PhDBCauberghe PDF

RSIT EIT
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
APPLIED FREQUENCY-DOMAIN SYSTEM

IDENTIFICATION IN THE FIELD OF EXPERIMENTAL
AND OPERATIONAL MODAL ANALYSIS
Bart CAUBERGHE
Promotor: Proefschrift ingediend tot

Prof. dr. ir. P. Guillaume het behalen van de academische
graad van doctor in de
toegepaste wetenschappen
May 2004
Bart CAUBERGHE – APPLIED FREQUENCY-DOMAIN SYSTEM IDENTIFICATION
IN THE FIELD OF EXPERIMENTAL AND OPERATIONAL MODAL ANALYSIS – May 2004

RSIT EIT
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
APPLIED FREQUENCY-DOMAIN SYSTEM

IDENTIFICATION IN THE FIELD OF EXPERIMENTAL
AND OPERATIONAL MODAL ANALYSIS
Bart CAUBERGHE
Jury: Proefschrift ingediend tot

Prof. dr. ir. G. Maggetto, voorzitter (ETEC, VUB) het behalen van de academische
Prof. dr. ir. J. Vereecken, vice-voorzitter (META, graad van doctor in de
VUB) toegepaste wetenschappen
Prof. dr. ir. P. Guillaume, promotor (WERK, VUB)
Prof. dr. ir. R. Pintelon, secretaris (ELEC, VUB)
Prof. dr. ir. W. Heylen (PMA, KULeuven)
Prof. dr. ir. B. De Moor (ESAT, KULeuven)
Dr. ir. B. Peeters (LMS International)
Prof. dr. ir. M. Van Overmeire (WERK, VUB),
Dr. ir. P. Verboven (WERK, VUB)
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.
Last but not least, I am greatly indebted to my family, especially my parents,

who have always supported and motivated me in an outstanding manner through-
out my studies and life. I will always be grateful for the inexhaustible faith they
had in me. And finally I would like to thank my girlfriend Sofie, who has more
contributed than she realizes, for her patience and valuable support. Thanks my
darling for all your warm love.
Brussels, May 2004
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
MSRE Mean Square Relative Error

OMA Operational Modal Analysis
OMAX Operational Modal Analysis with eXogenous inputs
PEM Prediction Error Method
RMS Root Mean Square
SDOF Single Degree of Freedom
SISO Single Input-Single Output
RMFD Right Matrix Fraction Description
SNR Signal to Noise Ratio
SVD Singular Value Decomposition
TLS Total Least Squares
UMPA Unified Matrix Polynomial Approach
WGTLS Weighted Generalized Total Least Squares
WNLLS Weighted Non-Linear Least Squares
XP Auto and Cross Power Spectra
Contents
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
2 Frequency-domain models for dynamical structures 11

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 The Modal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Common-denominator models . . . . . . . . . . . . . . . . . . . . . 13
2.4 State-space models . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5 Matrix fraction polynomial models . . . . . . . . . . . . . . . . . . 15
2.5.1 Left-Matrix Fraction Description (LMFD) models . . . . . 15
2.5.2 Right-Matrix Fraction Description (RMFD) models . . . . 17
2.6 Acceleration, Displacement, Velocity . . . . . . . . . . . . . . . . . 18
2.7 Continuous-time models and discrete-time models . . . . . . . . . . 18
2.8 Real versus complex models . . . . . . . . . . . . . . . . . . . . . . 20
2.9 Primary identification data . . . . . . . . . . . . . . . . . . . . . . 20
2.9.1 Input-Output measurements: a deterministic approach . . . 20
vii
viii Contents
2.9.2 Output-Only measurements: a stochastic approach . . . . . 23

2.9.3 Input-Output measurements: a combined deterministic-stochastic
approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3 Non-Parametric preprocessing steps 31

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2 Modal testing for EMA . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3 FRF estimation for EMA . . . . . . . . . . . . . . . . . . . . . . . 35
3.3.1 Arbitrary input signals . . . . . . . . . . . . . . . . . . . . . 35
3.3.2 Periodic input signals . . . . . . . . . . . . . . . . . . . . . 38
3.3.3 Decreasing the noise levels on FRFs . . . . . . . . . . . . . 38
3.4 Estimation of the power spectra for OMA . . . . . . . . . . . . . . 42
3.4.1 The correlogram approach . . . . . . . . . . . . . . . . . . . 42
3.4.2 The periodogram approach . . . . . . . . . . . . . . . . . . 44
3.4.3 The ’positive’ power spectra approach . . . . . . . . . . . . 45
3.5 Combined FRF and XP estimation for OMAX . . . . . . . . . . . 47
3.5.1 No structure-exciter interaction . . . . . . . . . . . . . . . . 48
3.5.2 Structure-Exciter interaction . . . . . . . . . . . . . . . . . 50
3.6 Simulations and measurement examples . . . . . . . . . . . . . . . 52
3.6.1 Experimental Modal Analysis . . . . . . . . . . . . . . . . . 53
3.6.2 Operational Modal Analysis . . . . . . . . . . . . . . . . . . 54
3.6.3 Combined Operational-Experimental Modal Analysis . . . . 60
3.6.4 Measurement examples . . . . . . . . . . . . . . . . . . . . 61
3.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.8 Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.9 Appendix 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4 Identification of common-denominator models 71

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.2 An extended parametric model for the H1 estimator . . . . . . . . 73
Contents ix
4.3 Weighted Least Squares Complex Frequency-domain (LSCF) esti-

mator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3.1 FRF driven identification . . . . . . . . . . . . . . . . . . . 75
4.3.2 Remarks on the extended LSCF estimator . . . . . . . . . . 81
4.3.3 IO data driven identification . . . . . . . . . . . . . . . . . 83
4.4 Maximum Likelihood identification . . . . . . . . . . . . . . . . . . 85
4.4.3 Noise intervals . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.5 Combined deterministic-stochastic identification . . . . . . . . . . . 89
4.6 Comparison between common denominator based algorithms . . . 92
4.7 Mixed LS-TLS, SK, IQML, WGTLS and BTLS estimators . . . . . 93
4.8 Simulation and Measurement examples . . . . . . . . . . . . . . . . 93
4.8.1 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.8.2 Measurement on an aluminium plate . . . . . . . . . . . . . 101
4.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.10 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5 Identification of right and left matrix fraction polynomial models107

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.2 Frequency-domain identification of RMFD
models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.2.1 Poly-reference Weighted Least Squares Complex
Frequency-domain estimator . . . . . . . . . . . . . . . . . 109
5.2.2 Poly-reference Maximum Likelihood Estimator . . . . . . . 111
5.2.3 Fast Poly-reference Maximum Likelihood Estimator . . . . 114
5.2.4 Fast Poly-reference IQML . . . . . . . . . . . . . . . . . . . 117
5.2.5 Poly-reference WGTLS and fast BTLS Estimator . . . . . . 117
5.2.6 RMF description for IO data . . . . . . . . . . . . . . . . . 119
5.2.7 From matrix coefficients to modal parameters . . . . . . . . 119
5.3 Left Matrix Fraction Description . . . . . . . . . . . . . . . . . . . 121
x Contents
5.3.1 Linear Least Squares estimator for IO data . . . . . . . . . 121

5.3.2 Linear Least Squares estimator for FRF data . . . . . . . . 123
5.3.3 From matrix coefficients to modal parameters . . . . . . . . 123
5.3.4 Data condensation . . . . . . . . . . . . . . . . . . . . . . . 124
5.4 Output-Only . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.5 Illustrating examples . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.5.1 Body-in-white . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.5.2 Fully trimmed car . . . . . . . . . . . . . . . . . . . . . . . 126
5.5.3 Villa Paso Bridge . . . . . . . . . . . . . . . . . . . . . . . . 128
5.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.7 Appendix: Confidence intervals on the estimated poles from the
p-ML estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6 Deterministic Frequency-domain Subspace Identification 135

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.2 Basic Frequency-Domain Projection Algorithm . . . . . . . . . . . 137
6.3 Starting from FRFs or power spectra . . . . . . . . . . . . . . . . . 139
6.4 A weighted frequency-domain projection algorithm . . . . . . . . . 139
6.5 Extended state-space model for initial and final conditions . . . . . 141
6.5.1 State-Space Model for Arbitrary Signals . . . . . . . . . . 141
6.5.2 Remarks on the extended state-space model . . . . . . . . . 142
6.5.3 A mixed non-parametric/parametric FRF estimator for val-
idation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.5.4 State-Space Model for FRF data . . . . . . . . . . . . . . . 144
6.6 Simulation and Measurement examples . . . . . . . . . . . . . . . . 145
6.6.1 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.6.2 Measurements on a subframe of a car . . . . . . . . . . . . 148
6.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.8 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7 Stochastic frequency-domain subspace identification 153

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
Contents xi
7.2 A first approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

7.3 A second approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
7.4 Connection to time domain stochastic subspace identification . . . 163
7.5 Geometrical Interpretation . . . . . . . . . . . . . . . . . . . . . . . 164
7.6 Final Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
7.7 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
7.8 Simulation and Measurement example . . . . . . . . . . . . . . . . 168
7.8.1 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.8.2 Measurement example: Subframe of an car . . . . . . . . . 170
7.8.3 Flight flutter testing . . . . . . . . . . . . . . . . . . . . . . 171
7.8.4 Villa Paso bridge . . . . . . . . . . . . . . . . . . . . . . . . 173
7.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
7.10 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
8 Combined frequency-domain subspace identification 177

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
8.2 Theoretical Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . 179
8.2.1 System description . . . . . . . . . . . . . . . . . . . . . . . 179
8.2.2 Main theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 180
8.2.3 From states to system matrices . . . . . . . . . . . . . . . . 182
8.2.4 Taking into account effects of transients and leakage . . . . 183
8.3 Practical Implementation . . . . . . . . . . . . . . . . . . . . . . . 183
8.4 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
8.5 Simulations and Real-life measurement examples . . . . . . . . . . 186
8.5.1 Monte Carlo Simulation . . . . . . . . . . . . . . . . . . . . 186
8.5.2 Flight flutter simulation . . . . . . . . . . . . . . . . . . . . 188
8.5.3 Flight flutter measurements . . . . . . . . . . . . . . . . . . 190
8.5.4 ABS-function driven identification . . . . . . . . . . . . . . 192
8.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
8.7 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
xii Contents
9 The secrets behind clear stabilization charts 199

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
9.2 Theoretical Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . 201
9.2.1 LS Identification . . . . . . . . . . . . . . . . . . . . . . . . 201
9.2.2 Influence of the constraint on an AR model . . . . . . . . . 201
9.2.3 Stochastic State-Space models . . . . . . . . . . . . . . . . 205
9.2.4 Extrapolation to ARX models . . . . . . . . . . . . . . . . 206
9.2.5 Combined deterministic-stochastic frequency domain sub-
space identification . . . . . . . . . . . . . . . . . . . . . . . 208
9.3 Application for Modal Parameter Estimation methods . . . . . . . 209
9.3.1 Least Squares Frequency-domain (LSCF) algorithm . . . . 210
9.3.2 Poly-reference Least Squares Frequency-domain (p-LSCF,
PolyMAX) algorithm . . . . . . . . . . . . . . . . . . . . . . 210
9.3.3 Least Squares Complex Exponential (LSCE) method . . . . 211
9.3.4 Frequency-domain subspace algorithms . . . . . . . . . . . 212
9.3.5 Coupled stochastic-deterministic dynamics . . . . . . . . . . 212
9.3.6 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
9.4 Application of experimental structural testing . . . . . . . . . . . . 214
9.4.1 Measurements on a car door . . . . . . . . . . . . . . . . . . 214
9.4.2 Measurements on a fully trimmed car . . . . . . . . . . . . 216
9.4.3 In-flight aircraft measurements . . . . . . . . . . . . . . . . 216
9.4.4 Villa Paso bridge . . . . . . . . . . . . . . . . . . . . . . . . 219
9.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
9.6 Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
9.7 Appendix 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
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
1.1 Research context
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.
More recently, system identification techniques were developed to identify the

modal model from the structure under its operational conditions from vibration
responses only. These techniques, referred as operational modal analysis (OMA)
or output-only modal analysis, take advantage of the ambient excitation as e.g.
wind, traffic and turbulence. During an EMA, the structure is often removed from
its operating environment and tested in laboratory conditions. The latter exper-
imental situation can differ significantly from the real-life operating conditions.
An important advantage of OMA is that the structure can remain in its normal
operating condition. This allows the identification of more realistic modal models
for in-operation structures.
In practice, performing a modal analysis typically consists of three basic steps:
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
excitation and eliminate the need for artificial excitation equipment.

2. Processing of the measured data and identification of a modal model : de-
pending on the length of measured time signals, the time signals can first
be converted into Averaged Based Spectral (ABS) functions, which serve
as primary data for the system identification algorithms. In the case of
frequency-domain system identification, the raw time domain data can be
transformed into Frequency Response Functions (FRFs) (EMA case) or auto
and cross power spectra (OMA case). Next, starting from the primary data,
the modal model can be estimated by using an appropriate parametric iden-
tification algorithm.
3. Validating the model : the extracted modal model must be assessed for its
physical representation of the dynamical behavior of the structure in the
studied frequency band. Therefore, the modal model must satisfy several
criteria, based on physical properties of the modal model.
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].
• Structural health monitoring and damage detection: given a reference model

of a healthy undamaged structure, a decision can be made on the struc-
tural integrity of that structure by comparing newly estimated models to
the reference one. The topic of structural health monitoring has received
a considerable amount of attention during the last few decades [30]. Apart
from the detection of damage, the information contained in the modal model
can also allow a localization and assessment of the structural defects. The
main disadvantage of these techniques is that other changes, such as bound-
ary conditions and/or environmental changes, can also produce changes in
the modal parameters that are of the same order of magnitude as those re-
sulting from the occurrence of damage. In this context, model analysis can
also be a tool for automated model based quality control [13].
• Load identification: is the inverse problem of response prediction. Given a

modal model of a structure, the idea consists in the identification of unknown
forces that gave rise to a specific measured response [84].
• Vibro-acoustics: the existence of vibro-acoustical coupling between the struc-

tural vibration and radiated noise is an important aspect during the design
process of for instance cars, airplanes, heavy machinery and control cabins,
in terms of comfort. Based on the modal parameters it is possible to com-
pute the sound intensity radiated from the vibrating structure without the
need of performing expensive acoustic experiments [104].
Nevertheless, the current evolution in mechanical engineering towards the use of

Computer Aided Design (CAD) like Finite Element Models (FEM) results in a
changing role for testing [113], [114], [112]. Today the optimization process in
product development is under strong pressure because of the competitive mar-
ket, increasing customers’ demands and by consequence the design cycle becomes
shorter in time. This results in an increasing use of simulations based on numerical
models to reduce the number of prototypes and expensive experiments. Still, test-
ing plays an important and evermore critical role, in every step of the development
process for target setting, bench-marking and model updating. The limitations of
FEM approaches lie in the long calculation times due to the need of huge model
sizes, required to properly describe complex structures (models with over 1 mil-
lion degrees of freedom) and accuracy limitations related to modelling damping,
non-homogeneous materials, structural joints, ... [114]. All this, together with the
decreasing expertise of the users, since EMA/OMA have been transferred from the
realm of the research experts to the product development workfloor [89], makes
that the demands for modal parameter estimation (MPE) algorithms still increase
in terms of accuracy, speed, automation and physical interpretation.
1.2. Focus, outline and organization of the thesis 5
1.2 Focus, outline and organization of the thesis
In this thesis, several dedicated extensions of existing frequency-domain modal

parameters estimation (MPE) algorithms and new frequency-domain algorithms
for the estimation of modal parameters from experimental measurements are pro-
posed. In general the proposed algorithms try to fulfill the requirements for new
MPE algorithms [113], i.e.
• Reduction of the test time.

• Allow maximal test data exploitation
• Increase the accuracy of the estimates
• Decrease the complexity of the analysis, allowing less-experienced staff to
process the data
• Extend the limits and ranges (large number of sensors, high damping, noisy
data, short data records, ...)
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
Instead of estimating the modal parameters directly from the measurements,

the proposed algorithms identify first mathematical polynomial or state-space
models. These mathematical models can be related to the modal parameters
in a next step. In chapter 2 different mathematical models, the modal model and
their relation is given. Next, the deterministic approach for the EMA case and
a stochastic approach for the OMA case are discussed, to introduce the OMAX
framework. Finally, a distinction is made between the data driven and Averaged
Based Spectral (ABS) function driven identification algorithms.
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
Modal Testing
Data driven ABS driven
Chapter 3
Mathematical models for EMA, OMA and OMAX
Chapter 2
Frequency-domain System Identification
Cost function based Subspace Algorithms
Common-denominator Deterministic identification

models Chapter 6
Chapter 4
Stochastic identification
Left and right matrix Chapter 7
fraction description
Combined deterministic-
Chapter 5 stochastic identification
Chapter 8
Construction of clear
stabilization charts
Chapter 9
Modal Applications
Figure 1.1: Organization of the text.
attention is paid to obtain the noise information on these ABS functions. It is

shown, that the use of a rectangular window, can reduce the noise levels and
errors introduced by leakage, if the final estimated parameters are corrected for
both time- and frequency-domain leakage. The idea of ’positive’ power spectra is
introduced to eliminate the 4-quadrant symmetry in case of an OMA. In this way,
MPE algorithms can start simultaneously from FRFs and positive power spectra
in an OMAX framework to estimate the modal parameters.
1.2. Focus, outline and organization of the thesis 7
During the last years, several MPE algorithms based on a common-denominator

(scalar polynomial fraction) model were developed in the Acoustics and Vibration
Research Group of the Vrije Universiteit Brussel for both EMA and OMA applica-
tions [131],[80]. Special attention was paid to the accuracy, calculation speed and
memory requirements. In chapter 4, these algorithms were generalized to start
from FRFs starting from arbitrary signals, without introducing bias errors from
leakage. Furthermore a combined stochastic-deterministic algorithm is proposed,
together with a discussion of the analogy between the algorithms starting from
FRFs as primary data and those starting directly from the input/output spectra
(data driven).
In chapter 5 a generalization of the fast common-denominator algorithms to

the identification of right and left matrix fraction description models is presented.
It is shown that for Multiple Input/Multiple Output (MIMO) measurements on
highly damped structures the new proposed maximum likelihood based algorithms
outperform their common-denominator counterparts. A least-squares based algo-
rithm and both a scalar and matrix implementation of a maximum likelihood
algorithm are developed, with special attention paid to speed up the algorithm
and to reduce the memory requirements.
Chapter 6 starts with the introduction of frequency-domain subspace algo-

rithms to identify state-space models from both FRFs and input/output spectra.
In this chapter an extension of the frequency-domain state-space model is discussed
to take into account the effect of initial and final conditions of the vibrating struc-
ture. In this way transient effects in the measured data are considered as an
extra input to estimate the modal parameters. Based on this extension a mixed
non-parametric/parametric estimation for FRFs is proposed for model validation.
Finally, this extension is applied to estimate state-space models from FRFs.
Until recently, frequency-domain subspace identification was limited to esti-

mate deterministic models from input/output or FRF data. In chapter 7, a
stochastic frequency-domain subspace algorithm is proposed starting directly from
output spectra. It is shown that is algorithm is closely related to the stochas-
tic time-domain subspace algorithms. This algorithm is very useful to estimate
stochastic models from a short data sequence in a frequency band of interest as it
is the case for operational in-flight aircraft tests.
Next, the extension to a combined deterministic-stochastic frequency domain

subspace algorithms is proposed in chapter 8. This algorithm results in consistent
estimates in presence of process and output noise on the primary data. Com-
pared to the cost function related algorithms proposed in chapters 4 and 5 the
combined deterministic-stochastic frequency-domain subspace algorithm is capa-
ble to estimate consistent parameters in a single step (no need for an optimization
algorithm), without requiring any a priori known noise information. Similar to
the inconsistent least-squares based estimators, the subspace algorithms can con-
struct a stabilization diagram in a fast way. This combined frequency-domain
subspace algorithm is closely related to its time-domain equivalent. In the case

input/output spectra serve as primary data the combined algorithm fits in the
OMAX framework, since the vibration responses are considered as the result of
both measurable and unmeasurable forces. Therefore, this algorithm gives good
results to process in-flight flutter measurements, where the airplane is simultane-
ously excited by both an artificially applied force and by atmospheric turbulence.
Nevertheless, the proposed algorithm also results in consistent estimates starting
from noisy FRFs and (’positive’) power spectra.
Finally, in chapter 9, the secret behind clear stabilization charts is revealed. By

many researchers it was noticed that the least-squares frequency-domain identifi-
cation algorithms results in very clear and easy-to-interpret stabilization diagrams.
It is now shown that this can be explained by the choice of the constraint on the
parameters to solve the least-squares problem. A smart choice of this constraint
results in the identification of stable physical poles modelling the deterministic con-
tribution and unstable mathematical poles modelling the stochastic contribution.
This property results in a simple distinction between the physical and mathemat-
ical poles based on the sign of the damping ratios and the start for an automatic
interpretation of the diagram. This key idea for a clear distinction between the
physical and mathematical poles based on the damping, is closely related to the
choice of the basis functions in frequency-domain models and to feed-forward/
feed-backward time-domain identification. It is shown that this property also can
be used for the implementations of frequency-domain subspace algorithms and
other well-known time-domain algorithms.
1.3 Original contributions in this work
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:
• The introduction of a combined EMA/OMA framework as Operational Modal

Analysis with eXogenous Inputs (OMAX) which considers the vibration re-
sponse as a combination of a contribution caused by measurable and unmea-
surable forces [22], [23].
• The use of ’positive’ power spectra to eliminate the 4-quadrant symmetry in
the poles of power spectra in combination with its compensation for time-
domain leakage. The noise and data reduction for FRF data based on the
use of a rectangular window without the introduction of bias errors due to
leakage [20].
• The extension of FRF based common-denominator models to take into ac-
1.3. Original contributions in this work 9
count the initial/final conditions to prevent the introduction of bias errors

caused by leakage [24], [14]. This work was rewarded with the ’Best paper
award’ of the VIII International Conference on Recent advances in Structural
Dynamics (Southampton, UK, 2003) organized by the Institute of Sound and
Vibration Research.
• The presentation of a combined deterministic-stochastic algorithm based on

a common-denominator model, together with the close analogy between in-
put/output spectra driven and FRF driven algorithms [22], [23], [132].
• The development of a scalar and matrix implementation of the maximum

likelihood algorithm to estimate right matrix fraction models, resulting in
an increased accuracy for MPE from noisy, highly damped, MIMO tests [21],
[89].
• The extension of frequency-domain state-space models to include transient

and initial/final conditions. This resulted in an increased accuracy and an
enhanced mixed non-parametric/parametric validation tool based on FRFs
[18], [17].
• The development of a consistent stochastic frequency-domain subspace al-

gorithm for operational test data [16].
• The development of frequency-domain counterparts of the combined determi-

nistic-stochastic time-domain subspace algorithms. This algorithm has the
specific advantages of frequency-domain identification and results in con-
sistent estimates starting from input/output spectra, FRFs or (’positive’)
power spectra [19].
• The mathematical explanation behind the construction of clear stabilization

diagrams. The explanation of the importance of the constraint on the pa-
rameters, the basis functions and the time axis in identification algorithms
resulted in the capability of constructing clear stabilization diagrams for
several well-known algorithms [25], [27].
• Application of all these contributions for several experimental test cases in

automotive, civil and aerospace engineering [26],[132].
Chapter 2
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.
2.2 The Modal Model
Newton’s equations of motion for a finite-dimensional linear structure are a set of

Nm second-order differential equations, where Nm is the number of independent
degrees-of-freedom, given by
M ÿ(t) + C1 ẏ(t) + Ky(t) = f (t) (2.1)
with M , C1 and K ∈ R Nm ×Nm respectively the mass, damping and stiffness

matrices, f (t) ∈ R Nm ×1 the applied force and y(t) ∈ R Nm ×1 the structure’s
displacement. Using the Laplace transform and neglecting the initial conditions
results in the frequency-domain equivalent given by
Z(s)Y (s) = F (s) (2.2)
with the dynamical stiffness Z(s) = M s2 + C1 s + K and s = jω. Inverting Eq.

2.2 yields
Y (s) = H(s)F (s) (2.3)
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)
with Nm the number of modes. The poles λr = σr + iωrpcontain the natural

frequencies fr = ωr /(2π) and the damping ratios dr = −σr / σr2 + ωr2 . In real-life
applications the number of modes Nm differs from the number of measured output
degrees of freedom No and the number of input forces Ni .
2.3 Common-denominator models
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)
with the elements of the residue matrix Rr given by
Roi,r = lim (s − λr )Hoi (s) (2.10)

s→λr
= φor Lir (2.11)
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.
2.4 State-space models
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)
results in a state-space formulation given by
sX(s) = AX(s) + BF (s) (2.13)

Y (s) = CX(s) + DF (s) (2.14)
with

0 I
A = ,C= I 0 (2.15)
−M −1 K −M −1 C1

0
B = and D = [0] (2.16)
M −1
By using the auxiliary state vector X, given by

Y (s)
X(s) = (2.17)
sY (s)
Eq. 2.2 is transformed in a first order differential expression. Identification algo-

rithms based on a state-space model identify the system matrices A, B, C and D
from the measurement data. The transfer function matrix between the outputs
and inputs is then given by
−1
H(s) = C [sI − A] B+D (2.18)
2.5. Matrix fraction polynomial models 15
By considering the righthand eigenvectors V of the system matrix A defined by
AV = V Λ (2.19)
the state-space equations are transformed to their modal form

−1
H(s) = CV [sI − Λs ] V −1 B (2.20)
−1 T
∗
Λ 0 L
= φ φ sI − (2.21)
0 Λ∗ LH
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 .
2.5 Matrix fraction polynomial models
The common-denominator model (scalar matrix fraction description) can be con-

sidered as a special case of multivariable transfer function models described using
a Matrix Fraction Description (MFD), i.e. the ratio of two matrix polynomials
[59], [42]. This set of models can be divided in two set of models i.e. a left MFD
(LMFD) and a right MFD (RMFD). Based on the relationship between the LMFD,
RMFD and the common-denominator model, a so-called Unified Matrix Polyno-
mial Approach (UMPA) was proposed in [4] for comparison of different estimation
algorithms using a common mathematical framework.
2.5.1 Left-Matrix Fraction Description (LMFD) models
The LMFD models consider all input-output measurements simultaneously by the

following model
H(s) = A−1 (s)B(s) (2.22)
with A(s) = Isn + An−1 sn−1 + . . . A0 a matrix polynomial with (No × No ) matrix
coefficients and B(s) = Bn sn +Bn−1 sn−1 +. . . B0 a matrix polynomial with square
(No × Ni ) matrix coefficients. This LMFD model can be obtained from the state-
space model. The denominator coefficients are obtained by solving the following
linear set of equations [2]:
CAn−1
 
 .. 
.  = −CAn

An−1 . . . A1 A0  (2.23)
 
 CA 
C
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)
Pre-multiplying by [ A0 . . . An−1 I ] yields

A0 . . . An−1 I Hn = A0 . . . An−1 I ΓIn (2.26)
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
denominator model with a denominator polynomial of high order n = 2Nm . For

the case of modal analysis the number of responses No is often larger than the
number Nm and therefore the number of response locations is often artificially
reduced by a data reduction steps [63].
Once the coefficients An are known, A(s) = 0 can be reformulated into a

generalized eigenvalue problem, resulting in nNo eigenvalues, yielding estimates
for the system poles λr and the corresponding left eigenvectors φr , corresponding
to the modal mode shapes [63]. The participation factors can be obtained from
the B coefficients [59] or from a least squares problem in a second step estimation
procedure by the Least Squares Frequency Domain (LSFD) method. Since both
the poles and participation factors are known, the FRFs are a linear function of
the mode shapes.
2.5.2 Right-Matrix Fraction Description (RMFD) models
The RMFD model is given by
H(s) = B(s)A−1 (s) (2.31)
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)
H T (s) = A−T (s)B T (s) (2.32)
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.
2.6 Acceleration, Displacement, Velocity
In a practical modal analysis experiment, the vibration response can be measured

as displacements, but also as velocities (laser vibrometer) or accelerations. Until
now, only displacement response measurements were considered. Consider the
general case were the No response measurements can be either accelerometers,
velocity or displacement transducers. In that case the observation equation is
given by
Y ′ (s) = Ca s2 Y (s) + Cv sY (s) + Cd Y (s) (2.33)
where Y ′ (s) are the spectra of the outputs; Ca , Cv , Cd ∈ N0 No ×No are the selection
matrices for the accelerations, velocities and displacement. These matrices contain
only zeros and a few ones and indicate which output is measured as an acceleration,
velocity or displacement. E.g. in the case that only accelerometers are used
Ca = I, Cv = 0 and Cd = 0. The corresponding state-space model is given by
sX(s) = AX(s) + BF (s) (2.34)
′
Y (s) = CX(s) + DF (s) (2.35)
with A and B defined by Eq. 2.15, 2.16 and C and D given by
C = Cd − Ca M −1 K Cv − Ca M −1 C1 and D = Ca M −1

(2.36)
Important to notice is that a modal analysis experiment, based on accelerometers,
requires a direct term D. For acceleration based tests, CB differs from the zero
matrix and thus by result the coefficients B2 and B1 in the LMFD model for
the case No = Nm differ from zero. Reducing the size of the matrix coefficients
and increasing the order of the polynomials finally results in a scalar common
denominator model. In the case one starts from accelerations both the numerators
and denominator in the common-denominator model have the same order.
2.7 Continuous-time models and discrete-time mod-

els
For time-domain models a distinction is made between continuous-time and discrete-

time models. Until now, all proposed models in this chapter are continuous-time
frequency-domain models. Unlike in time-domain, frequency-domain models allow
to identify continuous-time models from samples of the Fourier transforms of the
measured signals, where time-domain identification is restricted to discrete-time
models. The continuous-time state-space-model is given by
ẋ(t) = Ax(t) + Bf (t) (2.37)
y(t) = Cx(t) + Df (t) (2.38)
2.7. Continuous-time models and discrete-time models 19
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
sX(s) = AX(s) + BF (s) (2.39)

Y (s) = CX(s) + DF (s) (2.40)
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
with sk = i2πk/(N ∆t) and N the number of time samples.
Discrete-time models give the relation between discrete time samples yn =

y(n∆t) of continuous-time signals. In this thesis it is assumed that the sample
period ∆t is constant during the measurements. Under the Zero Order Hold
(ZOH) assumption, i.e. the input is piecewise constant over the sample period, a
continuous-time model converts to a discrete-time model [55]
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)
For this discrete-time model the frequency domain counterpart is given by
zk X k = Ad Xk + Bd Fk
Yk = Cd Yk + Bd Fk (2.45)
with zk = esk ∆t = ei2πk/N and Yk = Y (sk ), Fk = F (sk ), Xk = X(sk ). The

proposed continuous-time models i.e. the modal model, the common-denominator
model, the LMFD and RMFD models and state-space models all have their discrete-
time equivalent, which is of the same form but the basis functions sn are replaced
by z n . In the next chapters most of the proposed identification algorithms identify
discrete-time models in the frequency-domain.
2.8 Real versus complex models
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.
2.9 Primary identification data
In the previous paragraphs different types of mathematical models were proposed

and briefly discussed. In this section attention will be paid to the primary data
from which one starts to identify a mathematical model. A primary distinction
is made between the availability of input-output measurements or only output
measurements. Next, a distinction is made based on the amount of available data
samples resulting in data driven identification or averaged-based-spectral (ABS)
function driven identification. Finally, the concept of a combined operational-
experimental modal analysis approach is introduced as a generalization of input-
output and output-only based approaches.
2.9.1 Input-Output measurements: a deterministic approach
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
The application of system identification techniques for the identification of

modal models – natural frequencies, damping ratios, mode shapes and modal
participation factors – for linear time invariant mechanical structures is known as
Experimental Modal Analysis (EMA). Many textbooks give an extensive overview
of EMA and input-output modal parameter estimation method [31], [52], [72].
As mentioned in chapter 1, these experimentally determined modal models can
be used in a wide range of structural dynamics applications. Typical for modal
analysis, the system identification techniques must be able to deal with
• multiple inputs and multiple outputs (e.g. Ni = 3 and No = 100)

• high model orders (e.g. Nm = 100)
• very low and high damping (from 0.01% up to 10%)
• high modal density (close-coupled modes)
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
Figure 2.1: Deterministic Input-Output Model
I/O and FRF based Experimental Modal Analysis
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
of a typical modal analysis experiment a large amount of data is available and

some preprocessing of the data is recommended to reduce both the size of the
data set and the noise levels before starting the parametric identification of the
modal parameters.
Therefore in EMA applications it is common practice to reduce the amount of

data and the noise levels by using the Frequency Response Functions (FRFs) as
primary data instead of the input and output spectra. These FRFs are estimated in
a non-parametric preprocessing step. The measured forces and vibration responses
are divided in different data blocks in order to be averaged and estimate the FRFs.
In the case that only a limited data is available and averaging reduces the frequency
resolution below a critical value, it is advised to start the identification directly
from the raw input spectra and output spectra (IO data driven).
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.
• The common denominator model for FRFs is given by

B(Ωk )oi
Hoi (ωk ) = (2.46)
A(Ωk )
with Ωk = sk for continuous-time models and Ωk = zk for discrete-time
models and Hoi (ωk ) the FRF between output o and input i. The IO based
version is given by
Ni
X Boi (Ωk )
Yo (ωk ) = Fi (ωk ) (2.47)
i=1
A(Ωk )
• The state-space model for FRFs is given by

Ωk Xk = AXk + B (2.48)
H( ω k ) = CXk + D (2.49)
n×Ni No ×Ni
with X ∈ C and Hk ∈ C . The IO based version is given by
Ωk Xk = AXk + BF (ωk ) (2.50)
Y (ωk ) = CXk + DF (ωk ) (2.51)
n×1 No ×1 Ni ×1
with Xk ∈ C , Y (ωk ) ∈ C and F (ωk ) ∈ C
• The LMFD and RMFD are similar as the common denominator model for
e.g. the LMFD for FRFs is given by
H(ωk ) = A−1 (Ωk )B(Ωk ) (2.52)
while the IO based version is given by
Y (ωk ) = A−1 (Ωk )B(Ωk )F (ωk ) (2.53)
2.9.2 Output-Only measurements: a stochastic approach
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
Figure 2.2: Stochastic Output-Only Model
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 which directly start the

identification from the output time sequences or output spectra.
• Correlation or power spectral density driven stochastic identification algo-
rithms which estimate in a first step the power spectra between the outputs
and certain reference sensors. Next, from these functions a deterministic

model is extracted. This model is now related to the stochastic system.
Data-driven stochastic identification
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
Iyn + A1 yn−1 + . . . Ana yn−na = B0 en + B1 en−1 + . . . Bnb en−nb (2.54)
with Ai ∈ R No ×No and Bi ∈ R No ×No . The AR model can be considered as a

special case of an ARMA model with B1 = B2 = . . . = Bnb = 0. Their frequency-
domain counterpart is obtained by taking the Discrete Fourier Transform of Eq.
2.54
(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].
A second type of stochastic data driven models are stochastic state-space

models. Prominent references on stochastic state-space model identification by
subspace methods are [120] and their application to civil engineering [87]. The
stochastic time-domain state-space model is given by
xn+1 = Axn + wn (2.56)

yn = Cxn + vn (2.57)
In chapter 7 of this thesis a frequency-domain counterpart subspace algorithm is

developed to identify state-space models from output Fourier spectra only. The
frequency-domain data driven stochastic state-space model is given by
zk X k = AXk + Wk (2.58)
Yk = CXk + Vk (2.59)
with Wk and Vk correlated white noise, representing the unmeasurable ambi-

ent forces. All data driven stochastic identification algorithms are developed for
discrete-time models. In practice this implies that the frequency-domain coun-
terparts will only be theoretically consistent if the frequency band from DC to
Nyquist is processed simultaneously as will be discussed in more detail in chapter
7.
Correlation and auto/cross power density driven stochastic identifica-

tion
Power spectra driven stochastic identification algorithms can be considered as the

stochastic counterpart of the FRF-driven deterministic algorithms, since similar
as for FRF-driven identification, ABS functions are used as primary data. At the
same time the correlation driven algorithms can be considered as stochastic coun-
terparts of identification methods, which start from Impulse Response Functions
(IRF). In fact, both correlation and power spectra based identification methods
first estimate respectively correlation functions and power spectra between the
responses and certain reference response. In a next step a deterministic model,
related to the stochastic model, is fitted through these functions.
In this paragraph the modal decomposition of power densities and correlations is

briefly discussed and references for a more profound discussion are given. Accord-
ing to the modal theory of mechanical systems, the FRF matrix can be decomposed
as shown by Eq. 2.4, i.e.
−1 −1 H
H(ω) = φ iωINm − Λ LT + φ∗ iωINm − Λ∗

L
Nm
φr LTr φ∗ LH
X
= + r r∗ (2.60)
r=1
iω − λr iω − λr
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
be considered to be a constant matrix with respect to the frequencies. In [51], [80]

it is shown that by substituting Eq. 2.60 in Eq. 2.61 power spectra (XP) of the
outputs Syy (ω) evaluated at frequency ω can be modally decomposed as follows
Nm
φr KrT φ∗ K H Kr φTr Kr∗ φH
X
Syy (ω) = + r r∗ + + r
(2.62)
r=1
iω − λr iω − λr −iω − λr −iω − λ∗r
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
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.
2.9.3 Input-Output measurements: a combined deterministic-

stochastic approach
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
a combined approach, also known as a Operational Modal Analysis with eXoge-

nous inputs (OMAX) [45], was already introduced under the form of combined
deterministic-stochastic models in electrical and control engineering [122]. Never-
theless, in this engineering field, the noise model (i.e. the model for the stochastic
contribution) is usually not considered as a model describing the studied system
and thus no physical information is extracted from the noise model parameters.
Practical examples of OMAX situations are e.g. in-flight aircraft tests, where the
airplane is excited by both an artificially measurable excitation an unmeasurable
turbulent forces, civil structures excited simultaneously by a measurable drop-
mass or impact hammer and by unmeasurable ambient forces caused by traffic,
wind and seismic activity.
Similar to the purely deterministic and stochastic case combined algorithms can
be divided in data-driven methods and ABS function driven methods.
Data-driven based combined stochastic-deterministic identification
Data driven combined stochastic-deterministic identification can be considered

as data driven deterministic identification of which some inputs are replaced by
unknown white inputs. The combined frequency-domain state-space model is given
by
zk X k = AXk + BFk + Wk (2.64)
Yk = CXk + DFk + Vk (2.65)
From Eq. 2.66 the IO relationship is obtained:
h i
−1 −1
Yk = C (Izk − A) B + D Fk + C (Izk − A) Wk + Vk (2.66)
In chapters 4, 5 and 8 the data driven based combined identification is discussed

in more detail respectively for a common denominator model, a LMFD model and
a frequency-domain state-space model.
ABS-driven based combined stochastic-deterministic identification
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
This chapter presented different mathematical models to describe the dynamic

behavior of vibrating structures. Starting from the differential equations based
on the law of Newton, the modal model is introduced. The relationship between
the state-space, LMFD and RMFD, common denominator and the modal model
are briefly discussed. These models form the basis of the identification algorithms
presented in the next chapters of this thesis. A first distinction was made be-
tween deterministic models for EMA, stochastic models for OMA and combined
deterministic-stochastic models for an OMAX analysis. A second distinction is
made between data driven and ABS function driven identification algorithms. The
data driven methods for both EMA and OMA are preferred in case only a limited
amount of data is available, while the ABS function based algorithms require a
larger amount of data.
Small amount of data Large amount of data

MPE
(data driven) (ABS driven)
Inputs and outputs
Artificial forces
I/O data-driven FRF-driven EMA

EMA
OMAX
OMAX
Ambient forces
Outputs only
O-driven OMA XP-driven OMA
Figure 2.4: Overview of the EMA, OMA and OMAX approach in function of the
amount of data
Chapter 3
Non-Parametric
preprocessing steps
In this chapter, the non-parametric estimation of ABS functions is discussed in

more detail for the EMA, OMA and combined OMAX cases. Several methods
proposed in literature are presented. The use of a rectangular window to reduce
the noise levels on FRFs is introduced and the correction for time and frequency-
domain leakage is proposed. The concept ’positive’ power spectra is introduced to
eliminate the 4-quadrant symmetry in the poles of power spectra. The techniques
are illustrated and validated by a simulation example and real-life applications.
31
32 Chapter 3. Non-Parametric preprocessing steps
3.1 Introduction
In chapter 2, it was already mentioned that, depending on the time length of

the measured signals, the frequency identification algorithms start directly from
the spectra of the measured sequences i.e. data driven methods or from Aver-
aged Based Spectral (ABS) functions. In this chapter, more detailed attention is
paid to the non-parametric estimation of the ABS functions from sufficiently long
measured time sequences.
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.
Similar as for EMA, many OMA modal parameter identification algorithms

start from correlations or power density spectra between all the responses and a
few reference responses. In this chapter both the correlogram and periodogram
based power spectra estimators are briefly discussed [50], [80]. Next, it is shown
how the 4-quadrant symmetry of the poles in the power density spectra is reduced
to a 2-quadrant symmetry by the use of rectangular window on the correlations.
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
3.2 Modal testing for EMA
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
Modal analysis experiments are typically characterized by the large amounts of

data and therefore the use of FRFs as primary data is preferred for reasons of
data reduction, noise reduction and the combination of measurements in separated
patches. Furthermore, since the estimation of FRFs is based on an averaging pro-
cess the covariance matrix of the stochastic errors on the FRFs can be estimated
and be used as a weighting for the parametric identification. Depending on the ex-
citation device the EMA test can be categorized as a shaker or hammer excitation
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.
The vibration response is typically measured by accelerometers or by a scan-

ning laser Doppler vibrometer. Accelerometers are preferred to measure a three-
dimensional structure. Often the number of accelerometers or acquisition channels
is smaller than the number of response locations. In this case the structure is mea-
sured in several patches. Since every patch is measured separately, the noise on
the FRFs corresponding with different patches is uncorrelated. An advantage of
FRF-driven identification compared to IO-driven identification, is that the FRFs
corresponding to the different patches can easily be processed simultaneously to
estimate the modal parameters. Although one must be careful for data inconsis-
tencies between the FRFs caused by the mass-loading effect of the accelerometers
on the structure [118], [15]. The use of the scanning laser vibrometer avoids
mass loading problems, since this measurement is based on the Doppler effect of
a laser (light) beam and no contact with the structure is required. A scanning
laser vibrometer setup, can be considered as a multi-patch measurement, where
every patch consists of 1 response sensor. Since all the responses are measured
independently in time, the noise on the measurements for different responses is
uncorrelated. Special modal parameter identification procedures are proposed in
[126] to process high resolution scanning laser vibrometer measurements. Scan-
ning laser vibrometers are especially interested in the case of panel-like structures,
but less suitable to measure three-dimensional geometries.
3.3. FRF estimation for EMA 35
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.
3.3 FRF estimation for EMA
3.3.1 Arbitrary input signals
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
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
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.
3.3.2 Periodic input signals
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].
In the case a synchronized multi-excitation measurement setup is used i.e. the

external generator is triggered with the data-acquisition system, the so-called non-
parametric errors-in-variables estimator [38] Hev , based on cyclic averaging, is pro-
posed. Consider Ni linear independent inputs stacked as F = [F (1), . . . , F (Ni )] ∈
C Ni ×Ni and the resulting responses Y = [Y (1), . . . , Y (Ni )] ∈ C Ni ×Ni , for a Nb
number of periods. In practice, these signal assumptions can easily be realized by
applying Ni times, at Ni input locations periodic broadband excitation sequences,
e.g. multisines, by measuring Nb periods. The Hev FRF estimator is than given
by
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].
3.3.3 Decreasing the noise levels on FRFs
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
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:
• Since modal testing is characterized by a large data sets typically 100-1000

FRFs and large model orders the processing speed of the parametric identi-
fication is still an important issue in practice. A reduced data set certainly
results in a gain of process time for the parametric identification.
• A modal model is often validated by comparing the synthesized FRFs with
measured FRFs. Therefore, reducing the noise levels results in an easier
visual validation.
• Many well-known modal parameter estimators are not consistent and effi-
cient, e.g Least Squares Complex Exponential (LSCE), Least Squares Com-
plex Frequency (LSCF), PolyMAX and Eigenvalue Realization Algorithm
(ERA). Therefore, reducing the noise levels on the primary data is highly
desired in order to reduce the bias errors on the estimated parameters. More
advanced maximum likelihood identification algorithms use the covariance
matrix of the noise on the FRFs as a weighting in their cost function. Since
these algorithms require starting values, better convergence properties are
obtained if the initial values are estimated from FRFs with reduced noise
levels.
The reduced FRFs
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
(1 − λNr ) is introduced by so-called time-domain aliasing. Under the assumption

that H(ωk ) is obtained from periodic excitation and the use of the Hev estimator,
the FRF will be leakage free. In this case, identification methods formulated in
the time-domain (by starting from IRF, obtained by the IDFT) need to correct
the identified mode shapes or participation factors with this term 1 − λN r to avoid
the influence of time-domain aliasing.
Depending on the damping ratio and the frequency resolution of the FRFs, the
IRF contains low signal levels and high noise levels in its tail. Therefore, the DFT
taken from the first Nw samples of the IRF results in
NX
w −1
H r (ωk ) = hn zk−n (3.15)

n=0
Nm
φr LTr (1 − λN φ∗r LH ∗Nw

r ) r (1 − λr )
X w
= −1 N)
+ ∗z −1 ∗N )
(3.16)
r=1
(1 − λ z
r k )(1 − λ r (1 − λ r k )(1 − λ r
Nm
φr L′T φ∗r L′H
X
r r
= −1 + ∗ z −1
(3.17)
r=1
1 − λ z
r k 1 − λ r k
(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.
Comparing the exponential and rectangular window it must be noticed that

the use of the exponential window requires a correction of the damping values,
while the rectangular window requires a correction of the participation factors.
FRF
N m
φr Lr
H (ωk ) = + ...
Data and noise reduction

r =1 1 − λr zk
IDFT
IRF
Rectangular window
DFT
FRF
H (ωk ) = ∑
φr Lr (1 − λrN )
Nm w
+ ...
r =1 1 − λr zk (1 − λr )
N
Figure 3.2: Procedure to reduce the data and noise levels.
The covariances of the reduced FRFs
Consider the time window
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 zk = ei2πk/N , while the reduced FRF H r (ωk ) is given by

NX
w −1
H r (ωk ) = hn zkn (3.21)

n=0
with zk = ei2πk/Nw . For m = N/Nw ∈ N0 it follows that

′
H r (ωk ) = H (ωmk ) (3.22)
Since multiplying hn with wn corresponds with convoluting H(ωk ) with W (ωk ) =

PN −1 n ′
n=0 wn zk , the covariance on H is given by
CH ′ (ωk ) = |W (ωk )|2 ∗ CH (ωk ) (3.23)
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
CH r (ωk ) = CH ′ (mk) (3.24)
3.4 Estimation of the power spectra for OMA
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.
3.4.1 The correlogram approach
The correlogram approach starts by estimating the correlation functions. In the

next step, the correlation functions are transferred to the frequency-domain by
taking the DFT to obtain the power spectra. The unbiased discrete-time correla-
tion estimate between the response signals yn ∈ R No ×1 for (n = 0, . . . , N −1) and
3.4. Estimation of the power spectra for OMA 43
the reference output signals ynref ∈ R Nref ×1 is given by


N −n−1
1
yk+n ykref T
P
rn = for 0 ≤ n ≤ N − 1



 N −n

 k=0
(3.25)

 NP
−1
1
 rn = yk+|n| ynref T for − (N − 1) ≤ n < 0


N −|n|

k=|n|
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.
Nevertheless, the application of a time window introduces bias errors on the

final modal parameters [24]. In the case an exponential window is used, this bias on
the estimated parameters can be corrected which is not the case for other windows
e.g. Hanning, Hamming. The introduction of additional artificial damping by the
double sided exponential window given by
wn = e−β|n|∆t with − N ≤ n ≤ N (3.27)
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.
Another approach to estimate the correlation functions is based on dividing

the time sequences of the outputs in Nb equally sized adjacent blocks. In a next
step the correlation function can estimated for each block and finally averaged
over the different blocks. This has the advantage that the covariances of the noise
levels on the correlations can be obtained in a straightforward way.
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
additional block of N zeros. In other words, new blocks qb,n of length 2N are
created such that
qb,n = yb,n for 0 ≤ n ≤ N − 1 (3.28)

qb,n = 0 for N − 1 < n ≤ 2N − 1 (3.29)
Consider the DFT without applying a window

2N
X −1
Qb (ωk ) = qb,n zk−n (3.30)
n=0
and calculating the corresponding power spectra by

Nb
1 X
ref
Sqq (ωk ) = Qb (ωk )Qref,H
b (ωk ) ∈ C No ×Nref (3.31)
Nb
b=1
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.
3.4.2 The periodogram approach
A second widely-used method is the so-called periodogram power density estima-

tor, also known as the Welch estimator [102], [143]. The basic idea of the peri-
odogram estimator is to divide the response signals in Nb blocks of equal length.
Next, the Fourier spectra, for each block weighted with a time window wn , are
computed as
N
X −1
Yb (ωk ) = wn yb,n zk−n (3.34)
n=0
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
an overlap between the adjacent blocks in order to obtain a better contribution of

each raw time data sample is advisable. The power density spectra are given by
Nb
1 X ref H
ref
Syy (ωk ) = Yb,k Yb,k ∈ C No ×Nref (3.35)
Nb
b=1
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.
Variances on the Auto and Cross Power densities
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
When the stochastic process is not stationary, a technique to process operational

data is proposed in [83].
3.4.3 The ’positive’ power spectra approach
For the identification of modal parameters from output-only measurements several

frequency-domain identification methods start from the power spectra inspired by
expression Eq. 2.60 [50]. Nevertheless this technique has several disadvantages:
• 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.
• When using the periodogram approach to estimate power spectra a trade-

off must be made between the stochastic uncertainties and the bias errors
introduced by leakage.
Some of these disadvantages can be overcome be using time-domain identifica-

tion algorithms starting from the first Nw positive lags of the correlation function
rn given by
Nm
X
rn = φr KrT λnr + φ∗r KrH λnr for n ≥ 0 (3.39)
r=1
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 Kr′ = Kr (1 − λN r ) and zk = e

w i2πk/Nw
. Since Kr has no direct physical
interpretation and the mode shapes φr and poles λr are not affected, no correction
is needed. Similar as for the FRF Hr (ωk ), the variance is now given by
′
+ +
var(Syy (ωk )) = var(Syy (ωmk )) (3.43)
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:
1. Divide the time records in Nb adjacent blocks of N data samples. Extend

each block with an additional N zeros.
2. Based on these extended blocks, calculate the power spectra with the pe-
riodogram method. The inverse Fourier transform of these power spectra
results in the linear correlation functions.
+
3. The positive power spectra Syy and its variance is calculated by considering
only the first Nw positive lags of the correlation function.
3.5 Combined FRF and XP estimation for OMAX
An Operational Modal Analysis with eXogeneous (OMAX) inputs considers the

vibration response as a combination of a deterministic contribution caused by the
measurable forces and a stochastic contribution caused by unmeasurable forces.
As a result both the deterministic and stochastic contribution contain information
about the system.
In the classical EMA framework the stochastic contribution of the unmeasur-

able forces is considered as measurement noise. In fact no information from the
stochastic contribution in the vibration response is taken into account to estimate
the modal parameters. This is in contradiction with the approach followed in
the OMA framework, where all modal parameters are estimated from the purely
stochastic vibration responses. Therefore, in this paragraph both the contribution
of the deterministic part (i.e. the FRFs) and the stochastic part (i.e. the power
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.
Since an interaction between the structure and excitation device results in a

correlation between the measurable forces and unmeasurable forces a distinction is
made between experiments without structure-exciter interaction and experiments
with structure-exciter interaction.
m
For a set of measurable input forces F (ωk ) ∈ C Ni ×1 (Nim number of mea-
surable input forces) and a set of unmeasurable ambient random white forces
u
E(ωk ) ∈ C Ni ×1 (Niu number of unmeasurable input forces) the vibration response
spectra Y (ωk ) ∈ C No ×1 are given by

F (ωk )
Y (ωk ) = H(ωk ) (3.45)
E(ωk )
= H m (ωk )F (ωk ) + H u (ωk )E(ωk ) (3.46)
d s
= Y (ωk ) + Y (ωk ) (3.47)
m
with H m (ωk ) ∈ C No ×Ni the FRF matrix between the measurable inputs F (ωk )
u
and the outputs and H u (ωk ) ∈ C No ×Ni the FRF matrix between the unmeasur-
able forces E(ωk ) and the outputs. The deterministic and stochastic contribution
of the vibration response is respectively noted as Y d (ωk ) and Y s (ωk ). The mea-
surable forces F (ωk ) are given by
F (ωk ) = G(ωk )U (ωk ) + T (ωk )E(ωk ) (3.48)
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.
3.5.1 No structure-exciter interaction
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
H m (ωk ) = Syf (ωk )Sf−1

f (ωk ) (3.49)
1 −T
CH m (ωk ) = S (ωk ) ⊗ CY (ωk ) (3.50)
Nb f f
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
Sys yref,s (ωk ) = Syyref (ωk ) − H m Sf yref (3.51)
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.
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.
3.5.2 Structure-Exciter interaction
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
Sf f = GSuu GH + T See T H (3.53)

m H m u H

Syf = H GSuu G + (H T + H ) See T (3.54)
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.
An Instrumental Variable approach
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
generator has a linear behaviour)
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
H m (ωk ) = Syu (ωk )Sf u (ωk )−1 (3.57)
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
substituting Ybs = Yb − H m Fb results in
Sys yref,s = Syyref − H m,ref Sf yref − Syf H m,ref H

− H m Sf f H m,ref H
(3.59)
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.
Periodic input signals
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.
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
3.6 Simulations and measurement examples
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.
fe
f2 f7
f1
me
ms2 ms7
ms1 …
y1 y1 y7
Figure 3.4: Simulated 7 DOF system excited by an electrodynamical shaker
3.6.1 Experimental Modal Analysis
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
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.
3.6.2 Operational Modal Analysis
Traditional frequency-domain estimators for an OMA experiment start from auto

and cross power spectra. These power spectra are obtained by averaging over
different blocks and contain a 4-quadrant symmetry in the structures poles e.g.
both the stable and unstable poles appear in Eq. 2.63. This classic approach has
the disadvantage that it requires a double order. Furthermore, the periodogram
approach suffers from a bias introduced by leakage similar as for the H1 estimator
in the EMA case. During a simulation of 512s the structure was excited by two
uncorrelated gaussian distributed white noise sequences. The first force (normally
distributed noise N(0,1)) was directly applied to the 4th mass, while the second
force (N(0,0.4)) was applied directly to the 5th mass. Figure 3.8 gives a schematic
Nb=2 Nb=2, m=1

−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=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)
Nb=2 Nb=2, m=1

−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=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
standard deviation)
Nb=4, m=4 Nb=4, m=4

−60 −60
−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.
Block 1 Block 2 Block Nb
Zero padding
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
…
0 0 0
-0.2 -0.2 -0.2
-0.4 -0.4 -0.4
-0.6 -0.6 -0.6

10 20 30 40 50 60 10 20 30 40 50 60 10 20 30 40 50 60
time time time
DFT DFT DFT

50 50 50
Response
…
40 40 40
spectra
30 30 30
20 20 20
10 10 10
0 0 0
-10 -10 -10
-20 -20 -20

10 20 30 40 10 20 30 40 10 20 30 40
freq freq freq
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
0 1000 2000 3000 4000 5000
rectangular window
2000
1000
0
Window (first Nw
-1000 samples)
-2000
-3000
0 1000 2000 3000 4000 5000
DFT
100
90
80
‘Positive’ power spectra
Ampl (dB)
70
2-quadrant symmetry
60
50
40
0 10 20 30 40
freq
Figure 3.8: Procedure to determine the ’positive’ spectra

Nb=64 Nb=16, m=4

100 100
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=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).
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
3.6.3 Combined Operational-Experimental Modal Analysis
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:
• FRFs estimated by the classic H1 estimator with a Hanning window and

Nb = 128.
• FRFs estimated by the H1 estimator with Nb = 4 and m = 32.
• both FRFs and positive power spectra of the stochastic contribution with
Nb = 4 and m = 32.
• FRFs estimated by the H1 estimator with Nb = 8 and m = 16.
• both FRFs and positive power spectra of the stochastic contribution with
Nb = 8 and m = 16.
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.
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
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
3.6.4 Measurement examples
Subframe of car
The subframe shown in figure 3.1(b) was excited by an electrodynamic shaker

and the vibration responses are measured by 23 accelerometers during 64s with a
sample frequency of 2048Hz. Similar as for the simulation study the classic H1
approach is compared to the noise reduction obtained by the use of rectangular
window in figure 3.10. It is clear that the same reduction in data by the classical
approach results in biased FRF estimates and much higher uncertainty levels than
the FRF obtained from a rectangular window.
Villa Paso bridge
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
Nb=8 Nb=8, m=1
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
standard deviation)
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
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
495 500 505 510 515 520

Time (s)
(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
FRF Nb=8, m=8 FRF Nb=8, m=8

30 30
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
(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
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
Consider the sequence xn given by

xn = λn (3.76)
the DFT X(ωk ) is defined by

N
X −1
X(ωk ) = xn zk−n (3.77)
n=0
N
X −1
= (λzk−1 )n (3.78)
n=0
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
and thus it is proven that

N −1
λn 1 X zkn
= (3.84)
1−λ N N
k=0
1 − λzk−1
Substituting Eq. 3.84 in 3.75 results in

Nm n
λnr λ∗r
X
hn = φr Lr N
+ φ∗r LH
r (3.85)
r=1
(1 − λr ) (1 − λ∗N
r )
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
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 this chapter a generalization for common-denominator (see paragraph 2.3)

based algorithms is given by taking into account the initial and final conditions
for every block used by the H1 estimator [24]. Next, the strong analogy between
the FRF and IO data driven algorithms is illustrated [132] and finally a combined
deterministic-stochastic algorithm is proposed [22].
4.2. An extended parametric model for the H1 estimator 73
4.2 An extended parametric model for the H1 es-

timator
In chapter 3 several non-parametric FRF estimators were discussed with their

relation to the introduction of leakage errors. In practise special attention must
be paid to errors introduced by leakage especially for low-damped structures. It
is well-known that the use of a Hanning window reduces the effect of leakage by
suppressing the initial and final samples of each time block to zero. Although,
for short time blocks still significant errors are introduced, since no correction on
the final parameters is possible to compensate for the Hanning window. Therefore
the use of exponential and rectangular windows are proposed. These windows
introduce bias errors on the estimated FRFs, but the final parameters can be
corrected for this bias. Although both the use of an exponential and rectangular
window require an initial leakage free estimate of the FRFs e.g. from a small
number of blocks. In this chapter a different approach to deal with leakage is
proposed.
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
with zk = ej2πfk ∆t and

n
boi,j zkj
X
Boi (zk ) =
j=0
the numerator polynomial,

n
aj zkj
X
A(zk ) =
j=0
the common-denominator polynomial and

n−1
toj zkj
X
To (zk ) =
j=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].
For system identification algorithms starting from IO spectra the transient

polynomial can easily be modelled by considering an extra input. Eq. 4.1 can
be formulated for each data block b. The transient term coefficients tob,j will
depend on the initial and final conditions of the corresponding block b. Multi-
plying Eq. 4.2 by the hermitian of the input force spectra at frequency line k
[ F1 (zk ) . . . FNi (zk ) ]T and averaging over all blocks results in
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
Fbt (ωk ) = Fb (ωk )H SF−1F (ωk ) (4.5)
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
4.3 Weighted Least Squares Complex Frequency-

domain (LSCF) estimator
4.3.1 FRF driven identification
The model for the FRF between output o and input i estimated by the H1 esti-
mator is given by
Nb
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
The polynomial coefficients can be grouped together in one column vector θ

 
 η1 
 . 
 
 . 

.

θ= (4.10)
 ηNo 
 

 

α
 
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 denominator coefficients α are global parameters of the structure inde-

pendent of the output o, input i and data block b
• 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 )
Figure 4.1: LSCF FRF model (Nb = 1)
Equation Error and Cost Function:
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
which corresponds to solving
Jθ = 0 (4.12)
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
and the submatrices Γoi , ςoi and Ξoib

1
z1n
 
Woi (ω1 )
1 z1 ...
 
1
z2 . . . z2n


Woi (ω2 )
1 
Γoi = − (4.16)
 
.. 
.
 
 
1 n

Woi (ωN )
1 z N . . . zN
1
1 z1 . . . z1n
 
Woi (ω1 ) Hoi (ω1 )
 
1 n

 Woi (ω2 ) Hoi (ω2 ) 1 z2 . . . z2
 
ςoi =  (4.17)

.. 
.
 
 
1 n

Woi (ωN ) H oi (ω N ) 1 z N . . . zN
1
1 z1 . . . z1n
 t

Woi (ω1 ) Fbi (ω1 )
 
1 t n

 Woi (ω2 ) Fbi (ω2 ) 1 z2 . . . z2
 
Ξoib = − (4.18)

.. 
.
 
 
1 t n

F
Woi (ωN ) bi (ω N ) 1 z N . . . z N
Reduced normal equations
A modal analysis experiment is typically characterized by a large number of out-

puts and therefore special attention must be paid to reduce memory requirements
and processing time. The use of a common denominator model results in dedicated
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
with the submatrices Moi , Noi , Qoi , Ro , So and To given by
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
Forming the reduced normal equations
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
Further calculation and elimination of the parameters to leads to
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
Substituting Eq. 4.23 and Eq. 4.24 in Eq. 4.27 results in
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
the entries of these matrices equal

N
−2
X
Moi[r,s] = |Woi (ωk )| zks−r (4.30)
k=1
N
−2
X
Qoi[r,s] = − |Woi (ωk )| Hoi (ωk )zks−r (4.31)
k=1
N
−2 2
X X
To[r,s] = Ni |Woi (ωk )| |Hoi (ωk )| zks−r (4.32)
i=1 k=1
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.
Solving the reduced normal equations
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
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
To remove this parameter redundancy a constrained must be applied to the co-

efficients of the polynomials by e.g. fixing a coefficient to a constant value or
by fixing the L2 norm of the coefficients to 1. The least squares solution of the
reduced normal equations when fixing coefficient aj = 1 is given by
−1
α[1:j−1,j+1:n+1] = D[1:j−1,j+1:n+1],[1:j−1,j+1:n+1] D[1:j−1,j+1:n+1],1 (4.35)
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)
forces a L2 norm equal to 1 on the coefficients α. In chapter 9 the impact of the

constraint on the final solution is discussed in more detail with a special focus
on the quality of the stabilization diagram. In [137] it is shown that solving the
reduced normal equations by exploiting its predefined structure is much faster
than solving Eq. 4.12 directly (approximately No2 Ni2 ).
4.3.2 Remarks on the extended LSCF estimator
Continuous-time model
Instead of using a discrete-time domain model a continuous-time model can be

used by changing the basic functions zkj into sjk . A similar extended model to
take into account the influence of the initial and final conditions can be formu-
lated based on the results given in [96]. Although, special attention must be paid
to the numerical conditioning of the identification problem. Therefore the use of
orthogonal polynomials such as Forsythe and Chebyshev polynomials is advised
to improve the numerical conditioning [100], which were applied in the domain
of modal analysis by [117]. In [100] a numerical well-conditioned solution is dis-
cussed to transform the estimated coefficients in the orthogonal basis to the modal
parameters. The use of Forsythe polynomials has however the drawback that a
different set of polynomials must be generated for each individual FRF resulting in
a high computation time. This computation can be reduced by the use of Cheby-
shev polynomials, but for high model orders Nm = 50 the numerical condition
becomes still a problem since the polynomials are only approximately orthogonal.
A profound study and comparison between the use of continuous- and discrete-
time models for MIMO frequency domain identification is given in [137]. The use
of discrete-time models and an uniform frequency grid solves the numerical con-
ditioning problem, since the basic functions zkj are orthogonal with respect to the
unity circle. The use of the discrete-time model allows a fast and memory efficient
implementation based on the Toeplitz structure of the submatrices and the use of
the FFT algorithm to compute the matrix entries. Furthermore, it is shown in
chapter 9 that discrete-time models identified by a least squares algorithm under
the correct constraint result in a high quality of the stabilization diagram, which
is a major advantage in modal analysis applications. When using of discrete-time
model the frequencies in the frequency band of interest are mapped between 0Hz
and 1Hz. However, this introduces model errors which require a model order to
be chosen higher than the true one to obtain a good model fit. The use of bilinear
transformation allows the possibility to model a continuous-time system exactly by
a discrete-time model [73]. However, the transformation does not result anymore
in a uniform distribution of the zkj values over the unit circle and as consequence
the complex basic functions or not longer orthogonal. This results in both a worse
numerical conditioning and less clear stabilization charts. Although it should be
noted that for lower orders the bilinear transformation improves the quality of the
damping estimates.
With or without transient polynomials
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.
LSCF for operational and combined operational-experimental data
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
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:
• the mean of the estimated parameters from different experiments with N →

∞ differs from the exact parameters (asymptotically biased)
• the expected value of the estimated parameters for N → ∞ differs from the
exact one
• the uncertainty on the estimated parameters is larger the Cramr-Rao uncer-
tainty bounds
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].
4.3.3 IO data driven identification
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 )
which is schematically illustrated by Figure 4.2. The linearized least squares

equation error and cost function are then given by
Ni
X
LS
Eob (ωk ) = A(zk )Yo,b (ωk ) − Boi (zk )Fi,b (ωk ) − Tob (zk ) (4.39)
i=1
No X
X Nb
N X
LSCF LS
lIO (θ) = |Eob (ωk )|2 (4.40)
o=1 k=1 b=1
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
Eob (ωk )
?
1
A(zk )
Fib (ωk ) - Boi (zk )

A(zk )
@
R +?
@
@ h - Yob (ωk )

1 - Tob (zk )
A(zk )
Figure 4.2: LSCF IO model (Ni = 1)
2
Ni 1
0
0
5
Z
10
15
No 20
25
2 0
6 4
30 10 8
Nb
Figure 4.3: Data space visualization (e.g. Ni = 2, No = 30 and Nb = 10)
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.
4.4 Maximum Likelihood identification
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].
Equation Error and Cost Function:
Consider the common-denominator model for a measured FRF, disturbed by col-

2
ored independent circular complex noise with variance σH oi
(ωk ), given by
N
b
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
with the equation error given by

PNb
ML A(zk )Hoi (ωk ) − Boi (zk ) − b=1 Tob (zk )Fbit (ωk )
Eoi (ωk ) =
Woi (ωk )A(zk )
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 ML estimator is equivalent to a weighted non-linear least-squares (WNLLS)

with a weighting Woi (ωk ) = σHoi (ωk ). It is shown in [98] that independent of
the weighting Woi (ωk ) the WNLLS results in consistent estimates if the primary
data i.e. the FRFs are estimated in a consistent way (so no bias in presence of
noise). The choice of the weighting function only influences the efficiency of the
estimates i.e. the closer the weighting function agrees with the true variance on the
primary data the better the efficiency. To obtain the noise information sufficient
data must be available to average several blocks. However, in many cases the
engineer responsible for the modal analysis has only access to the FRFs. In this
case different alternatives for the weight Woi (ωk ) can be proposed:
• If no noise information is available, the assumption of white, additive noise

on the FRFs corresponds with a weight Woi (ωk ) = 1.
• 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
with SF−1ji defined by
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.
• The assumption of relative noise to the FRFs corresponds with a weight

Woi (ωk ) = |Hoi (ωk )|. This assumption gives in general good results in terms
of a fit of the FRFs considered in a logarithmic scale, since both resonances
and anti-resonances have the same weight in the cost function. In the case
of an operational modal analysis, the FRFs are replaced by power spectra
or ’positive’ power spectra and for these spectral functions as primary data
the assumption of relative noise is often in close agreement with true noise
conditions.
Gauss-Newton optimization
A Gauss-Newton optimization is applied to minimize the cost function given by

Eq. 4.42 by taking into account the quadratic form. The Gauss-Newton iterations
are given by
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
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
The normal equations J H J∆θ = −J H E have a similar the structure as in the

LSCF case. The left hand-side J H J is given by Eq. 4.20, Eq. 4.21 and Eq. 4.22,
while the right hand-side J H E is given by
" No
#T
X
H
J E = P1T ... PNTo WoT (4.49)
o=1
T T
T
VoT

Po = Uo,1 . . . Uo,Ni
(4.50)
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
and Eoi defined by

T
Eoi = Eoi (ω1 ) Eoi (ω2 ) . . . Eoi (ωN ) (4.54)
Finally, it can be shown that the submatrices in J H J have a Toeplitz structure

and the entries can be calculated in a time-efficient way using the FFT algorithm
4.5. Combined deterministic-stochastic identification 89
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.
4.4.3 Noise intervals
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.
4.5 Combined deterministic-stochastic identifica-

tion
The combined deterministic-stochastic common denominator estimator estimates

both the deterministic model and the color of the noise on the primary data i.e.
a polynomial describing the color of the noise. For IO data driven identification,
the algorithm can be interpreted in OMAX framework by assuming that the noise
present on the response measurements is caused by unmeasurable stochastic forces.
Eo (ωk )
?
Co (zk )
A(zk )
Fb (ωk ) - Bo (zk )
A(zk )
@
R +?
@
@ h - Yob (ωk )

1 - To (zk )
A(zk )
Figure 4.4: CLSF IO model
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
Boi (zk ) To,b (zk ) Co (zk )

Yo,b (ωk ) = Fb (ωk ) + + Eo,b (ωk ) (4.56)
A(zk ) A(zk ) A(zk )
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
A(zk )Yo,b (ωk ) − Bo (zk )Fb (ωk ) − To,b (zk )

Eo,b (ωk ) = (4.58)
Co (zk )
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.
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.
4.6 Comparison between common denominator based

algorithms
Different criteria are evaluated to draw a relative comparison between the different
frequency-domain estimators based on a common-denominator model.
• How fast are the estimators?

• How well can the estimators deal with process noise?
• How well can the estimators deal with measurement noise?
• How robust are the estimators?
• How do the estimators deal with short data sequences?
• Is the information in the data maximally exploited?
• How much data must be stored?
CRITERIUM LSCF IO MLCF IO CLSF IO LSCF FRF MLCF FRF CLSF FRF
computation speed +++ ++ + +++ ++ +

process noise −− ++ ++ − ++ ++
measurement noise −− +++ + − +++ ++
robustness (convergence) ++ ++ + ++ ++ ++
short data sequences ++ ++ ++ + + +
data exploitation − − ++ − − +
data storage (for Nb > Ni ) − − − + + +
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.
4.7 Mixed LS-TLS, SK, IQML, WGTLS and BTLS

estimators
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].
4.8 Simulation and Measurement examples
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.
4.8.1 Simulations
The extended LSCF estimator versus the classical LSCF estimator
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’.
• The second estimation procedure, denoted as ’LSCF Hanning’, is the most

often used approach. FRFs are derived by the H1 estimator with the use of
a Hanning window and the parameters are estimated by the classical LSCF.
• The third procedure, denoted as the ’Extended LSCF’ uses an H1 estimator

with no window and takes into account the initial and final conditions of
every block by considering the signals Fbit (ωk ).
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 (%)
Damping ratio (%)

−2
−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)
The combined stochastic-deterministic common denominator estimator:

Flight flutter testing
The benefits of using of combined stochastic-deterministic approach is shown in the

framework of the EUREKA-FLITE project, where new methods were developed
and applied to the analysis, validation and interpretation of aircraft structural
dynamics [132]. The main focus of the project is on the identification of modal
parameters from flight test data. Aircraft and winged-launch vehicles must be free
from aerodynamic instabilities such as flutter to ensure safe operation. Flutter is
a dynamic instability that results from the coupling of aerodynamic, elastic and
inertial forces acting on the structure. During flight the structure extracts energy
from the airstream. At speeds larger than the critical airspeed, the energy dissi-
pated by the available structural damping is less than the energy injected by the
airstream and the motion becomes divergent. Airworthiness regulations require
that a full-scale aircraft is demonstrated free from flutter by a flight flutter test.
In these tests, natural frequencies and modal damping ratios are estimated for
different flight conditions. Most common approaches track the damping ratios of
−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
(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
(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
Figure 4.7: Overview of modal analysis during flight flutter testing
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].
Today many research is focussed on operational modal analysis, where the

structure is excited by ambient noise excitation such as e.g. atmospheric turbu-
lence [90], [50] and [5]. Although these output-only methods have been used with
some success, there are several disadvantages. The turbulence is often not intense
enough and excites usually only the lower modal modes. Furthermore, output-only
methods need long data records to obtain accurate parameter estimates and flight
time is lost during the search for sufficient turbulence levels. [60].
Figure 4.8: Aircraft and the grid used for the GVT
To compare the CLSF IO estimator with other approaches simulations were

done to check for bias errors and to compare the efficiency. The simulations are
based on a modal model extracted from a ground vibration test (GVT). Figure
4.8 illustrates the actual airplane and the grid used for the GVT. The simulations
were done in the continuous-time with a state space model based on the first six
modes, visualized in Figure 4.9. In all the simulations one random noise force
f (t) is applied on a discrete point perpendicular to the surface of the left wing. A
sample frequency of 256Hz was used and 16K samples, corresponding to a total
measurement time of 64s, were simulated for each test.
To simulate the influence of the atmospheric turbulence, spatially correlated

white noise sources were acting on the nose, wings and tail of the aircraft. The level
of atmospheric turbulence is measured as the ratio of the RMS values between the
(a) (b)
(c) (d)
(e) (f)
Figure 4.9: Different mode shapes (from a to f: increasing natural frequency)
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
Damping ratio: mode 1 Damping ratio: mode 2

2 2.8
1.9
2.6
1.8
2.4
damping ratio (%)

damping ratio (%)
1.7
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

8 8
6 6
4
4
damping ratio (%)
damping ratio (%)

2
2
0
0
−2
−2
−4
−4
−6
−8 −6
−10 −8
1 2 3 4 5 6 0 1 2 3 4 5 6 7

2.7 3
2.6 2.5
2
2.5
damping ratio (%)
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
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
levels. The CLSF IO method is also compared to a ML IO data driven estimator,

under the assumption of white output noise. Although the ML IO estimator is
consistent under the assumption of only output noise, the results show a bias for
increasing noise levels. This can be explained by local minima. The CLSF IO
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
(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.
Other examples of the CLSF algorithm applied on measurements of a vibra-

ting plate excited by a measurable force and unmeasurable acoustic excitation are
presented in [22], while [109] illustrates the application for a coupled structural-
acoustical problem.
4.8.2 Measurement on an aluminium plate
An aluminium plate excited by a mini shaker was measured in 39 points with a

scanning laser vibrometer. To obtain very accurate reference measurements the
structure was excited by 10 periods of pseudo random noise (each period contained
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
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
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,
• leakage effects are present, since no periodic signals were used,

• only a very low frequency resolution is obtained
• the measurements are noisy, since only 2 blocks were averaged
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
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
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.
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
Taken the discrete fourier transform (DFT) defined by
N −1
1 X
Y (zk ) = DF T (y(r)) = √ y(r∆t)zk−r (4.60)
N r=0
with zk = ejk/N . Consider the DFT of y(r − 1)
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
since zkN = 1. The terms y(−1)−y(N√

N
+1)
are often neglected in frequency-domain
identification methods. It is clear that these terms appear due to the fact that
initial and final conditions are not equal. This is the case when non-periodic or
arbitrary excitation signals are used or when periodic measurements are corrupted
with transients. In the case purely periodic signals (y(−∆t) = y(N ∆t)) or time
limited signals (y(−∆t) = y(N ∆t) = 0) are considered these terms disappear.
Analog to Equation 4.61 we can show that for the more general case that
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
A(z)Y (z) = B(z)F (z) + T (z) (4.63)

4.10. Appendix 105
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].
Chapter 5
Identification of right and

left matrix fraction
polynomial models
This chapter presents several implementations of modal parameter estimation al-

gorithms based on a generalization of the fast common-denominator based algo-
rithms to identify right and left matrix fraction description models. It is shown
that, these algorithms results in more accurate modal parameters in the multiple
input case than their common-denominator counterparts. Both a poly-reference
least squares and poly-reference maximum likelihood estimator are proposed based
on a right matrix fraction model. Next, a fast maximum likelihood estimator is
developed, which uses a scalar frequency weighting. Finally a left matrix fraction
based least squares estimator is proposed for both IO data and FRF driven iden-
tification. It is shown that the IO data driven implementation can be interpreted
as a combined stochastic-deterministic estimator. The applicability of the matrix
fraction descriptions is shown by simulations and several real-life experiments and
a comparison is made with the common-denominator based algorithms.
107
108 Chapter 5. Identification of right and left matrix fraction polynomial models
5.1 Introduction
In the previous chapter the identification of common-denominator models was dis-

cussed starting from both FRF and IO data. In this chapter the identification of
both right and left matrix fraction polynomial models is introduced and discussed
resulting in several well-performing modal analysis estimators. It is shown that
a similar methodology to solve the identification problem in a time and memory
efficient way as developed in the previous chapter can be applied. A LS estimator
and mixed LS-TLS estimator is discussed, which fits a right matrix fraction de-
scription model trough the FRF data. This, so-called poly-reference LSCF is today
commercially known as the LMS PolyMAX estimator [91], [69]. Next, the ML es-
timator, which optimizes the poly-reference LSCF estimates by an Gauss-Newton
algorithm, is discussed. Unfortunately the initial ML implementation results in a
slower implementation and therefore a fast version is developed by using a scalar
frequency weighting. Finally, a LS estimator is proposed based on a left matrix
fraction description model and starting from IO data its combined interpretation
is discussed in the OMAX framework.
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
participation factors can be avoided. As a result, no loss of quality is encountered

from the conversion to the modal model and closely coupled poles can be sepa-
rated more easily. A disadvantage of a RMFD is that the model order can only
be increased in steps of Ni (the number of input forces) and that the method is
more sensitive to data inconsistencies due to e.g. mass loading effect over different
patches [15].
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.
5.2 Frequency-domain identification of RMFD

models
5.2.1 Poly-reference Weighted Least Squares Complex

Frequency-domain estimator
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
Ho (ωk ) = Bo (zk )A−1 (zk ) (5.1)
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
The linear LS equation error is obtained by right multiplication of Eq. 5.1 by

A(zk ) and taking the difference between the left and right part resulting in
Eo (ωk ) = Wo−1 (ωk ) (Bo (zk ) − Ho (ωk )A(zk )) (5.4)

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
which corresponds with solving

Jθ = 0 (5.7)
with J the Jacobian matrix given by
 
Γ1 0 ... 0 Υ1
 0 Γ2 ... 0 Υ2 
J = . .. .. .. (5.8)
 
 .. .. 
. . . . 
0 0 ... ΓNo ΥNo
with
Wo−1 (ω1 ) z1n
 
1 z1 ...
 Wo−1 (ω2 ) 1 z2 . . . z2n
 
 ∈ C N ×(n+1)

Γo = 

..  (5.9)
 . 
−1 n

Wo (ωN ) 1 zN . . . zN
−Wo−1 (ω1 ) 1 z1 . . . z1n ⊗ Ho (ω1 )
 
 −Wo−1 (ω2 ) 1 z2 . . . z2n ⊗ Ho (ω2 ) 

 
Υo = 

..

 (5.10)
 . 
−1 n

−Wo (ωN ) 1 zN . . . zN ⊗ Ho (ωN )
∈ C N ×Ni (n+1)
Note that the cost function Eq. is equivalent to
lp−LSCF (θ) = tr θH J H Jθ

(5.11)
Similar as in the common-denominator LSCF implementation, the number of fre-
quencies N can be eliminated from the equations dimensions by starting from the
normal equations given by
J H Jθ = 0 (5.12)
R1 ··· 0 S1
  
β1
 .. .. .. .. 
.

 . . . 
 = 0 (5.13)

 
 0 RNo S o βNo 
PNN

S1H ··· H
SNo
o
o=1 To α
(n+1)×(n+1) (n+1)×Ni (n+1)

with Ro = ΓH o Γo ∈ C , So = ΓH o Υo ∈ C and To =
H Ni (n+1)×Ni (n+1)
Υo Υo ∈ C . Elimination of the βo coefficients
βo = −Ro−1 So α (5.14)
results in the so-called reduced normal equations

No
X
To − SoH Ro−1 So α = M α = 0

(5.15)
o=1
with M a square Ni (n + 1) matrix. Similar as for the reduced normal equations

for the common-denominator model a constraint must be imposed to remove the
parameter redundancy. This can be done, for instance, by imposing that one of
the denominator coefficients is equal to the unity matrix INi e.g. for An = INi ,
for which the least squares solution is given by
−1
−M[1:nN ,1:nN ] M[1:nNi ,nNi +1:(n+1)Ni ]
α= i i (5.16)
INi
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
with i = [(r − 1)Ni + 1 : rNi ] j = [(s − 1)Ni + 1 : sNi ] for both r, s = 1, 2, . . . n + 1.

Taken into account the Toeplitz structure and using the FFT algorithm to calculate
the matrix entries reduces both the memory requirements and computation time.
The gain in calculation time, if 15log2 (N ) < 32n [131], by the use of the FFT
algorithm is typical less than for the common-denominator LSCF estimator, since
for the same number poles the model order n is Ni times smaller.
5.2.2 Poly-reference Maximum Likelihood Estimator
By taking the uncertainty information on the estimated FRF data into account
during the parametric identification, more accurate parameter estimates can be
obtained compared to the LS approach. The ML estimator minimizes the ML cost

function by a Gauss-Newton algorithm starting from initial parameters estimates
from the p-LSCF algorithm.
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 the equation error Eo (ωk ) defined by

Eo (ωk ) = Ho (ωk ) − Bo (zk )A−1 (zk ) (5.21)
and the covariance matrix of Ho (ωk ) defined by
Co (ωk ) = E ∆HoH (ωk )∆Ho (ωk )

(5.22)
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
 
and the Jacobian matrix J is defined by

 
Γ1 0 . . . 0 Υ1
 0 Γ2 . . . 0 Υ2 
J = . .. (5.24)
 
 .. .. 
. . 
0 0 ... ΓNo ΥNo
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)
The normal equations are given by

 
R1 ··· 0 S1 V1
  
 vec(∆β1 ) 
 .. .. .. .. ..

 

.
   
 . . . .
=  (5.27)
  
 
 0 RNo SNo  vec(∆βNo )  V
PNoNo
   
PNo  
S1H H
· · · SN T vec(∆α) o=1 Wo
 
o o=1 o
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 ))
Elimination of the βo coefficients
vec(∆βo ) = −Ro−1 (So vec(∆α) + Vo ) (5.29)
results in the so-called reduced normal equations

No
X No
X
To − SoH Ro−1 So vec(∆α) = − Wo − SoH Ro−1 Vo

(5.30)
o=1 o=1
Compared to the LS implementation, the dimensions of the reduced normal equa-

tions are increased by a factor Ni resulting in a Ni3 slower implementation for
each iteration. Similar as for the common-denominator ML estimator, the start-
ing values are generated by the poly-reference LSCF estimator and optimized by
the Gauss-Newton implementation of the poly-reference ML estimator. The use of
the predefined block Toeplitz structure of the submatrices Ro , So and To reduces
both the memory requirements and calculation time. Next, the uncertainties on
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.
Similar as for the common-denominator ML estimator a more robust logarith-

mic poly-reference ML estimator that is more suited to handle large dynamical
ranges is given by the following cost function
No
N X
lFp−ml,log
X
RF (θ) = Eo∗ (ωk )Clog,o (ωk )−1 EoT (ωk ) (5.31)
k=1 o=1
with the equation error Eo (ωk ) defined by

Eo (ωk ) = log (Ho (ωk )) − Bo (zk )A−1 (zk )

(5.32)
and with the covariance matrix defined by
 −H  −1
Ho1 0 ... 0 Ho1 0 ... 0
 0 H o2 ... 0   0 Ho2 ...0 
Clog,o =  . .. ..  Co  . .. .. (5.33)
   
. ..  .. .. 
 . . . .  . . . 
0 0 0 HoNi 0 0 0 HoNi
var(x)
in analogy with the scalar case (y = var(x) ⇒ var(y) = |x|2 and with the
submatrices of the Jacobian matrices given by
−Clog,o (ω1 )−1/2 1./ A−T (z1 )N T (z1 ) . ∗ [(Bo (z1 )A(z1 )−1 ) ⊗ A(z1 )−T ]Z 2 (1)
 
N

 i 

 . 
Υo = − . 
.
 
 
−Clog,o (ωN )−1/2 1./ A−T (z1 )N T (z1 ) . ∗ [(Bo (zN )A(zN )−1 ) ⊗ A(zN )−T ]Z 2 (N )
 
N
i
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)
IN 2 zkn INi zkn

with ZNi2 (k) = . . . IN 2 zk IN 2 and ZNi (k) = [
i i i
. . . INi zk INi ]
and the operators ./ and .∗ defined as respectively the element-wise division and
product. Based on this weighting and the vec operator both a poly-reference IQML
and BTLS can be proposed to solve the identification iteratively with a weighting
depending on the parameters estimated in the previous iteration.
5.2.3 Fast Poly-reference Maximum Likelihood Estimator
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
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
Eo (ωk ) = Wo (ωk )−1 (Bo (zk ) − Ho (ωk )A(zk )) (5.37)
with the parameter dependent scalar weighting Wo (ωk ) now defined by
Wo2 (ωk ) = E ∆Eo (ωk )∆EoH (ωk )

= E ∆Ho (ωk )A(zk )AH (zk )∆HoH (ωk )

= tr AH (zk )Co (ωk )A(zk )

(5.38)
Since the equation error is non-linear in the parameters a Gauss-Newton approach

is used to minimize the cost function. A necessary condition to be consistent is
that the expected value of the cost function is minimal in the exact parameters
[98], [94]. As a result, for the fast poly-reference ML this is shown by

H
No tr E
N X Boe (zk ) − HoH AH
e (zk )(ωk ) (Boe (zk ) − Ho (ωk )Ae (zk ))
p−f M L
X
E (lF RF (θe )) = (5.39)
k=1 o=1
Wo2 (ωk )
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
tr AH (zk )E ∆HoH (ωk )∆Ho (ωk ) A(zk )

+ (5.41)
Wo (ωk )2
No
N X H

X tr Eoe (ωk )Eoe (ωk )
= +1 (5.42)
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
Bo (z) the submatrix Γo is similar as for the poly-reference LSCF case

Wo−1 (f1 ) 1 z1 . . . z1n
 
 Wo−1 (ω2 ) 1 z2 . . . z2n

 
 ∈ C N ×(n+1)

Γo = − 

..  (5.44)
 . 
n
Wo−1 (fN ) 1 zN

. . . zN
The submatrices Υo need some more attention since the weighting depends on the
denominator polynomial A(z) and thus
 
1 z1 . . . z1n ⊗ ∂E∂A o (f1 )

 
1 z2 . . . z2n ⊗ ∂E∂A o (f2 )
 
Υo =   ∈ C N ×Ni (n+1) (5.45)
 
..
.
 
 
n
∂Eo (fN )
1 zN . . . zN ⊗ ∂A
with
∂Eo (ωk ) ∂W (ωk )−1 ∂(Ho (ωk )A(zk ) − Bo (zk ))
= (Ho (ωk )A(zk ) − Bo (zk )) + W −1 (5.46)
∂A ∂A ∂A
and
∂Wo (ωk )−1 1 ∂Wo (ωk )
= − Wo−3 (5.47)
∂A 2 ∂A
1 −3 tr ∂AH (zk )Co (ωk )A(zk )

= − Wo (5.48)
2 ∂A
1 −3
= − Wo Co (ωk )A(zk ) + CoH A(zk )

(5.49)
2
∂ (Ho (ωk )A(zk ) − Bo (zk ))

= Ho (ωk ) (5.50)
∂A
Specified for coefficient An the derivative is given by
∂Eo (ωk ) ∂Eo (ωk ) n
= zk (5.51)
∂An ∂A
The normal equations have a similar structure as for the matrix-weighted poly-
reference ML estimator given by Eq. 5.27 and as a result, they also can be reduced
to a compact formulation given by Eq. 5.29 and Eq. 5.30 with Ro = ΓH o Γo ∈
C (n+1)×(n+1) , So = ΓH o Υo ∈ C
(n+1)×(n+1)Ni
, To = ΥH o Υo ∈ C
Ni (n+1)×Ni (n+1)
,
(n+1)×1 Ni (n+1)×1
Vo = ΓHo E o ∈ C and W o = Υ H
o E o ∈ C with
EoT = [EoT (f1 ) EoT (f2 ) . . . EoT (fN )]. Compared to the poly-reference ML, the
submatrices become Ni times smaller in the dimension of the rows and columns
resulting in Ni3 faster algorithm, which can be significant in time for multiple input
test with e.g. Ni > 3. Similar as for the vec implementation, a logarithmic version
of the fast ML algorithm can be developed to increase the robustness.
5.2.4 Fast Poly-reference IQML
A fast IQML estimator based on the scalar weighting function

Wo2 (ωk ) = tr AH (zk )Co (ωk )A(zk ) works iteratively by minimizing in iteration j
the cost function given by
No X N
X (Ho (ωk )Aj (zk ) − Boj (zk )) (Ho (ωk )Aj (zk ) − Boj (zk ))H
lp−f IQM L = (5.52)
o=1 k=1
tr AH j−1 (zk )Co (ωk )Aj−1 (zk )
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.
5.2.5 Poly-reference WGTLS and fast BTLS Estimator
Consider the LS formulation in paragraph 5.2.1, which can be shortly denoted as

Jθ = 0. The (Weighted) Generalized Total Least Squares (WGTLS) solution for
this identification problem is then given by [54]
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
with Q = J H J. In practice the solution is found by means of a Generalized

Eigenvalue Decomposition (GEVD) that directly follows from the cost function
QV = λCJ V (5.56)
where the solution θ is given by the Ni eigenvectors corresponding to the Ni

smallest eigenvalues.
The poly-reference WGTLS implementation is in strong analogy with the

WGTLS implementation based on a common-denominator model [133], [138]. The
covariance matrix on the Jacobian matrix J has the same block structure as the
normal matrix M
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
and the subentries of the CTo are calculated by
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
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)
and substitution in last Ni (n + 1) rows of the GTLS problem results in a compact

GTLS formulation
No
X
To − SoH Ro−1 So α = λCT α

(5.62)
o=1
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
An improvement in efficiency is obtained by solving the WGTLS problem iter-

atively with the same frequency weighting Wo (ωk ) as the fast poly-reference IQML
solver based on the parameters α estimated in the previous iteration. This ap-
proach is called the fast poly-reference BTLS estimator and improves the accuracy
in each iteration, while each iteration results in consistent estimates.
5.2.6 RMF description for IO data
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)
Yo (ωk ) = Bo (zk )A−1 (zk )F (ωk ) (5.63)
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.
5.2.7 From matrix coefficients to modal parameters
The poles and participation factors are found from the denominator polynomial
coefficients Aj . The companion matrix Ac , build from the coefficients, is given by
[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 Vr is related to the rth modal participation vector according to

 n−1 
 λr L[:,r] 
..

 

 
Vr = . (5.66)
 λr L[:,r] 

 

L[:,r]
 
The poles λr , given by the eigenvalues of Ac , must be converted to the continuous

time domain and re-scaled according to the scaling procedure to cover the full
unit circle. The mode shapes Φ are obtained in a second step by considering the
frequency-domain formulation of the modal model. Once the physical poles are
distinguished from the mathematical ones by interpretation of the stabilization
chart, the mode shapes Φ, according to the a priori known poles and participation
vectors, can be estimated directly by the well-known Least Squares Frequency-
Domain (LSFD) estimator [118], [68]. This estimator estimates the mode shapes
and upper and lower residues in a linear least squares sense from
 
H1 (ω) Nm
φr LTr φr LH

..  X r LR
= + − 2 + UR (5.67)

 . jω − λ jω − λ ∗ ω
r=1 r r
HNo (ω)
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
5.3 Left Matrix Fraction Description
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.
5.3.1 Linear Least Squares estimator for IO data
The LMFD model between the outputs Yo and inputs Fi is given by
Y (ωk ) = A−1 (zk )B(zk )F (ωk ) (5.68)
with Y (ωk ) ∈ C No ×1 a column vector containing all outputs and F (ωk ) ∈ C Ni ×1 a

column vector containing all inputs. The low-order polynomials A(zk ) ∈ C No ×No
and B(zk ) ∈ C No ×Ni are defined by
n n
Aj zkj Boj zkj
X X
A(zk ) = Bo (zk ) = (5.69)
j=0 j=0
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
Y (ωk ) = A(zk )−1 B(zk )F (ωk ) + T (zk ) (5.70)
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
E(ωk ) = A(zk )Y (ωk ) − B(zk )F (zk ) (5.71)
and a corresponding cost function

N
X
lrmf d = tr E(ωk )E H (ωk )

(5.72)
k=1
The parameters are grouped together in θT = [β T αT ] with β T = [B0 . . . Bn ] and

αT = [A0 . . . An ]. The minimization of the cost function in a least squares sense
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
The normal equations J H Jθ = 0 can be solved by imposing a constraint to the

coefficients e.g. An = I. It should be noted that the number of estimated poles
equals Nm = nNo . In the case of a large number of outputs (i.e. No > Nm ),
this means that a huge number of mathematical poles are estimated. A possible
solution to prevent this is the use of a data condensation based on a SVD as
discussed in paragraph 5.3.4. A smart choice of the parameter constraint results
in stable physical poles and unstable mathematical poles as discussed in chapter
9.
In the theoretical case that the number of outputs equals the number of modes,
the Newton equation of motion is given by
Is2k + M −1 C1 sk + M −1 K Y (ωk ) = M −1 F (ωk )

(5.77)
and can be modelled in the frequency-domain with a discrete-time model as
A(zk )Y (ωk ) = B(zk )F (ωk ) + T (zk ) + E(ωk ) (5.78)
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.
A physical interpretation is that E(ωk ) represents unmeasurable forces, which

satisfy the noise assumption. This combined stochastic-deterministic point of view,
places the LMFD estimator in the OMAX framework as a combined IO data
driven estimator. Notice that in absence of measurable inputs, i.e. in the purely
5.3. Left Matrix Fraction Description 123
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.
5.3.2 Linear Least Squares estimator for FRF data
The LMFD model for FRFs is given by
H(ωk ) = A(zk )−1 B(zk ) (5.79)
with H(ω) ∈ C No ×Ni the FRF matrix. The the equation error is given by
E(ωk ) = A(zk )H(ωk ) − B(zk ) (5.80)
is minimized by solving Eq. 5.74 with
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 )
In theory an elimination procedure over the number of inputs can be used in a

similar way as the elimination over the number of outputs for the RMFD model
estimators. However, in practice the number of outputs exceeds the number of
inputs and the gain in calculation time by the elimination procedure in the case
of a LMFD is negligible.
5.3.3 From matrix coefficients to modal parameters
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.
5.3.4 Data condensation
In many industrial applications an experimental modal analysis is characterized

by several hundreds of output measurements, while the frequency band of interest
typically contains 10-30 modes. Direct identification using the LMFD LS estimator
would result in a long process time and many mathematical modes. Therefore, a
data condensation is used to reduce artificially the number of outputs. Consider
the output data matrix Yd = [Y (ω1 ) . . . Y (ωN )] ∈ C No ×N for the data driven
identification and Hd = [H(ω1 ) . . . H(ωN )] ∈ C No ×Ni N for the FRF driven,
the transformation matrix T is than given by the Nr first columns of U from
respectively the SVD of Yd = U SV H in the data driven case and of Hd = U SV H
in the FRF driven case. The reduced data matrices Yr and Hr are given by
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
This chapter focussed mainly on the identification of modal parameters starting

from FRF data. Only the LMFD LS estimators was discussed as a data driven
estimator. Nevertheless, all the FRF-estimators can start from power spectra or
’positive’ power spectra in the case of an operational modal analysis.
5.5 Illustrating examples
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
Figure 5.1: Modal test of a body-in-white
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
the rank of the residue matrices (The common-denominator model uses Ni × No

parameters to model each residue matrix, while the RMFD uses only No + Ni
parameters.). Nevertheless, it should be noticed that the overall differences for the
different algorithms are rather small for this data set. Nevertheless, transforming
Table 5.1: Comparison of model quality obtained by different algorithms for measure-
ments on a body-in-white
algorithm error correlation error correlation

polynomial model modal model
p-ML 1.4% 98.7% 2.1% 98.04%
ML 0.7% 99.3% 2.9% 97.4%
p-LSCF 2.7% 97.4% 2.5% 97.6%
LSCF 5.8% 95.7% 3.24% 96.9%
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.
5.5.2 Fully trimmed car
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 5.2: Natural frequencies and damping ratios estimated by p-ML
frequency (Hz) damping ratio (%)

40.17 1.36
41.46 0.60
43.31 0.65
43.69 0.45
44.66 0.43
45.91 0.59
49.96 0.65
50.55 0.72
53.73 0.18
54.77 0.42
58.84 1.27
60.12 0.11
60.63 0.25
61.27 0.29
ments on a fully trimmed car

p-MLFD 0.99% 99.0 % 1.41% 98.7%
MLFD 0.42% 99.6 % 21.48% 85.47%
p-LSCF 16.87% 86.9 % 9.92% 91.20%
LSCF 11.2 % 90.1 % 28.36% 81.94%
ML suffer an important loss in quality by converting the common-denominator

model to the modal model by reducing the residues to a rank-one matrix using an
SVD. Other examples [116] and [89] confirm this fact. This fact of loosing qual-
ity by transferring common-denominator models into modal models tends to be
more problematic for highly-damped cases (damping ratios > 2%), which can be
concluded from a comparison of the ’body-in-white’ with the ’fully trimmed car’
example. Table 5.4 gives the estimated natural frequencies and the corresponding
damping ratios. This proves that the p-ML method is suited to deal with high
modal densities and highly damped structures.
In fact, the p-ML optimizes the estimated model in a maximum likelihood

sense by optimizing the resonance frequencies, damping ratios, mode shapes and
Table 5.4: Natural frequencies and damping ratios estimated by p-ML
frequency (Hz) damping ratio (%)

4.05 5.80
4.29 7.49
4.72 6.55
6.04 4.29
8.59 6.80
14.69 5.75
15.74 8.30
17.07 5.54
18.34 5.09
20.79 4.01
21.82 3.07
22.43 4.29
25.15 3.03
25.93 3.03
26.93 6.00
participation factors. The common-denominator based ML algorithm optimizes

the model by optimizing the resonance frequencies, damping ratio and residues,
while there is no guarantee that the residue matrices are of rank one. Instead
of optimizing the modal parameters, the ML algorithm uses this extra freedom
in the model for fine tuning its parameters by using the freedom of not rank 1
residue matrices and by consequence quality is lost by transforming the residue
matrices to rank one matrices. Furthermore, it is clear for this example that a
part of the errors of the p-LSCF estimator are compensated in the second step
when estimating the mode shapes by the LSFD. The quality of the modal model
is improved compared to the polynomial model by a reduction of the mean error
from 16.9% to 9.9%. Figure 5.3 illustrates the improvement of the p-LSCF by the
second step LSFD algorithm by comparing the synthesized polynomial model with
the synthesized modal model. Nevertheless, the p-ML estimator outperforms the
p-LSCF algorithm for both the polynomial and the modal model.
5.5.3 Villa Paso Bridge
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
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.
ments on the Villa Paso bridge

p-MLFD 12.2% 88.2 % 11.4% 88.7%
MLFD 12.2% 88.3 % 14.5% 85.0%
p-LSCF 243.5% 18.9 % 39.8% 60.6%
LSCF 149.7 % 43.1 % 30.3% 69.6%
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.
5.7 Appendix: Confidence intervals on the esti-

mated poles from the p-ML estimator
The poly-reference ML estimator (the vec implementation with matrix weight-

ing) reaches the Cramer-Rao lower bound for the uncertainties on the estimated
parameters. The covariance matrix on the estimated parameters is given by

−1
cov(θM L ) = J H J (5.87)
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
A(λr )Lr = 0 (5.91)
Consider now a perturbation ∆Aj on the coefficient Aj and Ãj = Aj + ∆Aj .

This new set of coefficients A0 , . . . , Ãj , . . . , An defines a polynomial Ã(zk ) which
satisfies
Ã(zk ) = A(zk ) + ∆Aj zkj (5.92)
Furthermore, the new set of coefficients defines a perturbed companion matrix.

The EVD on this perturbed companion matrix results in the perturbed eigenvalues
λ̃r = λr + ∆λr and eigenvectors Ṽr = Vr + ∆Vr for which it holds that
Ã(λr + ∆λr )(Vr + ∆Vr ) = 0 (5.93)
Substitution of Eq. 5.92 results in
A(λr + ∆λr ) + ∆Aj (λr + ∆λr )j (Vr + ∆Vr ) = 0

(5.94)
5.7. Appendix: Confidence intervals on the estimated poles from the p-ML estimator131
Applying a Taylor expansion on the polynomial A(zk ) around λr , taking into

account that A(λr )Vr = 0 and neglecting all second order perturbations results in
A′ (λr )∆λr + ∆Aj λjr Vr = 0

(5.95)
Pn−1
with A′ (z) = j=1 nAn z n−1 the first derivative with respect to z. Elimination of
the perturbation on the eigenvalue λr resulting from the perturbation on coefficient
Aj is than given by
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
This sensitivity analysis to calculate the relation between the perturbation on a

matrix coefficient and the poles is generalization of the scalar case presented in
[44].
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
p−LSCF polynomial model p−LSCF 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)
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
p−LSCF polynomial model p−LSCF 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
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
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
p−MLFD MLFD
80 80
Ampl. (dB)
Ampl. (dB)
60
60
40
40
20
20
4 6 8 10 2 4 6 8 10
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
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
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
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
Chapter 6
Deterministic
Frequency-domain Subspace
Identification
Frequency-domain subspace algorithms estimate state-space models by means of ge-

ometrical projections and can be considered as the frequency-domain counterparts
of time-domain subspace algorithms. This chapter gives a short introduction in
frequency-domain subspace identification for structural engineering. Special atten-
tion is paid to an extended state-space model to consider the effect of the initial and
final conditions to avoid errors introduced by transients and leakage. This exten-
sion makes it possible to estimate state-space models in the frequency domain from
nonperiodic signals without any approximation and under the same assumptions
as in the time domain. Next, the extended state-space model is further used in a
mixed non-parametric/parametric frequency response function (FRF) estimation
procedure to eliminate bias errors introduced by leakage. In this way, the esti-
mated multiple input/ multiple output (MIMO) state space models can be validated
by accurate FRF estimates.
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].
6.2 Basic Frequency-Domain Projection Algorithm
Consider the state space model given by
zk X(k) = AX(k) + BU (ωk ) (6.1)

Y (ωk ) = CX(k) + DU (ωk ) (6.2)
with X(k) ∈ C n×1 , Y (k) ∈ C No ×1 and U (k) ∈ C Ni ×1 respectively the states,

responses and inputs for spectral line k. Recursive use of Eq. 6.1 and 6.2 gives
zkp Y (ωk ) = zkp−1 (Czk X(k) + zk DU (ωk )) (6.3)

= zkp−1 (CAX(k) + CBU (ωk ) + zk DU (ωk )) (6.4)
..
.
= CAp X(k) + (CAp−1 BU (ωk ) + CAp−2 BU (ωk )zk
+ . . . + CBU (ωk )zkp−1 + DU (ωk )zkp )U (ωk ) (6.5)
Stocking Eq. 6.5 for p = 0, 1, . . . , r − 1 yields to

   
Y (ωk ) U (ωk )
 zk Y (ωk )   zk U (ωk ) 
..  = Or X(k) + Γ  .. (6.6)
   
 
 .   . 
zkr−1 Y (ωk ) zkr−1 U (ωk )
with Or the extended observability matrix defined by
 
C
 CA 
Or =  .. (6.7)
 

 . 
CAr−1
and the lower triangular block Toeplitz matrix Γ given by
 
D 0 ... 0
 CB D ... 0 
Γ= .. .. .. ..  (6.8)
 
 . . . . 
CAr−2 B CAr−3 B . . . D
The matrix formulation of Eqs. 6.8 for k = 1, 2, . . . , N leads to
Y = Or X + ΓU (6.9)
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
Yre = Or X re + ΓUre (6.10)
where ()re represents a matrix with the real and imaginary parts aside each other,
for example
Yre = [Re(Y) Im(Y)] (6.11)
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)
with the orthogonal projection defined by
Yre /Ure,⊥ = Or X re /Ure,⊥ = R22

T
QT2 (6.14)
The second step uses a SVD to estimate the extended observability matrix Ôr .
For a model order n, Ôr is given by
Ôr = U[:,1:n] (6.15)

6.3. Starting from FRFs or power spectra 139
In a third step an estimate of A and C from Ôr is found in a LS sense

+
Â = Ô[1:No (r−1),:]
Ô[No +1:No r,:] and Ĉ = Ô[1:No ,:] (6.16)
In the fourth and last step, given the estimates Â 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].
6.3 Starting from FRFs or power spectra
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.
6.4 A weighted frequency-domain projection al-

gorithm
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)
Consider C(ωk ) as the covariance matrix of the output noise defined by
C(ωk ) = E N (k)N H (k) ∈ C No ×No

(6.20)
with N (k) ∈ C No ×1 the noise on the response measurements. The covariance

matrix of the Vandermonde matrix Y is then given by [77]
C = Re Z diag (C(ω1 ) . . . C(ωM )) ZH

(6.21)
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.
In [74] another consistent frequency-domain algorithm based on an instrumen-

tal variable approach is presented not requiring any a priori noise information.
Although, from several experiments, simulations and testing of different imple-
mentations from independent authors, this IV-base frequency-domain subspace
algorithm did not result in satisfactorily estimates. In fact, the classical projec-
tion algorithm outperformed this IV approach is most cases.
6.5. Extended state-space model for initial and final conditions 141
6.5 Extended state-space model for initial and fi-

nal conditions
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.
6.5.1 State-Space Model for Arbitrary Signals
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:
xn+1 = Axn + Bun (6.25)

yn = Cxn + Dun (6.26)
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
by Eq. 6.27. The discrete Fourier transform of xn+1 is given by
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)
with T = x0 − xN . These frequency domain state space equations describe the

system under the same assumption as their time-domain equivalents, although
in many identification methods the contribution of T is neglected. A sufficient
condition for T to be zero is that the initial and final conditions are equal, which
is the case for periodic and time-limited signals.
6.5.2 Remarks on the extended state-space model
• 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
6.5.3 A mixed non-parametric/parametric FRF estimator

for validation
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
for the initial/final conditions. As a model validation the synthesized FRFs from
the estimated system parameters
Ĥ(zk ) = Ĉ(zk I − Â)B̂ + D̂ (6.37)
can be compared with the FRFs obtained by the mixed non-parametric/parametric

estimation H1,e (ωk ).
6.5.4 State-Space Model for FRF data
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
zk X(k) = AX(k) + B ′ U ′ (k) (6.39)

H1 (ωk ) = CX(k) + D (6.40)
with B ′ = [B T1 . . . TNb ], U ′T = [IN i F1t,T . . . FNt,T

b
] and the state vector X(k) ∈
n×(Nb +1)
R . It is important to notice that this extended state space model fits
the FRFs calculated by the H1 method exactly if a rectangular window is used.
Furthermore, each extra block b used by the H1 estimator introduces n extra
parameters Tb in the model, while the equivalent common-denominator approach
in [14] and [24] introduces No n extra parameters in the model for each block like
discussed in chapter 4.
−4
x 10
response y(n)
2
−2
−4
−6
100 200 300 400 500

n (samples)
Figure 6.1: Response of a 6-order discrete time system excited by random noise.
6.6 Simulation and Measurement examples
6.6.1 Simulations
Extended frequency-domain state-space model
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.
Mixed non-parametric/parametric FRF estimation for validation
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
−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).
Extended model for FRFs
In a second set of simulations the response measurements yn are corrupted by

colored noise yr,n . For each of the 100 Monte Carlo runs the ’4096’ time samples
are processed by the identifying the classic state-space model (classic model)and
the extended state-space model (extended model) from FRFs, obtained from a
different numbers of blocks, as primary data. The mean square error (MSE) of
the damping ratios from the 100 Monte Carlo runs is calculated as the
100
1 X ˆ
M SE = (di − de )2 (6.41)
100 i=1
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
−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.
6.6.2 Measurements on a subframe of a car
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.
20 20
10 10
amplitude (dB) 0 0
amplitude (dB)
−10
−10
−20
−20
−30
−30 −40
−40 −50
−50 −60
340 360 380 400 320 340 360 380 400

(a) (b)
20 20
10 10
0 0
amplitude (dB)
amplitude (dB)
−10 −10
−20 −20
−30 −30
−40 −40
−50 −50
−60 −60
340 360 380 400 340 360 380 400

(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
This chapter introduces frequency-domain subspace algorithm to identify discrete-

time state-space models from both IO data and FRFs. Next, an extended frequency-
domain state-space model is introduced to allow the identification starting from
arbitrary signals without introducing systematic errors caused by leakage and tran-
sients. Therefore the initial and final conditions of the states have to be taken into
account. It is shown that this can be realized by using an additional input signal.
Moreover, it is shown how the state space model can be validated by using an ex-
tended FRF estimator based on a mixed non-parametric/parametric approach that
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
(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
Table 6.1: Natural frequencies and damping ratios of the subframe
fref fESSM fCSSM dref dESSM dCSSM

129.5 129.5 130.3 0.122 0.116 0.202
147.8 147.9 148.1 0.081 0.081 -0.002
175.9 176.3 176.0 1.138 1.118 0.208
205.6 205.6 204.8 0.106 0.112 0.110
241.0 241.0 241.4 0.115 0.110 0.160
252.9 252.9 253.0 0.095 0.091 0.102
286.3 286.3 286.7 0.221 0.224 0.291
291.2 291.5 1.006 0.801
329.3 329.3 329.6 0.164 0.164 0.214
331.3 331.3 331.3 0.130 0.134 0.277
356.1 356.1 355.8 0.099 0.099 0.212
359.7 359.7 359.5 0.101 0.101 0.306
379.7 379.7 379.8 0.142 0.143 0.061
402.6 402.6 402.3 0.098 0.097 0.148
6.8 Appendix
The system matrices used for the simulations are given by

 
0.6541 −0.8108 −0.0009 −0.0042 0.0003 0.0018
 0.6977 0.6512 0.0004 0.0031 −0.0001 −0.0013 
 
 −0.0007 0.0050 −0.0211 −1.0106 0.0131 −0.0436 
A =   0.0001 −0.0001

 0.9599 −0.0270 −0.0079 0.0165 

 −0.0003 0.0019 −0.0118 −0.0442 −0.4961 0.9079 
0.0002 −0.0016 0.0027 0.0176 −0.7944 −0.4636
 
−0.0018 −0.0025

 0.0010 0.0014 

 −0.0014 0.0000 
B =  

 0.0004 −0.0000 

 0.0007 −0.0010 
−0.0001 0.0002
 
−0.1735 −0.0301 −0.2569 0.0922 0.1920 0.0696
C =  −0.2453 −0.0435 −0.0054 −0.0114 −0.2753 −0.0786 
−0.1736 −0.0318 0.2650 −0.0744 0.1976 0.0408
 
0.1327 −0.0085
D =  −0.0085 0.1322  1.0e − 003
−0.0006 −0.0085
Chapter 7
Stochastic frequency-domain
subspace identification
Until now frequency-domain subspace algorithms are limited to identify determinis-

tic models from input/output or FRF measurements. In this chapter, a frequency-
domain subspace algorithm is presented to identify stochastic state-space models
in a consistent way from the spectra of the given output data. The relation to
and the analogy with time domain stochastic subspace models is established. The
applicability of the method is shown by both simulations and several measurement
examples.
153
154 Chapter 7. Stochastic frequency-domain subspace identification
7.1 Introduction
Identification methods which identify state-space models by geometrical operations

of the input and output sequences are commonly known as subspace methods
and have received much attention in the literature. In [124] a frequency-domain
subspace algorithm for the identification starting from power spectra is proposed
based on the deterministic basic projection algorithm.
In time-domain subspace identification, several algorithms were developed to

identify stochastic models from output-only measurements [120]. In [123] it is
shown how well know methods like the Principal Component algorithm (PC),
the Unweighted Principal Component algorithm (UPC) and the Canonical vari-
ate algorithm (CVA) fit in this framework of stochastic time-domain subspace
identification.
However, until today frequency-domain subspace identification is limited to

estimate deterministic models from FRFs, IO data or power spectra. In this
chapter a consistent stochastic frequency domain subspace identification method
is presented that directly starts from the output-only spectra, without forming the
power density matrices. The relation and analogy to the time domain stochastic
subspace identification [123] is established.
The main advantage of this stochastic frequency domain subspace algorithm

is that the identification problem can be solved in a single step, while frequency-
domain prediction error methods (PEM) result in a non-linear estimation ap-
proach, which require iterative optimization methods (e.g. Gauss-Newton) [97],
[98]. In that framework, the main problems arise from convergence difficulties
and/or the existence of local minima. Subspace algorithms typically use numeri-
cally robust and time efficient QR and SVD decompositions to optimize compu-
tation time and memory usage.
7.2 A first approach
Consider a proper (order of the numerator ≤ order of the denominator of the

equivalent polynomial transfer function), stable nth order multiple input-multiple
output stochastic system. The frequency domain state-space equations of this
discrete-time system are given by
zk X k = AXk + Wk (7.1)
Yk = CXk + Vk (7.2)
with Yk ∈ C No ×1 the vector of the output spectra at spectral line k, Xk ∈ C n×1

the state vector at frequency lines k and zk = ei2πk/N (k = 1, 2, . . . , N ) covering
7.2. A first approach 155
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.
Assumption 1: Wk and Vk are zero mean circular complex independent and

identically distributed noise sources with a covariance matrix

Wk Q S
WkH VkH

E = (7.3)
Vk SH R
where E is the expected value.
Recursive use of the Eqs. 7.1 and 7.2 results in
zkp Yk = zkp−1 (Czk Xk + Vk zk ) (7.4)

= zkp−1 (CAXk + CWk + Vk zk ) (7.5)
..
.
= CAp Xk + CAp−1 Wk + CAp−2 Wk zk + . . . + CWk zkp−1 + Vk zkp(7.6)
Stocking equation 7.6 for p = 0, 1, . . . , r − 1 gives

     
Yk Wk Vk
 zk Yk   z k Wk   zk V k 
..  = Or Xk + Γ  .. + .. (7.7)
     
 
 .   .   . 
zkr−1 Yk zkr−1 Wk zkr−1 Vk
with Or the extended observability matrix

 
C
 CA 
Or =  .. (7.8)
 

 . 
CAr−1
and Γ a lower triangular block Toeplitz matrix.

 
0 0 ... 0 0
 C 0 ... 0 0 
Γ= .. .. .. .. .. (7.9)
 

 . . . . . 
CAr−2 CAr−3 . . . C 0
Matrix formulation of Eq. 7.7 for k = 1, 2, . . . , N yields
Y = Or X + ΓW + V (7.10)
with Y ∈ C No r×N , W ∈ C nr×N and V ∈ C No r×N block Vandermonde matrices

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
and X ∈ C n×N matrix defined by

X = X1 X2 ... XN
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.
Consider the matrix

z1r Y1H z1r−1 Y1H z1 Y1H
 
...
 z2r Y2H z2r−1 Y2H ... z2 Y2H 
L= .. .. .. (7.11)
 
.. 
 . . . . 
r r−1 H
zN YNH zN YN ... zN YNH
with r > n/No , than the following theorem holds:
Theorem 1: Under Assumption 1

YL Or XL
→ w.p. 1 for N → ∞ (7.12)
N N
(w.p.1 stands for with probability one)
7.2. A first approach 157
Proof : To prove the strong consistency it needs to be shown that

1 1
a.s. lim YL − Or XL = 0 (7.13)
N →∞ N N
Taking into account equation (7.10) it is sufficient to show that
1
a.s. lim WL = 0
N
N →∞
1
a.s. lim VL = 0
N →∞ N
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
and elimination of Xk in equations (7.1) and (7.2) gives

−1 H
YkH = WkH Izk−1 − AH C + VkH
−1 H
= WkH zk I − AH zk C + VkH
2

= WkH zk I + AH zk + AH zk2 + . . . C H + VkH (7.15)
+∞
X
= WkH Al zkl + VkH
l=1
l−1 H 2
with Al = AH C . The Taylor expansion (I − AH z)−1 = I + AH z + AH z 2 +
3
AH z 3 + . . . holds since |z| = 1 if |eig(AH )| < 1. This is equivalent to |eig(A)| < 1
and hence the system must be stable. Substitution of (7.16) in (7.14) results in
+∞ X N N
!
zkr+i+l−j Wk WkH Al + zkr+i−j Wk VkH
X X
(WL)ij = (7.16)
l=1 k=1 k=1
Under Assumption 1, (W L)ij /N converges w.p.1 to its expected value for N → ∞

(strong law of large numbers for independent and identically distributed random
variables, [70]):
1 1
a.s. lim (WL)ij = lim E (WL)ij (7.17)
N →∞ N N →∞ N
+∞ N N
! !
1 X X r+i+l−j S X r+i−j
= lim QAl zk + z (7.18)
N →∞ N l=1 k=1
N k=1 k
= 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
k = 1, . . . N covers the full unit circle (i = 1, . . . , r and j = 1, . . . , r) and thus

PN r+i+l−j
k=1 zk = 0. Except for the values of l such that r + i + l − j = qN with
PN
q ∈ N0 the term k=1 zkr+i+l−j = N is different from zero. So it still needs to
be proven that the terms with r + i + l − j = qN in 7.18 are zero for N → ∞.
Consider thus r + i + l − j = qN and thus the corresponding term
N
1 X qN −r−i+j−1
lim QAl zkr+i+l−j = lim QAH (7.20)
N →∞ N N →∞
k=1
l−1 l
since Al = AH . Substituting the eigenvalue decomposition AH = T ΛT −1 , with
Λ a diagonal matrix containing the discrete poles, results in
lim QT ΛqN −r−i+j−1 T −1 = 0 (7.21)
N →∞
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.
In a next step a singular value decomposition of YL gives an estimate of the

extended observability matrix Or . This is a commonly-used step in subspace
algorithms
YL = U ΣV T (7.23)
An estimate of Ôr for model order n is then given by
Ôr = U[:,1:n] (7.24)
In a last step an estimate of A and C are obtained from Ôr in a LS sense

†
Â = Ô[1:No (r−1),:]
Ô[No +1:No r,:] and Ĉ = Ô[1:No ,:] (7.25)
7.3 A second approach
In a second approach a normalized multiplication is proposed to obtain a relation-

ship with the Kalman filter state estimate. The second and main theorem of this
chapter is now discussed and proven.
7.3. A second approach 159
Theorem 2 Under Assumption 1
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
Since the Taylor expansion Xk = (Izk − A)−1 Wk ≃ zk−1 (I + Azk−1 + A2 zk−2 +

. . .)Wk holds since |zk | = 1 for |eig(A)| < 1 (stable system) the expected value
PN
limN →∞ E ( N1 k=1 Xk WkH ) = 0 (follow the lines of the proof of Theorem 1).
PN
Similarly it can be shown that limN →∞ E ( N1 k=1 Wk XkH ) = 0 and as a result
the expected value of equation (7.28) is equal to the following Lyapunov equation
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.
To prove the strong consistency it needs to be shown that

1 1
a.s. lim YL( LH L)−1 LH f − O X̂
r f =0 (7.32)
N →∞ N N
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
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 →∞
Combining equations (7.31) and (7.33) results in

 
C
1  CA 
 Ar−1 G Ar−2 G

a.s. lim YL = lim  .. ... G (7.34)
 
N →∞ N N →∞  . 
r−1
CA
= lim Or ∆r (7.35)
N →∞
So far, it is proven that Y N1 L( N1 LH L)−1 LH 1 H

f → Or ∆r ( N L L)
−1 H
Lf w.p.1 for
H −1 H
N → ∞, so it still has to be shown that ∆r (L L) Lf → X̂f w.p.1 for N → ∞.
To prove this, a result from stochastic time domain subspace and Kalman filter
theory is used. In [123] it is proven that
 
yq
x̂i+q = ∆t Mt−1 
 .. 
(7.36)
. 
yi+q−1
with ∆t and M defined as
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
xi+q+1 = Axi+q + wi+q (7.39)

yi+q = Cxi+q + vi+q (7.40)
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
Since zk = ei2πk/N and zf = ei2πf /N
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)
This is an important observation for stochastic frequency domain subspace algo-

rithms, since it shows that the spectra of the forward Kalman filter state estimate
can be expressed as a function of the output spectra. The only thing left to
prove Eq. 7.26, is to show that ∆t and Mt−1 are respectively equal to ∆ and
a.s. limN →∞ ( N1 LH L)−1 . Since
∆ = 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
obtains
G′ = E (xn+1 ynH ) (7.48)

N
1 X
= E (xn+1 ynH ) (7.49)
N n=1
N N N
1 X 1 X 1 X H −n
= E √ Xk1 zkn+1 √ Yk zk (7.50)
N n=1 N 1
N k2 =1 2 2
k1 =1
N N
!!
1 1 X X
n+1 −n
= E Xk1 Yk2 zk 1 zk 2 (7.51)
N N n=1
k1 =1
N
1 X
= E( zk Xk YkH ) (7.52)
N
k=1
= G (7.53)
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
YL(LH L)−1 LH = U SV H (7.54)
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
7.4 Connection to time domain stochastic sub-

space identification
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
z1−r Y1 z2−r Y2 ... −r

zN YN
 
 z1−r+1 Y1 z2−r+1 Y2 ... −r+1
zN YN 
.. .. .. ..
 
 
 . . . . 
 def
z1−1 Y1 z2−1 Y2 −1

 ... zN YN  Y−
Y=  = (7.57)

 Y1 Y2 ... YN 
 Y+

 z1 Y1 z2 Y2 ... zN YN 

 .. .. .. .. 
 . . . . 
z1r−1 Y1 z2r−1 Y2 ... r−1
zN YN
and the equivalent main theorem is given by
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.
7.5 Geometrical Interpretation
It is well known that subspace algorithms are related to geometrical operations.

Basic equations (7.26) can be considered as the projection of the row space of Y+
onto the row space of Y− with Y− = LH .
H H −1
Y+ /Y− = Y+ Y− (Y− Y− ) Y− (7.59)
= Or X̂ (7.60)
In [123] a fast projection algorithm is proposed based on a QR decomposition.
Consider the following decomposition

B RB 0 QA
= (7.61)
A RAB RA QB
then the projection of A onto B is given by
A/B = RAB QA (7.62)
7.6 Final Algorithms
Basically, the numerical implementation of the frequency-domain stochastic sub-

space identification methods are similar to their time-domain counterparts.
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)
3. Calculate the SVD of projection

Yf /Yp = U SV H (7.65)
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.
+
Â = 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
(b) The second approach exists of estimating the matrices G, Σ and Λ0 .

N
1 X
Ĝ = zk Xk YkH (7.71)
N
k
N
1 X
Σ̂ = Xk XkH (7.72)
N
k
N
1 X
Λˆ0 = Yk YkH (7.73)
N
k
By using equations 7.28-7.30 the covariance matrices R, S and Q can

then be calculated.
The total identification procedure is schematically given by figure 7.1 for the case
of (5a) and (6a).
• Frequency band selection, scaling

• Construct the Vandermonde matrices
Estimate X by projection (QR and SVD)
Y+ / Y- = Or X
Least-squares

zk X k A W
= Xk + k

Yk C Vk
A,C natural frequencies, damping

ratios and mode shapes (EVD)
Figure 7.1: Schematic overview of the stochastic frequency-domain subspace algorithm
7.7 Remarks
• Equation 7.54 can be replaced by

W1 YL(LH L)−1 LH W2 = U SV H (7.74)
with W1 ∈ C No r×No r and W2 ∈ C N ×N such that W1 is of full rank and
rank(LH ) = rank(LH W2 ). For time domain stochastic subspace it is shown
in [123] that special choices of W1 and W2 correspond to well known algo-
rithms like the Principal Component algorithm (PC), the Unweighted Prin-
cipal Component algorithm (UPC) and the Canonical Variate Algorithm
(CVA). Analogously, the frequency counterpart of this weighing matrices
correspond to the frequency counterparts of the PC, UPC and CVA algo-
rithms.
• Working in the frequency-domain has some specific advantages compared to
time domain methods. One advantage is the easy pre-filtering of P the data
N
e.g. disturbing harmonics can simply be removed [97]. Although k zkr =
0 is not fulfilled, since the disturbing harmonics are eliminated, the bias
introduced is negligible if N is large w.r.t. the number of spectral DFT
lines that have been removed [97]. A second advantage is the easy frequency
7.7. Remarks 167
band selection, which is an important advantage for applications like modal

analysis. The chosen frequency band is re-scaled to cover the full unit circle
uniformly to ensure consistency and the numerical stability is improved since
the basis functions z r are orthogonal.
• 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
for l ∈ N0 , since Vk VkH = R, Wk WkH = Q and Wk VkH = S for the excitation

of a state space model with a flat multisine signal. As a result, the proposed
technique is exact for this case and no uncertainty is present on the esti-
mated parameters for a finite N . This can easily be understood since the
power spectra density matrix is exactly given by Syy (ωk ) = Yk YkH for this
deterministic excitation signal and as a result the model given by Eq. 2.62
exactly fits Yk YkH . In practice this means that the modal parameters i.e.
natural frequencies, damping ratios and mode shapes can be exactly deter-
mined from an output-only experiments from a finite amount of data by the
use of deterministic multisine excitation. A specific application is found in
laboratory testing of small and light structures which can not be excited by
classic hammer or shaker excitation. In this case an unmeasurable acoustic
excitation is often applied in the test setup.
• 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
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 Simulation and Measurement example
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
E (Y Y H ) = C(Iz−A)−1 Q(Iz ∗ −AH )−1 C H +R+C(Iz−A)−1 S+S H (Iz ∗ −AH )−1 C H

(7.79)
7.8. Simulation and Measurement example 169
−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)
−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
200 250 300 350

Freq. (Hz)
Figure 7.5: Subframe. Stabilization chart for the stochastic frequency-domain subspace
algorithm. Notice that all identified poles are stable (positive damping), ∗: stable pole.
7.8.2 Measurement example: Subframe of an car
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
Table 7.1: Natural frequencies and damping ratios of the subframe.
fIO (Hz) dIO (%) fO (Hz) dOf (%)

206.1 0.20 206.3 0.22
241.9 0.12 241.7 0.16
254.4 0.10 254.6 0.15
287.1 0.32 287.3 0.28
329.4 0.33 328.9 0.33
333.0 0.14 332.9 0.15
357.8 0.13 357.9 0.12
359.9 0.17 360.1 0.14
380.0 0.31 380.3 0.26
is consistent for a stable system. Since output-only spectra can be modelled by

both a stable and an unstable system (both resulting in the same fit of the power
spectra), the solution obtained by the proposed subspace algorithm always results
in stable poles for N → ∞. This is clearly illustrated by the stabilization diagram
in figure 7.5, which only shows stable poles. Figure 7.6 compares the synthesized
power spectra with the output spectra. Table 7.1 shows a good agreement between
the natural frequencies fIO and damping dIO ratios obtained from deterministic
modelling from input/output measurements (i.e. the 23 responses and the 2 elec-
trical generator signals) and the natural frequencies fO and damping ratios dO
obtained from the output-only analysis.
7.8.3 Flight flutter testing
The proposed algorithm can be used to process in-flight vibration measurements

without the use of the excitation signals. This is e.g. the case when the airplane
is excited by the natural turbulence only or even when excitation is applied but
unmeasurable. Nevertheless, sources from NASA report that the natural excita-
tion is often band limited [10] and many time is lost during flight to search for
sufficient turbulent force levels. Therefore, the use of artificial excitation is still
often used and nowadays a fly-by-wire system allows to inject perturbation signals
in the control loops of the flap mechanisms of the wings to excite the structure. In
this case the rotational vibration of the flaps is often measured as an input signal,
while the dynamical vibration behavior is already partially included in this rota-
tional vibration measurement. (One can discuss whether this rotational vibration
has to be considered as an input or output signal.) Therefore, it can be better
to start from the vibration measurements only, without such an input. In this
example, no use is made of the input signals and the algorithm starts only from
the response measurements. The ambient excitation forces are in this case both
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
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
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
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
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.
7.8.4 Villa Paso bridge
The vibrations on the Villa Paso bridge were measured by 10 accelerometers in

6 different patches during ambient excitation by the traffic. Transformation of
the measured acceleration sequences to the frequency domain by the DFT and
a frequency band selection from 1Hz to 6.5Hz provides the primary data for the
algorithm. Figure 7.9 illustrates the stabilization chart for a model order of 20
modes. The use of simple criteria such as the relative damping and frequency de-
viations between the poles estimated for subsequent orders guides the user trough
the selection of the physical modes. Figure 7.10 illustrates the synthesized power
spectral densities for 2 different outputs.
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
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
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
Figure 7.10: Comparison between the power spectra synthesized from the estimated
model (full line) and the output spectra (×)
7.9 Conclusion
In this chapter a frequency-domain subspace algorithm was presented to identify

stochastic state-space models from output measurements only. Different from
other frequency domain output-only algorithms is that the proposed algorithm
starts directly from the spectra, while other frequency-domain algorithms used
for OMA start from (’positive’) power spectra. Therefore, the main advantage
of the proposed algorithm is that it can operate from a limited amount of data.
Furthermore the relationship and analogy to the stochastic time-domain subspace
identification is made. Finally the algorithm is applied on both simulations and
real measurements such as in-flight flutter measurements and operational bridge
measurements. Compared to its time-domain counterpart the main advantage is
the physical interpretation of the frequency domain, the simple pre-filtering and
band selection.
7.10 Appendix
The model used for the simulation is given by

−0.8455 + 0.4839i −0.0217 + 0.0057i
A =
0.0056 + 0.0253i −0.1531 − 0.9131i

−0.2918 0.3204
C =
−0.2848 − 0.0068i −0.3364 + 0.0073i

0.5207 −0.0750 − 0.2295i
Q = 10−5
−0.0750 + 0.2295i 0.1120

0.0200 0.0682 − 0.0374i
R = 10−6
0.0682 + 0.0374i 0.3024

0.2966 − 0.1275i 0.7726 − 0.9887i
S = 10−6
0.0135 + 0.1491i 0.3246 + 0.4829i
Chapter 8
Combined frequency-domain
Until recently frequency-domain subspace algorithms were limited to identify deter-

ministic models from input/ouput measurements. In the previous chapter, a con-
sistent frequency-domain subspace identification method was proposed to identify
stochastic models from output-only spectra. In this chapter, a combined deterministic-
stochastic frequency-domain subspace algorithm is presented to estimate models
from input/output spectra, frequency response functions or power spectra. The re-
lation with time-domain subspace identification is elaborated. It is shown by both
simulations and real-life test examples that the presented method outperforms tra-
ditional frequency-domain subspace methods. The proposed algorithm outperforms
the classic projection frequency-domain subspace algorithms.
177
178 Chapter 8. Combined frequency-domain subspace identification
8.1 Introduction
In general three different types of subspace identification algorithms exist in the

time-domain [123]:
• Identification of deterministic models given by
xn+1 = Axn + Bun

yn = Cxn + Dun
Deterministic algorithms only estimate the dynamics between the measured

outputs yn and the measured inputs un . Their frequency-domain equivalents,
discussed in chapter 6, can be made consistent in the case of colored output
noise by introducing the covariances of output noise Ck = E Nk NkH as a

frequency weighting in the projection algorithm
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.
• Identification of stochastic models [120] given by
xn+1 = Axn + wn
yn = Cxn + vn
Stochastic models estimate the dynamics of a system without the use of

artificially applied forces un , but instead considers all unmeasured forces
to be white zero mean (random) noise wn and vn . These type of models
have their applications in the OMA framework and their frequency-domain
equivalents is presented in chapter 7.
• Identification of combined deterministic-stochastic models given by [122]
xn+1 = Axn + Bun + wn

yn = Cxn + Dun + 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
measured inputs, the dynamics excited by unmeasurable forces as well as

the coupled dynamics i.e. modes excited by both the measurable and un-
measurable forces. This combined interpretation of this subset of subspace
algorithms fits in the framework of IO-data driven OMAX identification. In
this chapter a frequency-domain counterpart for these combined stochastic-
deterministic algorithms is developed and its applicability is shown by seve-
ral examples and simulations. This combined algorithm can also be used to
obtain consistent estimates from noisy FRFs and (positive) power spectra.
8.2 Theoretical Aspects
8.2.1 System description
Consider a proper, stable nth order multiple-input/multiple-output combined deter-

ministic-stochastic system. The term combined means in the context of this the-
sis that the deterministic, the stochastic and the coupled deterministic-stochastic
dynamics are considered simultaneously. The frequency-domain state-space equa-
tions of a combined discrete-time system is given by
zk X k = AXk + BUk + Wk (8.2)

Yk = CXk + DUk + Vk (8.3)
with Yk ∈ C No ×1 and Uk ∈ C Ni ×1 respectively the vectors of the output and input

spectra at spectral line k, Xk ∈ C n×1 the state vector at frequency spectral lines
k and zk = ei2πk/N (k = 1, 2, . . . , N ) covering the full unit circle. The vectors
Wk ∈ C n×1 and Vk ∈ C No ×1 contain respectively the process and the output
measurement noise at spectral line k.
Assumption 1: Wk and Vk are zero mean circular complex independent and

identically distributed (over k) noise sources with covariance matrix

Wk Q S
WkH VkH

E = (8.4)
Vk SH R
where E is the expected value.
The goal of this chapter is to estimate the unknown matrices A, B, C, D, Q, R, S

(up to a similarity transformation) from the given spectra Yk and Uk and assuming
the data is generated by an unknown combined deterministic-stochastic system.
8.2.2 Main theorem
Consider the input and output Vandermonde matrices U ∈ C 2rNi ×N and Y ∈

C 2rNo ×N respectively given by
z1−r U1 z2−r U2 ... −r

zN UN
 
−r+1
 z1−r+1 U1 z2−r+1 U2 . . . zN UN 
.. .. .. ..
 
 
 . . . . 
 def
 z −1 U1 −1 −1

1 z2 U2 ... zN UN  U−
U=  = (8.5)

 U1 U2 ... UN 
 U+
 z1 U1 z2 U2 ... zN UN 
 
 .
.. .
.. .
.. .. 
 . 
z1r−1 U1 z2r−1 U2 ... r−1
zN UN
and similarly for
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
Theorem 1 Under Assumption 1

U−
Y+ /U+ → Or X for N → ∞ (8.7)
Y−
with Or the extended observability matrix given by

 
C
 CA 
Or =  .. (8.8)
 

 . 
CAr−1
and X = [ X1 X2 . . . XN ] the spectra of the forward Kalman filter state

estimates. The oblique projection A/B C is defined by [123]
†
CC H CB H

H H

A/B C = A C B C (8.9)
BC H BB H [:,1:r]
and the geometrical interpretation of this oblique projection is given in [123].
Proof : Consider the Inverse Discrete Fourier Transform (IDFT) of both the
output and input spectra given by
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
x(n + 1) = Ax(n) + Bu(n) + w(n) (8.12)

y(n) = Cx(n) + Du(n) + v(n) (8.13)
PN PN
and w(n) = √1N k=1 Wk zkn , v(n) = √1N k=1 Vk zkn zero mean, white noise vector
sequences. The input and output block Hankel matrices are then given by
 
u−r u−r+1 ... uN −r
 .. .. .. .. 
 .
 . . . 

 u−2 u−1 ... uN −2 
  def
 u−1 u0 ... uN −2  up
  = (8.14)
 u0 u1 ... uN −1  uf
 
 u1 u2 ... uN 
 
 . .. .. ..
 ..

. . . 
ur−1 ur ... uN +r−2

yp
and similarly for .
yf
The main theorem for combined deterministic-stochastic subspace identification

in the time-domain, presented in [123], states that

up
yf /uf → Or x for N → ∞ (8.15)
yp
with x = [x0 . . . xN ]T the forward Kalman filter state sequence estimate. Consider
the Discrete Fourier Transformation matrix
z10 z20 ... zN0

 
−1 −1 −1
1  z1 z2 zN 
F= √  .. .. .. (8.16)

.

N . . 
−(N −1) −(N −1) −(N −1)
z1 z2 ... zN
From the definition of the DFT and IDFT it follows that
Y− = yp F
Y+ = yf F
U− = up F
U+ = uf F (8.17)
and F is orthogonal, i.e. FFH = I. Applying the property A/B C = D ⇔

[AF ]/[BF ] [CF ] = DF if F F H = I, which can be proven by the definition of the
oblique projection, for the time domain theorem results in

U− [up F]
Y+ /U+ = [yf F]/[uf F] (8.18)
Y− [yp F]

up
= yf /uf F (8.19)
yp
= Ôr x̂F (8.20)
= Ôr X̂ (8.21)
which ends the proof.
The main theorem Eq. 8.15 in time-domain subspace is proven to be consistent,

when at least one of the following conditions is satisfied [122]:
• The system is purely deterministic, i.e. wn = 0 and vn = 0

• The input signal is white noise.
• r → ∞, since for r → ∞ the non-steady state Kalman filter converges to
a steady state Kalman filter. This is intuitively clear, since by the time
the Kalman filter is in steady state, the effect of the initial conditions has
died out. In practice r ≥ n/No + 8 already results in good estimates for
mechanical systems [122].
8.2.3 From states to system matrices
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
8.2.4 Taking into account effects of transients and leakage
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
zk X k = AXk + BUk + T zk + Wk (8.24)

Yk = CXk + DUk + Vk (8.25)
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
zk X k = AXk + BUk + Wk (8.26)

Yk = CXk + DUk + Vk (8.27)
and B represents the transient term T . As a conclusion the presented algorithm

in this chapter allows to identify combined deterministic-stochastic models from
IO data and purely stochastic models from output-only spectra, without suffering
from transient and leakage errors.
8.3 Practical Implementation
In this section the implementation of the different algorithm steps is discussed:

1. Form the Vandermonde matrices Y+ , Y− , U+ and U− from the measured

spectra Yk and Uk . The inputs can be extended by zk to model the tran-
sients in the measurements. In case one wants to estimate real system
matrices A, B, C and D, the Vandermonde matrices V are replaced by
[ Re(V) Im(V) ] (V = Y+ , Y− , U+ and U− ).
2. Calculate the RQ decomposition as
    
Y+ R11 0 0 0 Q1
 U+   R21 R22 0 0   Q2 
 U−  =  R31 R32 (8.28)
    
R33 0   Q3 
Y− R41 R42 R43 R44 Q4
then

U− R31 R32 R33 0
= Ra Q with Ra = (8.29)
Y− R41 R42 R43 R44

U+ = Rb Q with Rb = R21 R22 0 0 (8.30)

Y+ = Rc Q with Rc = R11 0 0 0 (8.31)
(8.32)

U−
with QT = [ QT1 QT2 QT3 QT4 ]. The oblique projection Y+ /U+
Y−
is then given by
†
U− H † H †
h i h i
H H
Y+ /U+ = Rc I − Rb Rb Rb Rb Ra I − Rb Rb Rb Rb Ra Q (8.33)
Y−
A profound discussion of this expression for the oblique projection is given

in [123].
3. In the third step the SVD of the projection is calculated as

U−
Y+ /U+ = U SV H (8.34)
Y−
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
The total identification procedure is schematically given by figure 8.1.

8.4. Remarks 185
• Frequency band selection, scaling

• Construct the Vandermonde matrices
Estimate X by the oblique projection

(QR and
SVD)
U-
Y+ /

Y- = O r X
U+ Y
-
Least-squares

zk X k A B X k Wk

=

+

Yk C D Yk Vk
A,C,B,C natural frequencies,

damping ratios, mode shapes and
modal participation factors (EVD)
Figure 8.1: Schematic overview of the stochastic frequency-domain subspace algorithm
8.4 Remarks
• Equation (8.34) can be replaced by

U−
W1 Y+ /U+ W2 = U SV H (8.37)
Y−
with W1 ∈ C No r×No r and W2 ∈ C N ×N such that W1 is of full rank and

rank([ U− Y− ]T ) = rank([ U− Y− ]T W2 ). For time-domain com-
bined deterministic stochastic subspace it is shown in [123] that special
choices of W1 and W2 correspond to well-known algorithms like the N4SID
(Numerical algorithms for Subspace State Space System Identification),
MOESP (Multivariable Output-Error State Space) and the Canonical vari-
ate algorithm (CVA). Analogously the frequency counterpart of this weighing
matrices correspond to the frequency counterparts of the N4SID, MOESP
and CVA algorithms.
• The presented algorithm can also be used as a combined deterministic-

stochastic frequency-domain subspace method processing Frequency Response
Functions (FRFs) or positive power spectra. By simple substitution of Yk =
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).
8.5 Simulations and Real-life measurement exam-

ples
8.5.1 Monte Carlo Simulation
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

+ C(Izk − A)−1 Wk + Vk (8.38)

with in this simulation Wk = B[:,2] U2,k and Vk = D[:,2] U2,k . To examine the
consistency properties (asymptotic unbiasedness), Monte Carlo simulations are
performed for 4 different cases of the total number of time samples Nt (N = Nt /2)
in order to check if the expected value of the estimated model is equal to the true
model for the number of frequency lines N going to infinity. As an error measure
the mean square relative error (MSRE) over the different frequencies between the
8.5. Simulations and Real-life measurement examples 187
−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
(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
N=150 N=250 N=400 N=600

A1 0.0097 0.0054 0.0042 0.0027
A2 0.0373 0.0074 0.0066 0.0041
A3 0.2810 0.2132 0.2760 0.3404
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
combined algorithm (A1,A2) outperforms the classical projection algorithm (A3).

Furthermore, the error is further reduced by taking into account the transients
effect T to compensate for leakage and the non-steady state behavior (A1 versus
A2). This is more important for small Nt , since for large Nt both leakage and the
¯
transient effect have less influence. In figure 8.2, the mean estimated system Ĝ and
2
the variance σĜ over the 100 simulations are shown for the different identification
cases for N = 250. Assume now that, both inputs are unknown and thus only
the spectra of the outputs are known. In this case the identification procedure
becomes a stochastic identification problem. The influence of the final and initial
conditions can still be modelled by considering the parameters T and the input
zk . The output can be considered as the sum of transient effects and a stochastic
contribution as shown in Eq. 8.40.
Yk = C(Izk − A)−1 T zk + C(Izk − A)−1 Wk + Vk (8.40)
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
N=300 N=600 N=900

R̄ 0.9826 0.9840 0.9851
σR 0.0061 0.0044 0.0034
θ̄ 1.5950 1.5944 1.5939
σθ 0.0070 0.0043 0.0039
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.
8.5.2 Flight flutter simulation
To show the applicability of the combined deterministic-stochastic algorithm, the

same in-flight simulation data based on a GVT test of an aircraft (as used in para-
graph 4.8.1) are processed to compare the combined common-denominator (CLSF)
estimator with both the deterministic and combined deterministic-stochastic
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)
frequency-domain subspace estimators. For each of the 6 increasing turbulence lev-

els 10 different runs are simulated and processed. Figure 8.4 illustrates the mean
values and standard deviations of the estimated damping ratios by the different al-
gorithms for the 6 turbulence levels. The classic deterministic projection algorithm
is less accurate than both combined algorithms. For the first two modes, i.e. wing
modes as shown in figure 4.9, the larger uncertainty and bias can be explained
by the influence of leakage. Both combined algorithms included an additional
polynomial T (z) or term T to model the initial and final conditions, while this
was not the case for the deterministic subspace algorithm. Furthermore it is clear
that the damping ratios of the tail modes 3 and 4 (which are only well excited by
the unknown turbulent forces), can only be accurately identified by the combined
deterministic-stochastic algorithms. Both the combined common-denominator al-
gorithm and the combined subspace algorithm result in a comparable accuracy of
the estimated damping ratios. Nevertheless, it should be noted that compared to
the combined common-denominator algorithm, the combined subspace algorithm:
• forces rank 1 residues on the measurements.
• no optimization procedure is required.
• the construction of stabilization diagram is possible in a fast way.
• 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.

1.8 1.95
1.9
1.7
1.85
1.6 1.8
damping ratio
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
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
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
Figure 8.4: Comparison of the damping ratio estimates (∗ : CLSF IO, ◦ : determin-
istic frequency-domain subspace , : combined frequency-domain subspace, – : exact
parameter)
8.5.3 Flight flutter measurements
To show the applicability of the combined deterministic-stochastic frequency do-

main the same real-life in-flight vibration measurements as introduced in para-
graph 7.8.3 are now analyzed. In this case the measured signal of the angle per-
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
(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
(c) (d)
Figure 8.5: Comparison between stabilization charts obtained by the deterministic

and combined subspace algorithm: (a) deterministic algorithm (b) combined algorithm
(c) deterministic algorithm after applying relative criteria (d) combined algorithm after
applying relative criteria.
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
Table 8.3: Natural frequencies and damping ratios identified by the deterministic and
combined deterministic-stochastic frequency-domain algorithm.
fdet (Hz) ddet (%) fcom (Hz) dcom (%)

4.08 1.95 4.09 3.06
4.60 4.07 4.69 3.39
5.27 3.46 5.16 5.59
/ / 5.35 2.39
6.06 0.17 6.01 5.19
8.07 2.95 8.07 3.27
8.80 2.20 8.71 3.25
9.07 2.91 9.17 3.60
/ / 9.87 3.81
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.
8.5.4 ABS-function driven identification
The combined deterministic-stochastic algorithm can also start from FRFs, posi-
tive power spectra or both simultaneously as primary data to estimate the system
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
−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
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
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-
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)
The accelerations were measured at 23 response locations and two random

inputs were applied by electrodynamic shakers. Before the actual identification
task started, the frequency band of interest between 210 Hz and 410 Hz is re-scaled
to cover 90% of the full unit circle (complex model). A model order of 30 was used.
Both the models identified by the classical projection algorithm and the combined
algorithm, are validated with a high quality validation data set by means of the
MSRE over all frequencies per FRF. Figure 8.7 shows the MSRE for the 46 FRFs
for both identification approaches. The presented combined approach resulted
in an average MSRE over all FRFs of 0.2, while the classical projection based
subspace method resulted in a mean MSRE of 5.4. Figure 8.8 compares some
identified FRFs for both identification approaches with a high quality validation
data set.
Laser vibrometer measurements on a car door
To illustrate that this combined frequency-domain subspace algorithm is capable

to process FRF measurements with high modal and high spatial density measure-
ments were performed by a scanning laser vibrometer on the door of a car. The
FRF matrix has 79 responses, 1 input and a frequency resolution of 0.31Hz. A
frequency band from 46Hz up to 120Hz is modelled by a real state-space model
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
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
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
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
to obtain clear stabilization charts based on the sign of the damping ratio.
8.6 Conclusion
This chapter presented a combined deterministic-stochastic frequency domain sub-

space algorithm resulting in consistent estimates of the system matrices and the
noise covariance matrix. Furthermore, it is shown that the proposed algorithm
is the frequency-domain counterpart of the well-known combined deterministic-
stochastic time-domain subspace algorithms. The proposed algorithm can be used
to process output-only measurements with a transient term, input-output measure-
ments, frequency response functions and (’positive’) power spectra. In the case
of input-output data, the algorithm can be considered in the OMAX framework.
The applicability is illustrated by means of both simulations and several real-life
measurements. The method identifies simultaneously the deterministic dynam-
ics, the stochastic dynamics and the coupled stochastic-deterministic dynamics
−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
50 60 70 80 90 100 110 50 60 70 80 90 100 110

−70
−80
−80
−90
−90
−100
Ampl. (dB)
Ampl. (dB)
−100
−110
−110
−120
−120
−130 −130
−140 −140
50 60 70 80 90 100 110 50 60 70 80 90 100 110

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
an therefore outperforms classical deterministic subspace projection algorithms.

Furthermore, the algorithm can take advantage of the frequency-domain by a
simple frequency band selection. The proposed combined stochastic-deterministic
frequency-domain subspace algorithm is very promising to process all types of
frequency-domain data.
8.7 Appendix
The 6th order model used for the simulation is given by

 
−0.0103 0.9928 0.0377 −0.1746 −0.0367 0.0365
 −0.9409 −0.0070 0.2141 0.2065 0.1868 −0.0467 
 
 0.0076 −0.1240 0.6504 −0.8057 −0.1547 0.0538 
A =   
 0.1601 0.0689 0.6680 0.6289 −0.1739 −0.0240  
 −0.0136 −0.0056 −0.0103 0.0168 −0.2689 0.7985 
0.0003 −0.0024 −0.0315 −0.0193 −0.9110 −0.6742

C = −0.3971 −0.0433 −0.3410 −0.0192 −0.3572 −0.0234
 
−1.2561 −1.0395
 0.1651 0.0782 
 
 −0.7939 −0.6635  −3
B =    10
 0.7051 0.7799  
 −0.1138 0.5930 
−0.6482 −1.4204
−0.3937 −0.5708 10−3

D =
(8.41)
Chapter 9
The secrets behind clear

stabilization charts
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.
This chapter starts with a short introduction to the LS identification problem,

next the influence of a mathematical constraint is briefly explained by means of
a simple single input-single output (SISO) Auto-Regressive model (AR) [67]. It
is shown that the stochastic model for output-only data can be described exactly
by stable, unstable or a mix of stable and unstable poles. Depending on the con-
straint applied to the coefficients, the stochastic data is modelled by a stable or
unstable model. The analogy with state-space models and stochastic frequency-
domain subspace identification algorithms is given. In a next step, the conclusions
for this AR model and stochastic subspace identification are extrapolated to Auto-
Regressive eXogeneous (ARX) and combined deterministic-stochastic state-space
models. It is shown how MPE methods like LSCE, LSCF, p-LSCF and the sub-
space algorithms fit in this framework. Finally, the demonstrated influence of the
constraint on the quality of the identified models and stabilization charts is dis-
cussed and demonstrated by several real life applications. The theory developed
in this chapter is closely related to the results given in [62], where a distinction
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 Theoretical Aspects
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)
with θ a set of unknown parameters, J ∈ Cn×m (n > m) the information ma-

trix and assume E ∈ C n×1 to be vector with complex circular independent and
identically distributed noise on the equations. To solve this set of equations in a
LS sense, a constraint must be applied on the coefficients in order to remove the
parameter redundancy. For example consider, the j-th parameter fixed to one i.e.
θj = 1. In that case the least squares solution is given by
J˜θ̃ = −Jj + E
θ̃ = −J˜† Jj + J˜† E (9.2)
with J˜ = J[:,1...j−1,j+1...m] , θ̃ = θ[1...j−1,j+1...m] , Jj the j-th column of J and J˜†

−1
the pseudo inverse of the matrix J˜ given by J˜† = J˜H J˜ J˜H (·H is the complex
conjugate transpose (Hermitian) of a matrix). It can simply be shown that the
estimate for the parameters is consistent if J˜† E → 0 for n → ∞. In the next
subsections a more profound discussion is given about the influence of the choice
of the constraint on the stability of the poles.
9.2.2 Influence of the constraint on an AR model
Consider a SISO autoregressive model given by
a0 y(n) + a1 y(n − 1) + . . . ar y(n − r) = e(n) (9.3)
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
frequency-domain counter part

(a0 + a1 zk−1 + . . . + ar zk−r )Yk = Ek (9.4)
N −1 N −1
with Yk = √1N n=0 y(n)zk−n , Ek = √1N n=0 e(n)zk−n and zk = ei2πk/N . Notice
P P
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 |
with λj = eσj +iωj , λ△j defined by e

−σj +iωj
and Rj a constant independent of zk .
△
(Applying the operator · to a pole changes the sign of the damping and thus the
stability, while the natural frequency remains unchanged.) As a consequence, every
term (zk − λj ) of A(z) can be replaced by (zk − λ△ j ) without any consequence on
the validity of Eq. 9.6. The conclusion is that Eq. 9.4 has several solutions, since
the sign of the damping σj does not change the amplitude of the polynomial. In
the context of this chapter, we are particularly interested in two special solutions:
• The solution with only stable poles (σj < 0) given by

As (z) = as z −r (z − λs1 )(z − λs2 ) . . . (z − λsr ) (9.11)
• The solution with only unstable poles (σj > 0) given by
Au (z) = au z −r (z − λu1 )(z − λu2 ) . . . (z − λur ) (9.12)
with λuj = (λsj )△ .
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.
Depending on the constraint applied, the LS solution of Eq. 9.13 results

in stable or unstable poles.
• The constraint a0 = 1 results in a consistent estimate of the poly-

nomial As (z) with only stable poles.
• The constraint ar = 1 results in a consistent estimate of the poly-

nomial Au (z) with only unstable poles.
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.
Example Consider a system given by the coefficients a0 = 1, a1 = 0.5589−0.9710i

and a2 = −0.6941 − 0.8036i. From the characteristic equation Eq. (9.5) the
poles in the discrete-time domain can be calculated, i.e. λ1 = 0.5243 + 0.8166i
and λ2 = −1.0832 + 0.1544i. The continuous time domain poles are given by
ln(λ1 ) = σ1 + iω1 = −0.03 + i and σ2 + iω2 = 0.09 + 3i (∆t = 1). The first pole is
stable |λ1 | < 1 or σ1 < 0 and the second pole is unstable. Consider the spectrum
Yk = A(zk )−1 Ek for k = 1, . . . , N (zk covers the full unit circle) and Ek zero
mean circular complex independent and identically distributed noise e.g. real and
complex part of Ek white independent Gaussian noise. From the spectrum Yk the
poles λ1 and λ2 are estimated (N=1024). Figure 9.1(a) illustrates one of the output
spectra Y . A Monte Carlo simulation was done for 100 runs for both constraints
a0 = 1 and a2 = 1. Figure 9.1(b) shows the estimated damping values σ for both
constraints. It is clear that both constraints estimate exactly the same magnitude
of the damping values, but different signs. The constraint a0 = 1 estimates stable
30
20 0.1
10
damping value sigma

0.05
amplitude (dB)
−10 0
−20
−0.05
−30
−0.1
−40
0.2 0.4 0.6 0.8 1 20 40 60 80 100

frequency (Hz) Monte Carlo runs
(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.
Changing the constraint from a0 = 1 to ar = 1 is equivalent with replacing

z −1 by z which can be seen from
(1 + a1 zk−1 + . . . + ar zk−r )Yk = Ek (9.14)
Replacing the basis function zk−1 by zk results in
(1 + a1 zk + . . . + ar zkr )Yk = Ek (9.15)
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
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
known that the constraint a0 = 1 results in a consistent estimation of the poles

of a stable system [67]. Consider the identification problem of the coefficients
a0 , a1 , . . . ar in Eq. 9.3 starting from the time responses y(n). The set LS of
equations is now given by
    
y(r) y(r − 1) . . . y(0) a0 e(r)
 y(r + 1) y(r − 2) . . . y(1)   a1   e(2) 
.. .. .   ..  =  ..  (9.17)
    
 . .. ..
 . .  .   . 
y(N ) y(N − 1) ... y(N − r) ar e(N )
The constraint a0 = 1 corresponds with
y(n) + a1 y(n − 1) + . . . ar y(n − r) = e(n) (9.18)
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
a0 y(n) + a1 y(n − 1) + . . . y(n − r) = e(n) (9.19)
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.
9.2.3 Stochastic State-Space models
Consider the stochastic state-space model from chapter 7 given by
zk X k = AXk + Wk (9.20)
Yk = CXk + Vk (9.21)
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.
Replacing now z by z −1 results in a Taylor expansion (I − Az −1 )−1 = I + Az −1 +

A2 z −2 + A3 z −3 + . . . which is only valid for unstable poles. Thus, replacing zk
by zk−1 in the block Vandermonde matrices Y+ and Y− results in a strongly
consistent estimate of an unstable state-space model. This observation for the
stochastic state-space identification can be summarized as:
Depending on the basis function, the stochastic state-space model iden-

tified by the stochastic frequency-domain subspace algorithm results in
stable or unstable poles.
• The basis function zk results in a consistent estimate of a state-

space model with only stable poles.
• The basis function zk−1 results in a consistent estimate of a state-
space model with only unstable poles.
9.2.4 Extrapolation to ARX models
In this section an autoregressive model with an exogenous input (ARX) is con-

sidered. Next, it will be shown how some MPE estimation methods fall into this
category of ARX models. An ARX process is described by the following equation
a0 y(n) + a1 y(n − 1) + . . . ar y(n − r) = b0 f (n) + b1 f (n − 1) + . . . br f (n − r) + e(n)(9.23)
and the frequency-domain equivalent is given by

(a0 + a1 zk−1 + . . . + ar zk−r )Yk = (b0 + b1 zk−1 + . . . + br zk−r )Fk + Ek (9.24)
Dividing both parts of this equation by the polynomial A(z) results in
B(zk ) 1
Yk = Fk + Ek (9.25)
A(zk ) A(zk )
The response Yk can be considered as the combination of a deterministic contribu-

tion and a stochastic contribution. Consider the identification of an ARX model
starting from the data Yk and Fk by solving the following overdetermined set of
equations
 
a0
 .  
Y1 . . . Y1 z1−r −F1 . . . −F1 z1−r

 .. 

E1
−r −r
 Y2 . . . Y2 z2 −F2 . . . −F2 z2    E2 
 
 ..

. . . . .
  ar  =   .. 

(9.26)
 . . . .
. .
. . . .
.

b
 0   . 
 
−r −r  . 
YN . . . YN zN −FN . . . −FN zN  ..  EN
br
Since the model order is not a priori known, a model of order s > r (s estimated
order, r order of the underlying model) is estimated and we can factorize A(z) as
A(z) = Ar (z)At (z), with r and t indicating the order (s = r + t). Furthermore,
1
we assume that the stochastic contribution Ar (z k)
Ek is of a lower amplitude level
r
than the deterministic contribution B (zk )
Ar (zk ) Fk (e.g. 30dB lower). It is clear that the
estimated model has t = s−r extraneous mathematical poles, besides the r system
poles. The r system poles are determined both in damping ratio (magnitude and
sign) and natural frequency, since not only the magnitude of B r (z)/Ar (z) but also
the phase plays a role. In [62] it is shown that the t extraneous poles are determined
by the stochastic AR part. So whether the t extraneous poles are stable or unstable
just depends on the choice of the constraint in solving the least squares problem
like shown in paragraph 9.2.2. This facilitates making the distinction between the
r deterministic poles from the t stochastic (also called mathematical) poles.
Example Consider a system given by the coefficients a0 = 1, a1 = 0.3805 −

0.9456i and a2 = −0.5797 − 0.6712i. From the characteristic equation Eq. 9.5
the poles in the discrete time-domain can be calculated λ1 = 0.5243 + 0.8166i
and λ2 = −0.9048 + 0.1290i. The continuous time-domain poles are given by
ln(λ1 ) = σ1 + iω1 = −0.03 + i and ln(λ2 ) = σ2 + iω2 = −0.09 + 3i.
Consider the response spectrum Yk = B(z k)

A(zk ) (1 + 0.05Vk )Fk for k = 1, . . . , N (zk
covers the full unit circle) and Fk the force as the excitation signal. Five per-
cent relative noise is added to the simulated model by Vk , which is zero mean
circular complex normally distributed noise (real and complex part of Nk white
independent normal distributed Gaussian noise). From the spectrum Yk and Fk
the polynomials A(z) and B(z) are estimated with model order n = 9. A total of
1024 spectral lines N was used for the simulation.
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
1.5 4
1 3
natural frequency omega

2
0.5
1
0
0
−0.5 −1
−1 −2
−1.5 −0.7 −0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0

−1.5 −1 −0.5 0 0.5 1 1.5 damping ratio sigma
(a) (b)
1.5
3
1
natural frequency omega

2
0.5
1
0
0
−0.5
−1
−1
−2
−1.5
−2 −0.1 0 0.1 0.2 0.3 0.4 0.5 0.6

−2 −1 0 1 2 damping ratio sigma
(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.
9.2.5 Combined deterministic-stochastic frequency domain

Similar as for the transition of an AR model to an ARX model, a deterministic

input can be added in the stochastic state-space model resulting in a combined
deterministic-stochastic state-space model. Under the assumption that the deter-
ministic and stochastic dynamics are uncoupled, the stability of the poles of the
stochastic dynamics depends on the choice of the basis function i.e. z or z −1 . Thus
9.3. Application for Modal Parameter Estimation methods 209
the use of z −1 as a basis function instead of z results in unstable mathematical

poles modelling the noise (stochastic contribution) and stable system poles mod-
elling the deterministic contribution i.e. the structure. Based on the sign of the
damping ratio, a clear stabilization chart can be constructed to easily distinguish
the mathematical from the physical poles.
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
(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.
9.3 Application for Modal Parameter Estimation

methods
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.
9.3.1 Least Squares Frequency-domain (LSCF) algorithm
The LSCF method curvefits a common-denominator model on the FRFs or I/O

data in a least squares sense as discussed in chapter 3. In [116] it was already
noticed that this method resulted in very clear stabilization diagrams. In this
chapter, the reason behind this remarkable result is investigated in more depth.
The identification problem is solved by minimizing the following cost function
N X
X Ni
No X
lFLSCF
RF = |Eoi,k |2 (9.27)
k=1 o=1 i=1
with the equation error given by

Eoi,k = A(zk )Hoi (ωk ) − Boi (zk ) (9.28)
Rewriting this equation as
Boi (zk ) 1
Hoi = − Eoi,k (9.29)
A(zk ) A(zk )
shows the structure of an ARX model with input Fk = 1. Since this equation error
is linear-in-the-parameters ,the unknown polynomial coefficients are estimated in
a (linear) LS sense. The equations to solve are then given by
Jθ ≃ 0 (9.30)
with J the Jacobian matrix. Elimination and substitution results in the compact
normal equations given by Eq. 4.28, i.e.
No
X
−So H Ro −1 So + To α = M α = 0

(9.31)
o=1
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)
for increasing model orders m from 1 to n.
9.3.2 Poly-reference Least Squares Frequency-domain (p-

LSCF, PolyMAX) algorithm
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
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.
9.3.3 Least Squares Complex Exponential (LSCE) method
A well-known and industrial standard is the poly-reference Least Squares Complex

Exponential algorithm (LSCE) [11],[52]. The poles and participation factors are
obtained from the solutions of the following matrix finite difference equation of
order r
r
Aj zk−j = 0
X
(9.33)
j=0
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.
9.3.4 Frequency-domain subspace algorithms
It was already shown in paragraph 9.2.5 that the frequency-domain state-space

models also use a constraint, which can simply be changed by replacing zk by zk−1 .
Using zk−1 as a basis function results in the stochastic contribution of the data to
be modelled by unstable poles and thus in this way clear stabilization charts can
be obtained.
9.3.5 Coupled stochastic-deterministic dynamics
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.
Therefore, if the combined stochastic-deterministic subspace algorithm on IO

data is applied in the OMAX framework, the basis function must always be zk
in order to avoid bias errors on the damping estimates. In this case other tech-
niques must be utilized to distinguish the physical from the mathematical poles
[135], [108], [35]. Also the LSCF, p-LSCF and LSCE algorithm tend to estimate
more accurate models in the case of noisy data, when the constraint is chosen
to estimate stable stochastic poles. Therefore, it is suggested to choose the one
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.
The only algorithm which is capable of constructing a clear stabilization dia-

gram starting from noisy data, without introducing a bias on the damping ratios,
is the weighted subspace algorithm. Since this algorithm uses the covariance ma-
trix of the noise on the data as a weighting to be consistent all the dynamics in
the stochastic contribution are cancelled by the weighting. This means that no
coupled stochastic-deterministic dynamics need to be modelled and thus the choice
of the basis function does not affect the estimates of the physical poles.
9.3.6 Remarks
• In this chapter it is shown that by choosing an optimal constraint/basis func-

tion, the mathematical poles (used for modelling the noise) become unstable,
while the physical poles are stable. Another possibility, is to transform the
data in such way that the physical poles become unstable and the mathemat-
ical poles remain stable. This can be done by taking the complex conjugate
of the FRFs or by reversing the time axis for time-domain data. In this case,
the constraint for stable mathematical poles leads to an easy interpretation
of the stabilization diagram, since the only unstable poles will be the phys-
ical ones. After the identification the sign of the damping of these unstable
poles can be easily reversed to obtain the true physical poles.
• The influence of the constraint/basis function on the damping ratio estimate

of the poles holds only for discrete models in the time- or frequency-domain.
Therefore, discrete-time models are preferred compared to continuous-time
models in terms of the quality and interpretation of the stabilization dia-
grams. Furthermore the basis in zk or zk−1 must cover uniformly the half
unit circle in the case that real system matrices are estimated and the full
unit circle in the case that complex matrices are estimated. The bi-linear
transformation which exactly maps a discrete-time model on a continuous-
time underlying model does not result in a uniform distribution and thus
does not result in clear stabilization diagrams.
9.4 Application of experimental structural test-

ing
To study the influence of the constraint/basis function on the stabilization diagram

several data sets were analyzed, ranging from ground vibration tests of commercial
and military aircrafts, modal tests of a body-in-white, automotive subparts, a fully
trimmed car, operational data from bridges, ... . In this chapter the influence of
the constraint is now illustrated by four data sets: a car door (low damped data), a
fully trimmed car (highly damped data), in-flight measurements (noisy data) and
the Villa Paso bridge measurement (operational data). In [135] an automated pole
selection procedure is presented based on the clear stabilization charts obtained
by a proper choice of the parameter constraint.
9.4.1 Measurements on a car door
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
(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
(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).
domain subspace identification by replacing z by z −1 in the state-space equations.

Both figures 9.4 and 9.6 show that for this high quality data set (SNR ∼ 30dB) the
synthesized FRFs (synthesized polynomial/state-space model for maximal order)
from the modal parameters are in close agreement with the measurements for both
constraints/basis functions.
Nevertheless, in common practice the LSCE method, subspace identification algo-

rithms and many MPE algorithms are still applied with a constraint/basis function
resulting in stable mathematical poles. In that case, a distinction is made by a
relative comparison between each pole and the corresponding poles for the lower
order estimate. This approach often results in more confusing diagrams and there-
fore an experienced user is often required to decide whether some poles are physical
or mathematical.
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
(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)
9.4.2 Measurements on a fully trimmed car
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.
9.4.3 In-flight aircraft measurements
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
(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
(c) (d)
−90 −90
−100 −100
Ampl. (dB)
−110
Ampl. (dB)
−110
−120
−120
−130
−130
−140
−140
−150
50 60 70 80 90 100 110 120 50 60 70 80 90 100 110

(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).
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
(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
model)
stabilization diagrams becomes difficult. In this case, it is advised to use the ML

and p-ML algorithms [46], [21], which optimize the LSCF and p-LSCF estimates in
a Maximum Likelihood sense. The combined subspace algorithm results in a clear
stabilization diagram for the basis function zk−1 resulting in 7 of the 9 physical
poles, but the damping ratios are largely underestimated as can be seen from the
synthesized FRF in figure 9.11.
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.
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
(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).
9.4.4 Villa Paso bridge
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
In this chapter it is shown that the parameter constraint/basis function in the

identification of modal parameters of the structure play an important role with
respect to the interpretation of the stabilization diagram. A proper choice of this
constraint/basis function forces the mathematical poles to be unstable, allowing a
simple distinction between physical and mathematical poles based on the sign of
the damping value. This methodology is applicable for different modal parameter
estimation methods based on discrete-time models in both frequency- and time-
domain identification. However, for data with large noise levels, the proposed
technique to obtain clear stabilization charts must be applied with caution, since
the damping values will be underestimated. Therefore, the clear stabilization
chart can still be used as a pole selection tool, but the modal parameter estimates
must be further optimized, e.g. in a maximum likelihood sense. The influence
of the constraint/basis function has been illustrated by several experimental case
examples and this for several MPE algorithms
stabilization chart − LSCF

45 45
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
(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
model).
9.6 Appendix 1
Consider a complex value λ = eσ+iω and its counterpart λ△ defined by λ△ =

e−σ+iω . The ratio of the distance between of a complex value zk = eiθk lying on
the unit circle and λ and the distance between zk and λ△ is then independent of
θk .
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
(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
(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)
with Q = Rei(ω−θk ) . This ends the proof.

20 20
model order
model order
15 15
10 10
5 5
4 5 6 7 8 9 10 11 4 6 8 10
(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
(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
This set of equations can be written as
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
(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)
It will be proven that the constraint a0 = 1 leads to a strongly consistent estimate

of a stable system, while the constraint ar = 1 leads to a strongly consistent
estimate of an unstable system. Under the noise assumption that Ek is zero mean
circular complex independent and identically distributed noise with covariance
E (Ek ) = R, it has to be proven that
J˜† Jj + J˜† E → θ w.p. 1 for N → ∞ (9.43)
(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
J˜† E → 0 w.p. 1 for N → ∞ (9.44)

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
(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
(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)
This means that
a.s. lim J˜H E = 0 (9.45)

N →∞
In the case of the constraint a0 = 1, the matrix J˜ is equal to
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
The terms in this

PN last Eq. 9.51 are zero for n + j ∈ N0 and n + j 6= qN with
q ∈ N0 0 , since k=1 zk = 0 (zk (k = 1, 2, . . . N ) covers the full unit circle). In the
case that n + j = qN the terms are also zero since
∞ N ∞
!
j+n
X X X
lim bn R zk = lim bqN −j RN (9.52)
N →∞ N →∞
n=0 k=1 n=0
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.
Chapter 10
Conclusions
10.1 Summary and main contributions
In this thesis, the applicability of frequency-domain identification methods in the

field of modal analysis has been improved and generalized. Special attention has
been paid to deal with noisy data from short measurement sequences and to ob-
tain a maximal data exploitation. The techniques developed in this thesis are
applicable for an Experimental Modal Analysis (EMA) and an Operational Modal
Analysis (OMA) and for a combination of both i.e. an Operational Modal Analysis
with eXogenous inputs (OMAX). Several illustrating examples from automotive,
aerospace and civil engineering are discussed to demonstrate the performance of
the developed techniques. It is now worthwhile to summarize the most important
results achieved in the different chapters.
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.
The non-parametric identification of ABS functions has been revised in chap-

ter 3. Special attention is paid to the reduction of the noise levels, without intro-
ducing bias errors due to leakage. The classic approach to estimate ABS functions
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.
Parametric identification is discussed starting from chapter 4. In chapter 4

both the LSCF and ML common-denominator estimators are extended in order to
take into account the initial/final conditions of every time block averaged by the
H1 FRF estimator to prevent errors caused by leakage. The close analogy between
the implementation of FRF driven and IO driven algorithms is demonstrated and a
combined deterministic-stochastic algorithm is proposed, which, starting from IO
data, can be considered in the OMAX framework. The ML estimators proposed
in this thesis are always consistent under their noise assumptions, however, taking
into account the correct variances of the noise on the data as a frequency-domain
weighting improves only the efficiency. In the case that, no a priori noise infor-
mation is available, different weighting functions are proposed. These new contri-
butions in Modal Parameter Estimation (MPE) based on common-denominator
models are illustrated by simulations and measurement examples. An introduction
to and the current challenges of in-flight flutter analysis on aircrafts are given.
In the multiple input case, common-denominator models do not force the

residues to be rank one, resulting in a loss of quality when decomposing the es-
timated residues into mode shapes and modal participation factors. Therefore,
chapter 5 has been devoted to the generalization of the identification of common-
denominator models to right and left matrix fraction description models. Doing
this, special attention was paid to obtain an optimal memory/computation effi-
ciency. Similar as for the common-denominator models, starting from the normal
equations results in significant reduction of the problem and the exploitation of
the structure of the normal equations results in a fast implementations. Since
the least-squares implementation of the right matrix fraction description model
identification, also referred to as poly-reference LSCF or PolyMAX algorithm, is
not consistent, a ML approach is proposed. Under the assumption of uncorrelated
noise over the different outputs (this assumption only influences the efficiency of
the algorithm and not its consistency) the ML implementation, which uses a the
covariance matrix of the noise on the primary data for each output as a weighting
function, improves the accuracy of the estimates. Furthermore, it is shown how the
uncertainty levels on the final poles can be estimated. Since this ML implementa-
10.1. Summary and main contributions 229
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.
Chapter 6 introduces frequency-domain subspace algorithms to estimate state-

space models from IO data, FRFs or powers spectra. State-space models have
the advantage that the number of outputs No , inputs Ni and order n can be
chosen independently and that they can be converted to a modal model without
introducing an extra error (residues of rank one). Subspace algorithms estimate
the parameters, without the need for minimizing a cost function in an optimization
procedure. By means of geometrical projections, calculated by the use of a QR
decomposition the system matrices are estimated. Novel in this chapter, is the
extension of the state-space model to take into account the intial/final conditions
to deal with transient effects and prevent errors caused by leakage. Based on this
principle a mixed non-parametric/parametric FRF estimator is proposed to remove
leakage errors. This improved measured FRFs can be used to validate an estimated
state-space model by comparison to the synthesized FRFs. Similar to the common-
denominator models, the state-space models can also be extended to start from
FRFs without suffering from transient effects and leakage. These extensions of
the IO and FRF model, together with the validation procedure, are evaluated
by simulations and an experimental example. The proposed frequency-domain
subspace algorithms in literature only estimate the deterministic contribution in
the data and they are not consistent, if no a priori noise information is taken into
account. By using the full covariance matrix on the primary data as a frequency
weighting, the identification of the deterministic model is consistent. However,
this approach becomes impractical for a larger number of outputs and still has no
combined interpretation in the OMAX framework. In chapter 8 this drawback is
solved by the development of a consistent combined frequency-domain subspace
algorithm, without the need for a priori noise information.
In chapter 7 a stochastic frequency-domain subspace algorithm is presented

and proven to be consistent. This estimator starts from output-only spectra to
estimate the natural frequencies, damping ratios and mode shapes. This approach
is able to process short data sequences for the case of in-flight flutter test data
and combines the advantages of frequency-domain algorithms (e.g. a simple band
selection and easy pre-filtering) with the advantages of subspace algorithms (no
need for an optimization procedure).
A combined deterministic-stochastic frequency-domain subspace algorithm is

developed in chapter 8 and proven to be consistent based on its relation to time-
domain subspace algorithms. This algorithm can deal with noisy IO data, FRF
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).
Finally, it is shown that clear stabilization charts are obtained by a proper

choice of the constraint on the parameters for the least-squares problem or by
a proper choice of the basis function of the state-space models for the subspace
algorithms. This allows to make a distinction between mathematical and physical
poles based on the sign of the damping. It is proven that a purely stochastic system
can be modelled by stable poles, unstable poles or by a combination of both.
Depending on the constraint/basis function, the stochastic algorithm identifies in
a consistent way only stable or only unstable poles. This principle, results in
an easy-to-interpret stabilization diagram, facilitating the manual pole selection
and forms the basis for several automated interpretations of the diagram. In this
chapter, it is shown how this principle can be applied for MPE methods such as the
LSCF, poly-reference LSCF, LSCE and subspace algorithms. However, it is shown
that in the case of coupled deterministic-stochastic dynamics the damping values
are underestimated and therefore this principle must be applied with caution.
Several illustrative examples demonstrate the differences between the choices of
the constraints and basis functions on the stabilization diagram and synthesized
spectral functions.
10.2 Future research
The identification approaches in the framework of the OMAX concept presented

in this thesis, consider the influence of the unmeasurable forces to be a stochas-
tic contribution in the vibration response. Nevertheless, in some applications the
unmeasurable forces are deterministic forces e.g. harmonics of rotating machin-
ery and impact forces. Therefore, the modal parameter estimation techniques can
be extended to take into account these contributions. A first approach, where
test measurements are corrupted with harmonics is presented in [127]. In a next
step, these unmeasurable forces can be identified in both magnitude and location
by applying a force identification process [84], simultaneously with the combined
identification technique [128]. New identification techniques can be developed to
consider, all three contributions i.e. unmeasurable deterministic forces, unmea-
surable stochastic forces and measurable forces in a single model. Based on this
combined identified model, the loads can then be estimated and located in a next
10.2. Future research 231
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.
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
[110] T. Söderström and P. Stoica. System Identification. Prentice-Hall, Hemel Hempstead,

Hertfordshire (UK), 1989.
[111] A. Teughels. Inverse modelling of civil engineering structures based on operational modal
data. PhD thesis, Dept. Bouwmechanica, Katholieke Universiteit Leuven, Leuven, Belgium,
2003.
[112] H. Van der Auweraer. Requirements and opportunities for structural testing in view of
hybrid and virtual modelling. In Proceedings of the 25th International Seminar on Modal
Analysis, Leuven (Belgium), Leuven (Belgium), September 2002.
[113] H. Van der Auweraer. Testing in the age of virtual protoyping. In Proceedings of the
International Conference on Structural Dynamics Modelling.Funchal, Funchal, Madeira
(Portugal), 2002.
[114] H. Van der Auweraer. System identification for structural dynamics and vibra-acoustics
design engineering. In Keynote, Proceedings of the 13th IFAC Symposium on System
Identification., Rotterdam (Netherlands), 2003.
[115] H. Van der Auweraer and P. Guillaume. A maximum likelihood parameter estimation
technique to analyse multiple input-multiple output flutter test data. In In Structures and
Materials Panel Specialists, Meeting on Advanced Aeroelastic Testing and Data Analyis,
Rotterdam (The Netherlands ), AGARD, 1995.
[116] H. Van der Auweraer, P. Guillaume, P. Verboven, and S. Vanlanduit. Application of a fast-
stabilizing frequency domain parameter estimation method. Journal of Dynamic Systems,
Measurement, and Control, 123:651–658, 2001.
[117] H. Van der Auweraer and J. Leuridan. Multiple input othogonal polynomial parameter
estimation. Mechanical Systems and Signal Processing, 1(3), 1987.
[118] H. Van der Auweraer, W. Leurs, P. Mas, and L. Hermans. Modal parameters estima-
tion from inconsistent data sets. In Proceedings of the 19th International Modal Analysis
Conference (IMAC20), Kissemmee, Florida (USA), 2000.
[119] H. Van der Auweraer, C. Liefooghe, K. Wyckaert, and J. Debille. Comparative study of
excitation and parameter estimation techniques on fully equipped car. Proceedings of the
11th International Modal Analysis Conference (IMAC11), Kissimmee (FL), USA, pages
627–633, 1993.
[120] P. Van Overschee and B. De Moor. Subspace algorithms for the stochastic identification
problem. Automatica, 29(3):649–660, 1993.
[121] P. Van Overschee and B. De Moor. Continuous time-frequency domain subspace system
identification. Signal Processing, 52(2):179–194, 1994.
[122] P. Van Overschee and B. De Moor. N4sid: Subspace algorithms for the identification of
combined deterministic-stochastic systems. Automatica, 30(1):75–93, 1994.
[123] P. Van Overschee and B. De Moor. Subspace Identification for Linear Systems: Theory-
Implemantation-Applications. Kluwer Academic Publishers, 1996.
[124] P. Van Overschee, B. De Moor, W. Dehandschutter, and J. Swevers. A subspace algo-
rithm for the identification of discrete time frequency domain power spectra. Automatica,
33(12):2147–2157, 1997.
[125] P. Vanhonacker. The Use of Modal Parameters of Mechanical Structures in Sensitivity
Analysis-, System Synthesis- and System Identification Mecthods. PhD thesis, Dept. of
Mechanical Engineering, Katholieke Universiteit Leuven, Belgium, 1988.
[126] S. Vanlanduit. High Spatial Resolution Experimental Modal Analysis. PhD thesis, Dept.
ELEC, Vrije Universiteit Brussel, Belgium, 2001.
[127] S. Vanlanduit, P. Guillaume, and B.Cauberghe. Elimination of background disturbances.
Measurement science and technology, 14:155–163, 2003.
[128] S. Vanlanduit, P. Guillaume, B.Cauberghe, E. Parloo, G. De Sitter, and P. Verboven. On-
line identification of operational loads using exogenous inputs. Under review, Journal of
Sound and Vibration.
240 Bibliography
[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.

PhDBCauberghe PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PhDBCauberghe PDF

Uploaded by

Copyright:

Available Formats

RSIT EIT

APPLIED FREQUENCY-DOMAIN SYSTEM

Promotor: Proefschrift ingediend tot

IN THE FIELD OF EXPERIMENTAL AND OPERATIONAL MODAL ANALYSIS – May 2004

APPLIED FREQUENCY-DOMAIN SYSTEM

Jury: Proefschrift ingediend tot

Last but not least, I am greatly indebted to my family, especially my parents,

Brussels, May 2004

MSRE Mean Square Relative Error

2 Frequency-domain models for dynamical structures 11

2.9.2 Output-Only measurements: a stochastic approach . . . . . 23

3 Non-Parametric preprocessing steps 31

4 Identification of common-denominator models 71

4.3 Weighted Least Squares Complex Frequency-domain (LSCF) esti-

5 Identification of right and left matrix fraction polynomial models107

5.3.1 Linear Least Squares estimator for IO data . . . . . . . . . 121

6 Deterministic Frequency-domain Subspace Identification 135

7 Stochastic frequency-domain subspace identification 153

7.2 A first approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

8 Combined frequency-domain subspace identification 177

9 The secrets behind clear stabilization charts 199

1.1 Research context

More recently, system identification techniques were developed to identify the

In practice, performing a modal analysis typically consists of three basic steps:

excitation and eliminate the need for artificial excitation equipment.

• Structural health monitoring and damage detection: given a reference model

• Load identification: is the inverse problem of response prediction. Given a

• Vibro-acoustics: the existence of vibro-acoustical coupling between the struc-

Nevertheless, the current evolution in mechanical engineering towards the use of

1.2 Focus, outline and organization of the thesis

In this thesis, several dedicated extensions of existing frequency-domain modal

• Reduction of the test time.

Instead of estimating the modal parameters directly from the measurements,

Data driven ABS driven

Mathematical models for EMA, OMA and OMAX

Frequency-domain System Identification

Cost function based Subspace Algorithms

Common-denominator Deterministic identification

Figure 1.1: Organization of the text.

attention is paid to obtain the noise information on these ABS functions. It is

During the last years, several MPE algorithms based on a common-denominator

In chapter 5 a generalization of the fast common-denominator algorithms to

Chapter 6 starts with the introduction of frequency-domain subspace algo-

Until recently, frequency-domain subspace identification was limited to esti-

Next, the extension to a combined deterministic-stochastic frequency domain

subspace algorithm is closely related to its time-domain equivalent. In the case

Finally, in chapter 9, the secret behind clear stabilization charts is revealed. By

1.3 Original contributions in this work

• The introduction of a combined EMA/OMA framework as Operational Modal

count the initial/final conditions to prevent the introduction of bias errors

• The presentation of a combined deterministic-stochastic algorithm based on

• The development of a scalar and matrix implementation of the maximum

• The extension of frequency-domain state-space models to include transient

• The development of a consistent stochastic frequency-domain subspace al-

• The development of frequency-domain counterparts of the combined determi-

• The mathematical explanation behind the construction of clear stabilization

• Application of all these contributions for several experimental test cases in

2.2 The Modal Model

Newton’s equations of motion for a finite-dimensional linear structure are a set of

M ÿ(t) + C1 ẏ(t) + Ky(t) = f (t) (2.1)

with M , C1 and K ∈ R Nm ×Nm respectively the mass, damping and stiffness

Z(s)Y (s) = F (s) (2.2)

with the dynamical stiffness Z(s) = M s2 + C1 s + K and s = jω. Inverting Eq.