating on its own. The excitation may or may
not be measured. It is also possible to excite the system with a known periodic
excitation, such as sinusoidal, or several periodic signals mixed to obtain a pseudo-
random periodic excitation. Finally, as an alternative to the deterministic excitation,
the excitation might also be a stationary stochastic process with either known or
unknown statistical properties.
The combination of linear system theory, time series analysis and asymptotic theory
is the foundation of modern system identification. System identification basically
means modelling of the dynamic systems from experimental data. This general
definition indicates that system identification is applicable in many different
engineering fields.
A common feature of all these examples is that their dynamic behaviour can be
conceptually described as in figure 1.1.
2 System Identification
Figure 1.1: A dynamic system with input u(t), output y(t) and disturbance v(t).
The system is driven by input u(t) and affected by disturbance v(t). In some cases the
user can control the input u(t) but not the disturbance v(t). It might also be that the
actual input is unknown and therefore uncontrollable in some applications. The
output y(t) describes how the system reacts or responds to the applied input and
disturbance. Therefore, the output will be a mixture of dynamic response of the
system and characteristics of the input and disturbance as well. In general, the input
at previous time instances will also affect the current output. In other words, the
dynamic system has memory.
/ Transient analysis.
/ Frequency analysis.
/ Correlation analysis.
/ Spectral analysis.
Transient analysis is applied when the system response is transient, i.e. generated on
the basis of impulse or step excitation. The dynamic behaviour of the system is then
identified on the basis of the impulse or step response. Frequency analysis is applied
when the excitation is deterministic and either periodic, or pseudo-random and
periodic. The measured excitation and corresponding system response is transformed
to frequency domain, and the frequency response function is obtained as the ratio of
the transformed response and excitation. Correlation and spectral analysis are
methods that are applied to a stationary stochastically excited system. In these cases,
the excitation and the system response can be characterized either by the correlation
functions in time domain or the spectral densities in frequency domain. Having
estimated the correlation functions of the excitation and the response the impulse
response function of the system can be obtained. On the other hand, if the spectral
densities of the excitation and response are estimated instead, it is possible to obtain
the frequency response function.
Introduction 3
The traditional non-parametric system identification techniques are primarily based
on Fourier transform techniques, see e.g. Bendat et al. [14]. The Fourier transform
techniques are elegant mathematical tools, which are ideal for theoretical analysis of
dynamic systems. The Fourier transform of uniformly sampled data is usually
performed by the Fast Fourier Transform (FFT). The FFT algorithms have been a key
in system identification and signal processing, since the arrival in the mid 1960's, see
Cooley et al. [19]. The FFT is still a popular algorithm, and is today the standard tool
in commercial Fourier analysers. The reason for this popularity is the speed and
reliability of the FFT algorithms. Also, the user does not have to interact to the same
extent as it is necessary in system identification using parametric model structures.
However, the FFT has some limitations, which have completely escaped the attention
of most practitioners. The most obvious limitation is that the FFT assumes periodic
data, which is certainly not the case for sampled response from stochastically excited
structures. For data records with such response, it is quite certain that:
In principle, the Fourier transform assumes that the amount of data is infinite. In this
case the frequency response functions or the spectral densities will have an infinite
frequency resolution. However, in practice the available data records have a finite
length, resulting in a finite frequency resolution, see e.g. Schmidt [100].
The FFT assumes periodicity, i.e. that the finite data record of length repeats itself in
both ends of the record. Since sampled stochastic signals in general exhibit non-
periodicity, errors will certainly be introduced. These kinds of errors are called
leakage, since the energy of the resonance frequencies leaks out. In frequency domain
the effect of leakage is a seemingly higher damping of the corresponding modes.
Further, in the case of closely spaced modes, it might be impossible to separate these
if one of the resonance frequencies has a small amplitude compared to the other. In
this case the resonance frequency with the smallest energy content may be masked
completely by the resonance frequency with highest energy content.
The leakage errors are compensated by windowing the data before the FFT is applied,
to secure periodicity of the data by damping the discontinuities at the ends of the data
record. The problem of windowing is, that it introduces an extra damping into the
system, and thus creates its own leakage problem. Leakage is a systematic error.
However, there are also the random measurement errors that contaminate the data.
To eliminate these random errors the data is averaged, either before or after the FFT
has been applied. Several forms of averaging exist. A typical approach, though, is
linear averaging of segments of the data records after the FFT has been applied to
each segment, see Kay [51]. This kind of averaging will in the limit eliminate the
random errors as a consequence of the central limit theorem, but not the leakage.
However, for a fixed amount of data, averaging will be performed on the expense of
the frequency resolution.
4 System Identification
The finite length of the data record used in the FFT, instead of the theoretically
infinite length assumed in the Fourier transform, limits the lowest frequency and the
frequency resolution and causes leakage when violated. On the other hand, the finite
sampling interval used in the FFT, instead of the zero interval assumed in the Fourier
transform, limits the highest frequency to the Nyquist frequency and causes aliasing
when violated. Aliasing superimposes the contributions of all frequencies beyond the
range of zero and twice the Nyquist frequency, by folding around limits as many
times as necessary. If aliasing is not taken into account, it might seriously distort any
data analysis. If the sampling frequency is taken at least twice the frequency beyond
which the energy is nearly zero, then the aliasing will be negligible. But for a fixed
amount of data, a high sampling frequency can only be obtained at the expense of the
frequency resolution. Thus, inadequate frequency resolution or leakage and aliasing
are closely connected.
So in conclusion:
The general input / output model structure used for modelling of linear and time-
invariant dynamic systems excited by deterministic input, is Auto-Regressive Moving
Average with eXternal input (ARMAX)
y( t )
G( q ) u( t ) H( q ) e( t ) (1.1)
Introduction 5
where G(q) and H(q) are the transfer functions of the deterministic part and the
stochastic part. The stochastic input e(t) are the innovations, which is an equivalent
process of the noise and prediction errors. If H(q) = I (1.1) is called an output error
(OE) model. In any case the dynamic properties of the system are modelled by G(q).
A parametric model structure is called multivariable when it includes several
variables. If there are several outputs, it is characterized as a multivariate model
structure. If it only has one output, it is termed an univariate model structure.
The dynamical properties as well as the noise are now modelled by the same transfer
function H(q). In the multivariate case the model structure is called an Auto-
Regressive Moving Average Vector (ARMAV) model. As observed the choice of
model structure depends on whether the input is deterministic or stochastic, i.e.
whether the excitation of the structural system is known and measured or unknown.
It also depends on whether the system is stationarily excited, or excited by an impulse
or step excitation.
Due to the complexity of the multivariate parametric models, a lot of research on the
special properties of multivariate models was performed in the early eighties, see
Gevers et al. [29] and Hannan et al. [34]. This complexity and the computational
effort needed to estimate these models is probably some of the reasons why the FFT-
based non-parametric methods are still preferred by a lot of civil engineers.
/ Modal analysis.
/ Vibration based inspection.
Modal analysis covers a variety of applications all based on the analysis of modal
parameters. These parameters describe specific dynamic characteristics of the
structure. One of the applications that uses the modal parameters as basis is vibration
based inspection.
/ Modal frequencies.
/ Modal damping.
/ Modal vectors.
/ Modal scaling.
Since the modal parameters are directly related to the impulse and frequency response
functions, as well as the correlation functions and spectral densities, they can be
extracted from the non-parametric system identification methods by applying
different curve fitting procedures. In case of parametric system identification methods
there are direct mathematical relationships between the modal parameters and the
estimated model parameters. When modal parameters are used as mathematical
model of the dynamic behaviour of a system, the derived model is called a modal
model. It is therefore common to use experimental modal analysis as a synonym for
system identification.
Introduction 7
Typically, a VBI programme uses the modal parameters to describe the dynamic
characteristics of a structure. A synonym for the dynamic characteristics used as basis
for the VBI programme is damage indicators. In other words:
Often VBI has at random been referred to as damage detection. However, in Rytter
[96] it has been put in a right perspective as a part of a VBI programme. VBI can be
divided into the following four levels:
/ Level 1 - Detection.
/ Level 2 - Localization.
/ Level 3 - Assessment.
/ Level 4 - Consequence.
Methods of the first level give a qualitative indication that a structure might be
damaged. Level two methods give information about the probable location of the
damage as well. Methods of the third level provide information about the size of the
damage, and finally the level four methods give information about the actual safety
of the structure given a certain damage state. The use of a damage indicator primarily
gives a qualitative indication of the existence of damage, and should therefore be
characterized as a level 1 method. However, some of the damage indicators will in
some cases give rough estimates for the locations of damage, which is equivalent to
a primitive level 2 method.
Changes in natural eigenfrequencies are no doubt the most used damage indicators.
One of the reasons for this is that the natural eigenfrequencies are rather easy to
determine with a high level of accuracy. Another reason is, that they are sensitive to
both global and local damages, see Rytter [96]. So comparison of estimates of natural
eigenfrequencies is usually an effective level 1 method. A local damage will cause
changes in the derivatives of the mode shapes at the position of the damage. This
means that a mode shape having many coordinates or measurement points will be a
fast way to locate the approximate position of a damage. They can therefore be
characterized as a simple level 2 method. The introduction of damage in a structure
will usually cause changes in the damping capacity of the structure. In Rytter [96],
it has been shown that the damping ratios are extremely sensitive to the introduction
of even small cracks in a cantilever beam. However, dealing with real structures, the
estimation of the damping ratios of the individual modes is highly sensitive to time-
varying and nonphysical sources. Thus, a satisfactory accuracy of the estimates of the
damping ratios will in general be impossible to obtain. Therefore, the damping is
applicable as a damage indicator, but it cannot and should not be used as the only
damage indicator.
So in conclusion:
In this context the computational effort spent in obtaining reliable estimates is not so
important. Further, the limitations and systematic errors of the traditional FFT-based
non-parametric system identification techniques motivates the use of other
techniques.
/ The need for a more accurate estimation of the modal parameters from
sampled data, compared to what traditional FFT-based non-parametric
techniques can provide.
This need is basically the reason for using the parametric models in the system
identification, since the physical knowledge about a dynamic system in this way is
incorporated into the system identification process.
In the case of civil engineering structures there will most likely be a natural excitation
of the structure such as wind or waves. These natural forms of excitation are
commonly called ambient excitation and the vibrations of the structure caused by
them are called ambient vibrations. System identification of structural dynamics on
the basis of ambient excitation is also referred to as ambient testing. From an
experimental point of view, the simplest approach to measure the dynamic
parameters of a structure is to detect the response due to ambient excitation. In the
case of very large structures this approach is the only practical way of performing
dynamic tests, it is simply impossible to excite such structures artificially. The
ambient excitation is stochastic in nature. Therefore, it cannot be described by an
explicit time-dependent function, but must be characterized by certain statistical
parameters, such as its mean and covariance function. Since the structural system can
be seen as a linear transformation of the applied input, this means that the response
will also be stochastic, and may as such also be represented by its statistical
characteristics. In Ibánes [31], Jensen [42], Rubin et al. [95] and Srinavasan et al.
[104], it has been shown that ambient excitation provides a quick, inexpensive and
reliable way for testing of large civil engineering structures, such as buildings and
offshore structures.
Introduction 9
An extensive survey of available literature concerning full-scale measurements on
offshore platforms has been performed in Jensen [42]. It is found that the typical
excitation of offshore platforms for system identification is ambient excitation.
Further, in Morgan et al. [80], it has been concluded that parameter estimates
obtained by ambient excitation are as good as parameter estimates obtained by
external excitation. This conclusion is based on a study of several published results
of ambient versus forced vibration tests of high-rise structures in the USA.
Because of the nature of dynamic testing under ambient excitation conditions, this
method has advantages over others, such as the impulse and periodic excitation.
Ambient excitation has a broad frequency range, and thus theoretically excites all
relevant modes of a structure. Also, the use of ambient excitation in dynamic testing
does not disturb the normal functioning of a structure and no excitation equipment
is required for ambient testing. However, the disadvantage of ambient excitation is
that its characteristics cannot be controlled and measured directly.
These statements imply that the parametric models of section 1.1.2 are applicable to
system identification of civil engineering structures when a high degree of accuracy
is needed.
Therefore, the main aim of this thesis is to investigate how to represent ambient
excited civil engineering structural systems by stochastic time-domain models, and
how to estimate these on the basis of sampled structural response data.
Since measurements of the true ambient excitation are not available, system
identification using standard multivariate input / output ARMAX model cannot be
applied. The focus is therefore put on the use of stochastic models of the Auto-
Regressive Moving Average Vector (ARMAV) type. These models will be shown to
be directly related to stochastic state space systems. Particular emphasis will be put
on computationally accurate system identification methods, since the intention is to
provide a more accurate alternative to the traditional non-parametric system
identification methods typically applied in the field of civil engineering.
Introduction 11
time-lags, a sensible choice of representation would probably be the ARMAV model,
because of its recursive structure. But in any case, it is important to notice, that the
both representations describe the same system.
Since modal analysis is one of the main reasons for system identification of civil
engineering structures, a thorough treatment of the modal decomposition and
extraction of modal parameters will be given. In addition, guidelines for estimation
of the uncertainties of the estimated modal parameters will be given.
During the Ph.D. project emphasis has been put on the practical implementation of
user-friendly system identification software. Since this work is an essential part of the
project, this software will be described. Two experimental cases are shown. Through
these the performance of the numerical algorithms are tested, and the application of
modal analysis as basis for damage detection is shown.
The purpose of this experimental case is to verify the asymptotic properties of the
estimated ARMAV model. However, since the estimated model parameters are of
minor importance than the estimated modal parameters, the analysis will be
performed on the basis of the estimated modal parameters. The analysis will show
how the estimated modal parameters, their bias, and their estimated standard
deviations depend on the following variables:
The choice of other variables such as the sampling interval will also influence the
parameter estimates, see Kirkegaard [53], but in this analysis the sampling interval
will be fixed. The analysis will be based on simulated response of a five degree-of-
freedom linear system. It is intended that the modal properties of the system should
match the modal properties of a typical civil engineering structure. The linear system
being simulated is a model of a two-dimensional simply supported five storey frame
structure. This structure is illustrated in figure 1.2. The system response to a
stationary Gaussian white noise excitation w(t) is assumed to have been measured at
each floor of the structure. The excitations and the displacements of all five storeys
will be used in the analysis. As seen in figure 1.2, the displacement of the fifth storey
will be first element of the displacement vector y(t).
Figure 1.3: The lattice steel test mast seen from the east and north sides.
Introduction 13
An elevation of the 20 m high steel lattice test mast seen from the west is shown in
figure 1.4a. In one of the lower diagonals, which is marked in figure 1.4a, a damage
has been simulated by introducing a crack and increasing its depth. The depth of the
crack has been increased four times, see figure 1.4b. Before the damage is introduced,
the state of the structure is referred to as the virgin state. After the introduction of the
damage the four different states of the structure are referred to as damage states. The
mast has been equipped with six accelerometers shown in figure 1.4a. Three of them
are mounted in the top of the mast and three approximately in the middle of it.
Unfortunately one of the accelerometers placed in the middle was damaged due to
heavy rain, which means that the analysis has been performed with the remaining five
accelerometers. Figure 1.4c show a sketch of where the accelerometers are mounted
and their sensitive directions are indicated with arrows. The damaged accelerometer
has the number 2.2.
Figure 1.4: a) Elevation of the mast, where the location of the introduced damage
and the six accelerometers are marked. b) The simulated damage is
made by increasing the crack depth in four steps. These steps are
referred to as damage states 1 to 4. c) The top and middle cross-section
where the accelerometers are mounted. The sensitive directions of the
accelerometers are marked with arrows. The damaged accelerometer has
the number 2.2.
The location and orientation of the accelerometers makes it possible to measure both
translational and rotational vibrations. The accelerometers have the channel numbers
shown in table 1.1 in the analysis in chapter 9.
The reductions of the cross-sectional area of the diagonal due to the increase of the
crack are shown in table 1.2 together with the actual crack depths.
The question is at what damage state the damage can be detected with a significant
confidence. This question will be answered in chapter 9. This chapter describes the
actual data acquisition and the subsequent signal processing of the data. It also
describes the modal analysis made in the virgin as well as the damaged states, and
how the damage is detected.
Introduction 15
1.5 Reader’s Guide
In this section the organization of the thesis is presented to give the reader an
overview.
/ Chapters 5-7 explain how to estimate stochastic models from sampled data
and how to apply these models to modal analysis in theory and practice.
Since the secondary purpose of this thesis is to make the theory of system identifica-
tion using stochastic time domain models more accessible to civil engineers, a brief
introduction into the relevant theory of multivariate stochastic time-domain models
is given in chapter 2. This chapter presents the necessary tools needed to handle
stochastic state space systems and ARMAV models in an efficient manner.
Chapter 6 concerns modal analysis based on the estimated stochastic models. In this
context, it is shown how to separate physical and nonphysical modes, and how to
estimate the uncertainties in the estimated modal parameters.
16 Reader’s Guide
The first experimental case is given in chapter 8. The purpose of this case is to test
the performance of the PEM identification algorithms through simulations. It is also
the intention to verify that reliable estimates of the modal parameters and their
uncertainties can be obtained.
The second case is given in chapter 9. The purpose of this case is to show how
system identification and modal parameter estimation can be used as a basis for
vibration based inspection. This case will only consider the first level of vibration
based inspection, which concerns the detection of damage.
Finally, in chapter 10, conclusions are made, and future perspectives are discussed.
Throughout this thesis definitions are used and theorems are stated. To separate these
important parts of the text from the rest, they are presented with indented text in
smaller types. To indicate the end of these parts a box = is shown. Authors will be
referenced by the last name of the first author and a reference number given in
brackets []. In cases where coauthor exist they will be indicated by et al.
Introduction 17
18