Professional Documents
Culture Documents
VE BR
NI
U VRIJE UNIVERSITEIT BRUSSEL
US
E
VRIJ
S
FACULTEIT TOEGEPASTE WETENSCHAPPEN
EL
SCI E
RAS
VAKGROEP WERKTUIGKUNDE
EB
NT
IA
Pleinlaan 2, B-1050 Brussels, Belgium
N
V IN TE
CERE
Bart CAUBERGHE
May 2004
Bart CAUBERGHE – APPLIED FREQUENCY-DOMAIN SYSTEM IDENTIFICATION
US
E
VRIJ
S
FACULTEIT TOEGEPASTE WETENSCHAPPEN
EL
SCI E
RAS
VAKGROEP WERKTUIGKUNDE
EB
NT
IA
Pleinlaan 2, B-1050 Brussels, Belgium
N
V IN TE
CERE
Bart CAUBERGHE
May 2004
c Vrije Universiteit Brussel – Faculteit Toegepaste Wetenschappen
Pleinlaan 2, B-1050 Brussel (Belgium)
Alle rechten voorbehouden. Niets uit deze uitgave mag vermenigvuldigd en/of
openbaar gemaakt worden door middel van druk, fotocopie, microfilm, elektronisch
of op welke andere wijze ook zonder voorafgaande schriftelijke toestemming van
de uitgever.
All rights reserved. No part of the publication may be reproduced in any form by
print, photoprint, microfilm or any other means without written permission from
the publisher.
Acknowledgements
It is very difficult to express in words the great feelings of gratitude I feel for
the people, who helped me making this work possible. First of all, I’m specially
grateful to my supervisor, prof. dr. ir. Patrick Guillaume, for his support, en-
couragement and guidance throughout the course of this work. I thank him for
giving me his time and advices, for sharing his scientific knowledge and for giving
me all resources needed to accomplish this work.
I would also like to thank all members of the jury for their time and interest
in my work. I’m specially grateful to prof. dr. ir. Rik Pintelon for the several
interesting discussions and to dr. ir. Peter Verboven for the careful reading of my
text.
I thank all me colleagues of the Acoustics and Vibration Research Group, Pe-
ter Verboven, Steve Vanlanduit, Eli Parloo, Gert De Sitter, Joris Vanherzeele and
Gunther Steenackers for the stimulating work environment, for their excellent co-
operation and many joint publications. Many colleagues became close friends and
together we spent nice times besides the daily work.
Furthermore I would like to thank the different partners from the FLITE project,
especially dr. ir. Bart Peeters and dr. ir. Herman Van der Auweraer from
i
ii
LMS Intl., the engineers from Airbus, Dassault and PZL, prof. Albert Benveniste
from INRIA, Ivan Goethals from K.U.Leuven and finally ir. Carlos Refinetti from
Embraer to give me insight in real-life applications of modal parameter estimation
methods and for the enthusiasm they have shown for my research work. Many of
our discussions at different meetings and conferences lead to new innovative ideas,
which contributed greatly to the successful accomplishment of this thesis.
I would like to express my gratitude to Prof. Francesco Benedettini and his re-
search team from the University of L’Aguilla for inviting my to join them in a test
campaign on bridges.
The Fund for Scientific Research Flanders (FWO) is gratefully acknowledged for
my research associate grant and the Research Council (OZR) of the VUB for their
financial support.
I would also like to thank Thierry Lenoir for his help with computer problems
over the years and the secretaries of our department and prof. dr. ir. Dirk
Lefeber, head of the mechanical engineering department, for all the given support.
Bart Cauberghe
Nomenclature
List of operators
i i2 = −1
T
(·) matrix transpose
−1
(·) matrix inverse
∗
(·) complex conjugate
H
(·) Hermitian transpose:
complex conjugate transpose of matrix
†
(·) Moore-Penrose pseudo-inverse
−T
(·) Transpose of the inverse matrix
−H
(·) Hermitian transpose of the inverse matrix
Re (·) real part of
Im (·) imaginary part of
|x| absolute value of a complex number x
E (X) mathematical expectation
of a stochastic variable X
var(x) variance of a scalar x
σx2 = var(x)
cov(x) covariance of a vector x
tr(x) trace of a matrix x
iii
iv
List of symbols
Ni number of inputs
No number of outputs
Nm number of modes
N number of frequencies
number of time samples
Nref number of references
Nb number of blocks
n model order
∆t sampling period
ω angular frequency
y(t) time-domain vibration response
x(t) time-domain state sequence
f (t) time-domain force signal
x(i), xi sample at time i∆t of a signal x(t)
Y (ω) response spectra, Fourier spectra of x(t)
F (ω) input force spectra, Fourier spectra of f (t)
X(ω) state spectra, Fourier spectra of y(t)
X(ωk ), Xk spectral line k of the Fourier spectra X(ω)
H(ω) frequency Response Function
λr system pole r
ωr natural frequency of mode r
σr damping of mode r
φr mode shape vector of mode r
Lr modal participation vector of mode r
M , K, C1 mass, stiffness and damping matrix
v
List of abbreviations
CMC Changes in Mode shape Curvature
CMF Changes in Modal Flexibility
ABS Averaged Based Spectral
AR Auto Regressive
ARMA Auto Regressive Moving Average
BTLS Bootstrapped Total Least Squares
CLSF Combined Least Squares Frequency-domain
DOF Degree Of Freedom
EMA Experimental Modal Analysis
ERA Eigenvalue Realization Algorithm
EVD Eigen Value Decomposition
DFT Discrete Fourier Transform
FDPI Frequency Domain Direct Parameter identification
FEM Finite Element Model
FFT Fast Fourier Transform
FRF Frequency Response Function
GEVD Generalized Eigen Value Decomposition
GTLS General Total Least Squares
IDFT Inverse Discrete Fourier Transform
ITD Ibrahim Time-Domain algorithm
IQML Iterative Quadratic Maximum Likelihood
LMFD Left Matrix Fraction Description
LS Least Squares
LSCE Least Squares Complex Exponential algorithm
LSCF Least Squares Complex Frequency-domain
LSFD Least Squares Frequency Domain
MDOF Multiple Degree of Freedom
ML Maximum Likelihood
MIMO Multiple Input-Multiple Output
MPE Modal Parameter Estimation
MSE Mean Square Error
vi
Acknowledgements i
Nomenclature iii
Contents vii
1 Introduction 1
1.1 Research context . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Focus, outline and organization of the thesis . . . . . . . . . . . . . 5
1.3 Original contributions in this work . . . . . . . . . . . . . . . . . . 8
vii
viii Contents
10 Conclusions 227
10.1 Summary and main contributions . . . . . . . . . . . . . . . . . . . 227
10.2 Future research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
Bibliography 232
Chapter 1
Introduction
This chapter contains the general introduction and motivation of the research that
was conducted in the frame work of this thesis. The research context is described in
paragraph 1.1. The focus of the thesis and the organization of the text are discussed
in paragraph. 1.2, while paragraph 1.3 gives an overview of the new contributions
in this work.
1
2 Chapter 1. Introduction
During the last decade modal analysis has become a key technology in structural
dynamics analysis. Starting from simple techniques for trouble shooting, it has
evolved to a ”standard” approach in mechanical product development. Beginning
from the modal model, design improvements can be predicted and the structure
can be optimized. Based on the academic principles of system identification, ex-
perimental modal analysis helps the engineers to get more physical insight from
the identified models. Continuously expanding its application base, modal anal-
ysis is today successfully applied in automotive engineering (engine, suspension,
body-in-white, fully trimmed cars, ...), aircraft engineering (ground vibration test,
landing gear, control surfaces, in-flight tests), spacecraft engineering (launchers,
antennas, solid panels, satellites,...), industrial machinery (pumps, compressors,
turbines, ...) and civil engineering (bridges, off-shore platforms, dams, ...).
Experimental modal analysis (EMA) identifies a modal model from the measured
forces applied to the test structure and the measured vibration responses. The
modal model expresses the dynamical behavior of the structure as a linear combi-
nation of different resonant modes. Each resonance mode is defined by a resonance
frequency, damping ratio, mode shape and participation vector. These modal pa-
rameters depend on the geometry, material properties and boundary conditions
of the structure. Vibrations of the structure originate from its resonance modes
that are inherent properties of the structure. Forces exciting the structure at one
or more of these resonance frequencies cause large vibration responses resulting in
possible damage, discomfort and malfunctioning.
1. Design of the test setup and conducting the experiments: concerns the ex-
perimental setup (e.g., placement of sensors (and actuators), boundary con-
ditions, . . . ) and the data acquisition parameters, (e.g., measurement time,
sampling frequency, . . . ). During an EMA, the test structure is excited by
means of artificial excitation devices (shakers, impact hammers). For OMA
testing, freely available ambient excitation sources, are usually considered as
1.1. Research context 3
For engineers, the modal model is often not the final goal, but only an inter-
mediate result that can be used for a wide range of applications:
• Response prediction: modal models are often used for the purpose of predict-
ing the response of the structure to a given dynamic loading. These loadings
usually corresponds to the forcing sequences encountered during the real-life
operating conditions. This way, designers can check the robustness of the
developed product under a variety of working conditions.
• Sensitivity analysis and structural modification: the modal model can be
used in order to predict the effect of structural modifications to a test struc-
ture [125]. For instance, if a structure (e.g., prototype) suffers from a vibra-
tional problem, a variety of structural modifications (that attempt to solve
the problem) can be evaluated without actually having to apply any high-
cost changes to the prototype. As long as the structural modifications can
be considered small, a (linear) sensitivity analysis can be used in order to
predict the most sensitive areas of the structure for applying a structural
modification that aims at solving the problem.
• Model updating: the initial values chosen for the material properties, geomet-
rical properties and boundary conditions of a FEM often do not guaranty a
reliable model of the structure under test. For this reason, the use of exper-
imental data is required to update (correlation, optimization, verification)
the initial model and produce an FEM which can more reliably predict the
(dynamical) behavior of the structure [33].
• Sub-structuring: given the modal model of different components of a complex
structure, the dynamical behavior of the complete structure can be computed
by using sub-structuring techniques [72].
4 Chapter 1. Introduction
Furthermore, this thesis extends the EMA and OMA concepts, to a so-called com-
bined EMA-OMA framework. In this combined framework, the vibration response
is considered as a result of both measured artificial applied forces and unmea-
surable ambient excitation, i.e. an Operational Modal Analysis with eXogenous
inputs (OMAX), resulting in a maximum data exploitation. In the OMAX frame-
work, the stochastic contribution, i.e. the part of the response which can not
be related to the measurable input forces, is considered as valuable information.
Under the assumption that the stochastic contribution in the response is related
to unmeasurable ambient forces, extra information about the system can be ex-
tracted from this contribution i.e. both modes excited by the measurable and/or
by the unmeasurable forces can be identified from the data.
The following gives a short overview of the content of the different chapters.
A chapter-by-chapter outline is also presented in figure 1.1
In chapter 3, an overview of ABS function (i.e. FRFs for EMA, power spectra
for OMA, and both simultaneously for OMAX) estimators is given. The differ-
ent non-parametric FRF and power spectra estimators are discussed and special
6 Chapter 1. Introduction
Modal Testing
Chapter 3
Chapter 2
Construction of clear
stabilization charts
Chapter 9
Modal Applications
The research presented in this paper is based on several presentations given at in-
ternational conferences, published and submitted in international journals. Next,
an overview is given of the most important contributions together with their ref-
erences:
Frequency-domain models
for dynamical structures
In this chapter, different mathematical models are introduced to describe the dy-
namical behavior of vibrating structures. Their relation with the physical param-
eters of the modal model of a structure is established. This chapter discusses
common-denominator models, left- and right fraction description models and state-
space models. Finally, the concepts of an experimental modal analysis, operational
modal analysis and combined experimental-operational modal analysis are intro-
duced
11
12 Chapter 2. Frequency-domain models for dynamical structures
2.1 Introduction
This chapter discusses several mathematical models that can be used to describe
the vibrational behaviour of a structure with a limited number of parameters.
From an engineering point of view the modal model of a structure provides the
best physical understanding. However, since this model is highly non-linear in its
parameters most identification algorithms do not directly identify the model pa-
rameters. Instead, the modal parameter estimation (MPE) methods proposed in
the next chapters identify scalar matrix fraction models, also known as common-
denominator models, left-and right matrix fraction description models and state-
space models from the experimental measurements. In the next sections, the rela-
tion between these models and the modal parameters are discussed. Furthermore
a distinction between continuous-time and discrete-time models is discussed. A
distinction is made between data driven and spectral function driven identification
algorithms. Finally, it is shown how these models can be used to identify model
parameters from both input-output measurement and from output-only measure-
ments in absence of the input measurements and the concepts of a Experimen-
tal Modal Analysis, a Operational Modal Analysis and Combined Operational-
Experimental Modal Analysis are introduced.
with H(s) = Z −1 (s) the transfer function matrix. The transfer function matrix
can be formulated in its modal form [52], [72]
−1 −1 H
H(s) = φ sINm − Λ LT + φ∗ sINm − Λ∗
L (2.4)
2.3. Common-denominator models 13
where the modal parameters λr , φr and Lr are respectively the pole, mode shape
and modal participation factor of mode r. The diagonal matrix Λ is given by
Λ = diag {λ1 , λ2 , . . . , λNm } (2.5)
The common-denominator, also called scalar matrix fraction model, considers the
relation between output o and input i as a rational fraction of two polynomials, of
which the denominator polynomial is common for all input-output relations. The
transfer function matrix H(s) can be expressed as
Zadj (s)
H(s) = (2.6)
|Z(s)|
with Zadj (s) the adjoint matrix, containing polynomials of order 2Nm − 1. The
common-denominator is then given by the characteristic equation |Z(s)|, a poly-
nomial in s of order 2Nm , which roots are the poles of the structure. In general
the common-denominator model can be expressed as
B11 (s) . . . B1Ni (s)
.. .. ..
. . .
BNo 1 (s) . . . BNo Ni (s)
H(s) = (2.7)
A(s)
The relation between the modal model and the common-denominator model is
obtained by considering the frequency response function between output o and
input i
Nm
φ∗ L∗
X φor Lir
Hoi (s) = + or ir∗
r=1
s − λr s − λr
Boi (s)
= (2.8)
A(s)
From this equality it is clear that the structure poles are given by the roots of the
denominator A(s), while the mode shapes and participation factors are obtained
from a singular value decomposition (SVD) of the residue matrix Rr ∈ C No ×Ni of
mode r
Rr = φr LTr (2.9)
14 Chapter 2. Frequency-domain models for dynamical structures
and φr = [φ1r φ2r . . . φNo r ]T , Lr = [L1r L2r . . . LNi r ]T . From modal analysis
theory it follows that this residue matrix is of rank 1. Nevertheless the common-
denominator model does not force rank 1 residue matrices on the measurements.
The implications on the modal model are discussed in chapter 5. For a common-
denominator model each relation between one output and one input can be con-
sidered separately, this turns out to be an advantage in terms of the optimization
of the calculation speed of the identification algorithms proposed in chapter 4.
An other type of models are the so-called state-space models, which introduce the
concept of the states of a dynamical system. Reformulating Eq. 2.2 as
sY (s) 0 I Y (s) 0
= + (2.12)
s2 Y (s) −M −1 K −M −1 C1 sY (s) M −1 F (s)
with
0 I
A = ,C= I 0 (2.15)
−M −1 K −M −1 C1
0
B = and D = [0] (2.16)
M −1
AV = V Λ (2.19)
This last expression is totally equal to the more commonly used expression of
the modal model given by Eq. 2.4. In the most general case the number of
responses No differs from the number of modes Nm and consequently A ∈ R n×n ,
C ∈ R No ×n , B ∈ R n×Ni and D ∈ R No ×Ni with the model order n = 2Nm .
CAn−1
..
. = −CAn
An−1 . . . A1 A0 (2.23)
CA
C
16 Chapter 2. Frequency-domain models for dynamical structures
The numerator coefficients are obtained by considering the following set of equa-
tions
H(s) C D 0 ... 0 I
sH(s) CA CB D 0 sI
.. = .. X(s) + .. . . (2.24)
..
. .. ..
. . .
sn H(s) CAn CAn−1 B CAn−2 B . . . D sn I
denoted in short as
Hn = On X + ΓIn (2.25)
The term in X is cancelled because of Eq. 2.23. The left-hand side of Eq. 2.26
is nothing else than A(s)H(s) and thus by comparing this equation with Eq. 2.22
the right-hand must be an expression for the polynomial B(s). As a result the Bn
coefficients can be expressed as
B0 . . . Bn−1 Bn = A0 . . . An−1 I Γ (2.27)
In the case that the number output observations No equals the number of modes
Nm , the order n of the polynomials will be 2. Using the expressions for the system
matrices A, B, C and D defined by Eq. 2.15 and 2.16 and the formulas 2.23 and
2.27 respectively for the Aj and Bj polynomial coefficients results in
A2 = I , A1 = M −1 C1 and A0 = M −1 K
B2 = 0 , B1 = 0 and B0 = M −1 (2.28)
where B2 and B1 are zero because D = 0 and CB = 0. This is not the case if
H(s) is considered as the transfer function matrix between the accelerations and
forces, instead of the displacements. Indeed, for the situation where Nm = No the
relation with the dynamic stiffness matrix Z(s) is directly given by
−1
H(s) = Is2 + M −1 C1 s + M −1 K M −1 (2.29)
−1
= Is2 + A1 s + A0 B0 (2.30)
and therefore methods which identify LMFD models with n = 2 are referred to
as direct identification methods. The system poles are given by the roots of the
characteristic equation det(A(s)), which is of order 2Nm . Nevertheless, in the case
that the number of modes Nm exceeds the number of response locations No , the
polynomials A(s) and B(s) must be expanded to higher orders in order to identify
Nm modes. The order n of A(s) must be larger or equal to 2Nm /No to capture
all dynamics. For No = 1 the limit of this expansion process results in a common
2.5. Matrix fraction polynomial models 17
with A(s) = Isn + An−1 sn−1 + . . . A0 a matrix polynomial with square (Ni × Ni )
matrix coefficients and B(s) = Bn sn + Bn−1 sn−1 + . . . B0 a matrix polynomial
with (No × Ni ) matrix coefficients. This RMFD model can be considered as a
LMFD model of the transposed transfer matrix H T (s)
and by consequence similar relations between the state-space model and the RMFD
model can be derived. In practice, since the number of input Ni is typically much
smaller than the number of modes Nm , RMFD models have a higher model order
n than LMFD models, but the denominator coefficients have smaller dimensions.
This fact has some implications on the performance of the system identification
algorithms based on LMFD and RMFD models as discussed in chapter 5.
Similar as for the LMFD model, the poles λr and modal participation vectors
are obtained from reformulating A(s) = 0 into a generalized eigenvalue problem,
resulting in nNi eigenvalues and the corresponding left eigenvectors. The mode
shapes can be obtained from the B coefficients or from a second step estimation
procedure by solving a linear least-squares problem, since the modal model is linear
in the mode shapes.
18 Chapter 2. Frequency-domain models for dynamical structures
and since the continuous-time signals are not available for identification, the
model must be converted to the frequency domain model for identification of a
continuous-time model. By using the discrete Fourier transformation, the frequency-
domain model is given by
where F (s) and Y (s) are only available for discrete values sk of s
N −1 N −1
1 X 1 X
Y (sk ) = √ y(n∆t)e−2πnk/N and F (sk ) = √ f (n∆t)e−2πnk/N (2.41)
N n=0 N n=0
xn+1 = Ad xn + Bd fn
yn = Cd xn + Dd fn (2.42)
with yn and fn the sampled time signals and where the discrete system matrices
have the following relation to the continuous system matrices under the ZOH
assumption
Z ∆t
Ad = eA∆t , Bd = eAζ dζB (2.43)
0
Cd = C , Dd = D (2.44)
zk X k = Ad Xk + Bd Fk
Yk = Cd Yk + Bd Fk (2.45)
In the modal model each mode appears twice: once with a positive frequency λr =
σr + iωr and once with a negative frequency λ∗r = σr − iωr . This symmetry in the
modal model results in a polynomial model with real coefficients and state-space
models with real system matrices. Nevertheless, in modal analysis applications,
one is often interested in a modal model in a frequency band of interest. In this
frequency band of interest the contribution of poles with negative frequencies is
often much smaller than the out of band poles. Therefore this frequency band
(which includes only poles with positive frequencies) can be modelled by complex
parameters. In this way, the model order can be reduced by a factor 2, resulting in
a better numerical conditioning. In practice a small over-modelling (higher model
order) is required to take care of the systematic errors introduced by modelling a
frequency band selection by a discrete-time model. To conclude each model i.e.
common-denominator model, state-space model, LMFD and RMFD can be chosen
to have real or complex parameters.
It is a well known fact that linear time invariant systems can be modelled by start-
ing from input-output measurements. The identified parametric model essentially
contains the same information of the studied system as the original non-parametric
data, but is often preferred because of its compact form and possible physical inter-
pretation. Several important references on the subject of system identification in-
clude [67], [110], [123] for time domain identification and [98] for frequency-domain
identification. Originating from electrical and control engineering, system identifi-
cation is today used in many different fields such as e.g. chemical engineering, civil
engineering, mechanical engineering, biomedical engineering, econometrics, ... .
2.9. Primary identification data 21
Experimental modal analysis starts from identifying modal models from the
measured applied forces and vibration responses of the structure, when artificially
excited in one or more locations (illustrated by figure. 2.1). These experiments
are performed under laboratory conditions to obtain high quality measurements
(typically SNR of 40dB). From these input-output measurements a mathematical
model is identified, which can be converted to a modal model of the structure.
F1 - - Y1
F2 - - Y2
Structure
. .
. .
. .
FNi - - YNo
In the field of system identification most identification algorithms start from the
measured input and output time histories or in the case of frequency-domain iden-
tification algorithms from the input and output spectra. Nevertheless in the case
22 Chapter 2. Frequency-domain models for dynamical structures
From the estimates of the transfer function matrix i.e. the FRFs or the IO
Fourier spectra, the discussed mathematical models e.g. common-denominator,
RMFD, LMFD or state-space can be identified.
In some applications (e.g. civil engineering [88], in-flight testing [10], [60]) one is
more interested to obtain modal models from structures during their operational
conditions to model the interaction between the structure and its environment e.g.
wind, traffic, boundary conditions, turbulence, ... on the structure. An other ad-
vantage of an in-operational modal analysis is that non-linear effects are linearized
around the operational working point. For these applications the structures are
naturally excited by ambient excitation forces e.g. wind, traffic, seismic activity
(micro-earthquakes) etc. [36, 32], which are difficult or even impossible to mea-
sure. Elimination of this ambient excitation is often impossible and applying an
artificial measurable force which exceeds the natural excitation is expensive and
sometimes difficult. In these cases, one only measures vibration responses (illus-
trated by figures. 2.2). From this output-only data only one can again estimate
- - Y1
unmeasurable forces - - Y2
Structure
. .
. .
. .
- - YNo
the natural frequencies, damping values and mode shapes. The knowledge of the
input signal is replaced by the assumption that the response is a realization of
a stochastic process with unknown white noise as an input. Identifying system
parameters from these responses only is referred to as stochastic system identi-
fication [123, 85]. More specific to the identification of vibrating structures the
terms output-only modal analysis and in-operation or Operational Modal Analysis
(OMA) are commonly used.
In the field of stochastic system identification one can generally divide the
identification techniques in two basic subcategories i.e.
Data-driven stochastic identification algorithms directly start from the output time
sequences and their frequency-domain counterparts use the output spectra as pri-
mary data. For the time-domain, Auto-Regressive (AR) and Auto-Regressive
Moving Average (ARMA) models are described in [67]. Both AR and ARMA can
be considered as time-domain counterpart of polynomial models in the frequency-
domain [98]. An ARMA model is written as
(Izkna + A1 zkna −1 + . . . Ana )Yk = (B0 zknb + B1 zknb −1 + . . . Bnb )Ek (2.55)
Notice that the influence of the initial and final conditions are neglected in this
expression. However, in chapters 4 and 6 it is shown how these initial/final con-
ditions are taken into account for common-denominator and state-space models.
A major drawback of the formulation given by Eq. 2.55 is that for the coefficients
Bi 6= 0 the identification problem becomes highly non-linear in the system param-
eters and for larger number of outputs No , which is typically the case for modal
testing, the algorithms become too slow to be used in practice. Nevertheless, AR
models and their frequency domain equivalent can be used to model the structure
with a model order na = 2 if No ≥ Nm (referred to as direct models). For the
case where No < Nm it can be shown that an AR model with infinite order is
equivalent of a finite-order ARMA model. Unfortunately, the theoretical assump-
tion of an infinite order to obtain a reasonable fit, practically means that many
mathematical poles are introduced [85].
zk X k = AXk + Wk (2.58)
Yk = CXk + Vk (2.59)
where λr , φr and Lr are respectively the pole, mode shape and modal participation
factor of mode r.
In [67] it is shown that for stationary stochastic processes the power spectra of
the outputs Syy (ω) ∈ C No ×No are given by
H
Syy (ω) = H(ω)Sf f (ω)H(ω) (2.61)
where Sf f (ω) ∈ C Ni ×Ni contains the cross power spectra of the (unknown) input
forces. Under the assumption that the forces are white noise sequences Sf f (ω) can
26 Chapter 2. Frequency-domain models for dynamical structures
where φr and Kr are respectively the mode shape and operational reference vec-
tor for mode r. This reference vector is a function of the modal parameters and
the cross power spectrum matrix of the unknown random input force(s). Un-
fortunately, the modal participation factors and by consequence the modal scale
factors can not be determined from an OMA test. Based on a sensitivity analysis
a technique to estimate the modal participation vectors L is proposed in [83]. To
use this technique a second set of measurements is required, where the structure
is modified with known modification e.g. adding a known mass. It should be
noticed that the modal decomposition of the power densities of the outputs has
a symmetry in the poles i.e. both the positive and negative poles are present in
the model. This symmetry is referred to as a 4-quadrant symmetry. Thanks to
the similarity between the modal decomposition of the Auto and Cross spectral
densities of the outputs and the modal decomposition of the FRFs the modal pa-
rameter estimation techniques for FRFs can be used to start from power spectra in
the output-only case. In practice only a limited number of reference sensors Nref
are used, by consequence Syyref ∈ C No ×Nref . In the case of multi-patch measure-
ments these reference output sensors remain fixed for the different patches.
Taking the Inverse Discrete Fourier Transform (IDFT) of Eq. 2.62 yields the
correlation functions matrix R(k) for positive and negative time lags k [51]
Nm
∗
φr KrT eλr kTs + φ∗r KrH eλr kTs
P
for k ≥ 0
r=1
rk = (2.63)
N
m
−λ∗
Kr φTr e−λr |k|Ts + Kr∗ φH r |k|Ts
P
r e for k < 0
r=1
Interesting to note is that the causal part (positive lags) of the correlation functions
contain the stable poles i.e. λr = σr + iωr and λ∗r = −σr − iωr , while the non-
causal part (negative lags) contains the unstable poles −λr = σr + iωr and −λ∗r =
−σr + iωr . Usually time-domain modal identification methods estimate the modal
parameters from the causal part only. In this way the number of modes is reduced
by a factor 2. Furthermore the time-domain modal decomposition of the causal
part of the correlations is similar to the modal decomposition of IRFs and hence
classical modal parameter estimators can still be used.
The estimation of the correlation and power spectra can be considered as the
estimation of smooth functions (the random character is reduced) by an averaging
2.9. Primary identification data 27
process, which also reduces the original amount of data. Since the parametric iden-
tification step starts from these smooth functions as primary data, a continuous-
time model can be used in contrary to the data driven stochastic identification.
Finally for the ABS function driven stochastic identification the stabilization dia-
gram turns out to be very useful to distinguish physical poles from mathematical
ones, as discussed in more detail in chapter 9.
In the previous two subsections it was shown that Experimental Modal Analy-
sis (EMA) and Operational Modal Analysis (OMA) differ in the fact that they
respectively consider the input forces as known and unknown. Consider a test ex-
ample were both measurable and unmeasurable forces are acting on the structure
as shown in figure 2.3. In a EMA one is only interested in the deterministic rela-
- - Y1
unmeasurable forces -
- Y2
F1 - Structure
F2 - .
.
. .
FN
.
.
i
- - YNo
Figure 2.3: Operational Modal Analysis with eXogenous inputs (OMAX), F measurable
inputs and E unmeasurable inputs
tion between the measured inputs and outputs and therefore the contribution in
the response resulting from the unmeasurable forces is considered as undesirable,
disturbing measurement noise. This is in contradiction with an OMA approach,
which considers the output-only of a stochastic process as the primary data to
estimate the modal parameters. Therefore the concept of a combined EMA-OMA
approach is now introduced, which considers the response as both a deterministic
contribution from the measurable inputs and a stochastic contribution from the
unmeasurable forces. Considering specific situations as a combined EMA-OMA
test results in a maximal data exploitation. For example a mode weakly excited
by the measurable forces and strongly excited by the unmeasurable forces is not or
inaccurately identified in a purely EMA framework, while in a combined approach
this mode is still well identified from the stochastic contribution. This concept of
28 Chapter 2. Frequency-domain models for dynamical structures
In the case that sufficient data is available, the deterministic and stochastic con-
tribution of the responses can be separated in a non-parametric way. This cor-
responds with estimating the FRFs or IRF between the outputs and measured
inputs and estimating the power densities or correlation function of the stochastic
contribution in the responses. In a next step, a model with common dynamics
(i.e. system poles) is fitted through the FRFs and power spectra. The variances
of the noise on the estimated ABS functions can be taken into account in the al-
gorithm to present the relative importance of their contributions in the vibration
responses. More details about the non-parametric estimation of the deterministic
and stochastic ABS functions is given in chapter 3.
2.10. Conclusions 29
2.10 Conclusions
Figure 2.4: Overview of the EMA, OMA and OMAX approach in function of the
amount of data
30 Chapter 2. Frequency-domain models for dynamical structures
Chapter 3
Non-Parametric
preprocessing steps
31
32 Chapter 3. Non-Parametric preprocessing steps
3.1 Introduction
Since a classical modal analysis experiment typically results in long data se-
quences, the classical modal parameter estimation (MPE) techniques start from
FRF data to reduce the noise levels and size of the data set. Furthermore, in case
the measurements took place in different patches, starting from FRF data simpli-
fies the parametric identification. In literature many attention has been paid to
non-parametric FRF estimators [67], [64], [131], [52]. This chapter starts with a
presentation of well-known estimators like the H1 , the H2 and Hev in the frame-
work of the Total Least Squares Hv estimator. Given the Multiple Input-Multiple
Output (MIMO) character of modal testing, the different estimators are proposed
for a multivariable system. Besides the influence of noise on the measured data,
the estimation of the FRFs is complicated by leakage in the case arbitrary in-
put signals are used. To deal with leakage and measurement noise, the use of a
rectangular window is proposed and compared to the classical approach.
Finally, the non-parametric estimators for the EMA and OMA case are com-
bined in the context of an OMAX process. The key idea of this non-parametric es-
timator is to estimate both the FRFs, as a result of the deterministic contributions
and the power spectra as a result of the stochastic contributions. Parametric iden-
tification starting from both FRFs and XP results in the identification of modes
excited by the artificially applied forces, modes excited by the ambient excitation
and modes excited by both simultaneously. For the combined identification, a dis-
tinction is made between structures, with exciter-structure (e.g. electrodynamical
shaker) and without any exciter-structure interaction.
This chapter ends by illustrating the techniques by simulations, and two ex-
perimental cases. The first experimental case illustrates the applicability for mea-
surements on a subframe. In the second case, the applicability for measurements
of a bridge is studied.
3.2. Modal testing for EMA 33
In a typically EMA test setup the structure is flexibly mounted to obtain so-called
free-free conditions. Figure 3.1 illustrates several laboratory test setups.
(a) (b)
(c) (d)
Figure 3.1: Examples of EMA tests. (a) Ground vibration shaker test of an aircraft;
(b) Shaker test of a subframe; (c) Hammer test of a frame; (d) scanning laser vibrometer
test of a slattrack
test.
Shaker-based testing
Shaker testing excites the structure in one or several locations through a fixed
shaker-stinger-structure connection. This type of testing is typically character-
ized by a large number of outputs and a small number of inputs. The stinger
ensures that the force from the shaker is solely applied in the longitudinal direc-
tion. Typically for this test setup is the interaction between the structure and the
shaker, which results in a drop of the signal amplitude of the transmitted force at
the resonant frequencies of the structure. Furthermore, this interaction causes a
significant difference between the resonances in the measured vibration responses
and the FRFs of the structure. Since the force is measured between the stinger
and structure, the ratio between the measured response spectra and force spectra
still results in the correct FRF describing the structures dynamical behavior. On
the contrary, if the spectra of the electrical signals, that are sent to the shaker
amplifier are used as the input signals and the vibration responses as the output
signals, the FRF matrix describes the dynamics of the structure including the
stingers, shakers and electrical drive system.
Hammer-based testing
A second class of modal testing is the so-called hammer excitation testing, where
the structure is excited in different locations by a roving hammer impact, while
the vibration response is measured in a limited fixed number of reference locations.
Different from shaker excitation testing, hammer excitation testing measures only
vibration responses at a few locations, while the structure is excited separately in
time at the different locations of interest. From the reciprocity property of the
modal model ( i.e. the transfer function between a force at location 1 and the
response at location 2 is equal to the transfer function between a force at location
2 and the response at location 1), it follows that exciting at Ni locations and mea-
suring the vibration response at No locations, results in the transpose of the FRF
matrix obtained by exciting in No locations and measuring in Ni response loca-
tions. Therefore, the FRF matrix measured by hammer testing can be considered
as a No × Ni matrix, where No are the number of excitation locations and Ni the
number of reference accelerometers. When an impact hammer test is used, there
is no fixed interaction between the structure and the excitation device. Further-
more, the noise on the signals for excitation at different locations is uncorrelated,
resulting in exact maximum likelihood (ML) estimates for the ML algorithms pro-
posed in chapters 4 and 5. Hammer testing is simple in use, does not suffer from
massloading, needs only a limited amount of sensors and acquisition channels and
therefore is still commonly used for modal analysis purpose. Drawbacks are the
limitation of the excitation signals (only impacts or repeated impacts), the fact
that exciting at many locations is time-consuming and exciting in all 3 directions
might be difficult or impossible.
Arbitrary input signals, such as random noise, are still commonly used for modal
testing because of the their general availability as well as the averaging effect they
have for possible non-linear effects. Of course the use of arbitrary signals compli-
cates the estimation of FRFs due to leakage problems. Therefore window tech-
niques, such as a Hanning window, are applied to reduce the influence of leakage.
However, the errors introduced by leakage can not be completely eliminated by
the use of a Hanning window, especially for the case of lightly damped structure.
In chapter 4 an extension of classical FRF based algorithms is proposed which
models the leakage exactly by considering Nb (the number of averaged blocks)
extra inputs.
The use of time-limited burst random excitation was introduced for modal
36 Chapter 3. Non-Parametric preprocessing steps
analysis. For this input signal no leakage is introduced since the signals decays to
zero within the length of the time window. Similar as for impact hammer testing,
an exponential window can be applied in order to amplify this decay by artificially
increasing the damping. In [131] it is shown that the effect of the exponential win-
dow can be totally compensated on the modal parameters. Disadvantages of burst
signals compared to random noise excitation is that less energy is injected in the
structure within the same time window, resulting in a lower signal-to-noise ratio
and in addition a higher crest factor. The use of periodic broadband excitation
avoids leakage problems [1]. The FRF estimators discussed in this section start
from the power spectra, i.e. the Auto Power spectra of the inputs and the outputs
and the Cross Power spectra between the inputs and outputs respectively given
by
Nb
1 X
Syy (ωk ) = H
Yb,k Yb,k ∈ C No ×No (3.1)
Nb
b=1
Nb
1 X
Sf f (ωk ) = H
Fb,k Fb,k ∈ C Ni ×Ni (3.2)
Nb
b=1
Nb
1 X
Sf y (ωk ) = H
Fb,k Yb,k ∈ C Ni ×No (3.3)
Nb
b=1
Nb
1 X
Syf (ωk ) = H
Yb,k Fb,k ∈ C Ni ×No (3.4)
Nb
b=1
= SfHy (3.5)
with Nb the number of blocks by which the time data is divided and the input
Fb,k and output spectra Yb,k given by the Discrete Fourier Transform (DFT) in
combination with a window wn
N −1
1 X
Fb,k = √ wn fb,n zk−n (3.6)
N n=0
N −1
1 X
Yb,k = √ wn yb,n zk−n (3.7)
N n=0
with zk = ei2πk/N and fb,n , yb,n the time samples of the input and output signals
for block b at simple time n∆t (∆t the sample period).
The Hv estimator
Under the assumption that the noise on the measured inputs and outputs is uncor-
related and of equal amplitude, the Total Least Squares (TLS) method Hv results
3.3. FRF estimation for EMA 37
in a consistent estimate of the FRFs. The Hv estimator [64], [52] estimates the
FRFs by solving the following eigenvalue problem
HvH HvH
Sf f Sf y
= Λ (3.8)
Syf Syy −INo −INo
(the indication for the spectral line k is dropped) In [131] it is shown that the Hv
estimator is not consistent if these noise assumptions are violated. In that case, a
generalized TLS estimator must be used, which uses the noise covariance matrix as
a weighting in order to be consistent [131]. However, in practice this is impossible
since this covariance matrix is not a priori known. Furthermore, it is shown in
[131] that the FRF matrix obtained by processing each output separately is only
equal to the one obtained from considering all outputs simultaneously if HH H is
diagonal. Since there is no reason for this, the MIMO and MISO Hv estimator
generally differ.
The H1 estimator
Under the assumption that only noise is present on the outputs (and this noise un-
correlated with the input signals), the Hv FRF estimator reduces to the commonly-
used H1 FRF estimator given by
H1 = Syf Sf−1
f (3.9)
where Sf f is non-singular if the input forces are not totally correlated and enough
blocks Nb are used to calculate the auto power spectra Sf f . Under the specified
noise assumptions the H1 estimator is consistent (if leakage is neglected). In
the case no exciter-structure interaction exists e.g. impact hammer excitation,
the noise assumptions for the H1 estimator are often realistic, since process noise
caused by unmeasurable forces results as output noise uncorrelated with the inputs.
The covariance matrix CH1 of the noise on the estimated FRFs is obtained from
a sensitivity analysis given in appendix 3.8.
The H2 estimator
Similar to the H1 estimator, the assumption of only noise on the input signals and
noise-free output signals results in the H2 estimator given by
H2 = Syy Sf−1
y (3.10)
In order to compute the inversion Sf−1 y the H2 estimator can only be used for
Ni = No . In case the FRFs are estimated output by output, the H2 estimator
is only applicable in the single input case. In [131] is it shown that both the H1
and H2 estimator belong to the class of maximum likelihood estimators and the
covariance matrix on the estimated FRFs results in the Cramer-Rao lower bound
for their specific noise assumptions.
38 Chapter 3. Non-Parametric preprocessing steps
Although the use of periodic input signals (deterministic input signals) avoids
problems of leakage and results in better signal-to-noise ratios (SNR) [1], [98], [47]
they are not always available in signal generators of commercial measurement de-
vices and therefore one might have to stick with random input excitation signals.
Special design of the deterministic input signals e.g. multisines and Schroeder sig-
nals result in optimal SNR and crest factors [43], [37]. Depending on the damping
ratios of the structure and the length of the period, the first periods of the mea-
sured outputs must be omitted to leave out the transient response. In [106], [107]
special designed input signals i.e. odd-multisines, odd-odd multisines and special
odd-multisines are proposed to characterize the possible non-linear behavior of the
structure [140].
Nb
! Nb
!−1
X X
Hev = Xb Fb (3.11)
b=1 b=1
where b indicates the Nb different periods. Applying the Hev output by output
results in the same FRF estimates as considering all outputs simultaneously. It
can easily be proven that this FRF estimator is both asymptotically unbiased and
consistent in presence of both input and output noise. In [131] an expression is
given for the covariance matrix of the noise on the estimated FRFs. In the case
of asynchronous periodic measurements several other FRF estimators based on
non-linear averaging techniques in an error-in-variables framework are proposed
in [39].
In many practical cases the engineer responsible for the modal parameter estima-
tion only has the FRFs and in the best case also their covariances to start from.
The measurements itself are often carried out by the test engineers and they have
already done the data reduction step by estimating the FRFs. In the case that
high frequency resolution measurements are given with large noise levels, the noise
3.3. FRF estimation for EMA 39
levels can be reduced by a rectangular window, but the price paid is the lower fre-
quency resolution. Nevertheless, for several practical reasons, one is often willing
to pay this price:
Consider the given FRF data H(ωk ) and the covariance matrix CH (ωk ) with the
discrete-time domain model given by
Nm
φ∗r LH
!
X φr LTr [:,r]
H(ωk ) = + (3.12)
r=1
1 − λr zk−1 1 − λ∗r zk−1
notice that the discrete modal model has discrete poles λr and discrete participa-
tion factors Lr which are related to their continuous-time equivalents. The Inverse
Discrete Fourier Transform (IDFT) of the FRF matrix is the Impulse Response
Function matrix (IRF), for time-lag n is given by
N
X
hn = H(ωk )zkn (3.13)
k=1
Nm n
λnr λ∗r
X
= φr LTr + φ ∗ H
L
r r (3.14)
r=1
(1 − λNr ) (1 − λ∗N
r )
the proof is given in the appendix (N the number of spectral lines). This means
that the IRF can be considered as a sum of exponentially damped sines. The term
40 Chapter 3. Non-Parametric preprocessing steps
NX
w −1
(1−λN w)
with L′r = Lr r
(1−λN
. The proof follows directly from
r )
N −1
X 1 − xN
xn = (3.18)
n=0
1−x
The FRFs given by H r (ωk ) are considered as the reduced FRFs. The poles λr
of Hr (ωk ) are not effected by considering only the first Nw samples of the IRF.
On the contrary the participation vectors Lr are multiplied by a correction factor,
(1−λNw )
which depends on the specific pole λr . Notice that this correction term (1−λrN ) is
r
equal to 1 for Nw = N or for both N and Nw → ∞. In fact, the correction term
is caused by both time-domain aliasing and frequency-domain leakage. This data
and noise reduction process is visually presented by figure 3.2.
Another approach to reduce the noise levels and leakage for the estimation
of FRFs is based on exponential windows [139]. This method weights the IRF
with an exponential decaying window to reduce the noise levels. The exponential
window artificially increases the damping in such a way that 1 − λN
r
w
≈ 1. In this
way the modal participation factors should not be corrected, but the poles should
be instead.
FRF
N m
φr Lr
H (ωk ) = + ...
IDFT
IRF
Rectangular window
DFT
FRF
H (ωk ) = ∑
φr Lr (1 − λrN )
Nm w
+ ...
r =1 1 − λr zk (1 − λr )
N
wn = 1 for 0 ≤ n ≤ Nw − 1
wn = 0 for Nw ≤ n ≤ N (3.19)
′ ′
and hn defined by hn = hn wn . The DFT of h′n is given by
N −1
′ X ′
H (ωk ) = hn zk−n (3.20)
n=0
with ∗ the convolution symbol. The validity of this expression can easily be
proved since
P the convolution is equivalent with a sum in the frequency-domain
2
P
and cov( i ai Yi ) = i |ai | cov(Yi ). As a result the covariance of the reduced
FRFs H r (ωk ) is given by
Section 3.2 discussed different FRF estimators for the EMA case. In this para-
graph, the estimation of the power spectra between the outputs and the reference
outputs is discussed. The auto and cross power spectra Syyr are the primary data
for model parameter identification based on Eq. 2.62. Similar as for the EMA case,
this ABS-function technique is in the case of long-in-time available data sequences
preferred to the data driven approach.
Basically, two classical approaches exist for the estimation of auto and cross
power spectra. The periodogram [102] estimator operates directly on the spectra
of different time blocks resulting from a division of the time sequences. The cor-
relogram [9] approach first estimates the correlation functions in the time-domain
and next the power spectra are obtained by transferring the correlations to the
frequency-domain.
A procedure is proposed to eliminate the unstable poles from the power spectra
by considering so-called ’positive’ power spectra. In this way a stable model can
be fitted in the frequency domain based on ABS-functions for OMA applications.
The biased correlation estimate uses 1/N rather than 1/(N − n). The power
spectra Syy (ωk ) are given by Fourier transforming the correlation functions
N
X
Syy (ωk ) = wn rn zk−n (3.26)
n=−N
with zk = ei2πk/N . To reduce the effect of leakage the use of an adequate (2N +
1)-point time window wn (e.g. Hanning, Hamming, ...) symmetric around the
origin is advised. This window reduces the effect of leakage and thus the bias
error in the power spectra. For instance, applying a Hanning window to the
correlation functions will force the correlation to zero at the higher lags. Moreover,
the application of such a window reduces the stochastic uncertainty on the cross
power estimate due to the presence of a higher stochastic uncertainty near the
higher lags of the correlation function estimate.
reduces both the influence of leakage and the influence of the stochastic uncertain-
ties in tails. The factor β is typically chosen such that the amplitude at time lag
N of the window is 1% of its initial amplitude [52]. The poles are finally corrected
by removing the artificially added damping β.
In [7] it is proven that a fast calculation of the linear correlation can be obtained
by the use of the discrete Fourier transform by a technique called zero-padding.
Each of the Nb adjacent data blocks, each N output samples, is extended by an
44 Chapter 3. Non-Parametric preprocessing steps
additional block of N zeros. In other words, new blocks qb,n of length 2N are
created such that
then it is shown in [7] that the IDFT of this power spectra followed by a correction,
given by Eq. 3.33, results in the linear correlation function rn :
2N −1
1 X ref
rns = Sqq (ωk )zkn (3.32)
N
k=0
N
rn = rs (3.33)
N −n n
In fact this correlation rn is exactly the same correlation function as the one by
averaging the correlation functions from the Nb adjacent blocks, but the use of the
DFT and IFT functions speeds up the calculation time. This procedure is shown
in figure 3.8.
with N the number of time samples within a block. The time window wn (e.g.
Hanning, Hamming) typically reduces the influence of leakage. Since the window
reduces the contribution of the data at the begin and end of the record, introducing
3.4. Estimation of the power spectra for OMA 45
which is a similar expression as for the power spectra used by the Hv , H1 and
H2 FRF estimators. The choice of the number of samples N within a block and
thus the number of blocks Nb is a trade-off between variance and bias on the esti-
mated power spectra. Choosing a higher amount of data samples N and a lower
amount of blocks Nb reduces the effect of leakage but increases the stochastic
uncertainty i.e. the variance on the estimated power spectra. The periodogram
method is a well-established technique, which is available as a tool in most com-
mercial software packages. In [7] it shown that the inverse Fourier transformation
of the periodogram estimates of the power functions results in the so-called circular
correlation.
Similar as for the FRFs, the variances on the estimated auto and cross power
densities can be used as a weighting in the parametric identification process in
order to improve the consistency and efficiency properties of the algorithms. In [7]
the following formulas for the variances on the estimated power spectra are given
(for each reference considered separately)
2 2
var(Syref yref ) = S (3.36)
Nb yref yref
2
cov(Syo yref ) ≤ Sy y Sy y (3.37)
Nb o o ref ref
Of course the variance can also be calculated from the sample mean variance i.e.
N
b
1
|Yo,b Ybref ∗
X
cov(Syo yref ) = − Syo yref |2 (3.38)
Nb − 1
b=1
• The power spectra have a 4-quadrant symmetry i.e. the OMA modal model
contains λ, λ∗ , −λ and −λ∗ as poles. This results in a model order, which is
twice the modal order needed to model FRFs. A higher model order results
for all identification methods in an increasing calculation time and in a less
good numerical conditioning.
• The power spectra contain both stable λ, λ∗ and unstable poles −λ, −λ∗
poles in its model. This results in less interesting properties for the interpre-
tation of stabilization diagrams, when distinguishing physical from mathe-
matical poles. A more detailed discussion is given in chapter 9.
• Power spectra estimated from a limited amount of data are typically char-
acterized by high noise levels compared to FRFs. Therefore, an additional
noise reduction would be preferable.
Where rn only contains the stable poles. However, the focus of this thesis lies on
frequency-domain identification for modal analysis and this for reasons like of e.g.
simple frequency band selection and the use of frequency weighting functions in
the identification procedure.
Since Eq. 3.39 is analog to the expression for the IRF given by Eq. 3.14, a
similar approach as for the reduction of the noise levels on the FRFs (paragraph
3.3.3), can be applied by taking the Fourier transform of the first Nw samples.
+
The ’positive’ power spectra Syy are defined by the DFT of the first Nw samples
of the positive lags of the correlation function:
NX
w −1
+
Syy (ωk ) = rn zk−n (3.40)
n=0
Nm
φr K T (1 − λNw ) φ∗r KrH (1 − (λ∗r )Nw )
X
r r
= + (3.41)
r=1
1− λr zk−1 1 − λ∗r zk−1
Nm ′ ′
!
X φr KrT φ∗r KrH
= + (3.42)
r=1
1 − λr zk−1 1 − λ∗r zk−1
3.5. Combined FRF and XP estimation for OMAX 47
with
′
+
var(Syy (ωk )) = |W (ωk )|2 ∗ var(Syy (ωk )) (3.44)
for W (ωk ) defined as the Fourier spectra of the time window given by Eq. 3.19
and m = N/Nw . Similar as for the IFT, the side lobe of the positive samples
contains most of the stochastic uncertainty. Since only the first Nw positive lags are
considered, the stochastic uncertainty on the positive power spectra will decrease.
Finally the modal parameters can be estimated starting from these positive power
spectra and their variances as a frequency domain weighting.
+
The estimation of the positive power spectra Syy can be summarized as follows
and is visually given by figure 3.8:
densities) are taken into account resulting in an optimal data exploitation i.e. both
modes excited by the measurable and unmeasurable forces can be identified.
where U (ωk ) are the spectra of the electrical signals, that are sent to the excitation
m m
devices, G(ωk ) ∈ C Ni ×Ni the transfer functions characterizing the excitation
devices and the interaction with the structure. The transfer path between the
unmeasurable forces and m
theu forces measured at the exciter-structure connection
is given by T (ωk ) ∈ C Ni ×Ni . The multivariable frequency-domain OMAX model
is illustrated in figure 3.3.
Impact hammer and acoustic excitation are all examples were no fixed connection
exists between the structure and excitation device and thus the transfer path
T (ωk ) is equal to zero T (ωk ) = 0. For some other applications, the influence
of the exciter is negligible and thus no interaction should be taken into account
e.g. exciter on a large bridge. For these types of test setups, the measured forces
F (ωk ) are uncorrelated with the unmeasurable forces E(ωk ) and thus the H1
estimate is consistent under the assumption that the measurable forces contain no
measurement noise. The FRFs and the covariances of the noise on the FRFs are
3.5. Combined FRF and XP estimation for OMAX 49
U - G - +l - Hm
6 F
?
T +l - Y
6
6
E - H u
Figure 3.3: Frequency-domain OMAX model with F the measurable input forces, E
the unmeasurable random forces and Y the vibration responses.
then given by
The power densities of the stochastic contribution Y s (ωk ) = H u (ωk )E(ωk ) be-
tween the outputs and the reference outputs Y ref,s are given by
Interesting to notice is that Sys yref,s (ωk ) = CY (ωk )[:,ref ] (i.e. the columns corre-
sponding to the reference outputs) with CY , defined by Eq. 3.65, the covariance
matrix of the stochastic contribution in the output signals. In an EMA approach
this covariance matrix CY is considered only in the expression for the covariance
matrix of the noise on the estimated FRFs. In practice, this means that the larger
the stochastic contribution, the larger the uncertainties on the deterministic con-
tribution (i.e. on the FRFs) will be. In contrast with the EMA approach, the
OMAX approach considers the variance on the responses CY as the power spectra
of the contribution of the ambient unmeasurable forces and these now serve as
primary data for the modal parameter identification.
50 Chapter 3. Non-Parametric preprocessing steps
Since the power spectra given by Eq. 2.61 have a 4-quadrant symmetry, while
the FRFs given by Eq. 2.4 have a 2-quadrant symmetry, identification based on
both spectral functions simultaneously would result in a highly non-linear problem.
In order to remove the unstable poles from the power spectra and to reduce the
noise levels, only the first Nw samples of the inverse Fourier transform of H m (ωk )
and Sys yr,s (ωk ) are used, resulting in Eq. 3.17 and Eq. 3.42 given by
Nm
φr QTr φ∗r QH
X
r
H m (ωk ) Syy
+
(ωk ) = + (3.52)
r=1
1 − λr zk−1 1 − λ∗r zk−1
with Qr = L′r Kr′ . Notice that Eq. 3.52 has exactly the same formulation
as the modal model and thus all classical EMA identification algorithms based
on state-space models, common-denominator models and matrix fraction models
can be used in the OMAX framework, when starting from both FRFs and the
’positive’ power spectra of the stochastic contribution. It should be remarked
that the FRFs in the EMA, the positive power spectra in the OMA and both
simultaneously FRFs and positive power spectra in the OMAX case are described
by the same model structure and only the participation vectors in the models differ
in terms of the physical interpretation.
In the case that a fixed connection exist between exciter and structure e.g. shaker-
structure connection by a stinger, the transmission path T (ωk ) differs from zero.
The auto power Sf f (ωk ) and cross power spectra Syf (ωk ) are then given by
under the assumption that U (ωk ) and E(ωk ) are uncorrelated (the notation for the
spectral lines ωk is dropped for simplicity of the expressions). For T 6= 0 the H1
estimator is biased and not consistent. The H1 FRF estimator fails because the
exciter-structure connection introduces a correlation between F (ωk ) and E(ωk ).
A consistent estimate for the FRF matrix H m (ωk ) is obtained by measuring addi-
tionally the electrical generator signals U (ωk ) as additional signals or by the use
of periodic electrical generator signals.
Instead of starting from the auto and cross power spectra Sf f (ωk ) and Syf (ωk )
lets consider the cross powers between the force spectra, response spectra and
the electrical generater spectra given by (under the assumption that the electrical
3.6. Simulations and measurement examples 51
Sf u = GSuu (3.55)
m
Syu = H GSuu (3.56)
since U (ωk ) is uncorrelated with the unmeasurable forces E(ωk ). In fact, this
electrical generator signal U is used as a instrumental variable [139]. An unbiased,
consistent estimate of the FRFs H m (ωk ) is then obtained by
which is known as the instrumental variables FRF estimator. Analogous as for the
H1 FRF estimator the covariance matrix of the noise on the FRF estimates can
be calculated by a sensitivity analysis [131]. The power spectra of the stochastic
contribution can be found by
Nb
1 X
Sys yref,s = Ybs Ybref,s (3.58)
Nb
b=1
with H m,ref ∈ C Nref ×Ni the part of the FRF matrix H m containing the FRFs
between the reference outputs and the measurable inputs. To eliminate the un-
stable poles from Sys yref,s the positive power spectra Sy+s yref,s are calculated and
the same model as Eq. 3.52 can be identified.
In the case that periodic excitation signals are used, the instrumental variables
approach reduces to the Hev FRF estimator discussed in paragraph 3.3.2. Thus
FRFs between the outputs and the measurable inputs are given by Eq. 3.60, i.e.
Nb
! Nb
!−1
X X
m
Hev = Xb Fb (3.60)
b=1 b=1
The use of periodic signals avoids the need for the electrical generator signals
to obtain consistent estimates for H1 . Similar as for the instrumental variable
approach the power spectra of the stochastic contribution are obtained by Eq.
3.59. This approach has the advantage that no extra signal must be measured, no
errors are introduced by leakage and no linear behaviour of the electrical generator
must be assumed.
52 Chapter 3. Non-Parametric preprocessing steps
Table 3.1: Mass, damping and stiffness characteristics of the 7 DOF system and the
exciter
exciter structure
me = 15 m1 = m2 = m5 = m6 = m7 =10
ke = 3000 m4 = 26, m3 = 9
ce = 75 k1 = . . . = k8 = 150000
c1 = . . . = c8 = 10
In order to validate the results presented in this chapter a 7 DOF system, shown in
figure 3.4, is used in order to simulate several types of modal analysis experiments.
All three cases i.e. an EMA, OMA and OMAX are investigated by means of
simulations. In some simulations the system is excited by an electrodynamic shaker
with its body grounded. The exciter coil has a mass me and is supported on a
flexure of stiffness ke and damping ce . The exciter is fixed to the system by a
rigid stinger. The system itself consists of 7 masses connected to each other by
springs and dampers. The system and exciter characteristics are given in table
3.1. The first and last masses are connected to the ground. This model of an
electrodynamic shaker is considered as a good approximation of a real shaker
used in a modal analysis experiment. The electrodynamic shaker results in an
interaction between the structure and the exciter leading to force signal drop at the
resonance frequencies. Obviously, the larger the shaker compared to the structure,
the larger the interaction and influence of the exciter on the measurement setup.
Given the M , C and K matrices, respectively the mass, damping and stiffness
matrices both the vibration response caused by a force f (t) and the exact modal
parameters of the structure can be calculated. The vibration response is obtained
by solving Newton’s differential equations in the time-domain, which also allows
to include the effects of initial and final conditions. This makes it possible to
investigate the effect of leakage on the ABS function estimators. The goal of the
simulations is to shown the effectiveness of a rectangular window to reduce the
noise levels on the FRFs and to eliminate the unstable system poles from the
auto and cross power density functions. Finally, the advantage of a combined
operational-experimental modal analysis i.e OMAX-analysis is illustrated.
3.6. Simulations and measurement examples 53
fe
f2 f7
f1
me
ms2 ms7
ms1 …
y1 y1 y7
The system is excited at the 5-th mass with a white gaussian noise (N(0,1)) signal
during 128s with a sample frequency of 1024 Hz. A first simulation was done to
study the effect of leakage on the estimation of the FRFs and therefore no noise was
added to the force and vibration response signals. All 7 modes of the structure are
excited and the modes are within the frequency band 5-40Hz. To illustrate the use
of a rectangular time window to reduce the noise levels of the FRFs, a comparison
is made between the classical H1 FRF estimator with a Hanning window and the
H1 estimator with Hanning window combined with a noise reduction procedure as
explained in paragraph 3.3.3. The FRF and its standard deviation estimated from
2k blocks is compared to the FRF and its standard deviation from 2 blocks followed
by rectangular time window with m = N/Nw for m = 1, 4, 16, 32, resulting in a
respectively frequency resolution of 0.0156, 0.0625, 0.25 and 0.50Hz. Figure 3.5
clearly shows the initial variance on the FRFs for Nb = 1 due to leakage errors
(even with the use of a Hanning window on the relative long records of 64s) and
small simulation errors. Increasing the number of blocks Nb averaged for the classic
H1 estimator results in low SNR ratios in the resonances and biased damping ratios
caused by the use of the Hanning window, which is necessary to reduce the leakage.
On the other hand, the approach where only 2 blocks are averaged followed by a
noise reduction by the use of a rectangular window clearly results in a increase
of the SNR in the resonance frequencies. Since in modal analysis often least
squares estimators are used for the reason of their process speed if they are proper
implemented, the quality of the FRFs in the resonances is of prior importance.
Comparing the FRF estimated for Nb = 64 and Nb = 2, m = 32 with the initial
FRFs for Nb = 2 it is clear that the modal parameter estimators starting from
the FRF H1 estimate for Nb = 64 results in overestimated damping ratios and
a tendency to identify a double pole around 6.5Hz. Finally, it should be noted
that although the SNR ratios on the estimated FRFs by the use of a rectangular
window are not always that high, the FRFs are very smooth. This is explained
54 Chapter 3. Non-Parametric preprocessing steps
by the high correlation of the noise on the FRFs over the different spectral lines.
This can be observed in figure 3.7 showing the FRFs, their mean and standard
deviation from 10 independent simulation runs.
The primary goal of an FRF estimation is to reduce the data set for both
memory requirements and process time of the parametric identification and to
increase the SNR. A second set of simulations were done with 30% relative noise
on the FRFs and the results are illustrated by figure 3.6. Increasing the number
of blocks Nb results in a better SNR for the classic H1 estimates, but compared
to the use of a rectangular window the SNR is still low in the natural frequencies
and the damping values seem to be biased.
From a comparison between the exact FRF and the estimated FRF it is difficult
to make statements about the quality of both procedures, since the rectangular
window introduces an error on the participation factors, which does not harm
the quality of the estimates, since it can be compensated. Therefore a Monte
Carlo simulation of 30 runs was done with 10% relative (in the frequency do-
main) colored noise added on the responses and the quality of the estimates of the
natural frequencies and damping ratios from the combined deterministic-stochastic
frequency-domain subspace algorithm (proposed in chapter 8) is compared to eval-
uate the quality of the primary FRF data. Tables 3.2 and 3.3 respectively compare
the mean estimated natural frequencies and damping ratios with their standard
deviation. Both the natural frequency and damping ratio estimates are unbiased
for the rectangular window approach, while the classic H1 approach results in
largely biased damping ratios and a much higher variance on both the estimated
natural frequencies and damping ratios. Especially the damping ratios of the low
frequency modes suffer from the bias introduced by the leakage and the use of the
Hanning window. From this, it is clear that the use of a rectangular window to re-
duce the noise and data sets outperforms the classically-used of the H1 estimator.
Finally, it should be noticed that the noise and data reduction by the rectangular
window can be used in combination with any FRF estimator.
−80 −80
−100 −100
Ampl (dB)
Ampl (dB)
−120 −120
−140 −140
−160 −160
−180 −180
0 10 20 30 40 0 10 20 30 40
Freq (Hz) Freq (Hz)
Nb=8 Nb=2, m=4
−60 −60
−80 −80
−100 −100
Ampl (dB)
Ampl (dB)
−120 −120
−140 −140
−160 −160
−180 −180
0 10 20 30 40 0 10 20 30 40
Freq (Hz) Freq (Hz)
Nb=32 Nb=2, m=16
−60 −60
−80 −80
−100 −100
Ampl (dB)
Ampl (dB)
−120 −120
−140 −140
−160 −160
−180 −180
0 10 20 30 40 0 10 20 30 40
Freq (Hz) Freq (Hz)
Nb=64 Nb=2, m=32
−60 −60
−80 −80
−100 −100
Ampl (dB)
Ampl (dB)
−120 −120
−140 −140
−160 −160
−180 −180
0 10 20 30 40 0 10 20 30 40
Freq (Hz) Freq (Hz)
Figure 3.5: No noise added, left: classic H1 FRF estimate from Nb blocks; right: H1
estimate from 2 blocks, followed by a rectangular window (full line: FRF, dotted line:
standard deviation)
56 Chapter 3. Non-Parametric preprocessing steps
−80 −80
−100 −100
Ampl (dB)
Ampl (dB)
−120 −120
−140 −140
−160 −160
−180 −180
0 10 20 30 40 0 10 20 30 40
Freq (Hz) Freq (Hz)
Nb=8 Nb=2, m=4
−60 −60
−80 −80
−100 −100
Ampl (dB)
Ampl (dB)
−120 −120
−140 −140
−160 −160
−180 −180
0 10 20 30 40 0 10 20 30 40
Freq (Hz) Freq (Hz)
Nb=32 Nb=2, m=16
−60 −60
−80 −80
−100 −100
Ampl (dB)
Ampl (dB)
−120 −120
−140 −140
−160 −160
−180 −180
0 10 20 30 40 0 10 20 30 40
Freq (Hz) Freq (Hz)
Nb=64 Nb=2, m=32
−60 −60
−80 −80
−100 −100
Ampl (dB)
Ampl (dB)
−120 −120
−140 −140
−160 −160
−180 −180
0 10 20 30 40 0 10 20 30 40
Freq (Hz) Freq (Hz)
Figure 3.6: 30% relative noise, left: classic H1 FRF estimate from Nb blocks; right: H1
estimate from 2 blocks, followed by a rectangular window (full line: FRF, dotted line:
standard deviation)
3.6. Simulations and measurement examples 57
−70
−70
−80
−80
−90
−90
−100
Ampl (dB)
Ampl (dB)
−110 −100
−120
−110
−130
−120
−140
−130
−150
−160 −140
0 20 40 0 20 40
Freq (Hz) Freq (Hz)
Figure 3.7: left: FRFs estimated for 10 independent runs (no noise added); right: mean
estimated FRF (full line) and the standard deviation (dotted line)
Table 3.2: Monte Carlo Simulation (30 runs). Mean value and standard deviation of
the estimated natural frequencies from primary FRF data obtained by the classic H1
technique (Nb = 16) and from the H1 followed by a rectangular window (RW) (Nb =
4, m = 4).
fexact (Hz) fˆh1 (Hz) σfh1 (Hz) fˆrw (Hz) σfrw (Hz)
6.411 6.458 0.134 6.411 0.001
15.012 14.921 0.011 15.011 0.001
18.672 18.656 0.002 18.673 0.001
27.858 27.846 0.002 27.860 0.001
29.593 29.585 0.003 29.595 0.001
36.141 36.002 0.038 36.147 0.004
36.913 36.936 0.101 36.929 0.005
overview for the estimation of the ’positive’ power spectra. Figure 3.9 compares
the power spectra estimated by the periodogram technique with the estimation of
the ’positive’ power spectra. Comparing the resonance peaks, it is clear that the
periodogram approach suffers from leakage, which will result in an overestimation
of the damping ratios.
58 Chapter 3. Non-Parametric preprocessing steps
Zero padding
0.6 0.6 0.6
…
0 0 0
Response
…
40 40 40
spectra
30 30 30
20 20 20
10 10 10
0 0 0
120
100
80
60 Power spectra
40
4-quadrant symmetry
20
0
0 10 20 30 40
freq
IDFT
2000
1000
0
Positive and negative
-1000
-2000
correlation
-3000
rectangular window
2000
1000
0
Window (first Nw
-1000 samples)
-2000
-3000
DFT
100
90
80
‘Positive’ power spectra
Ampl (dB)
70
2-quadrant symmetry
60
50
40
0 10 20 30 40
freq
50 80
Ampl (dB)
Ampl (dB)
0 60
−50 40
−100 20
10 20 30 40 0 10 20 30 40
Freq (Hz) Freq (Hz)
Nb=128 Nb=16, m=8
100
50 80
Ampl (dB)
Ampl (dB)
0 60
−50 40
−100 20
10 20 30 40 0 10 20 30 40
Freq (Hz) Freq (Hz)
Nb=256 Nb=16, m=16
100
90
50
80
Ampl (dB)
Ampl (dB)
70
0
60
−50 50
40
−100 30
10 20 30 40 0 10 20 30 40
Freq (Hz) Freq (Hz)
Nb=512 Nb=16, m=32
100 100
50 80
Ampl (dB)
Ampl (dB)
0 60
−50 40
−100 20
0 10 20 30 40 0 10 20 30 40
Freq (Hz) Freq (Hz)
Figure 3.9: Comparison between the power spectra (left) and the ’positive’ power
spectra (right) (full line) and their standard deviation (dotted line).
60 Chapter 3. Non-Parametric preprocessing steps
Table 3.3: Monte Carlo Simulation (30 runs). Mean value and standard deviation of the
estimated damping ratios from primary FRF data obtained by the classic H1 technique
(Nb = 16) and from the H1 followed by a rectangular window (RW) (Nb = 4, m = 4).
dexact (%) dˆh1 (%) σdh1 (%) dˆrw (%) σdrw (%)
0.201 1.214 0.154 0.234 0.007
0.472 0.924 0.054 0.467 0.007
0.587 0.890 0.009 0.593 0.006
0.875 1.041 0.014 0.877 0.006
0.930 1.055 0.012 0.934 0.006
1.135 1.325 0.096 1.139 0.012
1.160 0.986 0.057 1.155 0.013
Consider the same simulation as for the output-only case, but with a measurable
force f4 (t), while the force f5 (t) is unmeasurable. For this case no interaction
between the structure and exciter is assumed i.e. f4 (t) and f5 (t) are uncorrelated.
The unmeasurable force introduces so-called process noise on the FRFs. A Monte
Carlo simulation was done for 30 runs. Tables 3.4 and 3.5 compare the mean and
standard deviation of the natural frequencies and damping ratios estimated by the
combined frequency-domain subspace algorithm for several cases of primary data:
It is clear that the classic H1 approach results in inaccurate and biased estimates.
Furthermore, by using the combined approach, where the parametric identifica-
tion starts from both the FRFs and the positive power spectra of the stochastic
contribution, the accuracy of the estimation of the 2 and 4th mode is improved
by a factor 2 to 3. This is explained by the fact that these modes are less excited
by the measurable force than by the unmeasurable forces. Furthermore, it can be
seen that the damping ratios of the first modes are identified with higher accuracy
if the initial FRFs are estimated from 4 blocks instead of 8 blocks. This can be
explained by the effect of leakage on the initial respectively 4 and 8 blocks.
3.6. Simulations and measurement examples 61
Table 3.4: Monte Carlo Simulation (30 runs). Mean value and standard deviation of
the estimated natural frequencies for the OMAX case. Comparison between the classic
H1 technique (fh1 ) with Nb = 128, the H1 combined with rectangular window (fhrw ) for
both Nb = 4, m = 32 and Nb = 8, m = 16 and the combined FRF-positive power density
approach (fc ) for both Nb = 4, m = 32 and Nb = 8, m = 16
fexact fh σf fh σf fc σf fh σf fc σf
1 h1 rw hrw c rw hrw c
Nb = 32 Nb = 4 m = 32 Nb = 4 m = 32 Nb = 8 m = 16 Nb = 8 m = 16
6.41 6.40 0.02 6.41 0 6.41 0 6.41 0 6.41 0
15.01 15.10 0.11 15.03 0.05 15.02 0.02 15.03 0.05 15.02 0.01
18.67 18.68 0.02 18.67 0 18.67 0 18.67 0 18.67 0
27.86 27.86 0.08 27.86 0.07 27.87 0.04 27.87 0.06 27.87 0.02
29.59 29.60 0.01 29.60 0.01 29.59 0.01 29.60 0.01 29.60 0.01
36.14 36.16 0.15 36.16 0.11 36.17 0.11 36.18 0.16 36.19 0.15
36.91 36.91 0.02 36.91 0.03 36.91 0.03 36.91 0.04 36.91 0.04
Table 3.5: Monte Carlo Simulation (30 runs). Mean value and standard deviation of
the estimated damping ratios for the OMAX case. Comparison between the classic H1
technique (σh1 ) with Nb = 128, the H1 combined with rectangular window (σhrw ) for
both Nb = 4, m = 32 and Nb = 8, m = 16 and the combined FRF-positive power density
approach (σc ) for both Nb = 4, m = 32 and Nb = 8, m = 16
dexact dh σd dh σd dc σd dh σd dc σd
1 h1 rw hrw c rw hrw c
Nb = 32 Nb = 4 m = 32 Nb = 4 m = 32 Nb = 8 m = 16 Nb = 8 m = 16
0.201 1.601 0.236 0.209 0.015 0.207 0.015 0.222 0.027 0.216 0.026
0.472 0.916 0.365 0.501 0.319 0.435 0.109 0.588 0.269 0.451 0.091
0.587 1.032 0.126 0.587 0.004 0.587 0.004 0.588 0.004 0.588 0.004
0.875 1.011 0.237 0.937 0.241 0.860 0.109 0.984 0.274 0.886 0.151
0.930 1.100 0.042 0.938 0.047 0.937 0.046 0.933 0.036 0.931 0.034
1.135 1.059 0.352 1.102 0.277 1.092 0.256 1.157 0.378 1.154 0.366
1.160 1.196 0.088 1.111 0.066 1.112 0.063 1.111 0.065 1.115 0.065
Subframe of car
The Villa Paso bridge, situated in the southeast of Italy, is an arch bridge, recon-
structed after the second world war. Measurements were performed on the deck
of the bridge in cooperation with the university of L’Aguilla in the context of a
62 Chapter 3. Non-Parametric preprocessing steps
20 20
0 0
Ampl (dB)
Ampl (dB)
−20 −20
−40 −40
−60 −60
−80 −80
100 200 300 400 500 100 200 300 400 500
Freq (Hz) Freq (Hz)
Nb=32 Nb=8, m=4
20 20
0 0
Ampl (dB)
Ampl (dB)
−20 −20
−40 −40
−60 −60
−80 −80
100 200 300 400 500 100 200 300 400 500
Freq (Hz) Freq (Hz)
Nb=64 Nb=8, m=8
20 20
0 0
Ampl (dB)
Ampl (dB)
−20 −20
−40 −40
−60 −60
−80 −80
100 200 300 400 500 100 200 300 400 500
Freq (Hz) Freq (Hz)
Nb=128 Nb=8, m=16
20 20
0 0
Ampl (dB)
Ampl (dB)
−20 −20
−40 −40
−60 −60
−80 −80
100 200 300 400 500 100 200 300 400 500
Freq (Hz) Freq (Hz)
Figure 3.10: Subframe. left: classic H1 FRF estimate from Nb blocks; right: H1
estimate from 2 blocks, followed by a rectangular window (full line: FRF, dotted line:
standard deviation)
3.6. Simulations and measurement examples 63
structural health monitoring program for the local authorities. Figure 3.11 (a),(b)
shows the bridge and the hammer used for the force excitation. The span was di-
(a) (b)
8 9 10 11 12 13 14
1 2 3 4 5 6 7
(c)
Figure 3.11: (a) Villa Paso arch bridge; (b) Hammer exciter (c); Schematic overview
of the measured grid
vided in 28 measurement locations shown by figure 3.11 (c), which were measured
in both horizontal and vertical direction. During the modal analysis experiment
64 Chapter 3. Non-Parametric preprocessing steps
the bridge was open for traffic and a forced vibration experiment was performed
by excitation by a calibrated hammer in node 2.
Accel. m/s2
0
−2
−4
−6
0 100 200 300 400 500 600 700
Time (s)
Accel. m/s2
0
−2
(a) (b)
Figure 3.12: a) Autopower of the Hammer excitation force spectra b) Vertical vibration
response in node 2
The hammer excitation generated low force levels for the low frequencies and
higher force levels for the higher frequencies, which is illustrated by Figure 3.12
(a). The traffic typically excites the structure well for the lower frequencies. Fig-
ure 3.12 (b) illustrates the vibration response in the vertical direction in node
2 from the repeated hammer excitation and the traffic. An OMAX approach is
applied to estimate the FRFs and the positive power spectra of the stochastic
contribution together with their uncertainty levels. From Figure 3.13 it is clear
that the SNR for the positive power spectra is better for the low frequencies than
the SNR of the FRFs for the low frequencies, while the FRFs have better SNR
for the higher frequencies. This can be explained by the bad excitation of the
low frequencies by the hammer impacts and the low frequent excitation of the
traffic. Since the hammer excitation only excited the vertical modes by hitting
the bridge perpendicular to the deck and the traffic excited both the vertical and
horizontal modes, the positive power spectra contain several modes at the lower
frequencies which are not visible in the FRFs. Finally, it must be noticed that
the amplitudes of the positive power spectra of the horizontal measurement are
clearly higher than the vertical measurement, since the most important low fre-
quent modes are horizontal modes. Figure 3.14 shows 6 modes obtained from the
combined operational-experimental parametric analysis. The first two modes (a)
and (b) of respectively 1.6Hz and 2.6Hz are horizontal bending modes, the next
two modes (c) and (d) of respectively 3.6Hz and 5.7Hz are vertical bending modes,
while the last three modes are coupled horizontal-vertical torsion modes of respec-
tively 9.3Hz, 14.9Hz and 17.5Hz. The first 4 modes could only be extracted with
high quality from the stochastic contribution from the traffic excitation, while the
last three modes are mainly excited by the hammer excitation. In fact, even for
the case where the bridge was closed for any traffic and thus the FRFs are less
3.7. Conclusions 65
20 20
10 10
Ampl. (dB)
Ampl. (dB)
0 0
−10 −10
−20 −20
−30 −30
5 10 15 20 5 10 15 20
Freq. (Hz) Freq. (Hz)
Positive power spectra Nb=8, m=8 Positive power spectra Nb=8, m=8
30 50
20 40
10 30
Ampl. (dB)
Ampl. (dB)
0 20
−10 10
−20 0
−10
5 10 15 20 5 10 15 20
Freq. (Hz) Freq. (Hz)
(a) (b)
Figure 3.13: FRF versus Positive power spectrum for node 4. (a) vertical direction, (b)
horizontal direction (full line: FRF or Positive power spectrum, dotted line: standard
deviation)
disturbed by ambient vibration, the first 2 modes could not be extracted from the
FRF data and the 3th and 4th only with a very poor quality. Another example
of operational vibration test on a bridge can be found in [81], where sensitivity
based mode shape normalization is applied.
3.7 Conclusions
In this chapter the estimation of Averaged Based Spectral functions like FRFs
and power spectra has been discussed. Several FRF estimators from literature are
proposed and discussed together with the periodogram and correlogram estima-
tors for power spectra. The use of a rectangular window for noise reduction on the
FRFs was introduced and the corrections for time and frequency-domain leakage
are presented. Furthermore, the concept of ’positive’ power spectra is introduced
by the use of a rectangular window and its benefits are given. Finally, the esti-
mation of both the FRFs from the deterministic contribution and positive power
spectra from the stochastic contribution is proposed in the OMAX framework.
Simulations are performed to show the reduction in leakage errors (especially on
66 Chapter 3. Non-Parametric preprocessing steps
the damping) by the use of the rectangular window for both experimental and
operational modal analysis. OMAX simulations illustrated the gain in accuracy
if both the measurable and unmeasurable forces are considered in the vibration
response. Several techniques are applied to real-life vibration experiments on a
subframe of a car and an arch bridge.
3.8 Appendix 1
In this appendix an expression for the covariance matrix CH1 on the FRFs esti-
mated by the H1 estimator is derived from a sensitivity analysis. A perturbation
of the outputs ∆Yb results in a perturbation ∆H1 given by
Nb
!
1 X
∆H1 = H
∆Yb Fb Sf−1
f (3.61)
Nb
b=1
Applying the Vector operator and eliminating the perturbation of the perturbation
∆Yb by the use of the Kronecker product results in
Nb
1 X
vec(∆H1 ) = Sf−T ∗
f Fb ⊗ INo ∆Yb (3.62)
Nb
b=1
Assuming the output noise to be stationarity and uncorrelated over the different
blocks the covariance matrix is given by
CH1 = E vec(∆H1 )(vec(∆H1 )H )
(3.63)
Nb
1 X
= 2 Sf−T ∗ T −1
f Fb Fb S f f ⊗ CY (3.64)
Nb
b=1
(E the expected value) with the covariance matrix of the output noise ∆Y given
by
CY = E (∆Y ∆Y H ) = Syy − H1 Sf y (3.65)
H T ∗ T
since Y = HF + ∆Y . Applying the property (F F ) = F F and substituting
3.2 in 3.64 results in
1 −T
CH1 = S ⊗ CY (3.66)
Nb f f
The multiple input coherence function [52] mγo2 describes the amount of energy
of output o correlated with the Ni inputs and is given by
Syo f Sf−1
f Sf yo
mγo2 = (3.67)
Syyo
H
H1,o Sf f H1,o
= (3.68)
Syyo
3.9. Appendix 2 67
with Syyo , Syo f , Sf yo the auto and cross power for output o and H1,o ∈ C 1×Ni
the FRFs of the H1 FRF matrix estimate between output o and the inputs. The
covariance matrix of Ho is given in function of the multiple coherence function as
1
CH1,o = (1 − mγo2 )Syyo Sf−1
f (3.69)
Nb
In the single input case the expressions for the coherence function mγo2 and co-
variance matrix CH1,o reduces to
Syo f Sf yo
mγo2 = γo2 = (3.70)
Sf f Syy,o
(3.71)
and
(1 − γo2 )
CH1,o = varH1,o = |H1,o |2 (3.72)
Nb γo2
In [105] it is shown that for parametric identification algorithms that use the
covariances on the FRFs as a frequency-domain weighting, at minimum of inde-
pendent blocks must be averaged to result in consistent estimates of the system
parameters.
3.9 Appendix 2
In many cases frequency-domain data, i.e. FRFs, are given as primary data. In
the case that periodic input signals such as multisines are used, these measured
FRF are often of high quality and contain no errors caused by leakage. The
discrete-time domain modal model is given by
Nm
φ∗r LH
X φr Lr r
H(ωk ) = −1 + (3.73)
r=1
1 − λ r zk 1 − λ∗r zk−1
Since many identification methods start from IRF as primary data, the FRFs are
converted to IRF by considering the IDFT
N
1 X
hn = H(ωk )zkn (3.74)
N
k=1
Nm N
!
1 X X zkn ∗ H zkn
= φr Lr + φ L
r r (3.75)
N r=1 1 − λr zk−1
k=1
1 − λ∗r zk−1
1 − λN zk−N
= (3.79)
1 − λzk−1
1 − λN
= (3.80)
1 − λzk−1
with zk = ei2πk/N and thus zkN = 1. Since the IDFT and DFT operations cancel
each other it holds that
xn = λn (3.81)
N −1
1 X
= X(ωk )zkn (3.82)
N
k=0
N −1
1 X (1 − λN )zkn
= (3.83)
N
k=0
1 − λzk−1
Nm
X ′ ′ n
= φr LrT λnr + φ∗r LrH λ∗r (3.86)
r=1
Lr
with L′r = (1−λ N). Several discrete time-domain algorithm as e.g. the Least
r
Squares Complex Exponential (LSCE), Eigenvalue Realization Algorithm (ERA),
time-domain subspace methods estimate a modal model based on Eq. 3.86. Often
these IRF are obtained from the IDFT of the FRFs, but yet often the correction of
′
the estimated participation vectors Lr is neglected. Although this correction can
be essential in the case of lightly damped structures. In [77] a similar correction
for models obtained from IRF calculated from the DFT of the FRFs is proposed
for discrete-time state-space models.
3.9. Appendix 2 69
(a) (b)
(c) (d)
(e) (f)
(g)
Figure 3.14: Vibration mode shapes of the deck of the Villa Paso bridge. (a)-(b)
horizontal bending modes, (c)-(d) vertical bending modes, (e)-(f)-(g) torsion modes
70 Chapter 3. Non-Parametric preprocessing steps
Chapter 4
Identification of
common-denominator
models
This chapter starts with a short overview of modal parameter identification algo-
rithms. Next, the attention is focussed to the identification of common denom-
inator models starting from both ABS-functions and IO data. The models are
extended for dealing with leakage when starting from FRFs obtained by the H1
estimator. It is shown how the identification starting from IO data from differ-
ent experiments can be combined without suffering from transients and leakage.
The analogy between the FRF-driven and IO data-driven algorithms is illustrated.
Finally, a combined deterministic-stochastic identification algorithm is proposed,
which estimates a common-denominator model from the measurements. Finally
an introduction to in-flight tests is given, together with a comparison of different
methods based on in-flight simulations of an aircraft.
71
72 Chapter 4. Identification of common-denominator models
4.1 Introduction
During the last decades many researchers have devoted an extensive effort in the
development of reliable identification algorithms for the characterization of dy-
namical structures by its modal parameters. Starting from Single Degree Of Free-
dom (SDOF) SISO methods e.g. the Peak Picking method [8] and Circle Fitting
method [61], the algorithms have evolved to Multiple Degree Of Freedom (MDOF)
MIMO methods as a result of the introduction of multi-channel acquisition, pow-
erful computer memory and processors. Furthermore, since the introduction of the
Fast Fourier Transform (FFT) frequency-domain algorithms have been developed
as the time-domain counterparts for identification for modal analysis. MDOF
estimators are able to estimate several modes simultaneously, while MIMO are
capable to handle data from several vibrations responses and force inputs simulta-
neously. The first MDOF MIMO time-domain model parameter estimators were
the Ibrahim Time-Domain (ITD) method, the Eigensystem Realization Algorithm
(ERA) [56], [57] and the Least Squares Complex Exponential (LSCE) method.
Both the ITD and LSCE estimate polynomial coefficients. The LSCE algorithm
is closely related to the class of Prediction Error Methods (PEM) [67]. The ERA
algorithm starts from a state-space formulation and requires the singular value
decomposition (SVD) of the so-called block Hankel matrix. More recently the
state-space formulation has been the basis for subspace identification algorithms
[123]. Several MDOF MIMO frequency-domain algorithms are developed, like
e.g. the Least Squares Frequency Domain (LSFD), ERA frequency domain and
Frequency-domain Direct Parameter Identification algorithm (FDPI). The LSFD
directly estimates the modal parameters by solving a non-linear optimization. In
the case that the poles and/or participation factors are already estimated in a first
step by e.g. LSCE, ITD, the LSFD reduces to a linear problem in the unknown
modal parameters (mode shapes) [68]. The ERA frequency domain version starts
from a complex block matrix with the FRFs as primary data and estimates the
modal parameters by the use of SVD and eigenvalue decomposition (EVD) [58].
Based on a LMFD in the Laplace domain, the FDPI estimates the modal pa-
rameters from the FRFs [63]. More recently several frequency domain algorithms
based on a common-denominator description are proposed to handle large and
noisy data-sets from EMA [131] and OMA [80] tests. Several implementations
solve the identification problem in a least-squares (LS), total least squares (TLS),
generalized TLS (GTLS) and maximum likelihood (ML) sense, each of them cor-
responding to different asymptotic properties [93]. A more profound overview of
modal parameter estimation algorithms can be found in [131], [52].
In [99] it is shown for SISO systems that the initial and final conditions of the
response signal can be modelled by an additional transient polynomial. In the case
that the initial and final conditions are equal (i.e. the response is periodic) the
extra polynomial disappears. Consider the vibration response yo (t) in location o
and the input fi (t) in location i, their relation can be written as
n
X Ni X
X n
an−r yo ((n − r)∆t) = boi,n−r fi ((n − r)∆t) (4.1)
r=0 i=1 r=0
with ∆t the sample period. In the appendix it is shown that the discrete Fourier
transformation of equation 4.1 is given by
Ni
X
A(zk )Yo (zk ) = Boi (zk )Fi (zk ) + To (zk ) (4.2)
i=1
the transient polynomial for output o. In fact, this transient polynomial models
the non-steady state response of the structure and in this way errors introduced
by leakage are avoided. The use of the additional transient polynomial improves
the damping estimates and the overall quality of the estimated modal model. By
taking into account this extra polynomial it is possible to deal with both arbitrary
signals (non-periodic) and periodic signals corrupted by transients under the same
assumptions as time-domain methods [99].
N b
[Bo (zk )] X Tob (zk ) H
SYo F (ωk ) = SF F (ωk ) + F (ωk ) (4.3)
A(zk ) A(zk ) b
b=1
with Bo (zk ) = [ Bo1 (zk ) . . . BoNi (zk ) ]. Multiplication by the inverse of SF−1F
leads to an extended parametric model for the H1 FRF estimates. Generally
N
b
Boi (zk ) X Tob (zk ) t
Hoi (ωk ) = + F (ωk ) (4.4)
A(zk ) A(zk ) bi
b=1
with Fbit (ωk ) the ith element of the row vector Fbt (zk ) defined by
From these equations it is clear that generally the H1 estimator is not equal to
the exact model B(z k)
A(zk ) and that leakage and transient phenomena have some noisy
influence, if the responses are not periodic. Since the spectra Fbit (ωk ) can be cal-
culated from the force measurements the transient polynomials can be estimated.
This extended formulation of the H1 estimator makes FRF based identification
methods attractive to process shorter data sequences. The major advantage of
all identification methods based on FRF data is that the noise levels decrease by
averaging different data blocks, while this is not the case by identification meth-
ods starting from the raw input and output spectra (except if periodic signals are
used). It is important to remark that time domain methods e.g. LSCE [52] also
suffer from leakage, if they start from impulse response functions obtained by the
Inverse Fourier Transformation (IFT) of the FRFs.
4.3. Weighted Least Squares Complex Frequency-domain (LSCF) estimator 75
The model for the FRF between output o and input i estimated by the H1 esti-
mator is given by
Nb
Boi (zk ) X Tob (zk ) t
Hoi (ωk ) = + F (ωk ) (4.6)
A(Ωk ) A(zk ) bi
b=1
with Boi (zk ), A(zk ) and Tob (zk ) scalar polynomials in the basis functions zkj given
by
n
boi,j zkj
X
Boi (zk ) = (4.7)
j=0
n
aj zkj
X
A(zk ) = (4.8)
j=0
n
tob,j zkj
X
Tob (zk ) = (4.9)
j=1
with
βo1
..
a0 boi,0 tob,0
.
a
b
t
βoNi
1 oi,2 ob,2
ηo = , α= .. , βoi = .. , τob = ..
τo1
. . .
..
. an boi,n tob,n
τoNb
The unknown polynomial coefficients differ in their level of being global parame-
ters:
• the numerator coefficients βoi are local parameters depending on the output
o and input i location
• the transient coefficients τob depend on both the output o location and the
data block b.
Eoi (ωk )
?
Woi (ωk )
A(zk )
e
Fbi (ωk ) - Tob (zk )
A(zk )
@
R +?
@
@ h - Hoi (ωk )
1 - Boi (zk )
A(zk )
By multiplying Eq. 4.6 with A(zk ) and by taking the difference between the left
and right hand side an equation error is obtained that is linear in the parameters
Nb
!
LS 1 X
t
Eoi (ωk ) = A(zk )Hoi (ωk ) − Boi (zk ) − Tob (zk )Fbi (ωk ) (4.11)
Woi (ωk )
b=1
with Woi (ωk ) a frequency-domain weighting function. Figure 4.1 visually repre-
sents the different contributions in the weighted linear least-squares model. The
weighted linear least-squares problem is found by minimizing
Nf
Ni X
No X
X
lFLSCF
RF (θ) = LS
|Eoi (ωk )|2
o=1 b=1 k=1
Jθ = 0 (4.12)
4.3. Weighted Least Squares Complex Frequency-domain (LSCF) estimator 77
with J the Jacobian matrix (which is in this case independent of the parameters)
given by
κ1 0 . . . 0 Υ1
0 κ2 . . . 0 Υ2
J = . . . .. (4.13)
.. .. . .. .. .
0 0 ... κNo ΥNo
with
Γo1 0 ... 0 Ξo11 ... Ξo1Nb
0 Γo2 ... 0 Ξo21 ... Ξo2Nb
κo = .. .. .. .. .. (4.14)
.. ..
. . . . . . .
0 0 ... ΓoNi ΞoNi 1 ... ΞoNi Nb
ςo1
Υo =
.. (4.15)
.
ςoNi
algorithms if the structure of the matrices is optimally used. The Jacobian matrix
J of this least squares problem has N No Ni rows and (n + 1)(No Ni + No Nb + 1)
columns (with the number of rows larger than the number of columns to make
the problem identifiable). The number of frequencies N can be eliminated by
formulating the normal equations, i.e.
J H Jθ = 0 (4.19)
K1 ··· 0 L1
η1
.. .. .. ..
.
. . . = 0 (4.20)
0 KNo L o ηNo
PNN
LH ··· LH o
o=1 To
α
1 No
The specific structure in the normal equations result from the use of a common-
denominator model and will allow a fast implementation. All the parameters θ are
divided in output dependent parameters ηo and output independent parameters
α. By taking a closer look to the model it can be seen that the parameters ηo
can be split up in a set βoi dependent of the input location i corresponding with
the FRF Hoi (ωk ) and a set τob independent of the input location. As a result the
submatrices of Eq. 4.20 have a structure by itself given by the following equations.
Mo1 · · · 0 No1
.. .. ..
Ko = .
. .
0 MoNi NoNi
H H
No1 · · · NoN i
Ro
Qo1
..
Lo = . (4.21)
QoNi
So
Moi = ΓH
oi Γoi
Qoi = ΓH
oi ςoi
H
Noi = Γoi Ξoi1 ... ΞoiNb
Ni
X H
Ro = Ξoi1 ... ΞoiNb Ξoi1 ... ΞoiNb (4.22)
i=1
Ni
X H
So = Ξoi1 ... ΞoiNb ςoi
i=1
Ni
X
H
To = ςoi ςoi
i=1
4.3. Weighted Least Squares Complex Frequency-domain (LSCF) estimator 79
The maximum gain in calculation time is obtained by taking into account the
structure of the normal equations. By simple algebraic elimination procedures the
numerator polynomial coefficients boi can be eliminated from Eq. 4.20 and 4.21
yielding
−1
τo
βoi = −Moi Noi Qoi (4.23)
α
τo = −Ro −1 So α (4.24)
with
Ni
X
H −1
Ro = Ro − Noi Moi No,i (4.25)
i=1
Ni
X
H −1
So = So − Noi Moi Qo,i (4.26)
i=1
From the last n rows of the normal equations Eqs. 4.20 the following equations
can be formed
No
X
LH
o ηo + To α = 0 (4.27)
o=1
No
X
−So H Ro −1 So + To α = Dα = 0
(4.28)
o=1
PNo
−So H Ro −1 So + To and
with D = o=1
Ni
X
−1
To = To − QH
oi Moi Qo,i (4.29)
i=1
Eq. 4.28 can be solved by a simple matrix inversion, with dimension n (the order
of the polynomial A(zk )), leading to the denominator coefficients α. Backward
substitution in Eq. 4.24 and Eq. 4.23 provides the coefficients τob and βoi . Ex-
amining the submatrices Moi , Noi , Qoi , Ro , So and To in more detail reveals that
80 Chapter 4. Identification of common-denominator models
Since the entry [r, s] of the submatrices only depends on r − s, these matrices have
a so-called Toeplitz structure. In practice this means that only one column in case
of an Hermitian matrix e.g. Moi or one column and one row in the general case e.g.
Qoi must be calculated to construct the total matrix resulting in both a memory
and time efficient computation. Furthermore, an additional gain in computation
time can be obtained by the using the Fast Fourier Transform (FFT) to compute
the entries Eqs. 4.30-4.32 of the matrices. The matrices Nob , Rob and So can be
decomposed in submatrices with similar structures and expressions as the above
discussed matrices and thus the memory and time efficient computation can also
be applied in the same way.
To solve the identification problem, the parameter redundancy (and to avoid the
trivial solution with all coefficients equal to zero) of the transfer function model
must be removed. It can easily be seen that a transfer function model defined by
a set of parameters θ, can also exactly be defined by γθ
b N
Boi (zk ) X Tob (zk ) t
Hoi (ωk ) = + F (ωk ) (4.33)
A(Ωk ) A(zk ) bi
b=1
b N
γBoi (zk ) X γTob (zk ) t
= + F (ωk ) (4.34)
γA(Ωk ) γA(zk ) bi
b=1
The mixed LS-Total Least Squares (TLS) solution given α = V[:,n+1] with V the
right singular vectors obtained by an SVD of the matrix D
D = U SV H (4.36)
4.3. Weighted Least Squares Complex Frequency-domain (LSCF) estimator 81
Continuous-time model
damping estimates.
In the case that the transient polynomials are neglected, the extended LSCF es-
timator reduces to the original LSCF estimator as proposed in [46]. The original
LSCF estimator models a common denominator model, given by
Boi (zk )
Hoi (ωk ) = (4.37)
A(zk )
on the FRFs in a least squares sense. The use of a steady state response of
a system excited by periodic excitation signals eliminates this problem since no
leakage is introduced. In practice, even in the case of arbitrary excitation, the
influence of leakage can often be neglected in the case the non-parametric FRF
estimation has taken place with caution as shown in chapter 3. However, in the
case of very lightly damped structures and the use of arbitrary excitation signals
a significant improvement of the estimated parameters and stabilization diagram
can be observed by modelling the initial and final conditions of each block. The
LSCF estimator, initially proposed as a starting value generator for the Maximum
Likelihood estimator, forms the basis of several other algorithms. They can employ
the same procedures to obtain a fast and memory efficient parameter estimation
method as discussed in this section.
Thanks to the mathematical similarity between the modal model for FRFs and
power spectra in the case of an operational modal analysis, the estimator with-
out the estimation of the transient polynomials can also be applied to start from
power densities [40]. In case one is interested to process operational data with
transients one has to stick to data driven methods e.g. the combined deterministic-
stochastic frequency-subspace algorithm of chapter 8. In the case one starts from
power spectra the algorithms are not capable to make a distinction between tran-
sients (deterministic contribution) and the stochastic contribution (notice that the
transients i.e. intitial/final conditions differ from block to block and thus no dis-
tinction can be made based on an averaging procedure). Starting from the positive
power spectra results in more interesting properties concerning the construction
of stabilization diagrams as shown in chapter 9.
Asymptotic properties
Modal parameter estimation often requires that a trade-off must be made between
accuracy and computation speed, since the data sets are typically characterized
4.3. Weighted Least Squares Complex Frequency-domain (LSCF) estimator 83
by a huge amount of data. In the general case the LSCF is not (asymptotic)
unbiased, not consistent and not efficient, since the equation error Eoi (ωk ) is not
white uncorrelated noise with bounded second order moments. In practice this
means that in the presence of noise on the FRFs:
To improve these properties more advance estimators like the Maximum Likelihood
estimator are proposed (discussed in more detail in paragraph 4.4). A detailed
study and comparison of the different frequency-domain identification approaches
can be found in [93].
In some cases one is interested to start directly from the measured input and output
spectra to identify the system parameters e.g. if only a limited of time samples
are available. In the most general case, different blocks of data from different
experiments can be combined in one identification procedure. The output o for
block b can be modelled as
Ni
X Boi (zk ) Tob (zk )
Yob (ωk ) = Fib (ωk ) + (4.38)
i=1
A(zk ) A(zk )
Classical FRF based estimators do not take into account the influence of the
transient and leakage effects. The extension of the LSCF FRF driven estimator
to its extended formulation takes into account the leakage and transient effects.
Comparing these extended FRF estimators with the LSCF IO based estimators
reveals their strong analogy. Consider both the FRF and IO models respectively
84 Chapter 4. Identification of common-denominator models
Eob (ωk )
?
1
A(zk )
2
Ni 1
0
0
5
Z
10
15
No 20
25
2 0
6 4
30 10 8
Nb
given by Eqs. 4.6 and 4.38. It is clear that the role of the inputs i is swapped with
the blocks b. This strong analogy results in the same mathematical implementation
for the LSCF FRF driven and the LSCF IO data driven estimators. Only the
primary data to the estimator and the interpretation of the estimated parameters
are changed. This analogy can be visualized in a 3-dimensional input-output-block
space shown in figure 4.3. IO data driven estimators work in the No × Nb and
Ni × Nb planes and need as a result Nb (No + Ni )Nf complex data samples. For
H1 based FRF estimators the data is situated in the No × Ni and Nb × Ni planes.
Compared to IO estimators the block axis is swapped with the input axis. From
this point of view it is clear that in the case Nb > Ni FRF based estimators can
4.4. Maximum Likelihood identification 85
operate from a smaller amount of data. In case no leakage and transient effects are
present (e.g. periodic excitation) the block axis disappears, since both IO data and
FRF data can be averaged over the blocks and no transient polynomials have to
be estimated. Classical FRF driven estimators often neglect the block axis, while
this condition (no leakage and transient effects) is not fulfilled, resulting in bias
errors. The use of periodic excitation signals simplifies the estimation problem a
lot, since no transient polynomials have to be identified. The block dimension will
disappear, if cyclic averaging takes place over the different periods. Furthermore,
cyclic averaging reduces the noise levels on both the responses and forces and gives
the possibility to obtain noise information.
Linear least-squares based algorithms like the LSCF algorithm are often used in
modal analysis for their ability to handle large amounts of data in a reasonable
time. However, the LS approach is not consistent and not efficient. This results in
biased estimates and in the cases that a LS approach is not accurate enough more
sophisticated estimators must be used. In the following paragraph a consistent
algorithm is proposed, which estimates a common-denominator model starting
from FRFs or IO data in a Maximum Likelihood (ML) sense. This algorithm
was introduced in [46] and applied to both experimental and operational modal
analysis experiments [40].
N
b
Boi (zk ) X Tob (zk ) t
Hoi (ωk ) = + F (ωk ) + σHoi (ωk )Eoi (ωk ) (4.41)
A(zk ) A(zk ) bi
b=1
Minimizing the quadratic cost function results in the maximum likelihood solution
under the assumption that Eoi (ωk ) is uncorrelated over the different FRFs [37],
[98]. The (negative) log-likelihood function is then given by
Nf
Ni X
No X
X
lFMRF
L
(θ) = ML
|Eoi (ωk )|2 (4.42)
o=1 b=1 k=1
86 Chapter 4. Identification of common-denominator models
with Woi (ωk ) = σHoi (ωk ). Notice that the equation error is non-linear in the sys-
tem parameters and thus a non-linear optimization algorithm such as e.g. Gauss-
Newton is required to estimate the polynomial coefficients. The ML equation error
is equal to the LSCF equation error but weighted by its standard deviation result-
ing in an equation error, which is white uncorrelated noise over the spectral lines
k. Notice that if the transients are taken into account by means of the use of the
H1 estimator the inputs are assumed to be free from noise. In [41] it is shown that
the logarithmic equation error, given by
PNb
Boi (zk ) Tob (zk ) t
log(Hoi (ωk )) − log A(zk ) + b=1 A(zk ) Fbi (ωk )
Elog oi (ωk ) = Woi (ωk )
|Hoi (ωk )|
is more robust for a large dynamical range, outliers and incorrect noise information.
Weighting function
• The assumption of white additive noise on the responses Yob (ωk ) corresponds
to a weight
v
u Ni
uX −1
Woi (ωk ) = t |S (ωk )|2 σ 2
F ji Fj
j=1
4.4. Maximum Likelihood identification 87
SF−111 SF−11Ni
...
.. .. .. −1
= SF F (4.43)
. . .
SF−1Ni 1 ... SF−1Ni Ni
By weighting the equation error with this weighting function, frequency lines
which contain not much energy over the different blocks, are not taken into
account in the cost function.
Gauss-Newton optimization
H H
(a) solve Jm Jm ∆Θm = −Jm E (4.44)
(b) set Θm+1 = Θm + ∆Θm (4.45)
with the Jacobian matrix J defined as the derivative of the equation errors Eoi (k)
with respect to the parameters θ for all spectral lines k, outputs o and inputs i.
It can easily be shown that the Jacobian matrix has a similar structure as the one
(cf. Eqs. 4.13 and 4.14) in the LSCF case. The submatrices Γoi , ςoi and Ξoib are
88 Chapter 4. Identification of common-denominator models
now given by
1
z1n
A(z1 )Woi (ω1 )
1 z1 ...
1
1 z2 . . . z2n
A(z2 )Woi (ω2 )
Γoi = − (4.46)
..
.
1 n
A(zN )Woi (ωN )
1 zN . . . zN
P Nb t
Boi (z1 )+ b=1 Tob (z1 )Fbi (ω1 )
z1n
A(z1 )2 Woi (ω1 )
1 z1 ...
P Nb t
Boi (z2 )+ b=1 Tob (z2 )Fbi (ω2 )
z2n
1 z2 ...
ςoi = A(z2 )2 Woi (ω2 ) (4.47)
..
.
P Nb
Boi (zN )+ b=1 t
Tob (zN )Fbi (ωN ) n
A(zN )2 Woi (ωN )
1 zN ... zN
1
z1n
t
A(z1 )Woi (ω1 ) Fbi (ω1 )
1 z1 ...
1 t
z2n
A(z2 )Woi (ω2 ) Fbi (ω2 )
1 z2 ...
Ξoib = − (4.48)
..
.
1 t n
A(zN )Woi (ωN ) F bi (ω N ) 1 zN ... zN
with
Uo,i = ΓH
oi Eoi (4.51)
Ni
X
Vo = [Ξoi1 . . . ΞoiNb ] Eoi (4.52)
i=1
Ni
X
Wo = ςoi Eoi (4.53)
i=1
similar as in the LSCF case. The initial parameter estimates θ0 (starting values)
to construct the normal equations in the first iteration are estimated by the LSCF
estimator. In fact, the ML estimator optimizes the results obtained by the least
squares procedure. From practical studies, a serious improvement of the accuracy
is observed after even 10 iterations, while the use of a Levenberg-Marquard loop
forces the algorithm to converge by solving
J H J + λLM diag(J H J) θ = −J H E
(4.55)
Increasing the parameter λLM forces the cost function to decrease, but decreases
the convergence speed. Therefore this factor λLM is adapted in every iteration
depending on the evolution of the cost function.
The same analogy between the FRF driven ML and IO data driven ML is present
as it is the case between the FRF and IO driven LSCF. In [136] it is shown how
both the noise on the input measurements and output measurements can be taken
into account by the use of periodic excitation and cyclic averaging.
One of the advantages of the ML estimator is the capability to estimate the con-
fidence intervals on the estimated parameters by the knowledge of the noise infor-
mation on the primary data. In [46] this capability is discussed in detail, while
in [141], [130] these noise intervals are considered as a tool to distinguish physical
from mathematical poles. Even in the case that the noise assumptions are violated
this tool still seems to work well.
Eo (ωk )
?
Co (zk )
A(zk )
Fb (ωk ) - Bo (zk )
A(zk )
@
R +?
@
@ h - Yob (ωk )
1 - To (zk )
A(zk )
Consider the situation where Nb measured time blocks yob (t) and fb (t) are given
to start from. (All further developed formulas are written for Ni = 1 for reason of
simplicity). For each output o and block b, with Yo,b and Fb respectively the DFT
spectra of yo,b (t) and fb (t), the input-output relation is given
with Eo,b (ωk ) unknown Gaussian white noise. The combined least-squares frequency-
domain (CLSF) estimator will estimate the system parameters A(z), Bo (z) and
Co (z) simultaneously with the initial and final conditions Tob (z). This Gaussian
contribution in the vibration response CA(z
o (zk )
k)
Eo,b (ωk ) can be physically interpreted
as the contribution from unmeasurable random forces [22], [23]. The CLSF-IO
method, visualized in Figure 4.4, forces a parametric model on the input-output
measurements in such a way that the residues Eo,b (ωk ) (in our case the unmeasured
forces) will be white noise. The vibration response consists in three contributions
respectively from the measurable forces, from the noise and unmeasurable forces
and from the transients. Identification of this model is closely related to the iden-
tification process of ARMAX models in the time-domain where the minimization
of the so-called prediction error leads to the model parameters [67]. Minimizing
the cost function
No XNb X Nf
X
CLSF
lIO (θ) = |Eo,b (ωk )|2 (4.57)
o=1 b=1 k=1
4.5. Combined deterministic-stochastic identification 91
with
leads to the estimates. In [98] the theoretical background is discussed for SISO
systems. Special attention must be paid to ensure numerical stability during the
estimation process, since this CLSF IO cost function can decrease be increasing the
polynomials Co (zk ). By rejecting the zeros of Co (zk ) into the unity circle during
each iteration numerical stability can be enforced [22]. Since only the magnitude
of the polynomial Co (ωk ) matters in the cost function, only the magnitude of the
mode shapes can be extracted from the stochastic contribution in the vibration
response. As a result mode shapes of modes, only excited by the unmeasurable
forces can only by identified in magnitude.
Gauss-Newton optimization
Similar as the MLE cost function, the CLSF cost function is the sum of the
quadratic equation errors over the different outputs. A Gauss-Newton optimiza-
tion is required, since the equation error is again non-linear in the parameters. In
general the same procedures as explained for the LSCF and ML estimators can
be used to speed up the algorithm and to reduce the memory requirements. The
structure of the normal equations, the Toeplitz structure of the submatrices and
the fast calculation of the entries by the FFT algorithm can be used to obtain a
practical algorithm in term of memory requirements and calculation time. More
specific details about the implementation of the algorithm and some applications
can be found in [23], [109]. The starting values for the coefficients of the polynomi-
als A(z), Bo (z) and Tob (z) are obtained from the IO data driven LSCF algorithm
or in the more noisy case from the IO data driven ML algorithm. The starting
values for the coefficients of the polynomials are assumed to be equal to the de-
nominator coefficients corresponding to the case of white noise on the responses.
During the optimization, the model has the freedom to shape the white noise to
colored noise. In [98] it is shown that to be consistent the method requires a uni-
formPfrequency grid covering the halve unit circle in the case of real coefficients i.e.
N
Re( k=1 zk ) = 0 and the complete unit circle in the case of complex coefficients
PN
i.e. k=1 zk = 0. Therefore in practice, when processing a frequency band, the
frequencies are re-scaled to fulfill this requirement. Unfortunately this requirement
introduces model errors. Therefore, overmodelling is required and often only 90%
of the halve or complete unit circle is covered to make a trade-off between model
orders and consistency. Notice that the ML estimator does not require a uniform
frequency grid and full coverage to be consistent.
92 Chapter 4. Identification of common-denominator models
The strong analogy between FRF and IO data driven identification makes the
implementation of the IO CLSF algorithm also applicable to process FRF data.
In general the CLSF estimator is preferred to process noisy FRFs, if no a priori
known noise information is available.
Different criteria are evaluated to draw a relative comparison between the different
frequency-domain estimators based on a common-denominator model.
CRITERIUM LSCF IO MLCF IO CLSF IO LSCF FRF MLCF FRF CLSF FRF
The most important conclusions are that LSCF methods are fast, but fail in pres-
ence of large noise levels. MLE methods are robust and can deal with noise. An
important benefit is the availability of uncertainty bounds on the estimates, which
are useful to detect mathematical poles [141]. However, the major drawback is
that ML methods require long data sequences to determine the a priori needed
noise information. Nevertheless, even if no noise information is a priori known,
the ML algorithm reduces to a NLLS, which is still consistent. However, slower
than the ML algorithm, since more parameters have to be estimated, the CLSF
4.7. Mixed LS-TLS, SK, IQML, WGTLS and BTLS estimators 93
algorithm is still relative fast. Their major advantage is that they can deal with
noise even in the case that only short data sequences are available and their com-
bined interpretation in the OMAX framework. Important to mention is that all
methods can be further speed up in the case that the leakage and transient ef-
fects are neglected, although this can introduce small bias errors and increases the
uncertainty on the estimates.
Finally, it should be noted that many other variants exist to estimate a common-
denominator model starting from both FRF and IO data. The mixed LS-TLS
approach estimates the denominator coefficients from the compact normal equa-
tions forcing a L2 norm fixed to 1 as a constraint on the denominator coefficients.
The Sanathanan-Koerner approach solves the weighted LS equations iteratively by
taking the weighting function for iteration j Wn (ωk ) = Aj−1 (zk ) with Aj−1 (zk )
the denominator coefficients from the previous iteration j − 1 [103]. The Itera-
tive Quasi Maximum Likelihood (IQML) estimator differs from the Sanathanan-
Koerner approach since it takes into account the noise information in its weighting
Wj,oi (ωk ) = Aj−1 (zk )σHoi (ωk ). Finally, the consistent Generalized Total Least
Squares (GTLS) and the consistent iterative Bootstrapped Total Least Squares
(BTLS) solvers can be applied to estimate the common-denominator model [131].
Both the GTLS and BTLS are discussed in more detail in paragraph 5.2.5 for the
estimation of a RMFD model. A theoretical overview and comparison of the dif-
ferent solvers is given in [93] and more applied to estimate common-denominators
for modal analysis in [131].
Both the LSCF and ML common-denominator estimator have proven their appli-
cability for both experimental and operational modal analysis. In [46], [116], [115]
both the LSCF and ML estimators starting from FRF data are applied to several
structures like e.g. noisy flight flutter measurements and automotive parts. Their
IO data driven LSCF and ML counterparts are presented and applied in [136]. A
BTLS IO data driven implementation is used to estimate the modal parameters
from a time-varying system in [134]. In [40], [50], [82] the ML estimator is applied
to operational modal analysis starting from auto and cross power densities. The
simulations and examples in this chapter are mainly focussed on the extended FRF
models and the combined stochastic-deterministic estimator.
94 Chapter 4. Identification of common-denominator models
4.8.1 Simulations
The same 7-DOF system as explained in paragraph 3.6 is now used for the simu-
lations. All simulations are performed in the time domain in order to be able to
model the transient effects. The structure is excited by a Gaussian random noise
signal at mass 5. The vibration responses and force signals are divided in 4 blocks
to estimate the FRFs. In the next step, the modal parameters are estimated from
these FRFs. For this process a distinction is made between three different cases:
• The FRFs are estimated by the H1 estimator without the use of a window
and the modal parameters are estimated by the classic LSCF estimator (i.e.
no transient polynomials are taken into account). This procedure is denoted
as ’LSCF No window’.
Figure 4.5 clearly shows the quality of the estimated damping ratios for the three
approaches for a decreasing block size from 40s to 2s corresponding with a fre-
quency resolution from 0.025Hz to 0.5Hz. It is clear that the first approach results
in the worst results, while the second approach for long blocks in combination with
a Hanning window still ends up with acceptable estimates. Nevertheless, taking
into account the contribution of the initial and final conditions by modelling an
extra transient polynomial for each block results in accurate estimates for the
damping ratios. Figure 4.6 (a) and (b) show the FRF estimated from 4 blocks of
each 8s when using a Hanning and rectangular window, while (c) and (d) illustrate
the synthesized FRF and error (difference) with the exact FRF for respectively
the classic LSCF and the extended LSCF algorithm. It is clear that the extended
LSCF estimator results in more accurate results, although the primary FRF data
shown in figure 4.6 (b) for the extended LSCF looks more noisy than the primary
FRF data shown in figure 4.6 (a) for the classic LSCF. Nevertheless, the noisy
contribution on the FRFs can perfectly be described by taking into account the
contribution of the initial and final conditions.
4.8. Simulation and Measurement examples 95
1 2
−1 1.5
Damping ratio (%)
−3 1
−4
0.5
−5
−6
0
−7
40 32 24 16 8 4 2 40 32 24 16 8 4 2
Block length (s) Block length (s)
(a) (b)
1.5
Damping ratio (%)
0.5
40 32 24 16 8 4 2
Block length (s)
(c)
Figure 4.5: Damping ratios estimated by (a) the classic LSCF (rectangular window),
(b) by the classic LSCF (Hanning window) and (c) the extended LSCF as a function of
the block length. (∗: estimated damping ratios, dotted line: exact damping ratios)
−60 −60
−70 −70
−80 −80
−90 −90
Ampl. (dB)
Ampl. (dB)
−100 −100
−110 −110
−120 −120
−130 −130
−140 −140
5 10 15 20 25 30 35 40 5 10 15 20 25 30 35 40
Freq. (Hz) Freq. (Hz)
(a) (b)
−60 −60
−80 −80
−100 −100
−120 −120
Ampl. (dB)
Ampl. (dB)
−140 −140
−160 −160
−180 −180
−200 −200
−220 −220
−240 −240
5 10 15 20 25 30 35 40 5 10 15 20 25 30 35 40
Freq. (Hz) Freq. (Hz)
(c) (d)
Figure 4.6: FRF estimated by (a) the H1 estimator and Hanning window and (b) by
the H1 estimator and rectangular window. Synthesized FRF and error (difference) by
(c) the classic LSCF (d) by the extended LSCF. (full line: FRF, dotted line: error
the different flight conditions, which are then extrapolated in order to determine
whether it is save to proceed to the next flight point as shown in figure 4.7. Flut-
ter can occur when one of the damping values tends to become negative. The
speed at which such an instability takes place is called the flutter speed and is
one of the most important design parameters for an aircraft wing. Before start-
ing the flight tests, ground vibration tests as well as numerical simulations and
wind tunnel tests are used to get some prior insight in the problem. Flight flutter
testing continues to be a challenging research area because of the concerns with
costs, time and safety in expanding the envelope of new or modified aircrafts. The
aerospace industry desires to decrease the flight flutter testing time for practical,
economical and safety reasons partially by improving the accuracy and reliability
of the parameter estimation process. The costs of test-flights are enormous and
hence as many flight conditions as possible should be verified in one single flight.
Moreover flight flutter tests can be highly dangerous even when approached with
caution. An aircraft represents a huge investment in terms of time and money and
a flutter occurence can be radically destructive. Several fatal cases are reported
in literature. An important goal of the FLITE project is to reduce the required
4.8. Simulation and Measurement examples 97
test time and to improve the accuracy of the damping ratio estimates by means of
improved system identification techniques. There are numerous system identifica-
tion methods that allow the estimation of the modal parameters from a vibrating
system, but not all of them can deal with flight flutter data. In-flight test data
is typically characterized by short time records and noisy measurements due to
the unmeasured atmospheric turbulent forces. Many traditional methods used in
modal testing work well with clean data, but, as the noise increases and measure-
ment time decreases, the accuracy of the flutter parameter determination rapidly
degrades, especially for the damping ratio estimates. Increasing the level of the
artificial applied force to increase the signal-to-noise ratio is no option, since the
response level is limited to structural integrity or comfort reasons. Furthermore,
non-linear effects will appear and the measured working point will differ from
operational flight conditions. However, the use of specially designed broadband
signals as multisines and chirp signals can improve the quality of the measurements
[12], [47]. Many traditional flight flutter techniques start from averaged data like
frequency response functions (FRFs) or impulse response functions (IRFs), since
usually there is an attempt to reduce the influence of the noise level on the outputs
by collecting many data blocks. However, this increases the required flight time at
a fixed flight condition, which is adversely to the desired procedures. The state-of-
the-art of flight flutter testing techniques in aircraft industry has been extensively
reviewed in [29], [65] and [10].
98 Chapter 4. Identification of common-denominator models
Figure 4.8: Aircraft and the grid used for the GVT
(a) (b)
(c) (d)
(e) (f)
mean stochastic contribution in the vibration responses over the different outputs
and the mean deterministic contribution over all outputs. Simulation flight tests
were carried for levels of atmospheric turbulence of 7, 14, 21, 28, 35 and 42%
(10 different runs for each turbulence level). The ML FRF driven, ML IO driven
and CLSF IO driven algorithm were compared for the different simulations. Figure
4.10 shows the mean estimated value and 68% confidence bars of the damping ratio
estimates (for the 10 runs) of each mode for an increasing level of the turbulence.
Since only a limited amount of data is considered the FRFs estimates are very noisy
and biased due to leakage effects. Since a classical LSCF FRF estimator failed, the
100 Chapter 4. Identification of common-denominator models
1.9
2.6
1.8
2.4
1.6 2.2
1.5 2
1.4
1.8
1.3
1.6
1.2
1.1 1.4
1 2 3 4 5 6 1 2 3 4 5 6
turbulence level turbulence level
6 6
4
4
damping ratio (%)
−8 −6
−10 −8
1 2 3 4 5 6 0 1 2 3 4 5 6 7
turbulence level turbulence level
2.6 2.5
2
2.5
damping ratio (%)
1.5
2.4
1
2.3
0.5
2.2
0
2.1 −0.5
2 −1
1 2 3 4 5 6 1 2 3 4 5 6
turbulence level turbulence level
Figure 4.10: Comparison of the damping ratio estimates (∗ : CLSF IO, ◦ : NLLS IO,
: MLE FRF, – : exact parameter)
classic ML estimator starting from FRFs was used. This ML FRF method suffers
from a bias on the damping ratios of the low frequent modes due to the leakage
effects in the FRF estimates. Furthermore the variance on the ML FRF estimates
is large even in the case of low turbulence levels. (The FRF driven ML did not
take into account the initial and final conditions of each block). From these results
one can conclude that the FRF based methods are not recommended to use in the
case of short time records. Since the IO data driven methods easily model the
transient effects they do not suffer from bias errors in the case of low turbulence
4.8. Simulation and Measurement examples 101
10
0
0
−10
Ampl. (dB)
Ampl. (dB)
−10
−20
−20
−30
−30
−40
−40
−50
2 4 6 8 10 3 4 5 6 7 8 9 10
Freq. (Hz) Freq. (Hz)
(a) (b)
Figure 4.11: Synthesized energy spectra (o: measured output spectra, full line: stochas-
tic + deterministic energy, dashed line: deterministic contribution (artificial force), dot-
ted line: stochastic contribution (turbulent forces))
resulted in unbiased estimates with the smallest uncertainty for all modes even
in the presence of high levels of turbulent forces. Interesting to notice is that
the CLSF IO method estimates the damping ratios of the 3rd, 4th and 5th mode
very accurately, although these modes are weakly excited by the artificial force.
However, the combined approach takes advantage from a better data exploitation,
since its was noted that these modes are well excited by the turbulent forces (these
modes are mainly vibrating at the tail, nose and cabin as can be seen in Figure 4.9
and thus less good excited by the artificial force). Figure 4.11 gives a comparison
between the deterministic and stochastic contribution in the response spectra from
which it is clear that modes 3,4 and 6 at respectively 4.6Hz, 5.1Hz and 8.4Hz have
an important stochastic contribution from the turbulent excitation.
25 40 exact FRF
synthezised FRF
20 20
model order
Ampl. (dB)
15 0
10 −20
5 −40
0 −60
200 400 600 800 1000 200 400 600 800 1000
Freq. (Hz) Freq. (Hz)
Figure 4.12: H1 estimate (Hanning), stabilization diagram and synthesized FRF for
the classical LSCF method
50
exact FRF
25 40 synthezised FRF
30
20 20
model order
10
Ampl. (dB)
15 0
−10
10 −20
−30
5 −40
−50
0 −60
200 400 600 800 1000 200 400 600 800 1000
Freq. (Hz) Freq. (Hz)
Figure 4.13: H1 estimate (rectangular), stabilization diagram and synthesized FRF for
the Extended LSCF method
213 samples) and measured with a sample frequency of 213 Hz. This results in
leakage free FRFs with a frequency resolution of 1Hz and a total measurement
time of 10s for each measured point. These FRFs are considered as the ’exact’
FRFs. In a second set of measurements the plate was excited by 211 samples
random noise and a sample frequency of 213 Hz resulting in a measurement time
of 0.25s for each measured point. These measured sequences were divided into 2
blocks to estimate the FRFs by the H1 technique leading to frequency resolution
of only 8Hz. It is clear that these FRFs are ’difficult data’ to handle because,
The goal is to process this ’difficult’ data by both the classical LSCF and
extended LSCF and to compare the synthesized models from this data to the
’exact’ FRFs. Figures 4.12 and 4.13 show respectively the stabilization chart and
4.9. Conclusions 103
20 30
exact FRF exact FRF
15 Extended log MLE Extended log MLE
Extended LSCF 20 Extended LSCF
10
5
10
0
Ampl. (dB)
Ampl. (dB)
−5 0
−10
−15 −10
−20
−20
−25
−30
−30
400 500 600 700 500 550 600 650 700
Freq. (Hz) Freq. (Hz)
Figure 4.14: Comparison between the Extended LSCF and Extended ML algorithm
a synthesized FRF for the classic and extended LSCF algorithm. If the leakage
is taken into account in the model by Fbit (ωk ), the obtained results are better
shown by the synthesized FRF. Comparison of both stabilization diagrams clearly
indicates that the extended LSCF method stabilizes for each physical pole, which
was not the case for the Classical LSCF (*: stable pole, +: unstable pole).
The extended LSCF outperforms the classical LSCF for these short time mea-
surements, but the obtained models can still be improved by the extended ML
algorithms starting from the initial estimates by the extended LSCF. Figure 4.14
clearly shows the more accurate models obtained by the logarithmic Advanced ML
technique under the assumption of white noise on the outputs (i.e. logarithmic
estimator with σHoi (ωk ) = |Hoi (ωk )|).
4.9 Conclusions
In this chapter the extensions of the LSCF and ML algorithms were developed
to take into account the initial and final conditions of each block used by the H1
estimator. Special attention was paid to the implementation to results in a fast and
memory efficient algorithm by using the structure in the matrix equations. The
relationship between the implementation of FRF and IO data driven algorithm was
discussed. Finally, a combined deterministic-stochastic algorithm was proposed
and applied on simulations in the framework of in flight test on aircrafts. Other
examples illustrated the use of the extended implementation of the FRF driven
algorithms.
104 Chapter 4. Identification of common-denominator models
4.10 Appendix
Assume that the discrete-time signals y(r) and f (r) are exactly known in the time
interval [0, (N − 1)Ts ] and are unknown outside the interval with N the number of
samples and ∆t the sample period. The discrete input and output signals satisfy
the following difference equation:
n
X n
X
an−m y((r − n)∆t) = bn−m f ((r − m)∆t) (4.59)
m=0 m=0
N −1
1 X
Y (zk ) = DF T (y(r)) = √ y(r∆t)zk−r (4.60)
N r=0
N −1
1 X
DF T (x(r − 1)) = √ y((r − 1)∆t)z −r
N r=0
N −1
z −1 X
= √ y(r∆t)z −r + y(−∆t) − z N y(N ∆t)
N r=0
y(−∆t) − y(N ∆t)
= z −1 Y (z) + √ (4.61)
N
n
1 X
DF T (y(r − n)) = z −n Y (z) + √ (y(−r∆t) − y((N − 1 + r)∆t))z r−n (4.62)
N r=1
Taken the DFT of both sides of equation 4.59 and using the expression of Equation
4.62 results in
with
n
X
A(z) = am z −m (4.64)
m=0
Xn
B(z) = bm z −m (4.65)
m=0
n−1
XX m
T (z) = bm (f (−r∆t) − f ((N − 1 + r)∆t))z r−m (4.66)
m=1 r=1
n−1
XX m
− am (y(−r∆t) − y((N − 1 + r)∆t))z r−m (4.67)
m=1 r=1
n−1
X
= tm z −m (4.68)
m=0
The extra polynomial T (z) takes into account the influence of the final and initial
conditions. A more profound study can by found in [99].
106 Chapter 4. Identification of common-denominator models
Chapter 5
107
108 Chapter 5. Identification of right and left matrix fraction polynomial models
5.1 Introduction
The interest in RMFD models for modal analysis can be explained by the
growing use of multiple input test setups for higher damped structures like fully
trimmed cars and airplanes. In fact, the well-known Least Squares Complex Ex-
ponential (LSCE) modal parameter estimator can be seen as RMFD of which
only the denominator coefficients are estimated in the time-domain. The devel-
opment of frequency domain algorithms in the Acoustics and Vibration Research
Group of the Vrije Universiteit Brussel for modal analysis started with apply-
ing common-denominator models for MIMO test setups. It was found that the
identified common-denominator models closely fitted the measured frequency re-
sponse functions (FRFs). Nevertheless, when converting this model to a modal
model by reducing the residues to a rank-one matrix using the singular value
decomposition (SVD), the quality of the fit decreased [116]. This drawback of
common-denominator based models tends to be more important for highly damped
structures, since in this case the natural frequencies are less explicit. Another the-
oretically associated drawback of a common-denominator model is that closely
spaced poles will erroneously show up as a single pole. However, in practice this
problem is solved by increasing the model order, which is typically done in EMA.
It can be stated that the common-denominator model does not forces rank one
residue matrices on the data and uses this extra freedom, which is not present in
the modal model, to obtain a close fit of the data. These two reasons motivate a
poly-reference version of the LSCF method using a so-called RMFD. This poly-
reference LSCF was initially presented in [48] and applied on industrial examples
in [86], while [101] discusses a RMFD LS estimator in the continuous-time. In this
approach both the system poles and the participation factors are available from
the denominator coefficients. The main advantage of the poly-reference RMFD
implementations is that an SVD decomposing the residues in mode shapes and
5.2. Frequency-domain identification of RMFD models 109
Finally the LMFD, which forces also rank one residues on the measurements,
is presented as an IO data and FRF driven estimator. In [63] a profound study is
given for the LMFD model identification in the Laplace domain (i.e. a continuous-
time model in the frequency domain) starting from FRF data.
The relationship between output o and all inputs can be modelled in the frequency
domain by means of a right matrix fraction description given by
with Bo (z) ∈ C 1×Ni the numerator row-vector polynomial and A(z) ∈ C Ni ×Ni
the denominator matrix polynomial defined by
n n
Boj zkj Aj zkj
X X
Bo (zk ) = A(zk ) = (5.2)
j=0 j=0
The matrix coefficients Aj and Boj are the parameters to be estimated. These
polynomial coefficients are grouped together in one matrix
β1
.. Bo0 A0
θ= . βo = ... αo = ... (5.3)
βNo
Bon An
α
with Wo−1 (ωk ) an additional scalar frequency-weighting for each output. The
weighted linear LS problem is found by minimizing the cost function given by
No X
N
lFp−LSCF
X
tr EoH (ωk )Eo (ωk )
RF (θ) = (5.5)
o=1 k=1
No X
X N
= Eo (ωk )EoH (ωk ) (5.6)
o=1 k=1
βo = −Ro−1 So α (5.14)
and backwards substituting the α coefficients in Eq. 5.14 results in the β coeffi-
cients. By the use of a singular value decomposition the reduced normal equations
can be solved in TLS sense resulting in a mixed LS-TLS solution for the original
(full) normal equations Eq. 5.13. The submatrices Ro , So and To have a predefined
block Toeplitz structure given by
N
X
[Ro ]rs = |Wk (ωk )|−2 zks−r (5.17)
k=1
N
X
[So ]rj = |Wk (ωk )|−2 Ho (ωk )zks−r (5.18)
k=1
N
X
[To ]ij = |Wk (ωk )|−2 HoH (ωk )Ho (ωk )zks−r (5.19)
k=1
By taking the uncertainty information on the estimated FRF data into account
during the parametric identification, more accurate parameter estimates can be
112 Chapter 5. Identification of right and left matrix fraction polynomial models
In this paragraph, it is shown that the use of the ML weighting requires the
vec operator to transform the matrix of unknown parameters θ into a vector of
unknown parameters. This will increase the memory requirements and calculation
time, since the dimensions of the submatrices of the normal equations increase by
a factor Ni . This type of implementation is defined as a vec implementation using
a matrix weighting.
Under the assumption that the noise on the FRFs is complex normally dis-
tributed and the noise on the FRFs corresponding to different outputs is uncorre-
lated, the (negative) log-likelihood function is given by [21], [37]
No
N X
lFp−ml
X
RF (θ) = Eo∗ (ωk )Co−1 (ωk )EoT (ωk ) (5.20)
k=1 o=1
with ∆Ho (ωk ) ∈ C 1×Ni the noise on the FRFs corresponding to output o. This
assumption of uncorrelated noise over the different outputs is exactly true in the
case of scanning vibrometer measurements or a roving hammer test. In the case
this noise assumption is violated, the ML estimator is still consistent and only some
efficiency is lost. These covariance matrices for each output o can be obtained
from the non-parametric FRF identification. Similar as for the ML common-
denominator, the weighted equation error is non-linear in the system parameters
and a Gauss-Newton optimization algorithm is implemented to minimize the cost
function. The entries of the polynomial coefficients are grouped in a vector given
by
β1
vec(Bo0 ) vec(A0 )
..
.. ..
θ= . βo = α = (5.23)
. .
βNo
vec(Bon ) vec(An )
α
with
h i
n
−Co (ω1 )−1/2 [(Bo (z1 )A(z1 )−1 ) ⊗ A(z1 )−T ] INi2 z1 . . . INi2 z1 INi2
..
Υo = −
. h i
n
−Co (ωN )−1/2
[(Bo (zN )A(zN ) ) ⊗ A(zN )−T ] INi2 zN . . . INi2 zN INi2
−1
2 2
∈ C N Ni ×(n+1)Ni (5.25)
−Co (ω1 )−1/2 A(z1 )−1 INi z1n
... INi z1 INi
Γo =
..
.
n
−Co (ωN )−1/2 A(zN )−1 INi zN ... INi zN INi
N Ni ×(n+1)Ni
∈ C (5.26)
2
Ni (n+1)×Ni (n+1) Ni (n+1)×Ni (n+1)Ni
with Ro = ΓH o Γo ∈ C , So = ΓHo Υo ∈ C , To =
2 2
H Ni (n+1)×Ni (n+1) H Ni (n+1)×1 H
Υo Υo ∈ C , Vo = Γo Eo ∈ C and Wo = Υo Eo ∈
2
C Ni (n+1)×1 and Eo defined by
vec(Eo (ω1 ))
Eo =
..
(5.28)
.
vec(Eo (ωN ))
the estimated parameters can be estimated from the Cramr-Rao lower bound (the
ML implementation reaches this lower bound). In the appendix 5.7 it shown how
the uncertainty levels on the estimated poles can be obtained from the covariance
matrix on the estimated denominator coefficients.
N Ni2 ×(n+1)Ni2
∈ C (5.34)
−Co,log (ω1 )−1/2 1./ A−T (z1 )B T (z1 ) . ∗ A(z1 )−1 ZN (1)
i
.
Γo =
.
.
−Co,log (ωN )−1/2 1./ A−T (z1 )B T (z1 ) . ∗ A(zN )−1 ZN (N )
i
N Ni ×(n+1)Ni
∈ C (5.35)
The estimator proposed in this section uses a scalar weighting resulting in a matrix
implementation i.e. the parameter θ are grouped in a matrix with Ni columns
5.2. Frequency-domain identification of RMFD models 115
similar as for the poly-reference LSCF and the use of the vec operator is avoided.
This results in a Ni3 faster implementation, but also in a small loss of efficiency
compared to the scalar implementation of the poly-reference ML with a matrix
weighting. Consider the same cost function and equation error as for the poly-
reference LSCF
No X
N
lFp−f ML
X
tr Eo (ωk )H Eo (ωk )
RF (θ) =
o=1 k=1
(5.36)
and
where the subindex e indicates the polynomials with exact coefficients. Taking
into account that Ho (ωk ) = Hoe (ωk ) + ∆Ho (ωk ) and Hoe (ωk ) = Boe (zk )A−1
e (zk )
results in
No
N X H
X tr Eoe (ωk )Eoe (ωk )
E (lp−f M L (θe )) = (5.40)
o=1
Wo (ωk )2
k=1
= 1 (5.43)
and Eoe the equation error in the exact parameters (which is zero) and thus the
cost function is minimal in the exact parameters θe . The Jacobian matrix has the
same structure as for the poly-reference ML that uses a matrix weighting shown by
Eq.5.24. Since the weighting Wo (ωk ) is independent of the numerator polynomials
116 Chapter 5. Identification of right and left matrix fraction polynomial models
with Aj (zk ) and Boj (zk ) the polynomials estimated in iteration j. Each iteration
j can be considered as a weighted poly-reference LSCF with a frequency weight-
ing |Wo (ωk )|2 = tr AH
(z
j−1 k )C o (ω k )A (z
j−1 k ) . In [94], it is shown that the IQML
improves the estimates of the least squares estimator, since its cost function con-
verges to the fast ML cost function. Nevertheless, it can not be shown that the
estimator is consistent and efficient.
arg minJ,θ ˆ −1 2 ˆ H
ˆ k(J − J)SJ kF subject to Jθ = 0 and θ θ = I (5.53)
The weighting matrix SJ is the square root of the covariance matrix of J i.e.
CJ = SJH SJ = E ∆J H ∆J , with ∆J = J − Jˆ the noise contribution on the
Jacobian matrix caused by uncertainty on the FRF data. The GTLS solution θ of
this estimation problem is a consistent estimate. By elimination of Jˆ it is proven
that Eq. 5.53 is equivalent to minimizing the following cost function [95]
θH J H Jθ
lFp−W
RF
GT LS
= subject to θH θ = I (5.54)
θH SJH SJ θ
of which the solution in practice is found by the Generalized Singular Value De-
composition (GSVD) of the matrices J and SJ . Since J H J is equivalent to the
normal matrix given by Eq. 5.13 and CJ = SJH SJ , the minimization problem can
be rewritten as
θH Qθ
lFp−W
RF
GT LS
= subject to θH θ = I (5.55)
θH CJ θ
QV = λCJ V (5.56)
118 Chapter 5. Identification of right and left matrix fraction polynomial models
CR1 · · · 0 CS1
. .. ..
. .
CJ = E ∆J H ∆J = . . (5.57)
0 CRNo CSNo
No
CSH1 · · · CSHNo
P
o=1 CTo
under the assumption that the noise over the different output on the FRFs is un-
correlated. This is exactly true for both Hammer and Scanning Laser Vibrometer
measurements since for those test setups the different outputs are measured in-
dependent in time. For simultaneous accelerometer measurements in the case of
a shaker setup, important correlations can be introduced over the different out-
puts [129], [126]. However, in order to obtain a fast and practically applicable
implementation these correlations must be neglected. Nevertheless, the results
obtained by neglecting the correlations over the outputs, are in general more accu-
rate then the classic poly-reference LS solution discussed in paragraph 5.2.1. Since
the submatrices Γo of the Jacobian matrix defined by Eq. 5.9 are free from noise
(under the assumption that the frequency
weighting Wo (ωk ) is exact) the matrices
CRo = E ΓH H
o Γ o and CS o
= E Γ o Υo equal the zero matrix. The subentries of
the matrices To with i = [(r − 1)Ni + 1 : rNi ], j = [(s − 1)Ni + 1 : sNi ] for both
r, s = 1, 2, . . . N are given by
N
X
[To ]ij = ΥH |Wk (ωk )|−2 HoH (ωk )Ho (ωk )zks−r
o Υo ij = (5.58)
k=1
N
X
∆ΥH |Wk (ωk )|−2 Co (ωk )zks−r
[CTo ]ij = o ∆Υo rs = (5.59)
k=1
with Co (ωk ) = E ∆HoH (ωk )∆Ho (ωk ) the covariance matrix of the FRFs corre-
sponding to output o. It can visually be checked that the matrices CTo have a
block Toeplitz structure similar as the submatrices of the normal matrix Q. The
generalized total least squares problem is then explicitly written as
R1 ··· 0 S1
β1 0 ··· 0 0 β1
. .. . . . .. . .
. . . . . .
.
. .
. = λ .
. .
.
(5.60)
0 RNo SNo βNo
0 0 0
βNo
P No
S1H ··· H
SNo o=1 To α 0 ··· 0 CT α
5.2. Frequency-domain identification of RMFD models 119
PNo
with CT = o=1 CTo . As for the LS implementation, the normal equations
can be reformulated to a reduced set of equations by elimination the numerator
coefficients given by
βo = −Ro−1 So α (5.61)
from which the left-hand side is identical to the LS formulation. Once the coeffi-
cients α are estimated, backwards substitution yields the coefficients βo . The use
of this predefined block structure of the normal equations results in an No2 faster
implementation than solving Eq. 5.60 directly. A stabilization chart can be con-
structed by solving the GEVD for an increasing model order of the denominator
polynomial by increasing the size of M and CT with Ni (n + 1) column and rows
PNo
To − SoH Ro−1 So ).
(M = o=1
The RMFD model is less suitable for IO data driven identification since the model
can only be linearized for Ni = 1 (which is of course identical to the common-
denominator model)
For the multiple input case, only non-linear equation errors in the parameters
can be formulated, resulting in an optimization problem. This complicates the
identification of starting values to start the Gauss-Newton algorithm. Therefore
no further attention is paid to IO data driven RMFD model identification in this
thesis.
The poles and participation factors are found from the denominator polynomial
coefficients Aj . The companion matrix Ac , build from the coefficients, is given by
120 Chapter 5. Identification of right and left matrix fraction polynomial models
[52]
′ ′ ′
An−1 ... A1 A0
I 0 0 0
Ac = (5.64)
.. .. .. ..
. . . .
0 ... I 0
′
with Aj = −A−1 n Aj . By solving an eigenvalue decomposition of the companion
matrix the poles and participation factors of the system are determined
(Ac − λr I) Vr = 0 (5.65)
where the lower and upper residues respectively model the influence from the lower
and higher band poles. Under the assumption that the estimates of the poles λr
and the modal participation vectors Lr are consistent estimates, the linear least-
squares estimation of the mode shapes is also consistent. An additional weighting
in the LSFD improves the efficiency and results in an ML estimation of the mode
shapes. Therefore in combination with the inconsistent poly-reference LSCF, from
which the poles and participation vectors are extracted, the LSFD is advised to
estimate the mode shapes in a second step, instead of extracting the mode shapes
directly from the poly-reference LSCF numerator coefficients estimates Boi . In the
different examples of paragraph 5.5 is shown that the LSFD improves the quality
of the modal model in combination with first poly-reference LSCF step.
5.3. Left Matrix Fraction Description 121
The left matrix fraction description considers all response measurements simul-
taneously. Based on the primary data, a distinction can be made between the
identification of na LMFD starting from IO data or FRF data. In [63], [68] the
Frequency domain Direct Parameter Identification (FDPI) algorithm is presented,
which uses a Laplace domain model to estimate only a denominator polynomial
of order 2 and no numerator polynomial. In the next paragraph, a discrete-time
LMFD model identification algorithm is discussed identifying both a denominator
and numerator in the frequency-domain.
The matrix coefficients Aj and Boj are the parameters to be estimated. An addi-
tional matrix polynomial T (zk ) can be taken into account to model the transient
phenomena
Since the transient contribution can be modelled by an extra input signal FNi +1 (zk ) =
1, the following expressions do not consider the polynomial T (zk ) explicitly. A
linear-in-the-parameters equation error is found by multiplying of Eq.5.68 by A(zk )
resulting in
can be formulated as
Jθ = 0 (5.73)
β
[Γ Υ] = 0 (5.74)
α
with
z1n ⊗ F T (ω1 )
− 1 z1 ...
− 1 z2 . . . z2n ⊗ F T (ω2 )
Γ = ∈ C N ×Ni (n+1) (5.75)
..
.
n T
− 1 zN . . . zN ⊗ F (ωN )
1 z1 . . . z1n ⊗ Y T (ω1 )
1 z2 . . . z2n ⊗ Y T (ω2 )
∈ C N ×No (n+1)
Υ =
.. (5.76)
.
n
⊗ Y T (ωN )
1 zN . . . zN
In the theoretical case that the number of outputs equals the number of modes,
the Newton equation of motion is given by
with A(zk ) a second order matrix polynomial modelling the dynamic stiffness and
T (zk ) the vector modelling the transients. This can be fitted in LMFD framework
by choosing a parameter constraint B(zk ) = I. The LMFD LS estimator is consis-
tent with constraint An = I under the restriction that E(ωk ) is circular complex
normal distributed Gaussian noise over the different spectral lines.
output-only case, the LMFD estimator can still be used as a data-driven stochastic
estimator. Unfortunately, the use of the LMFD estimator as a stochastic and
combined deterministic-stochastic estimator is only applicable if the number of
modes in the frequency band of interest is smaller than the number of outputs.
with H(ω) ∈ C No ×Ni the FRF matrix. The the equation error is given by
z1n
− 1 z1 ... ⊗ IN i
− 1 z2 . . . z2n ⊗ IN i
Γ = ∈ C Ni N ×Ni (n+1) (5.81)
..
.
n
− 1 zN . . . zN ⊗ IN i
1 z1 . . . z1n ⊗ H T (ω1 )
1 z2 . . . z2n ⊗ H T (ω2 )
∈ C Ni N ×No (n+1)
Υ =
.. (5.82)
.
n T
1 zN . . . zN ⊗ H (ωN )
The poles and mode shapes are determined from the eigenvalues and eigenvectors
of the companion matrix defined by Eq. 5.64 in an analogous way as the determi-
nation of the poles and participation factors in case of a RMFD. The participation
factors can then be estimated in a second step LS based on Eq. 5.67.
124 Chapter 5. Identification of right and left matrix fraction polynomial models
Yr = T Yd (5.83)
Hr = T Yd (5.84)
from which can be seen that the transformation matrix T reduces the number of
outputs from No in the physical space to Nr in the condensed data space. After the
identification process the mode shapes must be re-transformed from the condensed
data space to the physical data space. Since the number of poles estimated by
the LMFD LS estimator equals nNr , a stabilization chart can be constructed by
solving the equations for an increasing Nr . Finally, it should be noticed that in
analogy to the poly-reference MLE, fast-MLE, GTLS, IQML, fast-BTLS different
estimators with similar schemes can be developed to identify LMFD models.
5.4 Output-Only
The following examples will make a comparison of the quality of the estimated
polynomials and modal models estimate by the common-denominator and poly-
reference versions of both the LSCF and ML estimator.
5.5. Illustrating examples 125
5.5.1 Body-in-white
An impact hammer test was performed on a body-in-white (see figure 5.1) with
3 fixed reference accelerometers, while the structure was excited in 28 locations
in 3 directions. By applying the reciprocity property the total measured FRF
matrix consists of 84 outputs and 3 inputs. A frequency band from 35Hz to
62Hz is processed by the LSCF, ML, p-LSCF and p-ML estimators. For the
common-denominator models the poles are estimated in the first step. Next, after
selection of the physical poles, the residues are estimated by the LSFD algorithm
and decomposed in participation vectors and mode shapes. The poly-reference
models estimate both the poles and participation factors in the first step and only
the mode shapes are estimated in the second step by the LSFD estimator. Table
5.1 compares the mean error and correlation between the measurements and the
estimated polynomial models and between the measurements and the estimated
modal models. The correlation C and error E are defined by
P 2
No XNi N ∗
k=1 Ĥoi (ωk , θ)Hoi (ωk )
1 X
C = P (5.85)
No Ni o=1 i=1 P N ∗ N ∗
k=1 Ĥoi (ωk , θ)Ĥoi (ωk , θ) k=1 Hoi (ωk )Hoi (ωk )
P N 2
No XNi
k=1 Ĥoi (ωk , θ) − Hoi (ωk )
1 X
E = 2
(5.86)
No Ni o=1 i=1 PN
k=1 Ĥoi (ωk , θ)
with Ĥoi (ωk , θ) the synthesized FRF and Hoi (ω − k) the measured FRF. Since
no noise information was available the logarithmic version of the ML estimators
was used under the assumption of relative noise. It is clear that on the level
of the polynomial model the common-denominator ML estimator results in the
best quality. This can be explained by its inherent optimization and the extra
freedom in the common-denominator model compared to a RMFD with regard to
126 Chapter 5. Identification of right and left matrix fraction polynomial models
Table 5.1: Comparison of model quality obtained by different algorithms for measure-
ments on a body-in-white
the polynomial model to the modal model results in a larger loss of accuracy for
the common-denominator ML estimator than for the poly-reference ML estimator.
More surprising is the small gain in accuracy for the LS estimators by switching
to the modal model. This can be explained, since the first LSCF step results
in biased and inconsistent estimates of the polynomials. These error are slightly
compensated in the second LSFD step, which is consistent under the assumption
that the already known modal parameters (obtained from in the first step) are
consistent. Table 5.2 shows the natural frequencies and damping ratios of the
modes present in the frequency band.
Using a MIMO test a fully trimmed Porsche was excited in 4 different locations by
shakers (more details about the test setup can be found in [119]). The accelerations
were measured in 154 locations distributed over the car. Since no covariances were
available in the data set, the logarithmic implementation of both the p-ML and the
ML was used under the assumption of relative noise. A model order corresponding
to 24 modes was used to identify the modal parameters in the frequency band from
3Hz to 30Hz. Table 5.3 compares the mean errors and the mean correlation for
the different algorithms with the correlation between the synthesized FRF and the
measured FRF.
It is clear that the p-ML gives the best results in terms of the fit of the modal model,
which can also be concluded from figure 5.2. For the least squares algorithms the
p-LSCF method outperforms the LSCF algorithm. It should be mentioned that
both the p-ML and the ML algorithm resulted in perfect fits of their polyno-
mial models. Nevertheless, the common-denominator based algorithms LSCF and
5.5. Illustrating examples 127
Table 5.3: Comparison of model quality obtained by different algorithms for measure-
ments on a fully trimmed car
To show the applicability of the presented LSCF, p-LSCF, ML and p-ML for
OMA the operational measurements (only ambient excitation) on the Villa Paso
bridge are processed and the parametric results are compared. Starting from
the ’positive’ power spectra for two reference sensors the polynomial models are
estimated and converted to the modal model. Since operational spectral density
5.6. Conclusions 129
functions are typically characterized by high noise levels both the LSCF and p-
LSCF estimators result in larger errors, that are partially reduced in the second
step by the LSFD estimator. Nevertheless in this case it is recommended to use
the ML estimator under the assumption of relative noise. Figure 5.4 illustrates
some of the synthesized ’positive’ power spectra for the different algorithms.
Table 5.5: Comparison of model quality obtained by different algorithms for measure-
ments on the Villa Paso bridge
5.6 Conclusions
In this chapter the use of a right and left matrix fraction description is proposed
for modal parameter estimation. The implementation of the poly-reference LSCF
is based on similar approaches as the common-denominator LSCF to speed up
the algorithm and to reduce the memory requirements. Next, a consistent poly-
reference maximum likelihood estimator is proposed to handle noisy data. To
speed up this algorithm, a fast version is presented based on a scalar frequency
weighting. By several experiments, it was illustrated that modal models extracted
from the poly-reference implementations outperform the ones extracted from the
common-denominator model for MIMO measurements on highly damped struc-
tures. Furthermore, the proposed poly-reference algorithms can also be applied to
process power spectra in case of an OMA test. Finally, it was noticed the (poly-
reference) LSCF estimates can be improved by estimating in a second step the
mode shapes (and participation vectors) by the LSFD.
with J the Jacobian matrix given by Eq. 5.24. Elimination of the covariance
matrix on the coefficients α of the denominator only results in
No No
!−1
X X
cov(α) = To − So Ro−1 SoH (5.88)
o=1 o=1
(in fact this is the covariance matrix on the vec operators of the denominator
coefficients) To calculate the uncertainties on the poles from the uncertainties on
the estimated matrix coefficients, consider the EVD of the companion matrix
n−1 n−1
λr V[:,r] λr V[:,r]
.. ..
Ac . − λr . =0 (5.89)
λr V[:,r]
λr V[:,r]
V[:,r] V[:,r]
and
′ ′ ′
An−1 ... A1 A0
I 0 0 0
Ac = (5.90)
.. .. .. ..
. . . .
0 ... I 0
′
with Aj = −A−1
n Aj . It holds for the denominator polynomial A(zk ) that
λjr
∆λr = − V H A′−1 (λr )∆Aj Vr (5.96)
VrH Vr r
λjr
VrT ⊗ VrH A′−1 (λr ) vec (∆Aj )
= − H
(5.97)
Vr Vr
The sensitivity between the perturbation ∆Aj and its influence on the pole λr is
λj
then defined by Srj = − V HrVr VrT ⊗ VrH A′−1 (λr ) . The variances or uncertainties
r
on the poles λr are then given by
H
Sr0
2
..
σλr = Sr0 . . . Srn cov(α) . (5.98)
H
Srn
p−MLFD MLFD
−50 −50
−60
Ampl (dB)
Ampl (dB)
−60
−70
−80 −70
−90
−80
−100
10 20 30 10 20 30
Freq (Hz) Freq (Hz)
p−LSCF LSCF
−50 −50
−60
Ampl (dB)
Ampl (dB)
−60
−70
−70
−80
−80
−90
10 20 30 10 20 30
Freq (Hz) Freq (Hz)
p−MLFD MLFD
−50 −50
−60 −60
Ampl (dB)
Ampl (dB)
−70 −70
−80 −80
−90 −90
−100 −100
0 10 20 30 0 10 20 30
Freq (Hz) Freq (Hz)
p−LSCF LSCF
−50 −50
−60 −60
Ampl (dB)
Ampl (dB)
−70 −70
−80 −80
−90 −90
−100 −100
0 10 20 30 0 10 20 30
Freq (Hz) Freq (Hz)
Figure 5.2: Fully trimmed car. Comparison between the modal model obtained by the
p-MLFD, MLFD, p-LSCF and LSCF algorithm for two FRFs. (cross: measurement, full
line: estimated modal model)
5.7. Appendix: Confidence intervals on the estimated poles from the p-ML estimator133
−50 −50
AMPL. (dB)
AMPL. (dB)
−60 −60
−70 −70
−80 −80
5 10 15 20 25 5 10 15 20 25
FREQ. (Hz) FREQ. (Hz)
p−ML polynomial model p−ML modal model
−50 −50
AMPL. (dB)
AMPL. (dB)
−60 −60
−70 −70
−80 −80
5 10 15 20 25 5 10 15 20 25
FREQ. (Hz) FREQ. (Hz)
−60 −60
AMPL. (dB)
AMPL. (dB)
−70 −70
−80 −80
−90 −90
5 10 15 20 25 5 10 15 20 25
FREQ. (Hz) FREQ. (Hz)
p−ML polynomial model p−ML modal model
−50 −50
−60 −60
AMPL. (dB)
AMPL. (dB)
−70 −70
−80 −80
−90 −90
5 10 15 20 25 5 10 15 20 25
FREQ. (Hz) FREQ. (Hz)
Figure 5.3: Fully trimmed car. Comparison between the polynomial and modal models
obtained by the p-ML and p-LSCF algorithm for two FRFs. (cross: measurement, full
line: estimated modal model)
134 Chapter 5. Identification of right and left matrix fraction polynomial models
p−MLFD MLFD
80 80
Ampl. (dB)
Ampl. (dB)
60
60
40
40
20
20
4 6 8 10 2 4 6 8 10
Freq. (Hz) Freq. (Hz)
p−LSCF LSCF
80 80
70
Ampl. (dB)
Ampl. (dB)
60 60
50
40 40
30
20
4 6 8 10 2 4 6 8
Freq. (Hz) Freq. (Hz)
p−MLFD MLFD
85 85
80 80
Ampl. (dB)
Ampl. (dB)
75 75
70 70
65 65
60 60
4 6 8 2 4 6 8
Freq. (Hz) Freq. (Hz)
p−LSCF LSCF
85
80 80
Ampl. (dB)
Ampl. (dB)
75
70
70
65
60
60
2 4 6 8 10 2 4 6 8 10
Freq. (Hz) Freq. (Hz)
Figure 5.4: Villa Paso bridge. Comparison between the modal model obtained by the
p-MLFD, MLFD, p-LSCF and LSCF algorithm for two FRFs. (cross: measurement, full
line: estimated modal model)
Chapter 6
Deterministic
Frequency-domain Subspace
Identification
135
136 Chapter 6. Deterministic Frequency-domain Subspace Identification
6.1 Introduction
The previous two chapters were devoted to system identification for modal ana-
lysis by means of polynomial models. These identification techniques such as e.g.
LSCF, ML, BTLS, ... can all be related to an equation error, which is minimized
by a quadratic cost function and are well documented in [67], [98]. In parallel with
frequency-domain cost function related techniques, a class of so-called realization
algorithms was developed to estimate state-space models by factorization of the
Hankel matrix build from the impulse response function [53], [144]. The ERA
estimator is an implementation of a realization algorithm applied for mechanical
engineering [55], [56]. Meanwhile, identification techniques were developed in a
stochastic realization framework for the identification from output-only measure-
ments [3]. These realization algorithms formed the basis for the development of
subspace algorithms, which estimate the system matrices by means of projections
of so-called input and output block Hankel matrices. Starting from the pure de-
terministic framework [79], stochastic subspace algorithms were developed [120]
and combined in so-called combined deterministic-stochastic framework [122]. An
overview of subspace algorithms can be found in [6], [71] , while [123] gives a pro-
found discussion in a general framework. Subspace algorithms have been applied
successfully for several applications like control engineering, electrical engineering,
financial engineering and mechanical engineering [28]. In the domain of modal
analysis, subspace identification has mainly been applied to modal parameter es-
timation from output-only data [5], [78], [87], [85], [49]. Compared to the tra-
ditional identification algorithms, subspace algorithms are non-iterative and as a
result they always yield a solution without the risk of convergence problems. Un-
til recently, research efforts were and still are focussed on time-domain subspace
methods, while linear systems often are characterized in the frequency domain.
Therefore, it is quite natural to consider subspace identification algorithms in the
frequency domain to identify models directly from input/output spectra, FRFs or
(’positive’) power spectra. In [66] a basic projection frequency-domain subspace
algorithm is proposed in a deterministic framework, which does not require uni-
form frequency grid and which allows to use a frequency weighting. By considering
the covariances on the primary data as a frequency weighting, it is shown that this
projection algorithm is strongly consistent [77], [92]. Another approach consists
in transforming the frequency-domain data to the time-domain by the IDFT and
combining this with classical time-domain techniques [77]. However, in this case
the estimated modal parameters must then be corrected to compensate for time-
domain aliasing. This can be done by a correction of the participation factors
on the level of the modal factors as shown in paragraph 3.9 or by a correction
of the system matrices [77]. In this chapter, the basic principles for frequency-
domain subspace identification are briefly presented. Next, an extended model
is proposed in order to take into account the initial and final conditions. Based
on this extended formulation, a mixed non-parametric/parametric FRF estimator
is proposed for validation purposes [18]. In addition, the FRF driven frequency-
6.2. Basic Frequency-Domain Projection Algorithm 137
domain subspace algorithm is extended to consider the initial and final conditions
for each block used by the H1 FRF estimator [17].
Y = Or X + ΓU (6.9)
138 Chapter 6. Deterministic Frequency-domain Subspace Identification
with X = [X(1) . . . X(N )] (with N the number of frequency lines in the chosen
frequency band), Y a complex No r × N block Vandermonde matrix and U a
complex Ni r × N block Vandermonde matrix each defined by
Y (ω1 ) Y (ω2 ) ... Y (ωN )
z1 Y (ω1 ) z2 Y (ω2 ) . . . zN Y (ωN )
Y= .. .. .. ..
. . . .
z1r−1 Y (ω1 ) z2r−1 Y (ω2 ) . . . r−1
zN Y (ωN )
U (ω1 ) U (ω2 ) ... U (ωN )
z1 U (ω1 ) z2 U (ω2 ) ... zN U (ωN )
U= .. .. .. ..
. . . .
z1r−1 U (ω1 ) z2r−1 U (ω2 ) ... r−1
zN U (ωN )
In the case that one is interested in estimating real system matrices, Eq. 6.9 is
converted in a set of real equations
where ()re represents a matrix with the real and imaginary parts aside each other,
for example
Eq. 6.10 with r larger than the model order divided by the number of outputs
n/No (to have an extended observability matrix Or which is of rank n to extract
the system matrices A and C) is the basic equation in frequency-domain sub-
space identification and illustrates that Y lies in the subspace spanned by the row
spaces defined by X and U. Frequency-domain subspace identification algorithms
basically consists of a four step procedure. By projection of Y in a space U⊥
orthogonal to the row space U the inputs are eliminated and the extended observ-
ability matrix can be determined. In the first step, a practical implementation of
the orthogonal projection is obtained by the use of the QR-factorization [142]
re T T
U R11 0 Q1
= (6.12)
Yre T
R12 T
R22 QT2
T
R22 = U ΣV T (6.13)
The second step uses a SVD to estimate the extended observability matrix Ôr .
For a model order n, Ôr is given by
In the fourth and last step, given the estimates  and Ĉ, the system matrixes B
and D are estimated in a least squares sense from
Y (ωk ) = Ĉ(Izk − Â)−1 BU (ωk ) + D (6.17)
This basic projection algorithm can also be used to estimate a continuous-time
state-space model by replacing the basic function zk by ωk [121]. For reasons of
numerical conditioning the use of orthogonal polynomials in ωk are recommended
[121].
Instead of starting from IO data the subspace algorithm can also start with FRF
or (’positive’) power spectra data. Depending on the form of the primary data,
the residues are forced to be of rank one or not. By replacing Y (ωk ) by H(ωk ) ∈
C No ×Ni and U (ωk ) by INi a rank one residue model is estimated. In the case
that the FRFs are stacked under each other, no rank one model is forced on the
measurements and thus the mode shapes and participation vectors are obtained
from an SVD decomposition reducing to rank to 1. The stacked FRFs are given
No Ni ×1
T
by H st (ωk ) = [H1T (ωk ) . . . HN T
i (ωk )] ∈ C with Hi (ωk ) ∈ C No ×1 the FRFs
corresponding to input i. Similar the power spectra, corresponding to the different
reference sensors, can be stacked in one column. This approach can be useful for
data from low damped structures, with small data inconsistencies in the different
patches e.g. caused by temperature effects, mass-loading effect.
In order to make the basic projection algorithm, presented in the previous para-
graph, consistent an additional frequency weighting should be taken into account
by considering the covariances on the primary data [77]. In [92] the asymptotic
properties of this weighted projection algorithms are studied when the true noise
covariance matrix is replaced by the sample noise covariance matrix obtained from
a small number of repeated experiments. The theory was developed assuming that
the input is exactly known and noise is only present on the outputs. The state-
space model with output noise is given by
zk X(k) = AX(k) + BU (ωk ) (6.18)
Y (ωk ) = CX(k) + DU (ωk ) + N (k) (6.19)
140 Chapter 6. Deterministic Frequency-domain Subspace Identification
with
INo INo ... INo
z1 INo z2 INo ... zN INo
Z= .. .. .. .. (6.22)
. . . .
z1r−1 INo z2r−1 INo ... r−1
zN INo
Given the QR decomposition by Eq. 6.12 a consistent estimate for the extended
observability matrix is given by [77]
C−1/2 R22
T
= U ΣV T (6.23)
1/2
Ôr = C U[:,1:n] (6.24)
The system matrices A, B, C and D are than determined in a similar way as for
the unweighted projection algorithm. However, in practice the use of deterministic
subspace algorithms with a frequency weighting to guarantee consistency is not
straight forward for a MIMO test setup with e.g. No = 100 and thus for each spec-
tral line k a No × No covariance matrix. This would result in both large memory
requirements and calculation times. Furthermore, to obtain a invertible covariance
matrix at least No averages should be taken into account in the estimation of the
covariance matrix, which requires a long measurement time. Notice also the dif-
ference with maximum likelihood identification where the frequency weighting has
no influence on the consistency property and only improves the efficiency. Hence,
for the ML implementations the correlations between different outputs could be
neglected to result in a fast implementation with only a small loss in efficiency.
Frequency-domain system identification generally assumes that the input and out-
put signals are periodic or time limited within the observation window. This is
necessary to guarantee leakage-free spectra calculated through the discrete Fourier
transform. In the case that an arbitrary input signal is used, often a window such
as a Hanning window, is used to reduce errors introduced by leakage phenomena.
In [99] and chapter 4 it is shown for single input/single output (SISO) systems
that for a rational fraction polynomial model formulation the ’transient’ polyno-
mial takes into account the initial and final conditions of the experiments and
eliminates the bias error due to leakage. This key idea can be generalized for
MIMO state space models [75], [76]. It will be shown that the frequency-domain
state-space identification methods e.g. frequency domain subspace identification
can easily be extended to take into account the initial and final conditions of the
states.
Consider a linear time-invariant state space model. Assume that the input and
output samples are exactly known at discrete time instants tn = n∆t (sampling
period ∆t) inside the time interval [0, (N − 1)∆t] and unknown outside this time
interval. The discrete input un ∈ R Ni ×1 and output yn ∈ RNo ×1 samples satisfy
following difference equation:
with xn ∈ Rn×1 (n the number of states (’n’ as a subindex stands for time sample
n), Ni number of inputs, No number of outputs). The discrete Fourier transfor-
mation X(k) = DF T (xn ) of state vector xn is defined as
N −1
1 X
X(k) = √ xn zk−n (6.27)
N n=0
with zk = ej2πk/N . Equivalent expressions can be obtained for the DFT of the
signals un and yn as U (k) = DF T (un ) and Y (k) = DF T (yn ) with DF T () defined
142 Chapter 6. Deterministic Frequency-domain Subspace Identification
N −1
1 X
DF T (xn+1 ) = √ xn+1 zk−n (6.28)
N n=0
N −1
z X zk
= √k xn zk−n + √ (xN − x0 ) (6.29)
N n=0 N
zk
= zk X(k) + √ (xN − x0 ) (6.30)
N
with X(k) defined by Eq. 6.27. Taking the discrete Fourier transform of Eqs. 6.25
and 6.2 leads to
zk
zk X(k) = AX(k) + BU (k) + T √ (6.31)
N
Y (k) = CX(k) + DU (k) (6.32)
• Since a transient is nothing else than the output of a system caused by initial
conditions of the states different from zero, the model proposed by equations
(6.31) and (6.32) has exactly the same form as the model valid for periodic
excitations corrupted with transients.
• The exponential decay (stable system) of the transients in the outputs de-
pends on the damping of the system. Therefore, lightly damped systems are
more sensitive for errors due to leakage and transients than highly damped
systems. So, in case of lightly damped systems, it is certainly advised to use
the extended state-space formulation to avoid large errors on the damping
ratios of the system.
• Notice that the extended formulation of the frequency domain state space
model introduces only one extra input in the first equation of the model, i.e.
Eq. 6.31. Existing identification methods, e.g. frequency-domain subspace
algorithms, can easily be adapted to take into account this extra input.
6.5. Extended state-space model for initial and final conditions 143
Consider a data set of measurements with arbitrary input signals. Different identi-
fications methods such as prediction error methods, time-domain subspace meth-
ods, frequency domain-subspace methods exist to estimate the state-space matrices
A, B, C, D. However, an important step in the system identification process is the
validation of the estimated model. A common approach consists in comparing the
estimated model with the experimentally obtained FRFs.
Assuming that only noise is present on the response measurement and the inputs
are noise free, the maximum likelihood estimate for the FRFs is given by the
H1 estimator. Since we consider arbitrary input signals errors are introduced by
leakage, even with the use of a window (e.g. Hanning window). Therefore, a
validation based on the comparison between synthesized transfer functions and
measured FRFs can be misleading and lead to erroneous conclusions about the
model quality.
The goal is now to validate the estimated model Â, B̂, Ĉ and D̂ using an FRF
estimate that is free from bias errors due to leakage. To estimate the FRFs from
the measured time histories, they are divided in Nb blocks with Nb ≥ Ni . The use
of Eqs. 6.31 and 6.32 leads to the following expression for the spectra of block b
zk
Yb (ωk ) = Ĉ(zk I − Â)−1 B̂ + D̂ Ub (ωk ) + Ĉ(zk I − Â)−1 Tb √ (6.33)
N
Right multiplying by UbH (ωk ) and summation over all Nb blocks results in
Nb SY U (ωk ) = Ĉ(zk I − Â)−1 B̂ + D̂ Nb SU U (ωk ) (6.34)
z
√kU1 (k)
N
+ Ĉ(zk I − Â) −1
[ T1 ... TNb
] ..
.
zk
√ UN (k)
N b
(6.35)
Since the matrices Â, B̂, Ĉ, D̂ are already identified in the first step the parameters
Tb for b = 1 . . . Nb can be estimated in a LS sense. In a next step the FRFs can
be estimated by taking into account the initial conditions in order to remove the
bias introduced by leakage and transients
z
√ k U1 (k)
N Nb
.
−1 −1
H1,e (ωk ) = SY U (ωk ) − Ĉ(zk I − Â) [ T1 ... TNb ] . SU U (ωk )
.
z
√ k UNb (k)
N Nb
(6.36)
with H1,e the extended H1 FRF estimator. Eq. 6.36 represents a mixed non-
parametric/parametric estimation of the FRFs, since the parametric compensation
144 Chapter 6. Deterministic Frequency-domain Subspace Identification
for the initial/final conditions. As a model validation the synthesized FRFs from
the estimated system parameters
Similar as for the common denominator model an extended state-space model can
be formulated to start from FRF data. Starting from FRFs has the advantage
that the size of the initial data set is reduced and the influence of the noise on the
measurements is reduced by the averaging process. Nevertheless, in practice one
has to deal with a trade-off between the variance and bias on the FRF estimate.
Given a data set with a limited amount of data, a choice must be made for the
number of blocks Nb . A large number of blocks will result in a large reduction of
the noise levels, but introduces a bias error caused by the leakage, since then only
a small number of data samples are present within a block. From Eq. 6.36 the
extended parametric model for the H1 estimator is given by
INi
F1t (k)
H1 (ωk ) = C(zk I − A)−1 [ B T1 T2 ... TNb ] .. +D (6.38)
.
FNt B (k)
with Fbt (k) = √N1N zk Ub (k)SU U (ωk )−1 and INi a unity matrix of dimension Ni .
b
Equation (6.38) can be rewritten as a state space model
−4
x 10
response y(n)
2
−2
−4
−6
Figure 6.1: Response of a 6-order discrete time system excited by random noise.
6.6.1 Simulations
The goal of the simulation is to illustrate that model ( Eqs.6.31 and 6.32) is
exactly true without any approximation. Therefore, no disturbing noise is added
to the signals. A sixth-order MIMO system (see Appendix) is excited by two
uncorrelated uniformly distributed noise sequences (see Fig. 6.1). Data samples
n = 0, 1, . . . 511 (N = 512) are used for the identification. The DFT spectra of
the 3 responses and 2 inputs are used as primary data for the identification of a
discrete-time state-space model by the basic projection. The state-space model is
identified by the basic projection subspace algorithm for 2 cases: once by using a
Hanning window and no extra term T and once with a rectangular window and
the additional term T to model the initial and final conditions. The differences
between both identified models and the true model are shown in figure 6.2. It can
be seen that the accuracy of the estimates for the second case including T in the
state space model, are at the level of the arithmetic precision of the calculations
(16 digits). Large errors remain if T is not included in the model.
In real-life applications one can not validate the identified model by comparing the
estimated model to the true model. Consider the model Â, B̂, Ĉ and D̂ identified
from the measurements (identification was done with the term T included in the
146 Chapter 6. Deterministic Frequency-domain Subspace Identification
−50
−100
−150
error (dB)
−200
−250
−300
−350
−400
0 0.1 0.2 0.3 0.4 0.5
frequency (Hz)
Figure 6.2: Magnitude of the complex difference between the true and the estimated
model using the classic projection frequency-domain subspace (dashed line) and using
the extended model (dotted line). The full line represents the exact FRF.
model). By comparing the synthesized transfer function with the FRFs estimated
by the H1 method (4 blocks, Hanning window) one concludes that still large errors
are present (see figure 6.3 (a)). This illustrates that the classic validation procedure
is not robust for leakage, since we know the estimated model is of the accuracy
at the level of the arithmetic precision of the calculations (15 digits) as shown by
figure 6.2. Eliminating the bias in the H1 estimator by estimating the Tb from
equation 6.35 and inserting in equation 6.36 leads to an unbiased H1,e estimate of
the FRFs. Validation of the model with these FRFs leads to correct conclusions
like as shown in figure 6.3 (b).
with dˆi the estimated damping ratio and de the exact damping ratio. The mean
square error is equal to the sum of the variance and square of the bias. Figure
6.4 clearly shows the trade-off between variance and bias for the CSSM approach.
The classical H1 approach reduces the noise levels on the FRFs by averaging, but
the bias due to leakage increases by reducing the block length. Based on this
6.6. Simulation and Measurement examples 147
−30 0
−40 −50
−50 −100
amplitude (dB)
amplitude (dB)
−60 −150
−70 −200
−80 −250
−90 −300
−100 −350
−110 −400
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
frequency (Hz) frequency (Hz)
(a) (b)
Figure 6.3: Validation by comparison between the synthesized transfer function (full
line) and the FRF (∗). The difference between both is indicated by the dashed line; (a)
Traditional H1 approach, (b) Mixed non-parametric/parametric approach H1,e
mode 1
mode 2 mode 3
0.7 0.12 0.05
0.6
0.1
0.04
0.5
0.08
0.4 0.03
MSE
MSE
MSE
0.06
0.3
0.02
0.04
0.2
0.01
0.1 0.02
0 0 0
3 5 7 9 11 13 15 3 5 7 9 11 13 15 3 5 7 9 11 13 15
Nb Nb Nb
Figure 6.4: MSE of the damping ratio estimates of the classic model approach (◦) and
extended model approach (∗) for different number of blocks.
trade-off the user has to the define the number of blocks. Notice that the optimal
number of blocks is not a priori known and depends on the total measurement
time and the modal density in the measurements. The extended modal approach
clearly improves the quality of the estimates, since the noise is reduced without
introducing bias errors caused by leakage. In this way the MSE reduces and
becomes much less dependent on the number of blocks.
148 Chapter 6. Deterministic Frequency-domain Subspace Identification
A modal analysis was performed on a subframe of car, which has to support the
engine (see figure 3.1 (b)). Two shakers excite the structure with random forces. A
sample rate of 2048Hz is used and 8K (1K=1024) samples are processed to extract
the model. The extended frequency domain state space model was estimated by
the frequency-domain projection subspace algorithm. To validate the estimated
model, the synthesized transfer functions are compared to the measured transfer
function. Large differences can be observed between the synthesized transfer func-
tions and the FRFs due to the spectral leakage errors in the calculation of the H1
estimator (see figure 6.5 (a) and (b)). When the FRFs are compensated for leak-
age and transients by equation 6.38, a comparison with the synthesized transfers
functions shows a good agreement up to a SNR of 40dB. Even when records of 64K
data samples are processed, a validation based on the classical H1 approach still
leads to misleading results, while the proposed validation technique shows again a
good agreement up to a SNR of 40dB (see Fig 6.5 (c) and (d)).
Starting from 256K time samples of both acceleration and force measurements
reference FRFs are estimated, by dividing the time records in 8 blocks of 64K
resulting in a frequency domain resolution of 0.0156Hz. These FRFs can be as-
sumed to be leakage free and are considered as the reference modal parameters.
Next, only the first 8K data samples are used and divided in 4 equal blocks of each
2K samples resulting in a frequency resolution of 1Hz. First, the FRFs are cal-
culated by the H1 method with a Hanning window. From these FRFs the classic
state-space model (A, B, C and D matrices) is estimated with the basic projection
subspace algorithm. Secondly the same 4 blocks of data are used to obtain the
FRFs with a rectangular window. These FRFs are processed to estimate the ex-
tended state-space model for FRFs (A, B, C, D and T1 , . . . , TNb . Table 6.1 shows
the good agreement between the natural frequencies and damping ratios obtained
from the reference FRFs and the estimates derived from the short data sequences
with the extended state-space model (ESSM) approach. The use of a Hanning
window and a classic state space identification clearly results in poor estimates,
especially for the damping estimates. Figure 6.6 shows the stabilization diagram
for both the classic state-space model (CSSM) approach and the ESSM approach.
Although, the ESSM approach does not use a Hanning window and as a result the
FRF looks much noisier, all the poles clearly appear in the stabilization diagram.
The CSSM approach suffers from estimating double poles, where only one physi-
cal pole is present e.g. around 129Hz, 205Hz, 286Hz and 380Hz. This is typically
caused by errors in the data (due to leakage) and makes the stabilization diagram
confusing for the end-user. Finally, figure 6.7 compares the synthesized FRFs of
both approaches with the reference FRF from the leakage free measurement. One
concludes that although the extended model approach started from only 8K data
samples, averaged in 4 blocks of 2K samples, still a very good agreement is found
between the model and the reference FRFs, while a classic approach clearly fails
due to errors introduced by leakage.
6.7. Conclusions 149
20 20
10 10
amplitude (dB) 0 0
amplitude (dB)
−10
−10
−20
−20
−30
−30 −40
−40 −50
−50 −60
(a) (b)
20 20
10 10
0 0
amplitude (dB)
amplitude (dB)
−10 −10
−20 −20
−30 −30
−40 −40
−50 −50
−60 −60
(c) (d)
Figure 6.5: Validation by comparison between the synthesized transfer function (full
line) and the FRF (∗). The difference between both is indicated by the dashed line;
(a) traditional H1 approach (8K), (b) mixed non-parametric/parametric approach H1,e
(8K), (c) traditional H1 approach (64K), (d) mixed non-parametric/parametric approach
H1,e (64K)
6.7 Conclusions
50
40 40
30
30
20
20
model order
10
model order
0 10
−10 0
−20
−10
−30
−20
−40
−50 −30
150 200 250 300 350 400 150 200 250 300 350 400
Freq. (Hz) Freq. (Hz)
(a) (b)
Figure 6.6: Stabilization diagrams (a) classic state-space model (b) extended state-
space model +: unstable pole, ∗: stable pole
10
−10
amplitude (dB)
−20
−30
−40
−50
−60
−70
150 200 250 300 350 400
frequency (Hz)
Figure 6.7: Comparison between the reference FRF (full line) , synthesized FRF from
the classic state-space model (◦) and from the extended state-space model (∗)
removes bias errors caused by transients and leakage. This validation procedure
is illustrated by both a simulation and a measurement example. In addition, an
extended formulation is proposed to estimate state-space models together with the
initial/final conditions of each data block used by the H1 estimator. Simulations
have shown, that this approach makes the quality of the estimated parameters less
dependent on the choice of the number of blocks to estimate the FRFs.
6.8. Appendix 151
6.8 Appendix
Stochastic frequency-domain
subspace identification
153
154 Chapter 7. Stochastic frequency-domain subspace identification
7.1 Introduction
zk X k = AXk + Wk (7.1)
Yk = CXk + Vk (7.2)
the unit circle. The vectors Wk ∈ C n×1 and Vk ∈ C No ×1 contain respectively the
spectral frequency lines of the process and the measurement noise.
Y = Or X + ΓW + V (7.10)
defined by
Y1 Y2 ... YN
z1 Y1 z2 Y2 ... zN YN
Y= .. .. .. ..
. . . .
z1r−1 Y1 z2r−1 Y2 ... r−1
zN YN
W1 W2 ... WN
z 1 W1 z 2 W2 ... zN WN
W= .. .. .. ..
. . . .
z1r−1 W1 z2r−1 W2 ... r−1
zN WN
V1 V2 ... VN
z1 V 1 z2 V 2 ... zN VN
V= .. .. .. ..
. . . .
z1r−1 V1 z2r−1 V2 ... r−1
zN VN
Eq. 7.10 with r larger than the model order divided by the number of outputs n/No
is the basic equation in frequency-domain subspace identification. In deterministic
frequency-domain subspace algorithms the basic equation Y = Or X + ΓU is in
close analogy to Eq. 7.10. The term ΓU can be removed by the use of (Π)⊥ U,
where (Π)⊥ U is the geometric projection onto the orthogonal complement of the
row space of the input Vandermonde matrix U. However, in the stochastic case
this projection can not be used since the matrices W and V are then unknown.
where a.s.lim stands for the as sure limit [98]) Multiplication of the ith row of W
with the jth column of L leads to
N
zki−1 Wk zkr−j+1 YkH
X
(WL)ij = (7.14)
k=1
= 0 (7.19)
PN
r+i+l−j r+i−j PN
This last equation is due to k=1 zk = 0 and k=1 zk = 0 since
respectively r + i + l − j ∈ N0 and r + i − j ∈ N0 and we assume that zk
158 Chapter 7. Stochastic frequency-domain subspace identification
under the assumption that all poles are stable. Similarly one can shown that
+∞ N N
!
1 1 X H r+i+l−j
zkr+i−j
X X
a.s. lim (VL)ij = S Al zk +R
N →∞ N N
l=1 k=1 k=1
= 0 (7.22)
which concludes the proof.
This observation forms a the basis for stochastic frequency domain subspace
identification. It shows that the terms ΓW and V in the basic equation (7.10) can
be removed by right multiplication with L and that a strongly consistent estimate
of Or is obtained.
YL(LH L)−1 LH
f → Or X̂f w.p.1 for N → ∞ (7.26)
with X̂f a vector of the spectra of the equivalent time domain Kalman filter state
estimate at spectral line f and Lf a part of L containing the information at spectral
line f .
Proof : From the stochastic frequency domain model given by Eq.7.1 one can
form the following relation
N N
1 X ∗ H 1 X H
zk Xk zk Xk = Xk Xk (7.27)
N k=1 N k=1
N N N N
A X H H A X H 1 X H H 1 X H
= Xk Xk A + Xk Wk + Wk Xk A + Wk Wk
N k=1 N k=1 N k=1 N k=1
Σ = AΣAH + Q (7.28)
PN PN
with Σ = E ( N1 k=1 Xk XkH ). Defining Λi = E ( N1 k=1 zki Yk YkH ),
N
G = E ( N1 k=1 zk Xk YkH ) it can be shown from the state-space model that
P
Λ0 = CΣC H + R (7.29)
H
G = AΣC +S (7.30)
i−1
Λi = CA G (7.31)
Note that the elements Λi form the Markov parameters of the deterministic state
space system given by A, G, C, Λ0 . This observation
PN is very closely related to the
stochastic time-domain theory, since Λi = E ( k=1 zki Yk YkH ) can be interpreted
as the Inverse Fourier Transformation (IFT) of the power spectra of the responses.
It is a well known result that the output correlation sequences can be considered
as Markov parameters.
with Lf = [zfr YfH zfr−1 YfH . . . zf YfH ] holds for each spectral line f and zf =
ei2πf /N with f = pN and p a fixed percentage between 0 and 100%. We start
160 Chapter 7. Stochastic frequency-domain subspace identification
with the first two products and apply the strong law of large numbers
1 1
a.s. lim (YL)ij = lim E (YL)ij
N →∞ N N →∞ N
N
!
1 X r+i−j H
= lim E zk Yk Yk
N →∞ N
k=1
= lim Λr+i−j (7.33)
N →∞
Ai−1 G′ Ai−2 G′ G′
∆t = ... (7.37)
Λ′0 Λ′−1 Λ′1−r
...
Λ′1 Λ′0 ... Λ′2−r
Mt = .. .. .. (7.38)
..
. . . .
Λ′r−1 Λ′r−2 ... Λ′0
with G′ = E (xn+1 ynH ) and Λ′i = E (yn+i ynH ). The sequence x̂i+q is the forward
Kalman filter state sequence and yi+q is generated by a stochastic time domain
process
Since Xk and Yk satisfy the frequency domain equivalent of equations (7.39) and
PN
(7.40), their Inverse Fourier Transform (IFT) xi+q = √1N k=1 zki+q Xk and yi+q =
7.3. A second approach 161
PN i+q
√1
N k=1 zk Yk fulfill (7.39), (7.40) and thus equation (7.36) holds. Equation
(7.36) indicates that the Kalman filter generating the estimate of x̂i+q only uses i
output measurements yq , . . . , yi+q−1 . Using the definition of the IFT leads to
PN q
N k=1 zk Yk
X
zki+q X̂k = ∆t Mt−1
..
(7.41)
.
k=1 PN q+i−1
k=1 zk Yk
with Yk and X̂k respectively the frequency spectra of output yi and the Kalman
state estimate x̂i . Multiplying equation (7.41) by zf−q−i , substituting i + q by m
and making the summation for m = 1, 2, . . . , N leads to
PN m−i
N N N k=1 zk Yk
!
X
zf−m
X
zkm X̂k = ∆t Mt−1
X −m ..
zf (7.42)
.
m=1 k=1 m=1 PN m−1
k=1 zk Yk
N
1 X −m m
z zk = δkf (7.43)
N m=1 f
N
1 X −m m−i
z zk = zk−i δkf (7.44)
N m=1 f
with δkf = 1 for k = f and δkf = 0 for k 6= f . Substituting equations (7.43) and
(7.44) in (7.42) results in
X̂f = ∆t Mt−1 LH
f (7.45)
∆ = Ai−1 G Ai−2 G . . . G
(7.46)
Λ0 Λ−1 . . . Λ1−r
1 Λ1 Λ0 . . . Λ2−r
a.s. lim LH L − . . . =0 (7.47)
N →∞ N .. .. .. ..
.
Λr−1 Λr−2 ... Λ0
the proof reduces to showing that G = G′ and Λi = Λ′i . (the proof of equation
(7.47) is similar that of equation (7.33)). Using the definition of the IFT one
162 Chapter 7. Stochastic frequency-domain subspace identification
obtains
since Eq. 7.44 holds and similar it can be shown that Λ′i = Λi . This concludes the
proof of the theorem 2.
The major difference between the first approach based on equation (7.12) and
the second approach based on (7.26), is that the second approach provides us an
estimate of the frequency spectra of the states (for f = 1, . . . N ). Given the
states, one can then easily obtain the covariance matrices Q, R and S as shown
in paragraph7.6. An estimate of the extended observability matrix and the state
estimate is obtained again from an SVD
In paragraph 7.6 it is shown how the observability matrix and the state spectrum
is obtained from U , S and V .
7.4. Connection to time domain stochastic subspace identification 163
In time domain stochastic subspace algorithms one starts from a Hankel matrix,
which is divided in a past and a future submatrix as follows
yo y1 ... yN −1
.. .. .. ..
. . . .
yr−2 yr−1 ... yN +r−3
def
yr−1 yr ... yN +r−2 yp
= (7.55)
yr yr+1 ... yN +r−1 yf
yr+1 yr+2 ... yN +r
.. .. .. ..
. . . .
y2r−1 y2r ... yN +2r−2
The main theorem of [123] for stochastic subspace identification states that
yf /yp = Or x̂ (7.56)
for N → ∞ with x̂ the forward Kalman filter state sequence. In frequency domain
stochastic subspace identification, the Hankel matrix is replaced by a Vandermonde
matrix
Y+ /Y− = Or X̂ (7.58)
for N → ∞ and with X̂ the DFT spectrum of the forward Kalman filterPN state se-
quence. In the case N → ∞, F(yn+j ) = zkj Yk with Yk = F(yn ) = √1N n=1 yn zk−n
holds and thus the Vandermonde matrix can be considered as the DFT of the rows
of the Hankel matrix.
164 Chapter 7. Stochastic frequency-domain subspace identification
1. Build the Vandermonde matrices Yf and Yp given the data Yk for a value
r > n/No .
2. Calculate the projection Yf /Yp by means of the QR-factorization
H
Yp RB 0 QA
= H H (7.63)
Yf RAB RA QB
H
Yf /Yp = RAB QA (7.64)
4. Determine the extended observability matrix Or and the state estimate X̂.
For a chosen order n (in theory the order n is equal to the rank of S)
1/2
Or = W1−1 U1 S1 (7.66)
1/2
X̂ = S1 V1H W2−1 (7.67)
with U1 = U[:,1:n] , S1 = S[1:n,1:n] and V1 = V[1:n,:] . In paragraph 7.7 different
choices for the weighting matrices are briefly discussed.
7.6. Final Algorithms 165
5. The estimates for the A and C matrices can be obtained in two different
ways (by not using the states or not)
(a) The A and C matrices can be obtained from the extended observability
matrix.
+
 = Ô[1:No (r−1),:]
Ô[No +1:No r,:] and Ĉ = Ô[1:No ,:] (7.68)
(b) Another possibility exists of solving the following equation for A and C
in a LS sense, since X̂k and Yk are known
zk X̂k A ρWk
= X̂k + (7.69)
Yk C ρVk
6. In the last step the covariance matrices Q, S and R are estimated. Two
possible approaches can be used.
(a) The first and most straight forward approach, results in a guaranteed
positive definite covariance matrix. If the A and C matrices are ob-
tained from the observability matrix (step 5a) the spectra Wk and Vk
can be calculated from equations 7.1 and 7.2 since X̂ is already known.
When the A and C matrices are obtained from the LS problem (step
5b) the residues ρWk and ρVk can be considered as Wk and Vk . The
covariance matrix is then estimated by
N
Q S 1 X Wk
WkH VkH
= (7.70)
SH R N Vk
k=1
The total identification procedure is schematically given by figure 7.1 for the case
of (5a) and (6a).
166 Chapter 7. Stochastic frequency-domain subspace identification
Y+ / Y- = Or X
Least-squares
zk X k A W
= Xk + k
Yk C Vk
7.7 Remarks
• In case that No r > N it follows from Eq. 7.59 that the projection Y+ /Y− =
Y+ , which makes the QR operation unnecessarily and the SVD decompo-
sition can start directly from Y+ . Physically it means that Y+ totally lies
in the row space spanned by Y− . In practice this occurs for measurements
with a high spatial density e.g. scanning laser vibrometer measurements and
a limited frequency band e.g. r = 10, No = 100 and N = 500.
• In the case that the excitation signal is a multisine with a constant am-
plitude spectrum and a random phase and only one input is used (e.g. a
loudspeaker), the assumption for the excitation signals Wk and Vk are still
valid. Notice that in this case no probability limit for N → ∞ is needed to
prove the main theorem 7.26. Indeed, in this case the following expressions
are valid, without the use of their expected value
N
X
zkl Vk VkH = 0 (7.75)
k=0
N
X
zkl Wk WkH = 0 (7.76)
k=0
N
X
zkl Wk VkH = 0 (7.77)
k=0
• The proof of the theorems in this chapter are given for a complex state-space
model. In the case that one is interested to fit a state-space model with real
valued system matrices on the measurements, the matrices Y+ and Y− must
168 Chapter 7. Stochastic frequency-domain subspace identification
re re
be converted respectively to Y+ and Y− in accordance with Eq. 6.11. The
main theorem is than given by
re re
Y+ /Y− → Or X̂re
f w.p.1 for N → ∞ (7.78)
with zk covering the half unit circle. In the case of real system matrices the
double order is required compared to a complex model for estimating the
same amount of poles, i.e. n = 2Nm .
• The algorithm can be slightly adapted to reduce the calculation time by
replacing the QR decomposition by a QR decomposition without the explicit
calculation of the Q.
• In some cases one is interested to use only a few outputs as references in-
stead of all outputs. This reduces the dimensions of the problems and by
consequence the calculation time. Furthermore, some sensors are of higher
quality in terms of the structural information than others (i.e. they are
not placed in nodes of certain modes). If these ’high quality’ sensors are
chosen as the reference sensors no information is lost and the identification
may even result in better estimates [87]. In fact the row space given by
Y− can be considered as a reference space on which the measurement Y+
are projected. So the use of only the reference outputs to construct the
r
reference subspace Y− transforms the proposed algorithm in a reference-
based stochastic frequency-domain subspace algorithm in analogy with the
time-domain counterpart [87]. This reference based subspace speeds up the
algorithm, without any loss of information (if the reference sensors are well
chosen).
• In the case that short time sequences are transformed to the frequency do-
main leakage errors are introduced resulting in biased estimates. Therefore,
an additional term T must be taken into account in the state-space model,
resulting in a combined deterministic-stochastic model, which is discussed in
the next chapter.
7.8.1 Simulation
The simulations are done in the frequency domain (N =1024) for a discrete stochas-
tic system described by the matrices given in appendix 7.10. The system contains
two poles and two outputs. Figure 7.2 shows a comparison of the estimated model
represented by the estimated power density function
−20 −20
−30 −30
−40 −40
−50 −50
amplitude (dB)
amplitude (dB)
−60 −60
−70 −70
−80 −80
−90 −90
−100 −100
−110 −110
−120 −120
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
frequency line (Hz) frequency line (Hz)
Figure 7.2: Comparison between the estimated model (full line) and the output spectra
(dots)
and the spectra of the outputs. Figure 7.3 compares the exact absolute value of
both poles with the estimates of 1000 Monte Carlo runs. The difference between
0.99
0.98
0.97
0.96
0.95
abs(pole)
0.94
0.93
0.92
0.91
0.9
0.89
0 200 400 600 800 1000
Monte Carlo run
Figure 7.3: Comparison between the exact magnitude of the poles and the estimated
magnitude of the poles of the 1000 Monte Carlo runs
the estimated power spectra and the true power spectra satisfies the 95% limit
given by 2σ, with σ the standard deviation of the 1000 runs (see Figure 7.4)
170 Chapter 7. Stochastic frequency-domain subspace identification
−20
−40
−60
Amplitude (dB)
−80
−100
−120
−140
−160
0 0.2 0.4 0.6 0.8 1
Frequency (Hz)
Figure 7.4: Comparison between the true system (full line), the error between the true
(dotted line) and estimated system and the 95% confidence limit (dashed line) from the
1000 Monte Carlo runs.
20
model order
15
10
Figure 7.5: Subframe. Stabilization chart for the stochastic frequency-domain subspace
algorithm. Notice that all identified poles are stable (positive damping), ∗: stable pole.
The proposed technique was applied to fit a stochastic model through the mea-
surements on the subframe of a car starting from the output-only spectra. The
accelerations at 23 different positions are measured at a sampling rate of 2048 Hz
during 10 s, while the structure was excited by two shakers with random forces. Be-
fore the actual modelling task starts, the frequency band of interest between 190Hz
and 390Hz is re-scaled to cover the half unit circle uniformly (real system matrices
A and C are estimated). Notice that the identification starts from output-only
spectra and the solution obtained from the stochastic state-space identification
7.8. Simulation and Measurement example 171
60 55
50
50
45
40
40
Ampl. (dB)
Ampl. (dB)
35
30 30
25
20
20
15
10
10
0 5
200 250 300 350 240 250 260 270 280 290
Freq. (Hz) Freq. (Hz)
60
60
50
50
40
Ampl. (dB)
Ampl. (dB)
30 40
20
30
10
20
0
10
200 250 300 350 200 210 220 230 240 250 260
Freq. (Hz) Freq. (Hz)
50 50
40
40
Ampl. (dB)
Ampl. (dB)
30
30
20
20
10
0 10
−10 0
200 250 300 350 325 330 335 340 345 350 355 360
Freq. (Hz) Freq. (Hz)
Figure 7.6: Subframe. Comparison between the synthesized power spectral density
from the estimated model (full line) and the output spectra (×).
the present turbulent forces and the artificial excitation. Figure 7.7 illustrates the
stabilization chart from the analysis of 12 acceleration measurements during flight
on a military aircraft. Similar as for the subframe all identified poles are stable. To
make a distinction between the physical and mathematical poles, one has to rely
7.8. Simulation and Measurement example 173
14
12
10
model order
8
4 5 6 7 8 9 10
Freq. (Hz)
Figure 7.7: In-flight test. Stabilization chart for the stochastic frequency-domain sub-
space algorithm. Notice that all identified poles are stable (positive damping), ∗: stable
pole.
on both a priori known insights in the dynamical behavior (from e.g. simulations
and ground vibration tests) and criteria based on both mathematics and physics.
In [141], [132] the use of pole-zero criteria combined with the uncertainty levels
on the estimated poles is presented to select the physical poles, while [135] pro-
poses an automated interpretation of the stabilization chart by the use of cluster
algorithms and physical parameters such as the mode complexity and the modal
assurance criterium [52]. Some different methods based on a heuristic approach
for an automatic interpretation of the stabilization chart are proposed in [108],
while in [35] some new techniques are presented for mode selection. Figure 7.8
illustrates the synthesized power spectra of 4 different acceleration measurements.
70
65
65
60
60
55
55
50
50
Ampl. (dB)
Ampl. (dB)
45
45
40
40
35
35
30
30
25
25
20
20
15
4 5 6 7 8 9 10 4 5 6 7 8 9 10
Freq. (Hz) Freq. (Hz)
65 70
60
60
55
50
50
Ampl. (dB)
Ampl. (dB)
45
40 40
35
30
30
25
20
20
15 10
4 5 6 7 8 9 10 4 5 6 7 8 9 10
Freq. (Hz) Freq. (Hz)
Figure 7.8: In-flight test. Comparison between the power spectral density synthesized
from the estimated model (full line) and the output spectra (×)
20
18
16
14
model order
12
10
2 3 4 5 6
Freq. (Hz)
Figure 7.9: Stabilization chart for the stochastic frequency-domain subspace algorithm.
Notice that all identified poles are stable (positive damping), ∗: stable pole.
7.9. Conclusion 175
80
70
70
60 60
Ampl. (dB)
Ampl. (dB)
50 50
40
40
30
30
20
20
10
1 2 3 4 5 6 1 2 3 4 5 6
Freq. (Hz) Freq. (Hz)
Figure 7.10: Comparison between the power spectra synthesized from the estimated
model (full line) and the output spectra (×)
7.9 Conclusion
7.10 Appendix
Combined frequency-domain
subspace identification
177
178 Chapter 8. Combined frequency-domain subspace identification
8.1 Introduction
zk X k = AXk + BUk
Yk = CXk + DUk + Nk (8.1)
with Nk zero mean complex circular independent noise. Remind that, al-
though this algorithm is consistent, dynamics excited by ambient forces are
not identified, since no information is extracted from Nk . Therefore, this
algorithm has no combined deterministic-stochastic meaning in the OMAX
framework.
xn+1 = Axn + wn
yn = Cxn + vn
white wn and vn white zero mean noise sources. Under the assumption that
the measured input forces un are uncorrelated with the unmeasurable in-
puts wn and vn combined stochastic-deterministic subspace algorithms are
consistent. The combined algorithms identify the dynamics excited by the
8.2. Theoretical Aspects 179
Wk Q S
WkH VkH
E = (8.4)
Vk SH R
def
Y−
Y= = (8.6)
Y+
with r > n/No . Both the input and output Vandermonde matrices are divided in
a ’positive’ U+ , Y+ and a ’negative’ U− , Y− Vandermonde matrices as defined
above. The main theorem of this chapter is then given by
Proof : Consider the Inverse Discrete Fourier Transform (IDFT) of both the
8.2. Theoretical Aspects 181
N
1 X
y(n) = IDF T (Yk ) = √ Yk zkn (8.10)
N k=1
N
1 X
u(n) = IDF T (Uk ) = √ Uk zkn (8.11)
N k=1
From the definition of the IDFT it follows that if y(n) = DF T (Y ) then y(n+m) =
DF T (z m Y ). Taking the IDFT from the basic model equations (8.2) and (8.3)
results in the following time-domain state space model
with x = [x0 . . . xN ]T the forward Kalman filter state sequence estimate. Consider
the Discrete Fourier Transformation matrix
Y− = yp F
Y+ = yf F
U− = up F
U+ = uf F (8.17)
Once the states Xk are estimated by the oblique projection according the main
theorem Eq. 8.7 the system matrices A, B, C and D can be estimated in a LS
sense
zk X k A B Xk ρWk
= + (8.22)
Yk C D Uk ρVk
Since it is assumed that no input noise is present and the state spectra Xk are
consistent estimates, the least squares estimate of the system matrices is consistent.
8.3. Practical Implementation 183
The covariance matrix of the stochastic sources Wk and Vk is then obtained from
the residuals ρWk and ρWk as
N
Q S 1 X ρWk
ρH ρH
= (8.23)
SH R N ρVk Wk Vk
k=1
Until now it was assumed that the given spectra Yk and Uk obey the frequency-
domain state-space model. Since the spectra of the outputs P and inputs are ob-
N −1
tained from the DFT of the measured time signals Yk = √1N n=0 y(n)zk−n and
PN −1
Uk = √1N n=0 u(n)zk−n the proposed model is only true if x(N ) = x(0) or
N = ∞. In all other cases the extra term T must be taken into account in
the state space model (according to paragraph 6.5.1). The extended combined
deterministic-stochastic frequency-domain state-space model is then given by
x(N )−x(0)
where T = √
N
models the influence of the initial and final conditions. These
extra parameters T (T ∈ C n×1 ) model the non-steady state response of the system.
Taking into account these additional parameters makes the frequency model robust
for leakage and transients effects in case a rectangular window (uniform window) is
used for the calculation of the spectra. This observation generalizes the stochastic
frequency-domain subspace identification method proposed in chapter 7, since the
influence of leakage and transients can be modelled by a combined deterministic-
stochastic model with input Uk = √1N zk
4. Determine the extended observability matrix Or and the state estimate X̂.
For a chosen order n (in theory the order n is equal to the rank of S)
1/2
Or = U1 S1 (8.35)
1/2
X̂ = S1 V1H (8.36)
with U1 = U[:,1:n] , S1 = S[1:n,1:n] and V1 = V[1:n,:] . In the case that real
1/2 ′
system matrices are estimated X̂ ′ = S1 V1H and X̂ = X̂[:,1:N ′
] +iX̂[:,N +1:2N ] .
5. Knowing Yk , Uk and X̂k the system matrices are obtained by solving the LS
problem 8.22 and the covariance matrix of the process and noise is obtained
by 8.23
Least-squares
zk X k A B X k Wk
=
+
Yk C D Yk Vk
8.4 Remarks
• Equation (8.34) can be replaced by
U−
W1 Y+ /U+ W2 = U SV H (8.37)
Y−
No ×Ni No ×Nref
Hk or Yk = G+ yy,k and Uk = Ik with Hk ∈ C and G+ yy,k ∈ C
respectively the FRF matrix and power spectra matrix at spectral line k and
Ik a unity matrix with dimensions Ni . Since the inputs Uk are considered
white in this case, the projection results in a consistent estimate of the state
spectra and the observability matrix. Compared to the consistent determin-
istic projection algorithm of paragraph 6.4 the advantage is that no a-priori
known covariance matrix of the noise has be taken into account.
• The same motivations to perform the identification in the frequency do-
main as demonstrated for the stochastic subspace identification are valid for
the combined identification: i.e. easy pre-filtering of the data and simple
frequency band selection. The chosen frequency band is re-scaled to cover
uniformly the full unit circle (half unit circle in case of real system matrices)
to ensure consistency and the numerical stability is improved, since the basis
functions z r are orthogonal. In practice, it is advised to cover only 90% to
make a trade-off between consistency and model errors (since the frequency
band of the underlying continuous-time domain model is approximated by a
discrete-time domain model).
Lets consider a discrete 6th order system excited by 2 Gaussian random sequences.
The A, B, C and D matrices are given in appendix 8.7. The first input is a stan-
dard gaussian distributed sequence with standard deviation 1, while the second
input has a standard deviation of 0.5. All simulations are performed in the time-
domain to include the non-steady state responses. During the identification process
the second input U2,k is considered as unknown and by consequence the identifi-
cation process can be considered as a combined deterministic-stochastic problem.
The response in the frequency-domain can be considered as a sum of three contri-
butions i.e. the deterministic part, the transient part and the stochastic part
Yk = C(Izk − A)−1 B[:,1] + D[:,1] U1,k + C(Izk − A)−1 T zk
−20 −20
−40
−40
−60
−60
−80
Ampltitude
Ampltitude
−80
−100
−100
−120
−120
−140
−140
−160
−180 −160
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
frequency (Hz) frequency (Hz)
(a) (b)
Figure 8.2: Mean and variance of the estimated system for N=250 (◦ exact system, full
¯ 2
line Ĝ, dashed line σĜ : a) Combined deterministic-stochastic frequency domain subspace
algorithm (A1) b) Classical frequency-domain projection subspace algorithm (A3). One
can clearly notice the bias of algorithm A3.
Table 8.1: MSRE for Monte Carlo simulations comparing the presented combined algo-
rithm including transient effects (A1), combined algorithm not including transient effect
(A2) and the classical frequency domain subspace projection algorithm (A3). From the
simulation results it is clear that the MSRE of A1 and A2 decreases if the number of
frequency lines N increases, while A3 remains biased
mean estimated model (for the 100 runs and 100 estimated models) and the true
model is calculated
N ¯
1 X Ĝk − Gk 2
|E|2 = | | (8.39)
N Gk
k=1
¯
with Gk the ’true’ system between the output and the first input and Ĝk the mean
of the 100 estimated models. This MSRE is calculated for N = 150, 250, 400 and
600 given in table 8.1. The presented subspace model is compared with the classical
projection frequency domain projection algorithm. The presented algorithm was
applied including the estimation of the transient T by considering an conditional
input zk and without estimating the transient effects. The identification starts
from the spectra of the response Yk and the first input U1,k . All spectra are
calculated using a rectangular window. From table 8.1 it is clear that the proposed
188 Chapter 8. Combined frequency-domain subspace identification
The presented algorithm can still estimate the system matrices C and A. This is
impossible by the classical frequency-domain projection subspace algorithm, since
it models only a deterministic model between the input and output spectra. From
the 100 simulations the mean value and the standard deviation of the estimates of
the magnitude R and angle θ of the second pole p2 = Reiθ with R = 0.9858 and θ =
1.5938 is shown in Table 8.2. (similar for the other poles). The simulations show
Table 8.2: Monte Carlo simulations for a stochastic identification for different numbers
of spectral lines N
clearly that the differences between the mean values of the 100 times estimated
magnitude and phase of the pole p2 and the exact value decreases for an increasing
number of spectral lines (consistent). The variances on the estimated R and θ
decrease for an increasing number of spectral lines (consistency). Figure 8.3 shows
the output spectra and the estimated stochastic model from this spectra for 1
simulation.
40
20
amplitude (dB)
−20
−40
−60
−80
−100
−120
0 0.1 0.2 0.3 0.4 0.5
frequency (Hz)
Figure 8.3: Output spectra (dots), estimated stochastic model (full line)
• both the amplitude and phase of the modes excited by the unknown forces
can be determined.
This makes the combined frequency-domain subspace algorithm in the general case
superior to the common-denominator based combined algorithm.
190 Chapter 8. Combined frequency-domain subspace identification
damping ratio
1.75
1.5
1.7
1.4 1.65
1.6
1.3 1.55
1.5
1.2
1.45
1 2 3 4 5 6 1 2 3 4 5 6
turbulence level turbulence level
2.8
2.6 2.5
2.4
2.2
damping ratio
damping ratio 2
2
1.8
1.6 1.5
1.4
1.2
1
1
0.8
1 2 3 4 5 6 1 2 3 4 5 6
turbulence level turbulence level
2.6
4
2.5 3.5
damping ratio
3
damping ratio
2.4
2.5
2.3
2
2.2
1.5
2.1 1
1 2 3 4 5 6 1 2 3 4 5 6
turbulence level turbulence level
Figure 8.4: Comparison of the damping ratio estimates (∗ : CLSF IO, ◦ : determin-
istic frequency-domain subspace , : combined frequency-domain subspace, – : exact
parameter)
14 14
12 12
10 10
model order
model order
8 8
6 6
4 4
2 2
4 5 6 7 8 9 10 4 5 6 7 8 9 10
Freq. (Hz) Freq. (Hz)
(a) (b)
f o f f f o f o f o f f f f s s ss s f sd fs s s
14 s o s sf f o o sf s f f o 14 s f sf s s o s fs s s
s f fo f f o s s f f s s s ss s o d s s s s
12 s f f f o o f f s of 12 o s s ss s fs s s s
f s s f so s s s f f s s ss s s s s s
10 s f ff f f f f f 10 s f ss s f f s d
model order
model order
f f ff f s s f o s s ss o s s s o
8 f o sf o s f s 8 f o fs f f s
s f fs f s s s o fs o s s
6 f ss o f f 6 f sf o f f
o sf o f d sf o s
4 f f o f 4 f f s
f f s ff s
2 2
4 5 6 7 8 9 10 4 5 6 7 8 9 10
Freq. (Hz) Freq. (Hz)
(c) (d)
turbation of the flaps is used as the input force signal. The stochastic contribution
in the responses is caused by the turbulent forces acting on the airplane during
flight. Starting from the input and output spectra in a frequency band from 3Hz to
11Hz, a state-space model with real system matrices is estimated by both the clas-
sical projection and combined deterministic-stochastic frequency-domain subspace
algorithms. Figure 8.5 compares the stabilization charts for both the (determin-
istic) projection algorithm and the combined deterministic stochastic approaches.
Based on the relative differences between the eigenfrequencies and damping ratios
estimated for subsequent model orders n, the poles are labelled in the stabilization
diagram. The symbols s, f , d and o respectively mean
• s: both relative damping ratio difference < 15% and relative eigenfrequency
192 Chapter 8. Combined frequency-domain subspace identification
Table 8.3: Natural frequencies and damping ratios identified by the deterministic and
combined deterministic-stochastic frequency-domain algorithm.
difference < 3%
• f : relative damping ratio difference ≥ 15% and relative eigenfrequency dif-
ference < 3%
• d: relative damping ratio difference < 15% and relative eigenfrequency dif-
ference ≥ 3%
• o: relative damping ratio difference ≥ 15% and relative eigenfrequency dif-
ference ≥ 3%
After applying these criteria it is clear that the stabilization diagram of the com-
bined algorithm results in a easier interpretation of the chart than the determin-
istic algorithm. Table 8.3 clearly shows that some of the natural frequencies and
damping ratios are not identified by the deterministic approach. Furthermore,
large deviations in the estimated damping ratios can be observed between the de-
terministic and combined identified parameters. From the aircraft manufacturer it
was confirmed that all the 9 poles identified with the combined are the only 9 poles
in the frequency band and the damping ratios were in very close agreement with
the manufacturers specification. From figure 8.6, which compares the synthesized
FRFs for both the deterministic and combined algorithm with the FRFs obtained
from the H1 estimator, it is clear that the deterministic algorithm results in much
larger model errors than the combined algorithm.
The combined deterministic-stochastic algorithm can also start from FRFs, posi-
tive power spectra or both simultaneously as primary data to estimate the system
8.5. Simulations and Real-life measurement examples 193
25
25
20
20
15
15
10
10
Ampl. (dB)
Ampl. (dB)
5
5 0
0 −5
−5 −10
−10 −15
−15 −20
−25
−20
4 6 8 10 4 5 6 7 8 9 10
Freq. (Hz) Freq. (Hz)
−5 −5
−10 −10
−15 −15
−20 −20
Ampl. (dB)
Ampl. (dB)
−25 −25
−30 −30
−35
−35
−40
−40
−45
−45
−50
−50
−55
−55
4 6 8 10 4 5 6 7 8 9 10 11
Freq. (Hz) Freq. (Hz)
15 15
10 10
5 5
0 0
Ampl. (dB)
Ampl. (dB)
−5 −5
−10
−10
−15
−15
−20
−20
−25
−25
−30
−30
4 5 6 7 8 9 10 11 4 5 6 7 8 9 10 11
Freq. (Hz) Freq. (Hz)
Figure 8.6: Comparison between the FRFs (×) and the synthesized FRFs obtained
by the classic projection algorithm (left) and the combined deterministic-stochastic fre-
quency domain subspace algorithm (right). The classical projection approach results in
large bias errors. (times: measured FRFs, full line: synthesized FRFs)
matrices and modal parameters. The combined algorithm was used in chapter 3
process both the simulation data and measurements obtained on the Villa Paso
bridge. The next two examples show that the algorithm is also capable to start
from FRF data from modal testing in the laboratory which illustrates its capabil-
194 Chapter 8. Combined frequency-domain subspace identification
ity to handle a large number of outputs, several inputs, frequency band selection
and large model orders.
Subframe of an engine
2
10
1
10
0
10
MSRE
−1
10
−2
10
10 20 30 40
FRF
Figure 8.7: Comparison between the mean square relative errors for the 46 FRFs.
(◦ classical projection frequency-domain subspace method, ∗ combined deterministic-
stochastic frequency domain subspace algorithm)
20 20
10 10
0 0
Ampl. (dB)
Ampl. (dB)
−10 −10
−20 −20
−30
−30
−40
−40
−50
−50
200 250 300 350 400 200 250 300 350 400
Freq. (Hz) Freq. (Hz)
30 30
20 20
10 10
Ampl. (dB)
0 Ampl. (dB) 0
−10 −10
−20 −20
−30 −30
200 250 300 350 400 200 250 300 350 400
Freq. (Hz) Freq. (Hz)
20 20
10 10
0 0
−10 −10
Ampl. (dB)
Ampl. (dB)
−20 −20
−30 −30
−40 −40
−50 −50
−60 −60
−70
200 250 300 350 400 200 250 300 350 400
Freq. (Hz) Freq. (Hz)
Figure 8.8: Comparison between the measured FRFs and the synthesized FRFs by the
classical projection algorithm (left) and the combined deterministic-stochastic frequency
domain subspace algorithm (right). The classical projection approach results in large
bias errors. (×: measured FRFs, full line: synthesized FRFs)
containing 50 modes, with the frequencies covering 90% of the half unit circle.
Figure 8.9 shows the high quality of the synthesized FRFs for 4 different FRF
measurements. Chapter 9 discusses in more detail how the physical poles can be
distinguished from the mathematical poles by a small adaption of the algorithm
196 Chapter 8. Combined frequency-domain subspace identification
to obtain clear stabilization charts based on the sign of the damping ratio.
8.6 Conclusion
−75
−85
−80
−90
−85
−95
−90
−100
Ampl. (dB)
Ampl. (dB)
−95
−105
−100
−110
−105
−115
−110
−120
−115
−125 −120
−130 −125
−70
−80
−80
−90
−90
−100
Ampl. (dB)
Ampl. (dB)
−100
−110
−110
−120
−120
−130 −130
−140 −140
Figure 8.9: Comparison between the measured FRFs and the synthesized FRFs by the
combined deterministic-stochastic frequency domain subspace algorithm. (×: measured
FRFs, full line: synthesized FRFs)
8.7. Appendix 197
8.7 Appendix
This chapter analyzes the influence of the parameter constraint and basis function
used in the implementations for Modal Parameter Estimation (MPE) methods on
the quality of the stabilization diagram. It is shown that by a proper choice of the
parameter constraint/basis function a distinction can be made between physical and
mathematical poles based on the sign of the real part of the pole. As a result, the
quality of stabilization diagrams can be improved. This approach can be applied for
several well-known modal parameter estimation methods and the user can benefit
from this. Several MPE applications in aerospace, automotive and civil engineering
are studied in this chapter, showing the implication of the constraint/basis function
and the specific advantages and disadvantages of obtaining an easy-to-interpret
stabilization diagram.
199
200 Chapter 9. The secrets behind clear stabilization charts
9.1 Introduction
In the previous chapters proposed several advanced modal parameter estimation al-
gorithms are proposed. In modal analysis, one typically uses an identification order
that is guaranteed to be larger than necessary to be sure all dynamics of the struc-
ture are described. Unfortunately, as the model order of the model is increased,
so will the number of identified poles. This results in the introduction of so-called
mathematical poles, which have no direct physical interpretation. Therefore, an
important step in modal parameter estimation (MPE), where user interaction is
required, is the interpretation of the stabilization diagram to make the distinction
between physical and mathematical modes. This chapter goes into the mathemat-
ical background of several MPE algorithms such as e.g. Least Squares Complex
Exponential (LSCE) method [52], [72], Least Squares Complex Frequency domain
(LSCF)[46] and frequency-domain subspace methods in order to discover the crit-
ical and underestimated importance of the parameter constraint/basis function in
the algorithms. It will be shown and explained how the choice of a constraint/basis
function in solving an identification problem has its influence on the separation
between physical and mathematical modes. In a next step, it is explained how this
mathematical key idea can be exploited in several MPE algorithms and its effect
on the stabilization diagram is illustrated. Both the advantages and disadvantages
of the choice of the constraints/basis function are discussed.
In several papers it was noticed that methods like the LSCF [116], [113], [114]
and the more recent poly-reference LSCF (also called PolyMAX algorithm) [46],
[89] produce very clear stabilization diagrams. This chapter explains why these
methods have such nice properties for EMA applications. Furthermore, this key
idea behind the construction of clear diagrams is generalized and applied for other
well known MPE algorithms such as e.g. the LSCE algorithm and frequency-
domain subspace algorithms.
between physical and mathematical poles for the estimation of damped sinusoids
is given based on the sign of the damping for an ARX time-domain model. In
this chapter a frequency-domain counterpart is developed and the influence of the
constraint or basis function is investigated. Finally the results are applied for se-
veral modal parameter estimation techniques and illustrated by several practical
examples.
9.2.1 LS Identification
The goal of this section is to show the influence of the constraint for a Least
Squares (LS) problem in identification algorithms. Consider an over-determined
set of equations
Jθ = E (9.1)
J˜θ̃ = −Jj + E
θ̃ = −J˜† Jj + J˜† E (9.2)
with r the order and aj complex coefficients (real coefficients are considered as
a special case of complex coefficients). Taking the DFT of Eq. 9.3 leads to the
202 Chapter 9. The secrets behind clear stabilization charts
that both leakage and transients are neglected. Taking into account these effects
would introduce an extra polynomial as discussed in chapter 3. The poles of this
AR-model are given by the solutions of the characteristic equation given by
ao z r + a1 z r−1 + . . . ar = 0 (9.5)
Frequency-domain solution
Since the only assumption on Ek in Eq. 9.4 is that Ek is circular complex in-
dependent and identically distributed noise, the phases of Eq. 9.4 contain no
information. Therefore, Eq. 9.4 is totally equivalent with
|â0 + â1 zk−1 + . . . + âr zk−r ||Yk | = |Ek | (9.6)
with |x| the absolute value of x and as a result, every set of estimated parameters
[â0 â1 . . . âr ] that satisfies Eq. 9.6 is a solution of equation 9.4. The polynomial
A(z) can be written as
A(z) = z −r (a0 z r + a1 z r−1 + . . . ar ) (9.7)
−r
= a0 z (z − λ1 )(z − λ2 ) . . . (z − λr ) (9.8)
with λj = eσj +iωj and σj , ωj respectively the damping and natural frequency.
The magnitude of the polynomial A(z) is given by
|A(z)| = |a0 ||z − λ1 ||z − λ2 | . . . |z − λr | (9.9)
In appendix 9.6, it is proven that
|zk − λj | |zk − eσj +iωj |
= = Rj (9.10)
|zk − λ△
j |
|zk − e−σj +iωj |
Consider the identification process of the coefficients aj of A(z) starting from the
spectra of the response Yk . This identification can be formulated as an over-
determined set of equations in the same way as in Eq. 9.1.
Y1 Y1 z1−1 . . . Y1 z1−r
a0 E1
Y2 Y2 z2−1 . . . Y2 z2−r a1 E2
.. .. .. .. = .. (9.13)
..
. . . . . .
−1 −r
YN YN zN ... YN zN ar EN
Eq. 9.13 can be solved in a least squares sense according to Eq. 9.2.
The proof of this statement is given in appendix 9.7. A stable, unstable or mixed
stable-unstable AR process in the frequency-domain can be identified as a stable
or unstable system that fulfils Eq. 9.4 depending on the choice of the constraint.
This observation plays a major role in the next sections.
30
20 0.1
10
−10 0
−20
−0.05
−30
−0.1
−40
(a) (b)
Figure 9.1: (a) Spectra of the response Y(k); (b) Monte Carlo runs: full line exact
damping values, estimated damping values σ for a0 = 1 (∗), estimated damping values
σ for a2 = 1 (◦)
poles (σ < 0) for every run, the constraint a2 = 1 estimates unstable poles for all
runs.
and multiplication by z −r does not change the assumption made for Ek (since the
phase of Ek is random) thus
′
(z −r + a1 zk−r+1 + . . . + ar )Yk = Ek (9.16)
′
with Ek = z −r Ek . This corresponds to a constraint for the highest order coefficient
in z −1 fixed to one or with ar = 1 in Eq. 9.4. Thus replacing z −1 by zk in the
identification process of the AR model results in a change of the sign of the damping
values σj . Finally, it was observed that solving the identification problem in a TLS
sense by imposing the norm-2 of the coefficients fixed to 1 results in damping values
σj = 0 and thus biased estimates.
Time-domain Solution
Under the assumption that the system is in steady-state, the time-domain identi-
fication has the same properties as the frequency-domain identification. It is well
9.2. Theoretical Aspects 205
Table 9.1: Stability of the poles in function of the constraint and basis function
z −1 z
a0 = 1 unstable stable
ar = 1 stable unstable
The solution of the least squares problem will be consistent if e(n) is only corre-
lated with y(n) and uncorrelated with y(n − j) with j ∈ N0 . These assumptions
corresponds to looking forward in time. On the other hand the constraint ar = 1
corresponds with
Solving the least squares problem with this constraint implies that e(n) can only
be correlated with y(n − r) and uncorrelated with y(n − i) with i = 0, . . . , r − 1
in order to be consistent and this corresponds to looking backward in time. Since
the difference between both constraints corresponds to reversing the time axis,
it is clear that the natural frequencies for both constraints are the same and the
damping values differ from sign. Table 9.1 gives an overview of the poles in function
of the constraint and basis function.
zk X k = AXk + Wk (9.20)
Yk = CXk + Vk (9.21)
206 Chapter 9. The secrets behind clear stabilization charts
with Wk and Vk are zero mean circular complex independent and identically dis-
tributed noise sources. The output spectra can also be written as
−1
Yk = C (Izk − A) Wk + V k (9.22)
in which a matrix denominator of order 1 can be recognized as (Izk − A), with the
highest order coefficient of the basis function z fixed to the unity matrix. Similar as
for the AR model, the stochastic spectra can be described by a state-space model
with stable, unstable or mixed stable-unstable system poles. In chapter 7, it is
shown that the developed stochastic subspace algorithm estimates a state-space
model with stable poles, since the Taylor expansion (I − Az)−1 = I + Az + A2 z 2 +
A3 z 3 + . . . holds for |z| = 1 if |eig(A)| < 1, which is equivalent to stable system
poles. This is in close analogy with the AR model. The stabilization charts of
figures 7.7 and 7.9 for the test examples explained in paragraphs 7.8.3 and 7.8.4
illustrate that stable poles are identified.
A Monte Carlo simulation is done for 100 runs for both constraints a0 = 1 and
a9 = 1. Figure 9.2 shows the estimated damping values σ for both constraints in
the discrete Z domain and the Laplace domain. For all 100 runs the estimates
with constraint a10 = 1 result in two stable poles, the system poles p1 and p2 and
eight unstable extraneous (mathematical) poles. While in the constraint a0 = 1
208 Chapter 9. The secrets behind clear stabilization charts
1.5 4
1 3
−0.5 −1
−1 −2
(a) (b)
1.5
3
1
1
0
0
−0.5
−1
−1
−2
−1.5
(c) (d)
Figure 9.2: The poles estimated in a Monte Carlo simulation of an ARX process: (a),(b):
results for constraint a0 = 1 in discrete Z-domain and continuous Laplace domain; (c),(d):
results for constraint a9 = 1 in discrete Z-domain and continuous Laplace domain. (∗
estimated poles, × exact pole). It is clear that the constraint a9 = 1 allows an easy
classification between the system poles and extraneous poles based on the damping.
results for every run in ten stable poles. From this example, it is clear that the
distinction between the deterministic system poles and the stochastic, extraneous
poles is very easy based on the sign of the damping ratio, if one chooses the right
constraint. Equivalent results are obtained by the estimation of the coefficients of
an ARX model in the time-domain.
30 30
25 25
20 20
model order
model order
15 15
10 10
5 5
0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Freq. (Hz) Freq. (Hz)
(a) (b)
Figure 9.3: Combined frequency-domain subspace: (a) basis function zk ; (b) basis
function zk−1 (∗: stable, ×: unstable)
Example Consider the FRFs, defined by the state-space matrices given in the ap-
pendix of chapter 8, as primary data for the combined frequency-domain subspace
algorithm. The simulated FRFs consist of 512 spectral lines and their correspond-
ing zk values cover the upper half unit circle and the zk−1 the lower half unit circle.
Since for real-life measurement the models order is not a priori known a model
with 30 modes is estimated. Figure 9.3 compares the stabilization charts for the
combined subspace algorithm with both basis function zk and zk−1 . This clearly
illustrates that in the second case a distinction between mathematical and physical
poles can be made easily based on the sign of the damping.
Today MPE methods for modal analysis applications must be able to process large
data sets (e.g. several 100 outputs and several inputs) with a high modal density
in a short time. In spite of the fact that LS based identification algorithms as
the LSCF, p-LSCF (PolyMAX), LSCE are not consistent, they are often used and
preferred for there speed. Since all these methods are basically LS algorithms, the
choice of the constraint has the same influence as for the ARX model.
210 Chapter 9. The secrets behind clear stabilization charts
Solving these compact normal equations with the constraint an = 1, i.e. the
coefficient corresponding with z n , results in unstable stochastic poles according to
table 9.1 (so a clear stabilization diagram), while the parameter constraint a0 = 1
results in stable stochastic poles. The easy to interpret stabilization charts are
constructed in a fast manner by solving the compact normal equations as
−1
α[1:m−1] = D[1:m−1],[1:m−1] D[1:m−1],m (9.32)
The p-LSCF (PolyMAX) algorithm and its application was presented as a new
future standard for modal testing [48], [89], because of its speed and the clear
9.3. Application for Modal Parameter Estimation methods 211
stabilization charts produced by the algorithm. The p-LSCF, discussed and pre-
sented in more detail in chapter 5, fits a right matrix fraction model on the FRFs
in a least squares sense. Similar as for the LSCF algorithm a constraint must be
applied to the matrix coefficients of the denominator polynomial to remove the
parameter redundancy. Eqs. 5.15 and 5.16 solve the compact normal equations.
With the constraint An , i.e. the matrix coefficient of zkn , fixed to the unity matrix
INi , the mathematical poles which model the noise on the FRFs become unstable
and again clear stabilization charts are obtained.
with Aj ∈ R Ni ×Ni . These matrix coefficients are obtained by solving the equations
given by
hh(n)io A0 + hh(n − 1)io A1 + . . . hh(n − r)io Ar = 0 (9.34)
for different n and with hh(n)io = [ho1 (n) . . . hoNi (n)] and hoi (n) the impulse
response function between output o and input i at time instant n. The parameters
are obtained by solving the following set of No N finite difference equations in a
least squares sense.
hh(r)i1 ... hh(1)i1 hh(0)i1
.. .. ..
. . .
hh(r)iNo . . .
A0
hh(1)iNo hh(0)iNo ..
.. ..
. . .
= 0 (9.35)
. . .
Ar−1
hh(N )i1 . . . hh(N − r + 1)i1 hh(N − r)i1
.. .. Ar
. ..
. .
hh(N )iNo . . . hh(N − r + 1)iNo hh(N − r)iNo
In literature this least squares problem is classically solved with the constraint
A0 = INi leading to both stable physical and mathematical poles according to the
theory explained in this chapter. However by changing the constraint into Ar =
INi the LSCE algorithm is transformed into a MPE leading to clear stabilization
diagrams, since the mathematical poles are identified as unstable poles. In fact,
the LSCE algorithm is the time-domain counterpart of the poly-reference LSCF.
Both methods have the same properties and result in very similar stabilization
diagrams if the same parameter constraint is applied.
212 Chapter 9. The secrets behind clear stabilization charts
The clear stabilization charts are obtained by the fact that an intelligent choice
of the constraint/basis function results in the fact that the stochastic contribu-
tion in the measurements is modelled by unstable poles. Since the system poles
are assumed to be stable, this facilitates the distinction between the physical and
mathematical poles. As long as the dynamics between the deterministic contri-
bution and the stochastic contribution are separated, this does not influence the
estimation of the poles of the deterministic contribution. Unfortunately, in prac-
tical modal analysis experiments the dynamics of the deterministic and stochastic
contribution are often coupled, i.e. the noise model contains some poles that are
also present in the deterministic contribution. Consider an OMAX experiment,
since the structure is excited by both measurable forces and unmeasurable forces,
the deterministic and stochastic contribution in the responses have coupled dynam-
ics. Even in the case that FRFs are estimated from the IO data of this OMAX
experiment, the dynamics of the noise on the FRFs will be coupled with the dy-
namics of the deterministic contribution. In all these cases, where the stochastic
and deterministic dynamics are coupled, a contradiction will appear in the algo-
rithm, if one is interested to build a clear stabilization chart based on the damping
ratios. The chosen constraint/basis function tries to force the poles in the stochas-
tic contribution to be unstable, while the deterministic contribution tries to force
these coupled poles to be stable. Depending on the relative contribution in data
of the stochastic versus the deterministic part of the coupled physical pole, the
estimated pole will be biased i.e. pulled to the unstable region and the damping
will be underestimated. For measurements with high SNR (30-40dB) this effect
is negligible. However, for low SNR as e.g. for in-flight flutter measurements or
power spectral density functions, this effect can seriously affect the quality of the
estimated model.
constraint/basis function to obtain clear stabilization diagrams and the other con-
straint/basis function to obtain more accurate damping estimates. Afterwards
both sets of poles must be correlated (in many cases the frequency of the physical
poles are almost identically for both constraints/basis functions). Another possi-
bility for noisy measurements is to construct the stabilization chart by means of
the LSCF or p-LSCF and next further optimize these estimates by the frequency-
domain ML estimators. Notice that the estimators based on an optimization
algorithm as the ML estimators are independent of the constraint and as a result
their damping estimates will be unbiased.
Data driven output-only algorithms always need to estimate stable poles, since
they model a stochastic process. The advantage of a clear stabilization diagram
for OMA can only be used in the case that power spectra are used as primary data.
However, power spectra are also characterized by high noise levels (compared to
FRFs) and a 4-quadrant symmetry (see Eq. 2.62) and thus they contain both
stable and unstable physical poles. Nevertheless, in the identification it often
happens that only one of both is identified depending on the constraint. If the
constraint is chosen to obtain clear stabilization diagrams, the algorithm tend to
give more weight to the unstable poles and only the modes that are very clearly
present in the data will be identified as both a stable and unstable. As a result, only
the modes that are clearly present will appear as physical poles in the diagram.
Therefore it is advised to transform the power spectra in ’positive’ power spectra,
since in that case the physical poles will only appear as stable poles.
9.3.6 Remarks
A modal test on the door of a car is performed by measuring the vibration response
in 79 outputs by a scanning laser Doppler vibrometer, while the door was excited by
an electrodynamical shaker(cfr. the laser vibrometer measurements processed in
paragraph 8.5.4). Since this test setup is a single input case, the LSCF and p-LSCF
result the same estimates. The LSCF, LSCE and combined subspace algorithms
are compared for both constraints/basis functions. A frequency band from 40Hz to
120Hz, with a high modal density, is modelled. The model order for the different
identification algorithms is chosen in order to estimate 50 modes. Figures 9.4 and
9.5 show the stabilization diagrams and a synthesized FRF for the LSCF an LSCE
algorithms. It is clear that for the constraint a49 = 1 it is very easy to distinguish
the physical from the mathematical poles based on the damping value. Subspace
methods have the tendency to place their mathematical poles close to the physical
poles. This fact complicates the interpretation of the diagram. Therefore, changing
the constraint in the state-space modal results in easy to interpreted stabilization
charts. Figure 9.6 illustrates the influence of changing the constraint in frequency-
9.4. Application of experimental structural testing 215
45 45
40 40
35 35
30
model order
30
model order
25 25
20 20
15 15
10 10
5 5
0
50 60 70 80 90 100 110 50 60 70 80 90 100 110
Freq. (Hz) Freq. (Hz)
(a) (b)
−90 −90
−100 −100
Ampl. (dB)
Ampl. (dB)
−110 −110
−120 −120
−130 −130
−140 −140
−150
50 60 70 80 90 100 110 120 50 60 70 80 90 100 110 120
Freq. (Hz) Freq. (Hz)
(c) (d)
Figure 9.4: LSCF processing car door data: (a) stabilization diagram for constraint
a0 = 1 (∗: stable, ×: unstable); (b) stabilization diagram for constraint a49 = 1 (∗
stable, × unstable); (c) synthesized FRF for constraint a0 = 1 (×: measurement, full
line: model); (d) synthesized FRF for constraint a49 = 1 (×: measurement, full line:
model).
50
45 45
40 40
35 35
model order
model order
30 30
25 25
20 20
15 15
10 10
5 5
0
50 60 70 80 90 100 110 50 60 70 80 90 100 110
Freq. (Hz) Freq. (Hz)
(a) (b)
Figure 9.5: LSCE processing car door data: (a) stabilization diagram for constraint
a0 = 1 (∗: stable, ×: unstable); (b) stabilization diagram for constraint a49 = 1 (∗
stable, × unstable)
During a MIMO modal test a fully trimmed Porsche was excited in 4 different lo-
cations by shakers. The accelerations were measured in 154 DOFs spread all over
the car. The FRFs are estimated by the H1 method [72]. A total of 616 FRFs is
processed by the LSCF, p-LSCF and LSCE method. Figures 9.7, 9.8 and 9.9 com-
pare respectively the stabilization diagrams obtained by different constraints for
respectively the p-LSCF, LSCE and LSCF algorithm. It is clear that the choice of
the constraint has a major influence on the interpretation of the stabilization dia-
gram. Furthermore, it can be observed that both the LSCE and the p-LSCF result
in similar stabilization diagrams. The effect of forcing rank 1 residue matrices on
the measurements by the poly-reference LSCF clearly results in the identification
of more stable poles compared to the common-denominator based LSCF.
Flutter data is typically characterized by high noise levels caused by turbulence and
the limited amount of data. Typical SNR on the FRFs is between 5−20 dB. During
flight , the airplane is artificially excited by the flaps by injecting an excitation
signal in the fly-by-wire system. In this case, the accelerations were measured
at 13 locations on a military aircraft. Figure 9.10 compares the LSCF for both
constraints. The constraint a25 = 1 results in a clear stabilization chart, but 3
of the 9 physical poles (given by table 8.6) do not appear as stable poles. From
the synthesized FRF it is also clear that some damping ratios are underestimated.
On the contrary, the constraint a0 = 1 results in a better fit of the model on
the data, but the distinction between the mathematical and physical poles in the
9.4. Application of experimental structural testing 217
45 45
40 40
35 35
model order
model order
30 30
25 25
20 20
15 15
10 10
5 5
0 0
50 60 70 80 90 100 110 50 60 70 80 90 100 110
Freq. (Hz) Freq. (Hz)
(a) (b)
50
45 45
40 40
35 35
model order
model order
30 30
25 25
20 20
15 15
10 10
5 5
0 0
50 55 60 65 50 55 60 65
Freq. (Hz) Freq. (Hz)
(c) (d)
−90 −90
−100 −100
Ampl. (dB)
−110
Ampl. (dB)
−110
−120
−120
−130
−130
−140
−140
−150
(e) (f)
Figure 9.6: Combined subspace algorithm processing car door data: (a) stabilization
diagram for basis function zk (∗: stable, ×: unstable); (b) stabilization diagram for basis
function zk−1 (∗ stable, × unstable); (c) zoom on figure (a); (d) zoom on figure (b); (e)
synthesized FRF for basis function zk (×: measurement, full line: model); (f) synthesized
FRF for basis function zk−1 (×: measurement, full line: model).
218 Chapter 9. The secrets behind clear stabilization charts
60 60
50 50
40 40
model order
model order
30 30
20 20
10 10
0
5 10 15 20 25 5 10 15 20 25 30
Freq (Hz) frequency (Hz)
(a) (b)
−15 −15
−20 −20
−25
−25
−30
Ampl. (dB)
−30
Ampl. (dB)
−35
−35
−40
−40
−45
−45
−50
−50
−55
−55 −60
5 10 15 20 25 5 10 15 20 25 30
Freq. (Hz) Freq. (Hz)
(c) (d)
Figure 9.7: p-LSCF processing data of a fully trimmed car: (a) stabilization diagram for
constraint a0 = 1 (∗: stable, ×: unstable); (b) stabilization diagram for constraint a12 = 1
(∗ stable, × unstable); (c) synthesized FRF for constraint a0 = 1 (×: measurement, full
line: model); (d) synthesized FRF for constraint a12 = 1 (×: measurement, full line:
model)
Similar conclusions can be made for the combined subspace algorithm applied
for both choices of the basis function. The clear stabilization chart misses two
stable poles and the synthesized FRF clearly illustrates the bias on the damping
estimates.
9.5. Conclusions 219
60 60
50 50
40 40
model order
model order
30 30
20 20
10 10
5 10 15 20 25 5 10 15 20 25
Freq. (Hz) Freq. (Hz)
(a) (b)
Figure 9.8: LSCE processing data of a fully trimmed car: (a) stabilization diagram
for constraint a0 = 1 (∗: stable, ×: unstable); (b) stabilization diagram for constraint
a12 = 1 (∗ stable, × unstable).
The last example illustrates the influence of the constraint/basis function for both
the combined subspace algorithm in figure 9.12 and the LSCF algorithm in figure
9.13 starting from ’positive’ power spectra (output-only processing) of the Villa
Paso bridge. Similar as for the other examples, an easy interpretation of the
stabilization chart is facilitated by the right choice of the constraint/basis function.
9.5 Conclusions
40 40
35 35
30 30
model order
Ampl. (dB)
25 25
20 20
15 15
10 10
5 5
0 0
5 10 15 20 25 30 5 10 15 20 25 30
frequency (Hz) Freq. (Hz)
(a) (b)
−15 −15
−20
−20
−25
−25
Ampl. (dB)
Ampl. (dB)
−30
−30
−35
−35
−40
−40
−45
−45
5 10 15 20 25 30 5 10 15 20 25 30
Freq. (Hz) Freq. (Hz)
(c) (d)
Figure 9.9: LSCF processing data of a fully trimmed car: (a) stabilization diagram for
constraint a0 = 1 (∗: stable, ×: unstable); (b) stabilization diagram for constraint a50 = 1
(∗ stable, × unstable); (c) synthesized FRF for constraint a0 = 1 (×: measurement, full
line: model); (d) synthesized FRF for constraint a50 = 1 (×: measurement, full line:
model).
9.6 Appendix 1
Proof :
|zk − λ| |1 − Rei(ω−θk ) |
= (9.36)
|zk − λ△ | |1 − R1 ei(ω−θk ) |
9.6. Appendix 1 221
22
20 20
18
16
15
model order
model order
14
12
10 10
8
6
5
4
2
0
4 6 8 10 4 6 8 10
Freq. (Hz) Freq. (Hz)
(a) (b)
20 20
15
10
10
model order
Ampl. (dB)
0 5
0
−10
−5
−20
−10
−30 −15
−20
4 6 8 10 12 4 6 8 10
Freq. (Hz) Freq. (Hz)
(c) (d)
Figure 9.10: LSCF processing in-flight aircraft data: a) stabilization diagram for con-
straint a0 = 1 (∗: stable, ×: unstable); b) stabilization diagram for constraint a24 = 1 (∗
stable, × unstable); c) synthesized FRF for constraint a0 = 1 (∗: measurement, full line:
model); d) synthesized FRF for constraint a49 = 1 (∗: measurement, full line: model)
with R = eσ . Taking the complex conjugate of the denominator does not change
the amplitude, thus
|zk − λ| |1 − Rei(ω−θk ) |
= 1 (9.37)
|zk − λ△ | |1 − Rei(ω−θ k)
|
|1 − Q|
= 1 (9.38)
|1 − Q |
= |Q| (9.39)
= R (9.40)
20 20
model order
model order
15 15
10 10
5 5
4 5 6 7 8 9 10 11 4 6 8 10
Freq. (Hz) Freq. (Hz)
(a) (b)
25
20
20
15
15
10
10
Ampl. (dB)
Ampl. (dB)
5
5
0
0
−5
−10 −5
−15 −10
−20 −15
4 5 6 7 8 9 10 11 4 5 6 7 8 9 10
Freq. (Hz) Freq. (Hz)
(c) (d)
Figure 9.11: Combined subspace algorithm processing in-flight aircraft data: (a) stabi-
lization diagram for basis function zk (∗: stable, ×: unstable); (b) stabilization diagram
for basis function zk−1 (∗ stable, × unstable); (c) synthesized FRF for basis function
zk (×: measurement, full line: model); (d) synthesized FRF for basis function zk−1 (×:
measurement, full line: model).
9.7 Appendix 2
The set of equations for estimating the coefficients aj from an AR process is given
by
Y1 Y1 z1−1 . . . Y1 z1−r
a0 E1
Y2 Y2 z2−1 . . . Y2 z2−r a1 E2
.. .. .. .. = .. (9.41)
..
. . . . . .
−1 −r
YN YN zN . . . YN zN ar EN
Jθ = E (9.42)
9.7. Appendix 2 223
20 20
15 15
model order
model order
10 10
5 5
0 0
2 3 4 5 6 7 8 2 3 4 5 6 7
Freq. (Hz) frequency (Hz)
(a) (b)
105 110
100 105
100
95
95
90
Ampl. (dB)
Ampl. (dB)
90
85
85
80
80
75
75
70 70
65 65
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
Freq. (Hz) Freq. (Hz)
(c) (d)
Figure 9.12: LSCF algorithm processing operational bridge data: (a) stabilization
diagram for the constraint a0 = 1 (∗: stable, ×: unstable); (b) stabilization diagram for
the constraint a24 = 1 (∗ stable, × unstable); (c) synthesized FRF for the constraint
a0 = 1 (×: measurement, full line: model); (d) synthesized FRF for the constraint
a24 = 1(×: measurement, full line: model)
(w.p.1 = with probability one and N the number of spectral lines) holds for a
stable system under the constraint a0 = 1 and for an unstable system for the
constraint ar = 1. To prove this it is sufficient to show that
25 25
20 20
model order
model order
15 15
10 10
5 5
0 0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
Freq. (Hz) Freq. (Hz)
(a) (b)
105 105
100 100
95 95
90 90
Ampl. (dB)
Ampl. (dB)
85 85
80 80
75 75
70 70
65 65
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
Freq. (Hz) Freq. (Hz)
(c) (d)
Figure 9.13: Combined subspace algorithm processing operational bridge data: (a)
stabilization diagram for basis function zk (∗: stable, ×: unstable); (b) stabilization
diagram for basis function zk−1 (∗ stable, × unstable); (c) synthesized FRF for basis
function zk (×: measurement, full line: model); (d) synthesized FRF for basis function
zk−1 (×: measurement, full line: model)
Y1 z1−1 Y1 z1−r
...
Y2 z2−1 ... Y2 z2−r
J˜ = .. .. (9.46)
..
. . .
−1 −r
YN zN ... YN zN
9.7. Appendix 2 225
Thus
N
J˜H E zkj Yk∗ Ek
X
a.s. lim = (9.47)
N →∞ j
k=1
N
zkj (A(zk )−1 )∗ Ek∗ Ek
X
= (9.48)
k=1
N ∞
zkj
X X
= bn zkn EkH Ek (9.49)
k=1 n=0
−1 P∞
where the Taylor expansion (A(z)∗ ) = n=0 bn z
n
holds on the unity circle
(|zk | = 1) for A(z) having stable poles. Under the noise assumption, J˜jH converges
w.p.1 to its expected value (strong law of large numbers for independent and
identically distributed random variables [70])
a.s. lim J˜H E = lim E J˜jH (9.50)
N →∞ j N →∞
∞ N
!!
j+n
X X
= lim bn R zk (9.51)
N →∞
n=0 k=1
qN −j
and the coefficients |bqN −j | ≤ K|λ|max . Thus it holds that
∞
X ∞
X
N −j
| bqN −j RN | ≤ RKN |λ|max |λ|mN
max (9.53)
n=0 q=0
N −j 1
≤ RKN |λ|max (9.54)
1 − |λ|N
max
N −j
≤ O(N |λ|max ) (9.55)
which converges to zero for A(z) containing stable poles and limN →∞ . This con-
cludes the proof that the constraint a0 = 1 results in a strongly consistent estimate
of a stable solution of 9.4. In a similar way one can show that the constraint ar = 1
results in a strongly consistent solution with only unstable poles.
226 Chapter 9. The secrets behind clear stabilization charts
Chapter 10
Conclusions
Chapter 2 discusses the different parametric models and their relation to the
modal model. These mathematical models form the basis of the identification
algorithms developed in this thesis. The OMAX concept is introduced to obtain
a maximal data exploitation by considering the vibration response as the sum
of both a deterministic contribution from the measurable forces and a stochastic
contribution from the unmeasurable ambient forces. Based on the length of of
the observation window and the desired data reduction, the identification can
start from the raw Output/Input spectra or from Averaged Based Spectral (ABS)
functions, i.e. FRFs in the case of an EMA, power spectra in the OMA case and
both simultaneously for an OMAX approach.
227
228 Chapter 10. Conclusions
is to average the spectra of several adjacent time blocks of the time signals in
combination with the use of a Hanning window. By using only a few time blocks,
an initial estimate of the FRFs is obtained, which suffers less from leakage errors.
Next, an additional noise reduction is obtained by the application of a rectangu-
lar window. This rectangular time window applied on the initial estimates of the
Impulse Response Functions (IRF) or correlations further reduces the noise levels,
while the leakage errors can exactly be compensated in a parametric way by a
correction factor on the finally estimated modal participation factors. Using both
simulations and measurements, it is demonstrated that this approach outperforms
the classic one. In particular, the accuracy of the estimated damping ratios is
greatly improved. Finally, the use of ’positive’ power spectra and their benefits
for the OMA case are discussed.
tion is rather slow for large amounts of data, a faster implementation is proposed
based on a scalar weighting. Several experimental examples, illustrated the gain
of accuracy by using a right matrix fraction description compared to the common-
denominator models. Especially, for applications characterized by high damping,
a large improvement is obtained. In the case, the measurements on highly damped
structures are contaminated with noise, the use of the poly-reference ML estimator
is advised. Finally, the use and interpretation of a left matrix fraction description
in the OMAX framework is discussed.
data or power spectra, without the need for a Gauss-Newton optimization. The
combined interpretation for IO data considers the vibration responses as the sum of
a deterministic and stochastic contribution respectively related to the measurable
inputs and unmeasurable inputs, which fits in the OMAX framework. Finally,
this algorithm also allows to start from output-only spectra in combination with
an additional term to model the initial/final conditions to remove leakage errors.
From simulations and several experimental examples it is shown that this algorithm
is applicable for different test cases (EMA, OMA, OMAX) and for all possible
primary data (IO data, Output-only spectra, FRFs and power spectra).
step.
In this thesis, only uncertainty intervals on the estimated parameters are con-
sidered for the ML estimators, which consider all stochastic contributions in the
measurements as measurement noise (and thus they do not extract system infor-
mation from the stochastic contribution on the primary data). A topic for further
research, is therefore the determination of uncertainties on the estimated model
parameters by the use of a combined deterministic-stochastic subspace algorithm.
The uncertainty levels on the estimated poles can be used as a tool to make a
distinction between physical and mathematical poles. A simplified approximated
approach to estimate uncertainty levels is to consider the states obtained by the
oblique projection as exact and next the noise levels on the estimated system ma-
trices can be obtained from the least-squares problem given by Eq. 8.22. Finally,
the noise levels on the system poles are obtained from a sensitivity analysis of the
eigenvalue decomposition of the A matrix. The uncertainty on the modal param-
eters can then be used in modal applications such as for a weighting in modal
updating procedures [111].
Other identification algorithms based on e.g. a left and right matrix fraction
descriptions can be extended to take into account unmeasurable forces. In the
broad field of system identification, the recent evolutions for modelling non-linear
and time-varying systems can be further investigated for their application in me-
chanical engineering in the context of an OMAX framework. In [34] a recursive
stochastic subspace algorithm is presented based on simulations for its application
to in-flight tests. Further research for its applicability for real-life testing must be
done to show the practical robustness.
Finally, the proposed identification algorithms can be applied for other appli-
cations in mechanical engineering such as e.g. acoustical modal analysis, transfer
path analysis, processing of large scale optical measurements and for other engi-
neering fields, where frequency-domain system identification in general is useful
with in particular electrical and control engineering. Furthermore the usefulness
of the OMAX approach in other engineering fields can be investigated.
232 Chapter 10. Conclusions
Bibliography
[1] K. Godfrey , Editor. Perturbation signals for System Identification. Prentice-Hall Inter-
national series in acoustics, speech and signal processing, 1993.
[2] H. Akaike. Markovian representation of stochastic processes and its application to the
analysis of autoregressive moving average processes. Anuals of the Institute of Statistical
Mathematics, 26:363–387, 1974.
[3] H. Akaike. Stochastic theory of minimal realization. IEEE Transactions on Automatic
Control, 19:667–674, 1974.
[4] R. Allemang, D. Brown, and W. Fladung. Modal parameter estimation: a unified matrix
polynomial approach. In Proceedings of the 14st International Modal Analysis Conference,
February 1994.
[5] M. Basseville, A. Benveniste, M. Goursat, L. Hermans, L. Mevel, and H. Van der Auweraer.
Output-only subspace-based structural identification: From theory to industrial testing
practice. Journal of Dynamic Systems, Measurement, and Control, 123(1), 2001.
[6] D. Bauer. Subspace algorithms. In Proceedings of the 13th symposium on System Identi-
fication (SYSID 2003), Rotterdam (The Netherlands ), 2003.
[7] J.S. Bendat and A.G. Piersol. Engineering Applications of Correlation and Spectral Ana-
lysis. John Wiley, New York, 1980.
[8] P. Bishop and G. Fladwell. An investigation into the theory of resonance testing. Philos-
ophycal Transactions of the Royal Society of London, 255 A:1055:241–280, 1963.
[9] R.B. Blackman and J.W. Tukey. The measurement of power spectra from point of view of
communication engineering. Dover Publications, Inc., New York, 1958.
[10] M.J. Brenner, R.C. Lind, and D.F. Vorafeck. Overview of recent flight flutter research at
nasa dryden. NASA Technical Memorandum 4792, 1997.
[11] D. Brown, R. Allemang, R. Zimmerman, and M. Mergaey. Parameter estimation techniques
for modal analysis. SAE paper 790221, page 19, 1979.
[12] A.P. Burrows and J.R. Wright. Optimal excitation for aircraft flutter testing. Journal of
Airospace Engineering, 209, 1995.
[13] B. Cauberghe. The use of scanning laser vibrometry for in-line quality control. In Poly-
tec Users Meeting organized by Techpro Engineering BVBA and K. Peraer BVBA, Vrije
Universiteit Brussel, Brussels, Belgium, October 2003.
[14] B. Cauberghe and P. Guillaume. Frequency response functions based parameter identifi-
cation from short data sequences. In Proceedings of the 8th International Conference on
Recent Advances in Structural Dynamics, Southampton (UK), July 2003.
[15] B. Cauberghe, P. Guillaume, and B. Dierckx. Identification of modal parameters from
inconsistent data. In Proceedings of the 20th International Modal Analysis Conference
(IMAC20), Los Angeles, CA (USA), 2002.
233
234 Bibliography
[16] B. Cauberghe, P. Guillaume, and R. Pintelon. Frequency domain subspace algorithms for
the stochastic identification problem. Under review in, Automatica, 2003.
[17] B. Cauberghe, P. Guillaume, R. Pintelon, and P. Verboven. Frequency domain subspace
identification using frf data from arbitrary signals. Under review in, Journal of Sound and
Vibration, 2003.
[18] B. Cauberghe, P. Guillaume, R. Pintelon, and P. Verboven. Mimo state space frequency
domain system identification and model validation from arbitrary signals. Under review
in, IEEE Transaction on Automatic Control, 2003.
[19] B. Cauberghe, P. Guillaume, P.Verboven, and R. Pintelon. Frequency domain subspace
algorithms for the combined deterministic-stochastic identification problem. Under review
in, Automatica, 2004.
[20] B. Cauberghe, P. Guillaume, and S. Vanlanduit. Operational modal analysis using the har-
monic excitation of rotating machinery. In 22nd Benelux Meeting on System and Control,
Lommel (Belgium), 2003.
[21] B. Cauberghe, P. Guillaume, and P. Verboven. A frequency domain polyreference maxium
likelihood implementation for modal analysis. In Proceedings of the 22th International
Modal Analysis Conference, Dearborn (Detroit), January 2004.
[22] B. Cauberghe, P. Guillaume, P. Verboven, and E. Parloo. Modal analysis in presence of
unmeasured forces and transient effects. Journal of Sound and Vibration (JSV), 265:p.609–
625, 2003.
[23] B. Cauberghe, P. Guillaume, P. Verboven, E. Parloo, and S. Vanlanduit. A combined
experimental-operational modal analysis approach in the frequency domain. In Proceedings
of the 21th International Modal Analysis Conference (IMAC21), Kissemmee, FL (USA),
2003.
[24] B. Cauberghe, P. Guillaume, P. Verboven, Steve Vanlanduit, and E. Parloo. Frequency
response function based parameter identification from short data sequences. Accepted,
Mechanical Systems and Signal Processing (MSSP).
[25] B. Cauberghe, P. Guillaume, P. Verboven, Steve Vanlanduit, and E. Parloo. The secret
behind clear stabilization diagrams: The influence of the parameter constraint on the
stability of poles. Under review, Mechanical Systems and Signal Processing (MSSP).
[26] B. Cauberghe, P. Verboven, P. Guillaume, and S. Vanlanduit. Frequency domain sys-
tem identification of modal parameters from flight-flutter data. In Proceedings of the 6th
National Congress about Theoretic and Apllied Mechanics, Gent (Belgium), May 2003.
[27] B. Cauberghe, P. Verboven, S.Vanlanduit, P. Guillaume, and E. Parloo. The secret behind
clear stabilization charts in experimental modal analysis. In Proceedings of the 10th SEM
International Conference on Experimental and Apllied Mechanics, Costa Mesa(US), June
2004.
[28] K. De Cock, B. Peeters, H. Van der Auweraer A. Vecchio, and B. De Moor. Subspace
system identification for mechanical engineering. In Proceedings of the 25th International
Seminar on Modal Analysis, Leuven (Belgium), September 2002.
[29] G. Dimitriadis and J.E. Cooper. Flutter prediction from flight flutter test data. Journal
of Aircraft, 38(2):355–367, 2001.
[30] S.W. Doebling, C.R. Farrar, M.B. Prime, and D.W. Shevitz. Damage identification and
health monitoring of structural and mechanical systems from changes in their vibrational
characteristics: A literature review. Technical Report LA-13070-MS, Los Alamos National
Laboratory, 1996.
[31] D.J. Ewins. Modal testing: theory and practice. Research Studies Press, 1985.
[32] C.R. Farrar, T.A. Duffey, P.J. Cornwell, and S.W. Doebling. Excitation methods for bridge
structures. In Proceedings of the 17th International Modal Analysis Conference (IMAC17),
pages 1063–1068, Kissemmee (FL), USA, 1999.
Bibliography 235
[33] M.I. Friswell and J.E. Mottershead. Finite Element Model Updating in Structural Dynam-
ics. Kluwer Academic Publishers, 1999.
[34] I. Goethals, L. Mevel, A. Benveniste, and B. De Moor. Recursive output-only subspace
identification for in-flight flutter monitoring. In Proceedings of the 22th International
Modal Analysis Conference, Dearborn (Detroit, US), January 2004.
[35] I. Goethals and B. De Moor. Subspace identification combined with new mode selection
techniques for modal analysis of an airplane. In Proceedings of 13th Symposium on System
Identification, Rotterdam (Netherlands), August 2003.
[36] M.F. Green. Modal test methods for bridges: A review. In Proceedings of the 14th
International Modal Analysis Conference (IMAC14), pages 552–558, 1995.
[37] P. Guillaume. Identification of multi-input multi-output systems using frequency-domain
models. PhD thesis, Dept. ELEC, Vrije Universiteit Brussel, Belgium, 1992.
[38] P. Guillaume. Errors-in-variables identification techniques applied to modal analysis. In
Proceedings of Design Engineering Technical Conference (DETC), CA (USA), ASME,
1997.
[39] P. Guillaume. Frequency response measurements of multivariable systems using non-
linear averaging techniques. IEEE Transactions on Instrumentation and measurements,
47(3):796–800, 1998.
[40] P. Guillaume, L. Hermans, and H. Van der Auweraer. Maximum Likelihood Identification
of Modal Parameters from Operational Data. In Proceedings of the 17th International
Modal Analysis Conference (IMAC17), pages 1887–1893, Kissimmee, FL, USA, February
1999.
[41] P. Guillaume, R. Pintelon, and J. Schoukens. Robust parametric transfer function estima-
tion using complex logarithmic frequency response data. IEEE Transactions on Automatic
Control, 40(7), 1995.
[42] P. Guillaume, R. Pintelon, and J. Schoukens. Parametric identification of multivariable
systems in the frequency domain - a survey. In Proceedings of the 21st International
Seminar on Modal Analysis, Leuven (Belgium), September 1996.
[43] P. Guillaume, R. Pintelon, J. Schoukens, and I. Kollar. Crest-factor minimization using
nonlinear chebyshev approximation methods. IEEE Transactions on Instrumentation and
measurements, 40(6), 1991.
[44] P. Guillaume, J. Schoukens, and R. Pintelon. Sensitivity of rooots to errors in the coefficient
of polynomials obtained by frequency-domain estimation methods. IEEE Transactions on
Instrumentation and Measurement, 38(6):1050–1056, 1989.
[45] P. Guillaume, P. Verboven, B. Cauberghe, and S. Vanlanduit. Frequency-domain system
identification techniques for experimental and operational modal analysis. In Proceedings
of 13th Symposium on System Identification, Rotterdam (Netherlands), August 2003.
[46] P. Guillaume, P. Verboven, and S. Vanlanduit. Frequency-Domain Maximum Likelihood
Estimation of Modal Parameters with Confidence Intervals. In Proceedings of the 23rd
International Seminar on Modal Analysis (ISMA23), pages 359–366, Leuven (Belgium),
September 1998.
[47] P. Guillaume, P. Verboven, S. Vanlanduit, and E. Parloo. Multisine excitations - new
developments and applications in modal analysis. In Proceedings of the 19th International
Modal Analysis Conference, Kissimmee (Florida), February 2001.
[48] P. Guillaume, Peter Verboven, S. Vanlanduit, H. Van der Auweraer, and B. Peeters. A
poly-reference implementation of the least-squares complex frequency domain-estimator.
In Proceedings of the 21th International Modal Analysis Conference, Kissimmee (Florida),
February 2003.
[49] L. Hermans, H. Van der Auweraer, M. Abdelghani, and P. Guillaume. Evaluation of
subspace identification techniques for the analysis of flight test data. In Proceedings of
the 16th International Modal Analysis Conference, Santa Barbara (California), February
1998.
236 Bibliography
[50] L. Hermans, H. Van der Auweraer, and P. Guillaume. A frequency-domain maximum like-
lihood approach for the extraction of modal parameters from output-only data. In Proceed-
ings of the 23th International Seminar on Modal Analysis, Leuven (Belgium), September
1998.
[51] L. Hermans, H. Van der Auweraer, and P. Guillaume. A frequency-domain maximum likeli-
hood approach for the extraction of modal parameters from output-only data. Proceedings
of the 23rd International Seminar on Modal Analysis (ISMA23), pages 367–376, 1998.
[52] W. Heylen, S. Lammens, and P. Sas. Modal Analysis Theory and Testing. K.U.Leuven,
1998.
[53] B. Ho and R. E. Kalman. Efficient construction of linear state variable models from
input/output functions. Regelungstechnik, 14:545–548, 1966.
[54] S. Van Huffel. Recent advances in total least squares techniques and errors-in-variables
modeling. SIAM, Philadephia, USA, 1997.
[55] Jer-Nan Juang. Applied System Identification. Prentice-Hall, 1994.
[56] Jer-Nan Juang and R.S. Pappa. An eigensystem realization algorithm for modal parameter
identification and model reduction. Journal of Guidance, Control and Dynamics, 8(5):620–
627, 1985.
[57] Jer-Nan Juang and R.S. Pappa. Effects of noise on modal parameter identified by the
eigensystem realization algorithm. Journal of Guidance, Control and Dynamics, 9(3):294–
303, 1986.
[58] Jer-Nan Juang and R.S. Pappa. An eigensystem realization algorithm in the frequency do-
main for modal analysis. ASME Transactions on Vibrations, Acoustics, Stress, 110(24):24–
29, 1988.
[59] T. Kailath. Linear systems. Prentice-Hall, New Jersey, 1980.
[60] M. W. Kehoe. A historical overview of flight flutter testing. NASA Technical Memorandum
4720, 1995.
[61] C. Kennedy and C. Pancu. Use of vectors in vibration measurement and analysis. Journal
of the Aeronautical Sciences, 14(11):603–625, 1947.
[62] R. Kumaresan and D W. Tufs. Estimating parameters of exponentially damped sinusoids.
IEEE Transactions on Acoustics, Speech and Signal Processing, 30(6):833–840, 1982.
[63] F. Lembregts. Frequency domain identification for experimental multiple inputs modal
analysis. PhD thesis, dept. PMA, K.U.Leuven, Belgium, 1988.
[64] J. Leuridan, D. De Vis, H. Van der Auweraer, and F. Lembregts. A comparison of some
frequency response function measurement techniques. In Proceedings of 4th International
Modal Analysis Conference. SEM, 1986.
[65] R.C. Lind and M.J. Brenner. Robust flutter margin analysis that incorporates flight data.
NASA/tp-1998-206543, NASA, Dryden Flight Research Center, Edwards, CA, USA, 1998.
[66] K. Liu, R.N. Jacques, and D.W.Miller. Frequency domains structural systm identification
by observability range space extraction. Proceeding of the American Control Conference,
Baltimore, 1:107–111, 1994.
[67] L. Ljung. System Identification: Theory for the User. Prentice-Hall, 1999.
[68] LMS Int. LMS CADA-X Fourrier Monitor Manual, 1997.
[69] LMS Int. LMS PolyMax: a revolution in modal parameter estimation, LMS Newsletter,
November 2003, 2003.
[70] E. Lukacs. Stochastic Convergence. academic Press, New York, 1975.
[71] Viberg M. Subspace methods in system identification. In Proceeding of SYSID 94, vol-
ume 1, Copenhagen, Denmark, 1994.
Bibliography 237
[72] N.M. Maia and J.M. Silva. Theoretical and Experimental Modal Analylis. John Wiley &
Sons, 1997.
[73] T. McKelvey. Identification of State-Space Models from Time and Frequency Data. PhD
thesis, PhD thesis, Linköping University, 1995.
[74] T. McKelvey. Frequency domain system identification with instrumental variable based
subspace algorithm. In Proceedings of ASME Design Engineering Technical Conferences,
Sacramento, CA, USA, September 1997.
[75] T. McKelvey. Frequency domain identification. In Preprints of the 12th IFAC Symposium
on System Identification, Santa Barbara, CA, USA, June 2000.
[76] T. McKelvey. Frequency domain identification methods. Circuits Systems Signal Process-
ing, 21(1):39–55, 2002.
[77] T. McKelvey, H. Akcay, and M. Ljung. Subspace-based multivariable system identification
from frequency response data. IEEE Transactions on Automatic Control, 41(7):960–978,
1996.
[78] L. Mevel, M. Goursat, and M. Basseville. Stochastic subspace-based structural identifi-
cation and damage detection and localisation - application to the z24 bridge benchmark.
Mechanical Systems and Signal Processing, 17(1):143–151, 2003.
[79] B. De Moor. Mathematical concepts and techniques for modeling of static and dynamical
systems. PhD thesis, Dept. of Electrical Engineering, Katholieke Universiteit Leuven,
Belgium, 1988.
[80] E. Parloo. Application of Frequency-domain System Identification in the Field of Opera-
tional Modal Analysis. Ph.D. Thesis, Vrije Universiteit Brussel, Departement of Mechanical
Engineering, 2003.
[81] E. Parloo, B. Cauberghe, F. Benedettini, R. Aloggio, and P. Guillaume. Sensitivity-based
operational mode shape normalization: Application to a bridge. accepted for publication,
Mechanical Systems and Signal Processing (MSSP), 2004.
[82] E. Parloo, P. Guillaume, and B. Cauberghe. Maximum likelihood identification of non-
stationary operational data. Journal of Sound and Vibration, 268, 2003.
[83] E. Parloo, P. Verboven, P. Guillaume, and M. Van Overmeire. Sensitivity-based operational
mode shape normalization. Mechanical Systems and Signal Processing (MSSP), 16(5):757–
767, 2002.
[84] E. Parloo, P. Verboven, P. Guillaume, and M. Van Overmeire. Force identification by
means of in-operation modal models. Journal of Sound and Vibration (JSV), 262(1):161–
173, 2003.
[85] B. Peeters. System Identification and Damage Detection in Civil Engineering. PhD thesis,
Dept. of Civil Engineering, Katholieke Universiteit Leuven, Belgium, 2000.
[86] B. Peeters, H. Van der Auweraer, and P. Guillaume. The polyreference least-squares
complex frequency domain mathod: a new standard for modal parameter estimation?
under review, Shock and Vibration, 2003.
[87] B. Peeters and G. De Roeck. Reference-based stochastic subspace identification for output-
only modal analysis. Mechanical Systems and Signal Processing, 13(6):855–878, 1999.
[88] B. Peeters and G. De Roeck. Stochastic system identification for operational modal ana-
lysis: a review. ASME Journal of Dynamic Systems, Measurement and Control, 123:659,
2001.
[89] B. Peeters, P. Guillaume, H. Van der Auweraer, B. Cauberghe, P. Verboven, and J. Leuri-
dan. Automotive and aerospace applications of the polymax modal parameter estimation
method. In Proceedings of the 22th International Modal Analysis Conference, Dearborn
(US), January 2004.
[90] B. Peeters and G. De Roeck. Stochastic system identification for operational modal ana-
lysis: A review. Journal of Dynamic Systems, Measurements, and Control, 2002.
238 Bibliography
[91] Bart Peeters, Geert Lowet, Herman Van der Auweraer, and Jan Leuridan. A new procedure
for modal parameter estimation. Sound and Vibration, Janary 2004, pages 24–29.
[92] R. Pintelon. Frequency-domain subspace system identification using non-parametric noise
models. Automatica, 38:1295–1311, 2002.
[93] R. Pintelon, P. Guillaume, Y. Rolain, J.Schoukens, and H. Van hamme. Parametric iden-
tification of transfer functions in the frequency domain: a survey. IEEE Transactions on
Automatic Control, 39(11):2245–2260, 1994.
[94] R. Pintelon, P. Guillaume, Y. Rolain, J. Schoukens, and H. Van Hamme. Parametric
identification of transfer function in the frequency domain - a survey. IEEE Transcations
on Automatic Control, 39(11):2245–2260, 1994.
[95] R. Pintelon, P. Guillaume, G. Vandersteen, and Y. Rolain. Analysis, development and
applications of tls algorithms in frequency domain system identification. Proceedings of the
Second International Workshop on Total Least Squares and Errors-in-Variables Modeling,
pages 341–358, 1996.
[96] R. Pintelon and J. Schoukens. Identification of continuous-time systems using arbitrary
signals. Automatical, 33(5):991–994, 1997.
[97] R. Pintelon and J. Schoukens. Time series analysis in the frequency domain. IEEE Trans-
actions on Signal Processing, 47(1), 1999.
[98] R. Pintelon and J. Schoukens. System identification: A frequency domain approach. IEEE
Press, 2001.
[99] R. Pintelon, J. Schoukens, and G. Vandersteen. Frequency domain system identification
using arbitrary signals. IEEE Transactions on Automatic Control, 43(12), 1997.
[100] Y. Rolain, R. Pintelon, K.Q. Xu, and H. Vold. Best conditioned parametric identification
of transfer function models in the frequency domain. IEEE Transactions on Automatic
Control, 40(11):1954–1960, 1995.
[101] R. Ruotolo. A multiple-input multiple-output smoothing technique: Theory and applica-
tion to aircraft data. Journal of Sound and Vibration, 247(3):453–469, 2001.
[102] Jr. S. Lawrence Marple. Digital Spectral Analysis. Prentice-Hall, 1987.
[103] C.K. Sanathanan and J Koerner. Transfer function synthesis as the ratio of two complex
polynomials. IEEE Transactions Automatic Control, 9(1):56–58, 1963.
[104] P. Sas. Relation between acoustic intensity and modal deformation patterns of vibrating
structures. Academiae Analecta, 46(4), 1984.
[105] J. Schoukens, R. Pintelon, G. Vandersteen, and P. Guillaume. Frequency-domain system
identification using non-parametric noise models estimated from a small number of data
sets. Automatica, 33(6):1073–1084, 1997.
[106] J. Schoukens, Y. Rolain, J. Swevers, and J. De Cuyper. Simple methods and insights
to deal with nonlinear distortions in frf measurements. Mechanical Systems and Signal
Processing, 14(4):657–666, 2000.
[107] J. Schoukens, J. Swevers, R. pintelon, and H. Van der Auweraer. Excitation design for
frf measurements in the presence of nonlinear distortions. In Proceedings of International
Conference on Noise and Vibration Engineering, K.U.Leuven, Leuven, Belgium, Septem-
ber 2002.
[108] M. Scionti, J. Lanslots, Ivan Goethals, A. Vecchio, H. Van der Auweraer, B. Peeters, and
B. De Moor. Tools to improve detection of structural changes from in-flight flutter data.
In Proceedings of the VIII International Conference on Recent Advances in Structural
Dynamics, Southampton, UK, July 2003.
[109] G. De Sitter, B. Cauberghe, and P. Guillaume. In operational vibro-acoustic modelling of
cavities. In Proceedings of the 21th International Modal Analysis Conference, Kissimmee
(Florida), February 2003.
Bibliography 239
[129] S. Vanlanduit, P. Guillaume, and J. Schoukens. High spatial resolution modal parameter
estimation using a parametric mle-like algorithm. In Proceedings of the 23rd International
Seminar on Modal Analysis (ISMA23), Leuven (Belgium), 1998.
[130] S. Vanlanduit, P. Verboven, P. Guillaume, and J. Schoukens. An automatic frequency
domain modal parameter estimation algorithm. Journal of Sound and Vibration, 265:647–
661, 2002.
[131] P. Verboven. Frequency-domain System Identification for Modal Analysis. PhD thesis,
Dept. of Mechanical Engineering, Vrije Universiteit Brussel, Belgium, 2002.
[132] P. Verboven, B. Cauberghe, and P. Guillaume. Flite: Research results and conclusions.
Internal Report Y03/001.
[133] P. Verboven, B. Cauberghe, and P. Guillaume. Improved structural dynamics using ad-
vanced frequency-domain total least squares identification. under review, Computers and
structures.
[134] P. Verboven, B. Cauberghe, P. Guillaume, S. Vanlanduit, and E. Parloo. Modal parameter
estimation and monitoring for online flight flutter analyis. Journal of Sound and Vibration,
18(3):587–610, 2003.
[135] P. Verboven, B. Cauberghe, E. Parloo, S. Vanlanduit, and P. Guillaume. User-assisting
tools for a fast frequency-domain modal parameter estimation method. Mechanical Systems
and Signal Processing, 18:759–780, 2003.
[136] P. Verboven, B. Cauberghe, S. Vanlanduit, E. Parloo, and P. Guillaume. Modal parameter
estimation from input/output fourier data using frequency-domain maxiumum likelihood
identificaion. in press, Journal of Sound and Vibration.
[137] P. Verboven, P. Guillaume, B. Cauberghe, S. Vanlanduit, and E. Parloo. A comparison of
frequency-domain transfer function model estimator formulations for structural dynamics.
In press, Journal of Sound and Vibration.
[138] P. Verboven, P. Guillaume, B. Cauberghe, S. Vanlanduit, and E. Parloo. Frequency-domain
tls and gtls identification for modal analysis. in press, Journal of Sound and Vibration.
[139] P. Verboven, P. Guillaume, and M. Van Overmeire. Improved modal parameter estima-
tion using exponential windowing and non-parametric istrumental variables technique. In
Proceedings of the 19th International Modal Analysis Conference, San Antonio (USA),
February 2000.
[140] P. Verboven, P. Guillaume, S. Vanlanduit, and B. Cauberghe. Assessment of non-linear
distortions in modal testing and analysis of vibrating automotive structures. In Proceedings
of the 22th International Modal Analysis Conference, Dearborn (USA), January 2004.
[141] P. Verboven, E. Parloo, P. Guillaume, and M. Van Overmeire. Autonomous structural
health monitoring - part i: Modal parameter estimation and tracking. Mechanical Systems
and Signal Processing, 2003.
[142] M. Verhaegen. Identification of the deterministic part of mimo state space models given
in innovations from input-output data. Automatica, 30(1):61–74, 1994.
[143] P.D. Welch. The use of fast fourier transform for the estimation of power spectra: A
method based on time averaging over short modified periodograms. IEEE Trans. Audio
Electracoust., AU-15:70–73, 1967.
[144] H. Zeiger and A.J McEwen. Approximate linear realizations of given dimension via Ho’s
algorithm. IEEE Transaction on Automatic Control, 19(153), 1974.