Frequency-Domain System Identification For Modal Analysis

BR
EL
VRIJ
E
RSIT EIT
S
US
VE
NI
U
V IN
NT
EB
RAS
SCI E
IA
CERE
TE
VRIJE UNIVERSITEIT BRUSSEL

FACULTEIT TOEGEPASTE WETENSCHAPPEN
VAKGROEP WERKTUIGKUNDE
Pleinlaan 2, B-1050 Brussels, Belgium
FREQUENCY-DOMAIN SYSTEM IDENTIFICATION

FOR MODAL ANALYSIS
Peter VERBOVEN
Promotor:
Prof. dr. ir. P. Guillaume
Copromotor:
Prof. dr. ir. M. Van Overmeire
Proefschrift ingediend tot

het behalen van de academische
graad van doctor in de
toegepaste wetenschappen
May 2002
BR
EL
VRIJ
E
RSIT EIT
S
US
VE
NI
U
V IN
NT
EB
RAS
SCI E
IA
CERE
TE
VRIJE UNIVERSITEIT BRUSSEL

FACULTEIT TOEGEPASTE WETENSCHAPPEN
VAKGROEP WERKTUIGKUNDE
Pleinlaan 2, B-1050 Brussels, Belgium
FREQUENCY-DOMAIN SYSTEM IDENTIFICATION

FOR MODAL ANALYSIS
Peter VERBOVEN
Jury:
Prof. dr. ir. D. Lefeber, voorzitter
Prof. dr. ir. J. Vereecken, vice-voorzitter
Prof. dr. ir. P. Guillaume, promotor
Prof. dr. ir. M. Van Overmeire, copromotor
Prof. dr. ir. B. De Moor (KULeuven)
Prof. dr. ir. J. Schoukens
Prof. dr. ir. J. Swevers (KULeuven)
Dr. ir. H. Van der Auweraer (LMS International)
Dr. lic. S. Vanlanduit
May 2002
Proefschrift ingediend tot

het behalen van de academische
graad van doctor in de
toegepaste wetenschappen
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.
Voorwoord
Tijdens mijn ruim vier jaren van onderzoek aan de vakgroep Werktuigkunde heb
ik de gelegenheid gehad met vele mensen samen te werken. Graag zou ik dan ook
iedereen willen bedanken die rechtstreeks of onrechtstreeks heeft bijgedragen tot
de realisatie van dit doctoraatswerk.
In de eerste plaats dank ik mijn promotor, prof. dr. ir. Patrick Guillaume en
copromotor prof. dr. ir. Marc Van Overmeire. Alleerst omdat beiden me motiveerden om na mijn ingenieursstudies het doctoraat te beginnen en me het nodige
vertrouwen en steun hebben gegeven om in een vrije en inspirerende omgeving
aan onderzoek te doen. De nauwe samenwerking met prof. dr. ir. Patrick Guillaume was een uitermate leerrijke en stimulerende ervaring en heeft me vele nieuwe
inzichten en methodes voor onderzoek bijgebracht. Niet alleen was hij een aangename bureaugenoot, maar bracht hij me ook de basis van multivariabele systeemidentificatie en modale analyse bij. De vele technische discussies en zijn assistentie
bij de experimenten hebben mede geleid tot het verhogen van het wetenschappelijk niveau van het geleverde werk. Tevens was het een voorrecht om zijn eerste
doctoraatsstudent te zijn.
Daarnaast dank ik de leden van de jury voor hun enthousiasme en interesse in
mijn onderzoek en hun kostbare tijd die zij besteedden aan het nalezen van dit
proefschrift. Ik hoop dat, door de realisatie van dit doctoraat, de reeds bestaande
contacten die ik met elk van hun heb kunnen leiden tot de verdere ontwikkeling
van nieuwe onderzoeksideeen en mogelijke samenwerking.
Mede doordat het onderwerp van dit onderzoek direct gerelateerd was met een
aantal onderzoeksprojecten binnen de onderzoeksgroep, heb ik de kans gehad om
met verschillende bedrijven in contact te komen. Dit liet toe van de toepasbaarheid
van de onderzoeksresultaten te toetsen voor vele practische probleemstellingen.
Daarom dank ik meer specifiek LMS International en ASCO aero industries voor
de nauwe en constructieve samenwerking en het verschaffen van de vele datasets
en teststructuren.
Prof. dr. ir. J. Schoukens en prof. dr. ir. Rik Pintelon van de vakgroep Al-
ii
gemene Elektriciteit en Instrumentatie ben ik erkentelijk voor het delen van hun
enorme kennis op het gebied van frequentiedomein systeemidentificatie en hun leidinggevende rol voor het initieren van nieuwe ideeen welke ik kon toepassen binnen
het domein van de modale analyse.
Verder wil ik Prof. dr. ir. D. Lefeber, vakgroepsvoorzitter en alle medewerkers
van de vakgroep Werktuigkunde bedanken voor de toffe en stimulerende werksfeer. Hierbij denk ik dan vooral aan de directe collegas van de onderzoeksgroep
Eli Parloo, Bart Cauberghe, Steve Vanlanduit en Gert De Sitter voor hun enorme
medewerking en het samen publiceren van onderzoekswerk alsmede de vele inspirerende discussies over de mogelijke toepassing van mijn resultaten binnen hun
onderzoek. Ook een speciaal woord van dank voor de secretaresse Carine Vaeremans voor de vele administratieve taken gedurende afgelopen jaren, Frank Daerden
voor het oplossen van vele van mijn LaTeX vragen, Andre Plasschaert en JeanPaul Schepens voor de uitmuntende technische bijstand in het labo en Thierry
Lenoir voor de steeds ongewenste informatica problemen. Vele collegas waren
eveneens vrienden waarmee ik ook naast het werk prettige tijden heb beleefd.
Tevens ben ik het Instituut voor de aanmoediging van Innovatie door Wetenschap en Technologie in Vlaanderen (IWT-Vlaanderen) en VUB Onderzoeksraad
erkentlijk voor de financiele steun.
Tenslotte wil ik benadrukken dat de realisatie van dit proefschrift slechts mogelijk was dankzij de onvoorwaardelijke steun, begrip en raad van mijn ouders
op elke mogelijke wijze. Helaas dat mijn vader niet meer kan genieten van dit
moment, waarop ik dit werk succesvol heb kunnen beeindigen. Aan hem en mijn
moeder draag ik dit werk op. Daarnaast wens ik ook mijn vriendin Els, hoewel
ze de laatste maanden niet veel aandacht van mij kreeg, te bedanken voor haar
voortreffelijke steun en begrip.
Peter Verboven
Brussel, Mei 2002
Abstract
Experimental modal analysis has become a commonly-used technique for studying
the dynamical behaviour of mechanical and civil structures, such as for example
cars, aircrafts, bridges, offshore platforms and industrial machinery. During a
modal test, both the applied forces and vibration responses of the structure are
measured when excited in one or more locations. Based on this data, a modal
model of the structure, that essentially contains the same information as the original vibration data, is derived by means of system identification. The appropriateness of the modal model for physical interpretation makes that scientists and
engineers most often prefer this type of model for their structural dynamics research. Recent advances in the field of frequency-domain system identification are
the development of the so-called stochastic Total Least Squares and Maximum
Likelihood methods. Since the noise on the measured data is taken into account
during the estimation process, these methods yield very accurate parameter estimates with confidence bounds, even in the case of a poor quality of the data.
Recently, the need for improved system identification methods within the domain of modal analysis results from the broadening field of applications, e.g. damage detection and vibro-acoustics, and the increased complexity of todays tested
structures. The aim of this thesis was to use the stochastic methods as a starting
point to develop more advanced modal parameter identification techniques. This
however required an extensive revision of both the initial formulation of these
methods with respect to computational efficiency and the modal analysis process
with respect to the integration of the noise on the measured data.
First of all, an improved nonparametric processing has been proposed in order
to obtain accurate frequency response function (FRF) estimates as well as the
uncertainty on this data, required by the stochastic methods. Based on an errorsin-variables framework a methodology was derived that yields FRF estimates with
maximum likelihood properties and at the same time the noise covariance matrix.
Comparing this so-called instrumental variables approach with the current stateof-the art, pinpoints the important benefits of this method, combining accuracy
and computational efficiency during the nonparametric data processing. At the
same time, by using signal windowing techniques, this approach has been optimized
iii
iv
to reduce the effects of leakage resulting from the use of random noise excitation
signals.
The adaptation of the stochastic Total Least Squares and Maximum Likelihood methods for modal analysis starts from the numerical optimization of the
linear Least Squares formulation. Different types of parameterization were assessed to obtained an improved numerical conditioning for a high modal density
as well as computational efficiency for large data sets. A normal equations formulation using a discrete-time common-denominator model with complex coefficients
demonstrates an important robustness for high model orders up to say 100 modes
and more, while different representations for a stabilization diagram are possible.
Based on the fast formulation and the knowledge of the noise on the FRF data,
the extension of the optimized Least Squares formulation to the stochastic Total
Least Squares and Maximum Likelihood methods is possible. In the case that
only very limited data sequences are available, it is preferable to start from the
Input/Output Fourier data for reasons related to both leakage and frequency resolution. It has been demonstrated that in this case, the efficient formulation of the
Least Squares problem is still possible, although the generalization to the stochastic Total Least Squares methods is not straightforward anymore. Moreover, the
Input/Output methods allow the parametric compensation for effects of leakage,
possibly present in the short data sequences.
Based on the improved accuracy of the parameter estimates and the availability
of confidence bounds, stochastic model validation criteria and the automation
of the modal parameter extraction are developed. This results in an important
reduction of the user-time as well as the possibility for non-expert users to apply
modal analysis techniques. At the same time, a method for tracking the modal
parameters over subsequent monitoring instances is proposed, providing a basic
tool for structural health monitoring applications.
The application of the proposed system identification methods to many reallife and benchmark case studies clearly illustrates that these techniques perform
very well and indicates their great potential within the domain of modal analysis.
Korte Inhoud
Experimentele modale analyse is geevolueerd tot een algemeen gebruikte techniek
voor het bestuderen van het dynamisch gedrag van mechanische en bouwkundige
structuren, zoals bijvoorbeeld autos, vliegtuigen, bruggen, boorplatformen en industriele machines. Tijdens een modale test worden beide de krachten, aangelegd
in een of meerdere lokaties, en de trillingsresponses opgemeten. Op basis van deze
data wordt vervolgens een modal model van de structuur, dat in essentie dezelfde
informatie bevat als de oorspronkelijke data, bekomen aan de hand van systeemidentificatie. De uitermate geschiktheid van dit model voor fysiche interpretatie
maakt dat wetenschappers en ingenieurs dit model meestal verkiezen voor structuurdynamica onderzoek. Recente vooruitgang binnen het domein van de frequentiedomein systeemidentificatie zijn de ontwikkeling van de zogenaamde stochastische
Total Least Squares en Maximum Likelihood methodes. Aangezien de onzekerheid op de meetdata in beschouwing wordt genomen tijdens het schattingsproces, resulteren deze methodes in een zeer nauwkeurige schatting van de parameters
samen met onzekerheidsintervallen, zelfs in het geval dat de meetdata van slechte
kwaliteit is.
Gedurende de laatse jaren is er een toenemende behoefte aan verbeterde systeemidentificatie methods binnen het domein van modale analyse als gevolg van het
verruimde toepassingsgebied, zoals bijvoorbeeld schadedetectie en vibro-akoestiek,
en de toenemende complexiteit van de huidig bestudeerde structuren. Het doel van
deze thesis was, vertrekkende van de stochatische methodes, geavanceerde modale
parameter identificatietechnieken te ontwikellen. Dit vereiste echter een uitgebreide doorlichting van zowel de oorspronkelijke formulering van deze methodes
met nadruk op de numerieke aspecten als het modal analyse proces met betrekking
tot het integreren van de meetruis.
Alleerst is een verbeterde niet-parametrische verwerking voorgesteld voor de
nauwkeurige schatting van frequenctie respons functies (FRF) alsmede de onzekerheid op deze data, vereist voor de stochastische methodes. In het errors-invariables kader is het mogelijk een methodologie af te leiden voor het bepalen van
FRF schattingen met maximum likelihood eigenschappen en terzelfdertijd een
ruisanalyse voor de meetdata. Vergelijking van deze zogenaamde instrumentele
v
vi
variabelen aanpak met de huidige state-of-the-art benadrukt de belangrijke
voordelen van deze methode welke nauwkeurigheid en numerieke efficientie combineert gedurende de niet-parametrische verwerking. Verder is deze aanpak, door de
toepassing van signaal vensters, geoptimaliseert voor het reduceren van leakage
effecten als gevolg van het gebruik van ruis excitatiesignalen.
De aanpassing van de stochastische Total Least Squares en Maximum Likelihood methodes voor modale analyse start van de numerieke optimizatie van de
lineaire Least Squares formulatie op basis van FRF data. Verscheidene parameterizaties zijn geevalueerd voor enerzijds een verbeterde numerieke conditionering
voor een hoge modale densiteit en anderzijds een gereduceerde rekentijd voor uitgebreide datasets. Een formulering vertrekkende van de normaal vergelijkingen
gebruikmakende van een discreet-tijdsdomein common denominator model met
complexe coefficienten blijkt zeer robuust voor een hoge model orde van 100 modes
en meer. Op basis van deze formulatie en de kennis van onzekerheid op de data,
volgt eveneens de numerieke optimizatie van de stochastische methodes. In geval
dat slechts zeer korte datasequenties zijn opgemeten, is het echter verkiesbaar
de modale parameters te schatten rechstreeks vanaf de ingang/uitgang Fourier
sequenties. Hoewel een efficiente formulatie van het Least Squares probleem
nog steeds mogelijk is, is de veralgemening naar de stochastische Total Least
Squares niet meer vanzelfsprekend. Verder, maken de ingang/uitgang methodes
een parametrische compensatie voor leakage effecten, vaak aanwezig in de korte
datasequenties, mogelijk.
Op basis van de verbeterde nauwkeurigheid van de geschatte parameters en
de bijhorende onzekerheidsintervallen, zijn stochastische mode validatie criteria
and een geautomatiseerde parameter schatting ontwikkeld. Dit resulteert in een
belangrijke reductie van de gebruikertijd alsmede de mogelijkheid voor niet-expert
gebruikers modale analyse toe te passen. Terzelfdertijd, is een methode voorgesteld
voor de opvolging van de modale parameters als functie van de tijd.
In deze thesis worden het nut en potentieel van de voorgestelde systeemidentificatie methodes voor modale analyse duidelijk gellustreerd door de toepassing
van deze algoritmen op talrijke simulatie en praktische datasets.
Nomenclature
List of Operators
i
Y
()T
()1
()
()H
()
()T
()H
()
Im
ij
B[i]
A[i,j]
A[:,j]
A[i,:]
X [m] (k)
vec(A)
Pn
trace(A) = i=1 A[i,i]
det(A)
diag(A1 , A2 , . . . , AL )
rank(A)
kAkF =
trace(AH A)
i2 = 1
convolution operator
outline uppercase font denotes a set, e.g R, C
are, respectively, real and complex numbers
matrix transpose
matrix inverse
complex conjugate
Hermitian transpose
complex conjugate transpose of matrix
Moore-Penrose pseudo-inverse
transpose of the inverse matrix
Hermitian transpose of the inverse matrix
orthogonal complement of a subspace or matrix
(m m) identity matrix
Kronecker matrix product
Kronecker delta
ith entry of vector B
i, jth entry of matrix A
jth column of matrix A
ith row of matrix A
mth realization of a stochastic process X(k)
a column vector formed by stacking the columns
of the matrix A on top of each other
trace of an (n n) matrix A
determinant of matrix A
block diagonal matrix with blocks Al , l = 1, . . . , L
rank of (n m) matrix A, maximum number of
linear independent rows (columns) of A
Frobenius norm of an (n m) matrix A
vii
viii
(A)
(A)
(A)
Re()
Im()
herm(A) = (A + AH )/2
ARE
p
|x| = (Re(x))2 + Im(x))2
0
E{X}
bX = X E{X}
Cov(X, Y )
covar(x, y)
eigenvalue of a square matrix A

singular value of an (n m) matrix A
condition number of an (n m) matrix A
(A) = (maxi i (A)/mini i (A))
real part of
imaginary part of
Hermitian
symmetric
part of an (n m) matrix A
Re(A)
ARE =
Im(A)
absolute value of a complex number x
true value of
estimated value of
mathematical expectation of stochastic variable X
bias of the estimate X
cross-covariance matrix of X and Y
Cov(X, Y ) = E{(X E{X})(Y E{Y })H }
covariance of scalars x and y
covar(x, y) = E{(x E{x})(y E{y}) }
variance of a scalar x
var(x) = E{|x E{x}|2 }

CX = Cov(X) = Cov(X, X)
CXY = Cov(X, Y )
x2 = var(x)
x = E{x}
mean value of a scalar x
MSE()
mean square error of the estimate X
= E{( 0 )( 0 )H }
MSE()
H
Rxx ( ) = E{x(t)x (t )} auto-correlation function of x(t)
Rxy ( ) = E{x(t)y H (t )} cross-correlation function of x(t) and y(t)
GXX (i) = GX (i)
Fourier transform of Rxx ( )
(auto-power spectrum of x(t))
GXY (i)
Fourier transform of Rxy ( )
(cross-power spectrum of x(t) and y(t))
a.s.lim
almost sure limit, limit with probability one
p.lim
limit in probability
N (, )
Gaussian distribution with
mean and standard deviation
ix
List of Symbols
Ni
No
n
Nm
t
Ts
tn
a[n]
Ntot
Ns
M
fs
Nf
f
= 2f
f
U (eiTs )
s = i
sf
z = esTs
zf
h(t)
(t)
H(i)
H(s)
H1
H2
Hv
Hiv
Hev
Hgtls
r(t), s(t), f (t), x(t)
ns (t), nf (t), nx (t)
R(i), S(i), F (i), X(i)
NS (i), NF (i), NX (i)
Z(f ) = [F (f )H X(f )H ]H
M ,C,K
number of inputs
number of outputs
number of parameters
number of modes
continuous-time variable
sampling period
discrete time variable tn = nTs
nth sample of the discrete-time signal a(tn )
total number of time samples
number of time samples within one record
number of experiments
sampling frequency
number of DFT frequencies
frequency
angular frequency
angular frequency at frequency f
Discrete Fourier transform of u(f Ts )
Laplace transform variable
Laplace transform variable at DFT frequency f
z-domain transform variable
z-domain transform variable at DFT frequency f
zf = eif Ts
impulse response function
Dirac delta
frequency response function
transfer function in the Laplace domain
noise-on-output estimator
noise-on-input estimator
Total Least Squares estimator
Instrumental Variables estimator
Errors-in-Variables estimator
or Hiv using periodic excitation
Generalized Total Least Squares estimator
generator, shaker, input, output time signals
disturbing time-domain noise on
the shaker, input and output time signals
Fourier transform of r(t), s(t), f (t) and x(t)
Fourier transform of ns (t), nf (t), nx (t)
data vector containing the measured input
and output DFT spectra at frequency f
mass, damping and stiffness matrices
column vector of model parameters
f
Pn a
A(, ) = j=1
a j
Pn b j j
B(, ) = j=1 bj
n
p
zoi
fd
L
UR
LR
generalized transform variable

Laplace domain = s, Z-domain = z 1
evaluated at DFT frequency f
f = if (Laplace), f = eif Ts (Z)
denominator polynomial expanded in basis
numerator polynomial expanded in basis
order of the system
system pole
transfer function zero
damped natural frequency fd = Im(p)
2
damping ratio = Re(p)
|p|
residue matrix R = LT
mode shape vector
participation factor vector
Upper Residual matrix
Lower Residual matrix
xi
List of Abbreviations
ARMA
ARMAX
ARX
BL
BTLS
CMIF
CRB
DFT
DOF
ERA
EV
EVD
FEM
FFT
FRF
FIR
FDPI
GSVD
GTLS
IFFT
I/O
IRF
ITD
IV
IWLS
IQML
LOG
LS
LSCE
LSFD
LTI
MAC
MDOF
MIMO
MISO
ML
MOV
MPC
MPD
MPE
MQI
NLS
PDF
AutoRegressive Moving Average

AutoRegressive Moving Average with eXternal input
AutoRegressive with eXternal input
Band Limited
Bootstrapped Total Least Squares
Complex Mode Indicator Function
Cramer Rao lower Bound
Discrete Fourier Transform
Degree Of Freedom
Eigensystem Realization Algorithm
Errors-in-Variables
Eigen Value Decomposition
Finite Element Method
Fast Fourier Transform
Frequency Response Function
Finite Impulse Response
Frequency-domain Direct Parameter Identification
Generalized Singular Value Decomposition
Generalized Total Least Squares
Inverse Fast Fourier Transform
Input/Output
Impulse Response Function
Ibrahim Time-Domain
Instrumental Variables
Iterative Weighted Least Squares
Iterative Quadratic Maximum Likelihood
LOGarithmic
Least Squares
Least Squares Complex Exponential
Least Squares Frequency-Domain
Linear Time Invariant
Modal Assurance Criterium
Multiple Degree of Freedom
Multiple Input Multiple Output
Multiple Input Single Output
Maximum Likelihood
Mode OVercomplexity
Modal Phase Collinearity
Mean Phase Deviation
Modal Parameter Estimation
Mode Quality index
Nonlinear Least Squares
Probability Density Function
xii
PEM
PLSCE
PSD
PS
rms
RLDS
SDOF
SISO
SLDV
SNR
SNS
SVD
TDPI
TLS
WGTLS
WLS
ZOH
Prediction Error Method

Polyreference Least Squares Complex Exponential
Power Spectral Density
Power Spectrum
Root Mean Square
Related Linear Dynamic System
Single Degree of Freedom
Single Input Single Output
Scanning Laser Doppler Vibrometer
Signal-to-Noise Ratio
Stochastic Noise Source
Singular Value Decomposition
Time-domain Direct Parameter Algorithm
Total Least Squares
Weighted Generalized Total Least Squares
Weighted Least Squares
Zero Order Hold
Contents
Voorwoord
Abstract
iii
Korte Inhoud
Nomenclature
vii
Contents
xiii
1 Introduction
1.1
Research Context . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2
Focus and Outline of the Thesis . . . . . . . . . . . . . . . . . . . .
2 FRF Identification in the Presence of Measurement Noise
11
2.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
2.2
Errors-in-Variables Stochastic Framework . . . . . . . . . . . . . .
13
2.3
Hgtls a Maximum Likelihood Approach . . . . . . . . . . . . . . .
16
2.4
Practical Issues of EMA . . . . . . . . . . . . . . . . . . . . . . . .
18
2.4.1
Modal Testing . . . . . . . . . . . . . . . . . . . . . . . . .
18
2.4.2
Effects of Environmental Noise . . . . . . . . . . . . . . . .
20
FRF identification for EMA . . . . . . . . . . . . . . . . . . . . . .
22
2.5.1
Hv TLS estimator . . . . . . . . . . . . . . . . . . . . . .
22
2.5.2
H1 and H2 LS estimators . . . . . . . . . . . . . . . . . .
23
2.5.3
Hiv estimator . . . . . . . . . . . . . . . . . . . . . . . . . .
25
2.5.4
Hev estimator Hiv Using Periodic Excitation . . . . . . .
28
2.5
xiii
xiv
Contents
2.6
Effects of Non-linear Distortions on FRF Identification . . . . . . .
31
2.7
Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
2.8
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
2.9
Appendix A The GTLS Estimator per Output . . . . . . . . . .
51
E . . . . . . . . . . . . . .
2.10 Appendix B Unbiased Estimate of G
X
51
3 FRF Identification in the Presence of Leakage
53
3.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
3.2
Simulated Vibration Experiment . . . . . . . . . . . . . . . . . . .
55
3.3
Power Spectra Estimators . . . . . . . . . . . . . . . . . . . . . . .
56
3.3.1
Periodogram PS Estimator . . . . . . . . . . . . . . . . . .
57
3.3.2
Correlogram PS Estimator . . . . . . . . . . . . . . . . . .
59
Exponential Windowing of FRF Estimates . . . . . . . . . . . . . .
64
3.4.1
Hexp FRF Estimator . . . . . . . . . . . . . . . . . . . . . .
64
3.4.2
Nonparametric Least Squares IRF Estimator . . . . . . . .
67
3.5
Iterative Nonparametric FRF Estimator . . . . . . . . . . . . . . .
69
3.6
Cyclic Averaging Approach . . . . . . . . . . . . . . . . . . . . . .
70
3.7
Comparison of Methods . . . . . . . . . . . . . . . . . . . . . . . .
72
3.8
Experimental Case Study . . . . . . . . . . . . . . . . . . . . . . .
73
3.9
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
75
3.4
4 Introduction to Modal Parameter Estimation
85
4.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
86
4.2
Parametric Models . . . . . . . . . . . . . . . . . . . . . . . . . . .
86
4.3
Overview of MPE Algorithms . . . . . . . . . . . . . . . . . . . . .
90
4.4
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
96
5 Frequency-domain MPE from FRF Data
99
5.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.2
Parametric model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.3
(Weighted) Linear Least Squares . . . . . . . . . . . . . . . . . . . 102

5.3.1
LS Formulation based on Jacobian Matrix . . . . . . . . . . 102
Contents
5.4
5.5
xv
5.3.2
LS Formulation based on Normal Matrix . . . . . . . . . . 107
5.3.3
Optimal LS Implementation for MPE . . . . . . . . . . . . 109
5.3.4
Transformation to Modal Model . . . . . . . . . . . . . . . 124
5.3.5
Constructing Stabilization Charts
5.3.6
LSCF Estimator . . . . . . . . . . . . . . . . . . . . . . . . 129
5.3.7
LS Residue Estimator . . . . . . . . . . . . . . . . . . . . . 130
5.3.8
Comparison of LSCF with LSCE . . . . . . . . . . . . . . . 131
. . . . . . . . . . . . . . 125
Weighted Generalized Total Least Squares . . . . . . . . . . . . . . 133

5.4.1
Cost Function Formulation . . . . . . . . . . . . . . . . . . 133
5.4.2
Generalized Eigenvalue Problem Formulation . . . . . . . . 135
Maximum Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . 137

5.5.1
ML Equations . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.5.2
Parameter Uncertainty Bounds . . . . . . . . . . . . . . . . 140
5.5.3
ML Stabilization Chart . . . . . . . . . . . . . . . . . . . . 141
5.5.4
Logarithmic ML . . . . . . . . . . . . . . . . . . . . . . . . 141
5.6
Choice of Frequency Weighting . . . . . . . . . . . . . . . . . . . . 142
5.7
Validation of Stochastic Estimators for Experimental Data . . . . . 148
5.8
Asymptotical Properties of Stochastic Estimators . . . . . . . . . . 148
5.9
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5.10 Appendix A Jacobian and Normal Matrix based LS Algorithms . 152

5.11 Appendix B Fast Calculation of Normal Matrix Entries . . . . . 153
5.12 Appendix C Equivalency of WGTLS Solutions . . . . . . . . . . 154
6 Frequency-domain MPE from Input/Ouput Data
155
6.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6.2
Parametric Model
6.3
Errors-in-Variables Noise Model . . . . . . . . . . . . . . . . . . . . 158
6.4
(Weighted) Linear Least Squares . . . . . . . . . . . . . . . . . . . 158
6.5
Weighted Generalized Total Least Squares . . . . . . . . . . . . . . 161
. . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.5.1
Cost Function Formulation . . . . . . . . . . . . . . . . . . 161
6.5.2
Generalized Eigenvalue Problem Formulation . . . . . . . . 161
xvi
Contents
6.5.3
Validation of Linear Approximation . . . . . . . . . . . . . 164
6.6
Validation of (WG)TLS for Experimental Data . . . . . . . . . . . 167
6.7
Maximum Likelihood Estimator . . . . . . . . . . . . . . . . . . . . 171
6.8
MPE in the Presence of Leakage Phenomena . . . . . . . . . . . . 173
6.9
MPE from Auto and Cross Power Spectra . . . . . . . . . . . . . . 175
6.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

7 Automated Modal Parameter Estimation and Tracking
181
7.1
Overview of State-of-the-Art . . . . . . . . . . . . . . . . . . . . 182
7.2
Autonomous Modal Parameter Identification . . . . . . . . . . . . 183
7.3
7.4
7.5
7.6
7.2.1
Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 184
7.2.2
Mode Validation Criteria . . . . . . . . . . . . . . . . . . . 184
7.2.3
Physical Mode Selection . . . . . . . . . . . . . . . . . . . . 187
Case Study I: Airbus A320 Slat Track . . . . . . . . . . . . . . . . 189

7.3.1
Test Structure and Experiments . . . . . . . . . . . . . . . 189
7.3.2
Autonomous Estimation Results for Slat Track . . . . . . . 192
7.3.3
Case Study II: I40 Bridge . . . . . . . . . . . . . . . . . . . . . . . 199

7.4.1
Test Structure and Experiments . . . . . . . . . . . . . . . 199
7.4.2
Model Order Reduction . . . . . . . . . . . . . . . . . . . . 200
7.4.3
Autonomous Estimation Results for I40 Bridge . . . . . . . 203
7.4.4
Mode Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

7.5.1
Tracking Approach . . . . . . . . . . . . . . . . . . . . . . . 206
7.5.2
Tracking Algorithm Problem of Missing Modes . . . . . . 208
7.5.3
Tracking Results for Slat Track . . . . . . . . . . . . . . . . 209
7.5.4
Damage Assessment during Tracking . . . . . . . . . . . . . 212
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
8 Conclusions
8.1
217
Thesis Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Contents
8.2
xvii
Ideas for Further Research . . . . . . . . . . . . . . . . . . . . . . . 219
Bibliography
221
Chapter 1
Introduction
The first chapter contains the general introduction of and motivation for this thesis.
First the research context experimental modal analysis and its applications are
discussed in Section 1.1. The focus and main contributions as well as the outline
of the thesis are presented in Section 1.2.
1.1
Chapter 1. Introduction
Research Context
During the last two decades, there has been a growing interest for the domain
of modal analysis. Evolved from a simple technique for trouble shooting, modal
analysis has become an established technique to analyze the dynamical behaviour
of complex mechanical structures. Important examples are found in the automotive (cars, trucks, motorcycles), railway, maritime, aerospace (aircrafts, satellites,
space shuttle), civil (bridges, buildings, offshore platforms) and heavy equipment
industry.
The vibration of a structure originates from its resonance modes that are inherent properties of the structure. Small forces exciting one or more of these
modes can result in an important deformation and possibly damage in the structure. These modes are determined by the material properties (mass, damping
and stiffness) and the boundary conditions of the structure. If either the material
properties or boundary conditions change, the structural modes change.
The modal model expresses the behaviour of a linear time-invariant mechanical
system as a linear combination of these different resonance modes. Each mode is
defined by a damped resonance frequency, modal damping and mode shape, i.e.
the so-called modal parameters. Given the physical interpretation of the modes,
engineers often prefer this mathematical model.
When the material properties and boundary conditions of a structure are fairly
known, the finite element method (FEM) can be used to compute the modal
parameters. The structure is discretized into a finite number of geometric elements
that are defined by the known material properties and boundary conditions. Based
on this information, the mass and stiffness matrices can be calculated [93, 27, 152],
from which the modal parameters are derived. During the last decade, finite
element analysis has been integrated in the design process since it can significantly
reduce the number of prototypes and expensive experiments and at the same time
reduce the time-to-market. This was mainly possible thanks to the availability
of extensive computation power. However, for complex structures the material
properties and boundary conditions are often not well-known, especially in the
case multiple materials with anisotropic properties such as composites. In addition,
including general damping in the finite element analysis is still a main topic for
research.
Another approach to obtain the modal model is the so-called Experimental
Modal Analysis (EMA), which is based on the measured forces and vibration
responses of the structure when excited in one or more locations. Since this generally results in large amounts of data, there is the need to compress the amount
of data by estimating an experimental parametric model of the studied structure
that essentially contains the same information as the original vibration data. More
general, the process of finding a model from the data is called system identification,
1.1. Research Context
originating from the domain of electrical engineering, where [85] is a prominent

reference. Recent advances in the field are the development of subspace methods
[141] and frequency-domain maximum likelihood methods [109]. The application of
system identification to identify the experimental modal model of vibrating structures, has been important research domain in mechanical engineering, especially
during the last two decades. Moreover, as a direct result of the emerging information technology significant advances have been available to both modal test and
analysis equipment, explaining the important gain of interest for EMA. A number of textbooks gives a good overview of the theory and current practice in the
domain of experimental modal analysis [62, 36, 88].
Consider a modal experiment, where the structure is excited at Ni locations
with a predefined signal while measuring the excitation forces f (t) at the input
locations and the structural responses (displacements, velocities, accelerations)
x(t) at No output locations. For a linear time-invariant system the input-output
relation is then given by
x(t) = h(t) f (t)
(1.1)
In the frequency-domain the convolution is replaced by a matrix multiplication
X() = H()F ()
(1.2)
where h(t) and H() are respectively the No Ni impulse response and frequency
response matrices, which can be decomposed as [62]
h(t) = et L
H() = (jI )
(1.3)
1
(1.4)
The 2Nm columns (Nm number of modes in the model) of the modal matrix
contain the mode shapes and the complex conjugates of the mode shapes, while
the modal participation matrix L is a measure for the contribution of the different
input forces to the total vibration. The diagonal matrix contains the complex
system poles from which damped resonance frequencies fn and damping ratios n
can be computed using (n = 1, . . . , Nm )
fn =
1
Im((n, n))
2
(1.5)
n =
Re((n, n))
| (n, n) |
(1.6)
In practice, an experimental modal analysis typically consists of three phases:

1. Performing the experiments: concerns the experimental setup (i.e. actuators
and sensors, boundary conditions,...) and the acquisition parameters, (i.e.
instrumentation, excitation signal, frequency band and resolution, number
of records,...).
2. Processing of the measured data: the measured time signals are first converted into Impulse Response (IRF) or Frequency Response Functions (FRF)
using signal processing and nonparametric identification techniques. Next,
based on this IRF or FRF data, the modal parameters , and L can be
extracted in the time-domain (cf. Eq. 1.3) or the frequency-domain (cf.
Eq. 1.4) using a parametric identification approach.
3. Validating the model : the extracted modal model must be assessed for its
physical representation of the dynamical behaviour of the structure in the
studied frequency band.
For engineers, the modal modal is very often not the final goal, but only a
means to get condensed experimental data used for a wide range of applications
based on a modal model:
Model Updating: since the reliability of initial finite element models is often
not guaranteed, modal updating (correlation, correction and verification) by
means of experimental data is necessary. This results in a finite element
model that is more reliable for predicting the dynamical behavior of the
structure under various loading and boundary conditions.
Damage Detection and Structural Health Monitoring: since the dynamical
behaviour of structures is influenced by damage, one is able to detect the occurrence and evolution of damage by monitoring the modal parameters [37].
Besides the detection of structural damage, modal analysis also enables the
localization and quantification of the damage since also spatial information
is present in the model (mode shapes). Compared to nondestructive testing
(ultrasound techniques, analysis of magnetic fields, radiology, thermal methods) vibration-based analysis allows a more global analysis of the structure
with the potential of being applied in situ. However, varying environmental
and operational conditions affect the modal parameters as well which can
seriously complicate in situ structural health monitoring.
Vibro-Acoustics: the existence of a vibro-acoustical coupling between structural vibrations and radiated noise is an important aspect for the design
process in terms of comfort. The noise radiated in an aircrafts fuselage
due to the vibrations induced by the engines is one of the many examples.
Based on the modal parameters it is possible to calculate the sound intensity
radiated from the vibrating structure [120] without the need for expensive
acoustical holography experiments.
Forced Response Analysis: is mainly applied to study the behavior of the
structure under operating conditions such as for example, a bridge with
road traffic and wind load, a car on the road or an aircraft during flight.
This way designers can check the system under normal and extreme loading
conditions and detect structural weaknesses subject to design modification.
1.2. Focus and Outline of the Thesis
Sensitivity Analysis and Structural Modification Prediction: computing the

sensitivity of the modal parameters for a certain change of mass, stiffness or
damping at each response location assists the design engineer by indicating
the type and location for structural changes that mostly affects the mode
that is subject for modification [142]. The actual quantitative change of the
modal model induced by the most promising modifications is next estimated
during the modification prediction.
Substructuring: given the modal model of different components of a structure, the dynamical behaviour of the complete structure can be computed
using substructuring techniques. This application is quickly gaining great
interest for studying large complex structures such as cars, aircrafts and
satellites.
Obviously, a successful applicability of the above discussed engineering methods
greatly depends on the quality of the modal model as well as on computational
efficiency of the identification algorithms used.
1.2
Focus and Outline of the Thesis
Due to the advances in measurement technology and new EMA applications, extensive data sets, obtained by multiple input excitation and measuring at a large
number of response locations, are common practice today. Given the increasing
complexity of the structures under test, this data is also often characterized by a
high frequency resolution, high dynamical range and close-coupled modes. Moreover, the measured signals f (t) and x(t) (in Eq. 1.1) are disturbed by errors that
interfere with the physical content of the system in the signals. These errors originate from different sources in the modal experiment: mechanical setup errors (mass
loading, interfering boundary conditions, structure-shaker interaction,...), excitation and acquisition errors (excitation in node, spatial aliasing, calibration sensors,
...), signal processing errors (leakage, temporal aliasing, imperfect compensation
for time windows), non-linear distortions and measurement noise (environmental noise, instrumentation noise, digitalization noise). These errors, can seriously
hamper the extraction of the modal parameters and introduce both systematic and
stochastic errors (i.e. a certain degree of uncertainty) in the final model. From
this it must be clear that the estimation of the modal model is not an obvious
problem at all.
As a result, the need for improved system identification methods within the
domain of modal analysis increases under impulse of the broadening field of applications, e.g. damage detection and vibro-acoustics, and the increased complexity
of todays tested structures. Recent advances in the field of frequency-domain
system identification are the development of the so-called stochastic Total Least
Squares and Maximum Likelihood methods [109, 41]. Since the noise on the measured data is taken into account during the estimation process, these methods
yield very accurate parameter estimates with confidence bounds, even in the case
of a poor quality of the data.
The aim of this thesis is to use the stochastic methods as a starting point to
develop more advanced modal parameter identification techniques. This however
requires an extensive revision of both the initial formulation of these methods with
respect to computational efficiency and the modal analysis process with respect
to the integration of the noise on the measured data. So, this thesis focusses on
the current shortcomings related to the application of these stochastic methods for
modal parameter identification. Three main parts are considered and discussed
in more detail in the next three sections, while a chapter-by-chapter outline is
presented in Figure 1.1.
Nonparametric Identification of FRF Data

The first part of this thesis deals with the nonparametric processing of modal data.
Most often frequency response functions (FRFs) are estimated by averaging the
measurements using a nonparametric FRF estimator. The accuracy (in terms of
systematic and stochastic errors) of these algorithms greatly depends the measurement noise as well as the excitation signal used (random or burst random noise,
multisine, periodic random,...). In the domain of experimental modal analysis, it is
common practice to consider only the errors on the measured response signals x(t),
although this assumption is often not satisfied, resulting in systematic errors on
the FRF estimates. However, the application of the accurate stochastic methods
for modal parameter estimation is only meaningful when starting from accurate
FRF data. Therefore, an approach for the identification of the FRF matrix based
on an errors-in-variables (EV) framework will be presented in Chapter 2. An
EV model considers the measurement noise on both the input and output signals,
yielding a better noise model for a practical modal test setup compared to the
classical noise assumptions. Besides obtaining unbiased FRF estimates, it is just
as important to know the uncertainty (noise) on the measured data, required by
the stochastic methods. A methodology for deriving the noise covariance matrix of
the FRF estimates without the need for additional measurements will be presented
as well. This uncertainty information can be used to check the quality of the data
and will be taken into account during the modal parameter estimation process.
On the other hand, leakage effects can as well result in important systematic errors on the FRF estimates. When applying arbitrary excitation signals, such as
random noise, effects of leakage occur as a result of the finite measurement time.
Although periodic signals can simply eliminate leakage errors, arbitrary signals are
still frequently used for modal testing purposes. Through the use of appropriate
signal windowing techniques, the effects of leakage on the FRF estimation can be
reduced as will be discussed in Chapter 3.
Parametric identification of Modal Parameters

Although significant research efforts during the last two decades have resulted in
an extensive number of parametric identification algorithms, most of them are
certainly not directly applicable for modal parameter extraction. Chapter 4 introduces the parametric identification part of this thesis by giving an overview of
the most important parametric identification methods with specific attention for
their potential adequacy for modal parameter estimation. Although estimators
such as the LSCE, LSFD, FDPI, ERA, and ITD were specifically developed for
EMA [62, 36, 88], these estimators tend to fail in cases where the modal data is
characterized by low signal-to-noise ratios and acquired from systems with high
dynamical range and modal density. In this thesis, the focus will be on the class
of stochastic frequency-domain methods that estimate the parameters of rational
fraction transfer function models using a (Total) Least Squares [108, 109] or Maximum Likelihood approach [41, 109]. These methods were initially developed in a
errors-in-variables framework, under the assumption that the measured data are
the input and output Fourier sequences acquired using periodic excitation. The
results presented in Chapter 2, already facilitate FRF and noise identification in
the same EV framework. Nevertheless, the initial formulation of these stochastic
methods requires a thoroughly revision with respect to the numerical performance
for analyzing typical modal data sets.
Hence, adapting these algorithms for modal parameter estimation represents
the main subject of the second part of this thesis. Typical for modal analysis is the
large number of output measurements and a high modal density. As will be shown
in Chapter 5, the extensive amount of data requires optimized algorithms that
balance between accuracy and memory/computation efficiency. A multivariable
implementation for frequency-domain estimators, based on a common denominator
transfer function model, will be given. First, a numerical improved implementation
of the frequency-domain linear Least Squares (LS) approach will be presented for
extracting the modal parameters from FRF data. Aspects, such as robustness
for high modal density and numerical performance (computation time, memory
usage) are studied for different possible parameterizations. The final result is
a fast LS solver, i.e. the so-called Least Squares Complex Frequency (LSCF)
method. This estimator also offers the user, in analogy with the well-known LSCE
method, a stabilization chart. Furthermore, it is shown in this chapter how the
fast implementation can also be extended to the class of Total Least Squares (TLS)
estimators and the Maximum Likelihood (ML) method.
For some applications, such as flight flutter testing for instance, only short
data sequences are measured. In that case, nonparametric FRF identification
based on averaging techniques becomes less applicable. Sufficient time samples
must be available to have an acceptable frequency resolution to detect closely
spaced modes. Therefore an alternative approach consists of estimating the parameters based on the input and output (I/O) Fourier sequences for which a fast
implementation of the frequency-domain TLS and ML methods will be discussed

in Chapter 6. Moreover, by generalizing the results presented in [111] for modal
parameter estimation, the effects of leakage (or transients in general) can be taking
into account by estimating the system and transient parameters simultaneously.
This approach can also be applied to estimate the parameters from auto and cross
power spectra, which can be preferable when sufficient data is available and a
parametric compensation for possible transient effects is desired.
Automation of Modal Parameter Estimation
Traditional experimental modal identification can still be user-time intensive and
as a result costly. Most important reasons for this are related to the constraints in
analyzing large, complex structures, which are discussed in [4, 136]. More specific
for the modal identification step, noisy data, a high modal density and the lack
of error bounds on the estimated parameters, can result in a very difficult model
order and mode selection process. Based on the availability of the advanced system identification techniques, the third and last part of this thesis, presented in
Chapter 7, focuses on the development of improved model validation criteria and
the automation of the modal parameter extraction. An automated approach will
certainly result in an important reduction of the user-time as well as the possibility for non-expert users to apply modal analysis techniques. At the same time, it
is a basic tool for structural health monitoring applications, which however, also
require a method for tracking the modal parameters over subsequent monitoring
instances. Hence, a method for an automated estimation and tracking of the modal
parameters was developed as well.
A summary and the main conclusions as well as indications for further research
are given in Chapter 8.
Modal Testing
FRF & Noise
FRF & Leakage
Chapter 2
Overview MPE
Chapter 4
Chapter 3
FRF-based MPE
I/O-based MPE
Chapter 5
Chapter 6
Autonomous MPE
Chapter 7
Modal Applications
Figure 1.1: Organization of the text. Three main parts can be distinguished: the nonparametric identification of FRF data, the parametric identification of modal parameters
and the automation of parameter estimation process.
10
Chapter 2
FRF Identification in the

Presence of Measurement
Noise
Estimating frequency response functions from the measured data is most often
the first step of experimental modal analysis. In this chapter, nonparametric FRF
identification has been studied for a maximum likelihood approach based the errorsin-variables framework, introduced in 2.2 and 2.3. Based on this result and
considering the practical issues related to experimental modal analysis discussed in
Sections 2.4, the approximations and possible errors made by the FRF estimators
used in practice are studied in Section 2.5. Next, a methodology for the accurate
estimation of the FRF and corresponding noise covariance matrix from the same
set of measurements is derived for the class of arbitrary excitation signals. It
is shown that this so-called Instrumental Variables nonparametric estimator has
maximum likelihood properties, while remaining consistent in the case of nonlinear
distortions originating from the excitation devices (cf. 2.6). In order to illustrate
the results of this chapter, the simulation of modal test setup is introduced.
11
12
2.1
Chapter 2. FRF Identification in the Presence of Measurement Noise
Introduction
As for many scientific investigations, traditional modal analysis typically starts

from frequency response functions (or impulse response functions). For this reason,
much attention has been paid to nonparametric estimators in the literature [7, 85,
14, 21, 80]. Although most of these estimators were originally developed for singleinput, single-output (SISO) systems, some of them where later generalized for the
multi-input, multi-output (MIMO) case.
As was already explained in Chapter 1, the nonparametric FRF identification
plays a crucial role in the overall quality of an experimental modal analysis (EMA).
Two important sources of errors, encountered during the nonparametric identification process, are noise on the measured data and spectral leakage due to a finite
measurement time in the case of random excitation. This chapter concentrates
on the effects of measurement noise and possible nonlinear distortions due to the
excitation devices on the estimated frequency response functions. Although, until
today, the H1 FRF estimator is still most often used, it is well-known that this estimator generally yields erroneous results due to much simplified noise assumptions,
i.e. only the output errors are considered.
For this purpose, first the stochastic errors-in-variables (EV) framework is introduced for a multivariable (MIMO) system. Based on this it is possible to study
and compare nonparametric FRF estimators as a function of the assumptions
regarding the measurement noise and the type of excitation signal (random or
periodic).
Given the MIMO character of EMA, multivariable expressions are derived for
the FRF matrix estimate and its corresponding covariance matrix. The uncertainty on the measured data is an important element for the evaluation of the test
setup and the measured data (i.e check of SNR, sensor working, ...) prior to the
parametric estimation step. Moreover, knowledge about the measurement noise
will be taken into account when using the parametric frequency-domain estimators developed in a stochastic framework [109]. This can significantly improve the
quality of the parameter estimates compared to e.g. a classical Least Squares approach, while confidence intervals for parameter estimates can be computed based
on the knowledge of the noise covariance matrix of the FRF estimates.
Based on the multivariable expressions, the asymptotic (statistical) properties
of the different FRF estimators are studied. It is shown under which assumptions
for the noise model and excitation conditions it is possible to derive consistent
FRF estimates in a statistically efficient manner (i.e. with maximum likelihood
properties).
Considering the excitation-side of traditional modal testing, a final part of this
chapter is related to the influence of the exciter setup (i.e. the power amplifier
and electro-dynamic shaker) on the estimated FRF and covariance matrix.
2.2. Errors-in-Variables Stochastic Framework
[m]
ns[1] (t)
[m]
r(t)[1]
[m]
?
- +l
s0[1] (t)
[m]
f0[1] (t)
nx[1] (t)
[m]
[m]
ns[2] (t)
[m]
r(t)[2]
[m]
?
- +l
s0[2] (t)
[m]
f0[2] (t)
nx[2] (t)
[m]
[h0 (t)]No Ni
.
.
.
.
.
.
[m]
ns[N ] (t)
i
[m]
r(t)[N ]
i
- s0[Ni ] (t)
[m]
[m]
f[N ] (t)
i
[m]
?
- +l
nf [N ] (t)
[2]
[m]
x0[2] (t)
+l
f[2] (t)
?
- +l- x[m] (t)
[m]
nf [2] (t)
[1]
[m]
x0[1] (t)
+l
f[1] (t)
?
- +l- x[m] (t)
[m]
nf [1] (t)
13
[m]
f0[N ] (t)
i
+l
nx[N ] (t)
o
?
- +l- x[m]
[No ]
[m]
(t)
x0[N ] (t)
o
Figure 2.1: Time-domain errors-in-variables noise model.
It should be noticed however that for some specific applications, it can be

preferable to estimate the modal parameters directly from the Input/Output Fourier
data. This can be the case when only short data sequences are available and a
sufficient frequency resolution is required in order to have sufficient nonparametric
information, especially for low damped modes (cf. Chapter 6). Examples are flight
flutter testing and condition monitoring of a rocket during the launch.
2.2
Errors-in-Variables Stochastic Framework
Consider a linear time-invariant multivariable system with Ni inputs and No outputs. In order to account for the most important sources of errors encountered during the non-parametric identification process, the modal test setup is schematically
represented by the time-domain errors-in-variables noise model shown in Figure
2.1. Following time-domain signals are observed during M measured records
14

[m]
[m]
f [m] (tn )
= f0 (tn ) + ef (tn )
x[m] (tn )
= x0 (tn ) + e[m]
x (tn )
[m]
(2.1)
with tn = nTs , n = 0, 1, . . . , Ns 1 and Ts the sampling period. The vectors

f [m] (tn ) R Ni 1 and x[m] (tn ) R No 1 respectively stand for the measured
forces (inputs) and accelerations, velocities or displacements (outputs) signals of
the multivariable system h0 (t) observed at time t = tn during the mth measured
[m]
[m]
record. f0 (tn ) and x0 (tn ) are the corresponding true but unknown values.
[m]
Using the notations as defined in the Nomenclature, f[i] (tn ) denotes the sample
[m]
n measured at input location i during record m. The noise sources ef (tn ) and
[m]
ex (tn ) on the measurement signals comprise of the disturbing time-domain noise

[m]
[m]
[m]
sources nf (tn ), nx (tn ), ns (tn ).
Using the Discrete-time Fourier transform (DFT), the time-domain vector sequences f [m] (tn ) and x[m] (tn ) , n = 0, 1, . . . , Ns 1 are transformed into the
frequency-domain
F [m] (f )
NX
s 1
f [m] (tn )e(i2nf /Ns )
n=0
X [m] (f )
NX
s 1
x[m] (tn )e(i2nf /Ns )
(2.2)
n=0
with angular frequency f = 2f /Ns Ts and f = 0, 1, . . . , Ns /2 1.

The relation in the frequency-domain between the noisy input vector F [m] (f )
C
and noisy output vector X(f ) C No 1 , and the true ones is given by the
following frequency-domain errors-in-variables stochastic model (Figure 2.2)
Ni 1
[m]
[m]
[m]
[m]
F [m] (f )
F0 (f ) + EF (f )
X [m] (f )
X0 (f ) + EX (f )
(2.3)
[m]
The relation between the trueinput and output Fourier vectors, F0 (f )

[m]
C Ni 1 and X0 (f ) C No 1 , respectively, is given by
[m]
[m]
X0 (f ) = H0 (f )F0 (f )
(2.4)
where the complex valued matrix H0 (f ) C No Ni stands for the Frequency

Response Function Matrix of the system at spectral line f .
[m]
[m]
The vectors EF (f ) C Ni 1 and EX (f ) C No 1 occurring in the

frequency-domain errors-in-variables (EV) model (2.3) are a general representation of all possible disturbing spectral noise sources. Referring to Figure 2.2,
2.2. Errors-in-Variables Stochastic Framework
15
[m]
[m]
NS[1] ()
[m]
R[1] ()
?
- +l
S0[1] ()
[m]
F0[1] ()
NX[1] ()
X0[1] ()
+l
F[1] ()
[m]
[m]
[m]
NS[2] ()
[m]
R[2] ()
?
- +l
S0[2] ()
[m]
F0[2] ()
NX[2] ()
[m]
[H0 ()]No Ni
.
.
.
.
.
.
[m]
NS[N ] ()
i
[m]
R[N ] ()
i
- S0[Ni ] ()
[m]
[m]
?
- +l
NX[N ] ()
o
[m]
F0[N ] ()
i
?
- +l- X [m]
[No ]
[m]
()
X0[N ] ()
NF [N ] ()
[m]
F[N ] ()
i
[2]
[m]
X0[2] ()
+l
F[2] ()
?
- +l- X [m] ()
[m]
NF [2] ()
[1]
[m]
[m]
NF [1] ()
?
- +l- X [m] ()
+l
Figure 2.2: Frequency-domain errors-in-variables noise model.

[m]
[m]
NF (f ) and NX (f ) represent the measurement channel noise mainly due to

instrumentation and digitizer noise. In the case of modal testing, the structure is
often excited by a shaker. The excitation units (i.e. power amplifier, shaker) are
characterized by the transfer functions S0[1] (), . . . , S0[Ni ] () denoted by the diagonal matrix S0 () C Ni Ni . Instrumentation noise due to these excitation units,
[m]
adds random errors, NS (f ) C Ni 1 , resulting in the trueFourier vectors
[m]
[m]
F0 (f ) C Ni 1 with disturbing shaker noise NS (f ).
Hence, the frequency-domain EV stochastic model can also be written under
the form
[m]
F [m] (f )
S0 (f )R[m] (f ) + EF (f )
X [m] (f )
H0 (f )S0 (f )R[m] (f ) + EX (f )
[m]
with
[m]
NS (f ) + NF (f )
[m]
H0 (f )NS (f ) + NX (f )
EF (f )
EX (f )
[m]
[m]
[m]
[m]
(2.5)
16

[m]
[m]
[m]
Assuming that the random errors EZ (f ) = [EF (f )H , EX (f )H ]H are

complex normally distributed, with
[m]
[n]
(2.6)
Cov EZ (f ), EZ (f ) = CEZ (f )mn , m, n = 1, . . . , M
where mn the Kronecker delta and CEZ C (Ni +No )(Ni +No ) an a priori known
Hermitian symmetric noise covariance matrix. Hence, the disturbing noise is assumed to be (second-order) stationary and uncorrelated for different measured
records (m 6= n). Also notice that CEZ accounts for possible correlations among
[m]
[m]
the noise sources EF (f ) and EX (f ) on the input/output Fourier coefficients,
[m]
[m]
while E{EF (f )} = 0 and E{EX (f )} = 0.
Furthermore, it is proven by Brillinger [21] that for a wide class of probability

[m]
[m]
density functions of the time-domain noise sequences ef (nTs ) and ex (nTs ),
[m]
[m]
n = 0, . . . , Ns 1, the Fourier coefficients of the noise EF (f ) and EX (f ),

f = 1, . . . , Nf , are asymptotically complex multivariate normally distributed and
independent, as the number of time samples Ns used to calculate the Discrete
Fourier Transform goes to infinity for a fixed number of DFT frequencies. In [124]
it is demonstrated that these properties are reached for practical values of N s (e.g.,
Ns = 512). Since the measurements for different spectral lines are asymptotically
uncorrelated, (f ) is omitted in all the following equations of this chapter.
2.3
Hgtls a Maximum Likelihood Approach
Given the errors-in-variables framework, defined in Section 2.2, accurate nonparametric FRF identification can be derived starting from a maximum likelihood
approach. The assumptions concerning the measurement noise in the EV model,
motivate the use of complex Gaussian probability density function to construct
a maximum likelihood FRF estimator (MLE). The (asymptotical) independency
over the frequencies allows to write the ML equations for a single frequency without the loss of (statistical) efficiency. This results in the (negative) log-likelihood
function
"
( ) #"
( ) #H
I
I Ni
N
i
F ) = trace C 1 Z
`(H,
(2.7)
F Z
F
EZ
H
H
and F = [F [1] , . . . , F [M ] ] are independent variables, while Z =

The matrices H
[1]
[M ]
[Z , . . . , Z ] represent the measurements, where Z [m] = [F [m]H , X [m]H ]H . In
[50] it has been shown that F can be eliminated from (2.7) using the method of
the Lagrangian multipliers, resulting in the following cost function
H 1 B
G
Z B
H
E B
= trace BC
(2.8)
`(H)
Z
2.3. Hgtls a Maximum Likelihood Approach
17
Z
Z = 1 ZZ H = 1 PM Z [m] Z [m]H and B
= [H,
IN ]. The matrix G
with G
o
m=1
M
M
contains the cross and auto power spectra between the measured input (force) and
output (response) signals.
is equivalent to the generalized
The minimization of (2.8) with respect to B
total least squares (GTLS) approach [138] (p.347), which boils down to solving a
generalized eigenvalue problem
Z V = CE V
G
Z
(2.9)
where represents a diagonal matrix containing the eigenvalues and V the matrix
of corresponding eigenvectors. This can be written more explicitly under the
formulation
" H #
" H #
F
F X
H
H
CEF EX
CEF
G
G
T
(2.10)
T =
CEX
CEX EF
GXF GX
INo
INo
H T where T represents a transformation matrix with dimensions
since V = B
(No No ). Notice that Eqs. (2.8) and (2.9) are scale invariant. As a result it is
sufficient to know the covariance matrix CEZ up to a certain scaling factor.
gtls is obtained by choosing the columns of B
H to be
The GTLS solution B
linear combinations of the No eigenvectors (contained by V ) of GZ in the metric

CEZ that are associated with the No smallest eigenvalues of . Finally, the FRF
gtls is given by
matrix estimate H
gtls (:, 1 : Ni )
gtls = B
gtls (:, Ni + 1 : Ni + No )1 B
H
(2.11)
gtls ) = PNo N +o . The generalized singular value decomposition prowhile `(H

i
o=1
vides a numerical stable way for computing the Hgtls [40]. In the modal analysis
society this estimator is better known as the scaled total least squares estimator
[82].
It has been proven that this estimator, which belongs to the class of the maximum likelihood estimators, is consistent and asymptotically efficient with the
Cramer-Rao lower bound given by
CCR (vec(H))
=
1 T
G C
M F0
(2.12)
with C = BCEZ B H (with B = [H0 , INo ]) and vec and the Vector and Kronecker product operators.
H or
E B
In Appendix A it is proven that under the assumption that either BC
Z
H
B GZ B in Eq. (2.8) are diagonal matrices, the FRF matrix obtained by solving
a GTLS problems for each output separately is the same as the one derived from
Eq. (2.11).
18
Notice however, that CEZ (i.e. the noise covariance matrix of the input and
output errors) should be a priori known to get consistent FRF estimates. As for
many real-life applications this represents one of the practical constraints for the
application of the Hgtls for modal analysis.
2.4
2.4.1
Practical Issues of EMA

Modal Testing
Nowadays, both the requirements of new EMA applications (cf. 1.1) and the
availability of multi-channel data acquisition units (typically 48 up to 128 measurement channels) often result in extensive amounts of measurement data from which
the modal parameters have to be derived. Structures being excited by 4 shakers and measured in 500 response locations by means of a multi-patch (tri-axial)
accelerometer (MPA) setup are not uncommon today. Also novel optical measurement techniques, such as Scanning Laser Doppler Vibrometer (SLDV) and
Electronic Speckle Pattern Interferometry (ESPI) can be applied for plate-like
structures.
However, the large number of output measurements introduces two sorts of
constraints on the nonparametric identification process:
although information technology has entered a new era, computation time
and memory requirements often remain important issues in practice since often only a limited time is available for processing and analyzing an increasing
amount of test data.
a consistent (i.e. full rank) estimate of the complete (No No ) output covariance matrix requires at least No response measurements at each output location. Obviously, this would drastically increase the time needed for modal
testing, which is often not feasibly in practice. Depending on the type of
excitation signal and the signal-to-noise ratios (SNR) the number of records
M measured at each location does not often exceeds 10 25 records.
Considering the configuration of the modal test setup, it is possible to take
a priori knowledge about the noise model into account, as discussed now for two
often used test setups
Multi-patch accelerometer (MPA) setup: When the number of available accelerometers or acquisition channels is smaller than the number of
response locations, the structure is measured in P different patches, where
each patch consists of one configuration of the accelerometers. In this case
2.4. Practical Issues of EMA
19
(N
+ No ) P
~
No
Ni
Ni
(N
+ No ) P
Ni
~
No
Ni
~
No
~
No
Figure 2.3: Noise covariance matrix for a multi-patch accelerometer setup.
o responses belonging to a same patch

it is clear that only the noise on the N
can be correlated to a certain extend as shown in Figure 2.3. At the same
time this also implies that a different set of input measurements corresponds
to each patch of responses. As a result, in the nonparametric case, it is possible to process the measured data for each patch separately, which results
o is typically 10 50). However,
in significant smaller matrix dimensions (N
o N
o ) output covariance maeven for this reduced problem, a full rank (N
o measurements at each response location. In
trix still requires at least N
practice, the time available for testing is often too limited to measure a large
number of records M and as a result only the output variances (diagonal
elements) are considered. This leads to a further reduction of the computational load since then each output can be processed separately. On the other
hand, since the number of input measurements is small, typically not more
than 5, the full input covariance matrix can be determined. For the FRF
matrix, this boils down to estimating the FRFs between the inputs and each
output, which corresponds to each row of this matrix. The noise covariance
matrix CEZ is then constructed by estimating the input covariance matrix
for each patch, and for each output its variance as well as the covariances
between the particular output and the corresponding set of inputs. This
20

results in a much faster approach than estimating the full covariance matrix.
In general, the price paid by omitting possible correlations between the outputs most often does not countervail against the gain in computation time,
especially when the number of measurements is large. A possible drawback
of a multi-patch setup relies in the mass-loading influence of the sensors on
the structure. Because the accelerometers are placed in different locations
for each patch the dynamic characteristics of the measured structure are also
slightly differently from patch to patch. The data of the different patches are
therefore inconsistent to each other [136, 24]. An optimal sensor positioning
(i.e. in practice usually a random spread) can significantly minimize the influence of mass-loading on the structure, which, in general, is also beneficial
for minimizing the level of correlation between the output measurements.
This avoids that a complete patch of sensors would be located near DOFs of
maximum amplitude for some modes resulting in important correlations and
data inconsistencies. Mass-loading inconsistencies can be avoided by using
for example a laser measurement setup. However, these systems are expensive and more difficult to use for measuring three-dimensional structures.
Scanning laser doppler vibrometer (SLDV) setup: can be considered
as a special case of the MPA setup where every patch consists of 1 response
sensor. Each response DOF and its corresponding set of input measurements
is measured separately, which is equivalent to measuring No times a MISO
system. Acquiring each output separately, implies that no correlation exists
o = 1) and consequently the EV
between any of the measured responses (N
noise model can be completely satisfied by this type of modal test setup.
In [143] adapted system identification techniques were proposed to process
SLDV or ESPI data sets, requiring data reduction prior to the parametric
estimation of the modal parameters.
2.4.2
Effects of Environmental Noise
Depending on the application and test environment, another source for stochastic
errors on the measured data originates from a stochastic and unmeasured excitation of the system by the environment, called process or environmental noise.
Important examples in the case of modal testing are the wind excitation on a bridge
or the influence of the road traffic nearby a lab. Another example is found in flight
flutter testing, where an unmeasured part of the total excitation of the wings of
an airplane stems from the wind and turbulence effects. Since this excitation is
not measured it is also observed as stochastic disturbing noise.
Although not explicitly shown in Figure 2.2, the general formulation of the
errors-in-variables model Eq. (2.3) also includes process noise when present in the
measurement setup. The general EV setup in the case of process noise is presented
2.4. Practical Issues of EMA
[m]
NS
[m]
()
S0 ()
21
[m]
()
?
- +l
NP
- +l
-6
[m]
F0
[m]
NF
()
H0 ()
[m]
()
X0
()
?
- +l
()
G0 ()
- +l
[m]
NX ()
F [m] ()
- +l
X [m] ()
Figure 2.4: Frequency-domain errors-in-variables noise model for a system operating in

a closed loop.
in Figure 2.4. Depending on the type of excitation device, the effect of this process
noise differs:
When an impact hammer is used, there is no fixed interaction between the
structure and the excitation device. In this case, the effect of process noise
will only result in some mutually correlated noise on the output measurements since there is no feedback to the input side (open loop case).
In the case of shaker excitation, the EV setup operates in a closed loop. The
structure-shaker interactions will result in a feedback of the process noise
to the input side through the system G0 characterized by the dynamical
stiffness of the excitation devices. As a result, mutually correlated noise
is, in addition to the output measurements, also introduced on each of the
measured input signals.
For the case that process noise is present in a shaker setup, the multivariable frequency-domain EV stochastic model (2.5) is then given, for the angular
frequency f , as
F [m]
X [m]
[m]
[INi G0 H0 ]1 S0 R[m] + EF
[m]
H0 [INi G0 H0 ]1 S0 R[m] + EX
(2.13)
22
with
[m]
= NS
[m]
= H 0 NS
EF
EX
[m]
[m]
+ NF
[m]
[m]
[INi G0 H0 ]1 G0 NP
[m]
[m]
+ NX H0 [INi G0 H0 ]1 G0 NP
[m]
where NP (f ) C No 1 represents the process noise vector and G0 C Ni No

the dynamical stiffness matrix of the excitation devices.
In 2.7, the effects of process noise on the correlations between the FRF estimates, will be illustrated by means of a simulated multivariable modal experiment.
In practice, both input channel noise and process noise are most often not considered, which however can result in important errors depending on the type of
application.
2.5
FRF identification for EMA
In this section, different nonparametric FRF estimators will be discussed for the
multivariable case and expressions for the noise covariance information are derived.
In addition, the estimators are also assessed for the specific case of modal analysis
which often requires a MISO approach, for which the effect on the asymptotical
properties is studied as well.
2.5.1
Hv TLS estimator
Assuming the input and output noise sources to be uncorrelated and of equal
amplitude results in the so-called Hv FRF estimator [80]. This estimator does not
require prior knowledge of the noise covariance matrix CEZ and yields the total
least squares (TLS) estimate for the FRF matrix by solving following eigenvalue
problem (cf. Eq. 2.10)
" H #
" H #
F
F X
H
H
0
I Ni
G
G
T
(2.14)
T =
0 INo IN
GXF GX
INo
o
In general however, this estimator is not consistent since the specific noise assumptions are not satisfied. Indeed, the noise covariance matrix can only be scaled to a
unity matrix, when all the noise sources are uncorrelated and of equal amplitude.
In practice, however, the signal-to-noise ratios for the different signals generally
differ depending on the input and response locations, while the actual noise sources
themselves can be of equal amplitude. As a result, this would require the proper
scaling of each signal in order to obtain a unity noise covariance matrix, which
obviously is not feasible in practice since the scaling factors are not a priori known.
2.5. FRF identification for EMA
23
Referring to Appendix A, it follows that the FRF matrix yielded by processing

each output separately is equal to the one obtained from solving Eq. (2.14) under
H = H
H
H + IN is diagonal, which is satisfied in
E B
the assumption that BC
o
Z
H
H
H to be
the case that H H is diagonal. However, as there is no reason for H
diagonal, the MIMO and MISO Hv estimator generally differ.
2.5.2
H1 and H2 LS estimators
H1 estimator
In the domain of modal analysis, the most commonly-used FRF estimator is the
so-called H1 estimator [62, 88], which considers only the errors EX on the output
measurements. The inputs are assumed to be error-free. The TLS problem (2.14)
now reduces to a classical least squares (LS) problem
" H #
" H #
F
F X
H
H
G
G
0
0
T =
T
(2.15)
0
I
GXF GX
No
INo
INo
for which the estimate of the FRF matrix is derived from the first Ni rows given
by
1 = G
XF G
1
H
(2.16)
F
Under the specific noise assumptions, the H1 estimator is consistent without the
need for the a priori knowledge of the noise covariance matrix CEZ . Since in
H = IN is diagonal by construction, the MIMO LS
E B
this case the matrix BC
o
Z
problem (2.15) and solving the MISO problem for each output separately always
yield the same FRF matrix.
Covariance matrix of H1 estimates
An estimate of the output noise covariance matrix CEX is derived from the last
No rows of (2.15)
X G
XF G
1 G
F X
E = G
T T 1 = G
X
F
(2.17)
An expression for the covariance matrix of the H1 FRF estimates can be obtained
by means of a sensitivity analysis (with Fm = 0 since no input noise)
!
M
X
1
1
1 =
X [m] F [m]H G
(2.18)
H
F
M m=1
Using the Vector and the Kronecker operator [20], equation (2.18) can be rewritten
as
M
X
1) = 1
Q[m] X [m]
(2.19)
vec( H
M m=1
24
with
T F [m] IN
Q[m] = G
o
F
(2.20)
Assuming stationarity and no correlation over the different measurements, the

covariance matrix becomes
1) =
CH1 = Cov(vec H
M
1 X [m] [m] [m]H [m]H
Q E X X
Q
M 2 m=1
(2.21)
and since for a Hermitian matrix the Transpose and Complex Conjugate operators
are equivalent, this can be rewritten as
CH1
=
=
M
1 X T [m] [m]T T
GF F
F
GF
CX
M 2 m=1
1 T
G CX
M F
(2.22)
with an estimate of CX given by expression (2.17), i.e

X G
XF G
1 G
E = G
F X
G
X
F
(2.23)
For M Eq. (2.22) becomes the Cramer Rao lower bound (2.12), which proves
that the H1 estimator is asymptotically efficient when there is no input noise.
Consequently, this estimator, under this specific noise assumption, also belongs to
the class of maximum likelihood estimators.
unbiased for a finite
Remark : As proven in appendix B, an unbiased estimate of G
EX
number of experiments M is given by
unbiased =
G
EX
M
E
G
(M Ni ) X
(2.24)
In practice however, the input measurements are never free of noise and to cope
with these errors, the EV model presented in Section 2.2 has to be considered. The
model equations are then given by Eq. (2.5) and the probability limit (plim) for
the number of measurements going tot infinity (M ) results in biased H1
estimates
1 = H0 1 + GN (S0 GR S0H + GN )1 1 = H0 1 + GN G1 1
plim H
F
S
F
F0
M
(2.25)
where the bias term is related to the signal-to-noise ratios of the input signals.
H2 estimator
As for the H1 estimator, similar equations can be derived for the H2 estimator,
which assumes that errors are only present on the input measurements. The
25
outputs are now assumed to be error-free. The TLS problem (2.14) again reduces
to a classical least squares (LS) problem
" H #
" H #
F
F X
H
H
I Ni 0
G
G
T =
T
(2.26)
XF G
X
0
0
G
I
I
No
No
for which the H2 estimate of the FRF matrix is derived from the last No rows
given by
2 = G
X G
1
H
(2.27)
FX
Again in practice, however, this generally results in biased FRF estimates
2 = H0 1 + GN G1
(2.28)
plim H
X
X0
M
where the bias term is now related to the signal-to-noise ratios of the output
signals. Furthermore, the H2 technique is constrained in practice to the case
1 . When analyzing the FRF data per
where No = Ni in order to compute G
FX
output, the H2 can only be applied in the case of a single excitation source.
Since, in many practical cases, especially when accuracy is important, the noise
sources on the input and output signals can not be neglected, both the H 1 and H2
techniques generally yield biased FRF estimates, which will also be illustrated in
2.7.
2.5.3
Hiv estimator
Unlike the H1 or Hv estimators, a nonparametric approach based on the instrumental variables (IV) principle has been proven to be unbiased for SISO systems and
is referred to in the literature as Hs or Hc or 3-channel FRF estimator [95, 26]. In
general, the instruments can be constructed from the measured force signal based
upon the technique of delayed observations. In practice, however, it is preferable
to use the exactly known generator (electrical) signals that are sent through the
power amplifier-shaker units. In that case, the assumption that the noise on the
reference signals has to be mutually uncorrelated with the noise on input and output signals is obviously satisfied. Considering this approach in the framework of
the MIMO errors-in-variables model (cf. 2.2) this estimator is defined as
FR1
1 G
RG
1 = H
XR H
1 = G
XR G
XR G
XF = G
H
1
1
iv
R
FR
FR
(2.29)
RR C Ni Ni .
F R C Ni Ni and G
XR C No Ni , G
with G
XR matrix estimate can be computed for each
Notice that in practice, the H
1
FR matrix estimate
output or a patch of outputs separately (cf. 2.5.2), while the H
1
is computed for all corresponding inputs.
26
Consistency of Hiv estimates

It can be proven under mild conditions that this Hiv FRF estimator yields consistent FRF estimates without the need for an a priori known noise covariance matrix
CEZ . To prove asymptotic consistency, consider the almost sure limit (a.s.lim) of
Eq (2.29)
XF
G
G
(2.30)
=
a.s.lim
a.s.lim H
XR F R
iv
M
XF (M ) the Hiv FRF estimate for M measured records (the number of

with H
iv
averages). Since a continuous function and a.s.lim may be interchanged this can
be rewritten as
XF
a.s.lim H
iv (M )
(2.31)
F R
XR a.s.lim G
a.s.lim G
M
1
M
M
X
X
1
1
a.s.lim
X [m] R[m]H a.s.lim
F [m] R[m]H
M M m=1
M M m=1
Under the Assumptions that the random variables X [m] , F [m] independent and
identically distributed [87] and uniformly bounded [16], the a.s.lim and the expected value operator E may be interchanged. Applying the expected value operator to equation (2.31) after substituting (2.5), assuming the noise sources N F ,
NX and NS to be uncorrelated with the reference signal R up to the second order
moments with these moments being bounded, yields
lim
M
XF }
E{H
iv
(2.32)
1
H0 S0 E R[m] R[m]H S0 E R[m] R[m]H
For the number of averages going to infinity this results in

lim
M
1 1
XF
E{H
iv } = H0 S0 GR GR S0 = H0
which proves that the Hiv estimator is consistent.
(2.33)
27
Covariance matrix of Hiv estimates

XF ) of this consistent H
XF estimate
An expression for the covariance matrix C(H
iv
iv
can be found again by starting from a sensitivity analysis
XF ) = (H
XR )H
FR
(H
1
iv
1
XR (H
FR )
+H
1
1
(2.34)
FR )H
FR . Using the Vector and Kronecker operator
FR (H
FR ) = H
with (H
1
1
1
1
this is written as
FR )
XR ) H
XR H
FR1 vec( H
FRT H
XF ) = H
FRT IN vec( H
vec( H
1
1
1
1
1
iv
1
o
(2.35)
or
XF ) = M
vec( H
iv
with
FR )
vec( H
1
XR )
vec( H
FR1
1FRT H
XR
H
M= H
1
1
"
,
1FRT IN
K= H
o
(2.36)
(2.37)
Assuming all signals to be stationary and the noise to be uncorrelated for different
measurements, the covariance matrix is given by
XF
iv
Cov H
n
o
XF H
XF
E vec( H
iv )vec( Hiv )
(2.38)
H
XR H
FR
MCov(H
1 )M + KCov(H1 )K
FR
XR H
2herm MCov(H
1 , H1 )K
Given that
ZR
H
1
FR
H
1
=
XR
H
"
and
Z=
( )
F
X
FR ) and vec( H
XR ) can be determined from
the covariance matrix of vec( H
1
1
ZR
Cov H
=
1
FR )
Cov(H
1
XR FR
Cov(H1 , H1 )
XR )
FR , H
Cov(H
1
1
XR
Cov(H1 )
1 T
G C EZ
M R
(2.39)
where an estimate of the covariance matrix CEZ of the signals X and F is easily
determined using again Eq. (2.17)
Z G
ZR G
1 G
E = G
RZ
G
Z
R
(2.40)
This equation illustrates that in practice the data covariance matrix is determined
by calculating the Auto and Cross Power Spectra of the input and output signals
obtained from multiple measurements under arbitrary excitation.
28
Remark : In modal analysis, the so-called coherence function is better known than
the variance as a measure for the quality of the FRF data, although both functions
are related as will now be shown. Reformulating Eq. (2.40) as
h
i
Z IN +N G
E = G
1 G
ZR G
1 G
RZ
G
Z
i
o
Z
R
(2.41)
1 G
ZR G
1 G
RZ are the multiple coherence
the diagonal elements of the matrix G
Z
R
functions for respectively the coherence between each force signal and all the generator signals
1
1 G
G
(2.42)
G
mi2 =
F F R[i,:] R RF[:,i]
G
i
and the coherence between each response signal and all the generator signals
mo2 =
1
1 G
G
G
X XR[o,:] R RX[:,o]
G
o
(2.43)
The multiple coherence function is a correlation coefficient bounded between 0 and

1. A value of unity indicates that the output signal is perfectly predicted by all
the reference signals. A coherence less than one can be due to one or more of the
following reasons:
uncorrelated noise on the measured signals, e.g. channel noise
non-linear distortions introduced by the system under test
leakage errors in the signal processing (cf. Ch. 3)
2.5.4
Hev estimator Hiv Using Periodic Excitation
The use of random excitation signals is still most common in the domain of experimental modal analysis since it has the advantage of averaging possible non-linear
behavior of the structure. Although this is also an advantage of periodic random
signals, such as for example a random multisine, these signals are not generally
available in the signal generator of commercial measurement equipment. However,
random excitation signals also have some drawbacks, such as spectral leakage errors and lower signal-to-noise ratios compared to periodic excitation signals.
For the case of deterministic excitation, the Instrumental Variables approach
boils down to a special case, i.e. the Hev estimator. Under the condition that
repeated observations of the same deterministic excitation signals are available,
it is possible to obtain consistent estimates of the FRFs without requiring any a
priori noise information as well as the need for additional instrumental variables.
29
Equation (2.29) gives the expression for the Hiv estimator, which can also be
written as
!
!1
M
M
X
X
1
1
XF
(2.44)
X [m] R[m]
F [m] R[m]H
H
iv =
M m=1
M m=1
For the case that a synchronized multi-excitation measurement setup is used with
Ni linear independent stimuli, R0 = [R0 (1), . . . , R0 (Ni )] C Ni Ni , the resulting input F C Ni Ni and output X C No Ni Fourier data are measured
= M times using the same excitation sequence R0 (i). In practice, these signal
M
Ni
assumptions can easily be realized by applying Ni periodic broadband excitation
periods of the signals.
sequences, e.g. a multisine, by measuring M
Consequently, Eq. (2.44) changes to
M
M
X
X
H 1 1
XF = 1
H
X [m] RH
F [m]
iv
0 R0
M
M
m=1
(2.45)
m=1
resulting in the so-called nonparametric errors-in-variables estimator [42]
M
M
X
X
1
1
ev =
X [m]
F [m]
H
M
M
m=1
m=1
(2.46)
It is clear from this equation that this estimator can also be processed for different
output patches or each output separately.
Like the Hgtls , this estimator belongs to the class of the maximum likelihood
estimators [75] (i.e. consistent and asymptotically efficient). Notice, however,
that the covariance matrix of the errors on the measured signals is not a priori
required to obtain consistent FRF estimates (cf. Eq. 2.46). Even in the case that
the errors on the measured signals are not complex normally distributed, the H ev
estimator is still proven to be consistent (but not efficient anymore) [42]. The use
of this estimator avoids the need for the reference signals R [m] (i.e. instrumental
variables) but requires the use of periodic excitation signals.
Defining X =
written as
PM
k=1
X [k] and F =
PM
l=1
F [l] , equation (2.46) can also be
ev = X F 1
H
(2.47)
ev ) = (X )F 1 + X (F 1 )
(H
(2.48)
Again, starting from a sensitivity analysis, an expression for the covariance matrix
of the Hev FRF estimates can be found
1
F . Using the Vector and Kronecker operators, this can

with (F ) = F (F)
also be written as
ev ) = F T IN vec( X ) F T X F 1 vec( F)
(2.49)
vec( H
o
30
ev FRF matrix is then given by

the covariance matrix of the H
n
o
ev ) = E vec( H
ev )vec( H
ev )H
Cov(H
(2.50)
n
o
H
F T X F 1 E vec( F)vec(
=
F)
F F H X H
n
o
+ F T INo E vec( X )vec( X )H

F INo
o
2herm F T X F 1 E vec( F)vec(

X )H
F INo
where for example
n
E vec( X )vec( X )H
M
M
1 X
X
1
vec(X [k] )
vec(X [l] )H
=E
M
M
k=1
l=1
(2.51)
under the assumptions that over the different measurements all signals and noise
are stationary and the noise is not correlated, reduces to
o
n
1
(I CX )
E vec( X )vec( X )H =
Ni
M
(2.52)
Consequently, the expression for the covariance matrix becomes

ev )
Cov(H
1 T
CX
F F
1 T 1
F F
+
X F CF F H X H
1 T 1
X F Cov(F, X )
F F
1 T
Cov(X , F)F H X H
F F
which, given equation (2.47), finally results in

T 1
M
M
X
X
1
ev ) = 1
1
Cov(H
F [k]
F [l] C
M
M
M
k=1
with
H
E B
C = BC
Z
(2.53)
l=1
i
h
= H
ev , IN
B
o
The probability limit (p.lim) for M results in the Cramer Rao lower bound
ev is asymptotically efficient. In practice,
(cf. Eq. 2.12), which proves that the H
the data covariance matrix CEZ is now determined from repeated measurements
under synchronous periodic excitation. Referring to Eq. (2.40), the calculation of
2.6. Effects of Non-linear Distortions on FRF Identification
31
the Power Spectra now reduces to the calculation of the sample covariance matrix
of the synchronized deterministic measurements.
In the case of asynchronous periodic measurements, it is possible to use, besides the Hiv estimator, FRF estimators based on non-linear averaging techniques.
ari , H
har and the H
log are also able to derive the FRF
FRF estimators such as the H
matrix with its covariance matrix under the EV noise model assumptions [48, 43].
However, nowadays, modal testing equipment has built-in generators facilitating
the synchronizing of the measurements automatically. Only in the case that an
external generator, which cannot be triggered with the data-acquisition system
is used, the input and output signals can not be synchronized. Recently, an automated spectral analysis of periodic signals was proposed in [127], for which no
synchronization between the generator and the data acquisition is needed. This
approach only requires that more than 2 periods of the periodic signal are present
in the acquired data .
2.6
Effects of Non-linear Distortions on FRF Identification
In order to examine the effects of non-linear distortions, introduced by the excitation devices (i.e. power-amplifiers and shakers), these non-linear contributions
have to be included in the errors-in-variables model. Although different model
assumptions are made depending on the type of excitation signal, the derived
conclusions are valid for the following types of excitation signals
Class of stochastic signals: i.e. random and burst random
Class of deterministic and periodic signals: i.e. sine, stepped sine, periodic
chirp and deterministic multisine, periodic random and random multisine.
Model for Nonlinear Distortions
In [13] it shown that, for the class of stochastic signals, the effects of non-linear
distortions on the FRF identification can be described by a so-called Optimum
linear system (OLS) and Revised nonlinear system (RNS). A similar approach
was presented in [121] for the case of a pseudo random excitation (i.e. multisines
with deterministic amplitudes and random phases, also called random multisine),
where the model is called the RLDS (Related Linear Dynamic System) and SNS
(Stochastic Noise Source) model. The latter terminology will now be used for the
remaining part of this section.
Using the same approach as Bendat, the non-linear distortions, introduced by
32
the excitation devices, can be included in the EV model (2.5) (cf. Figure 2.2). The
transfer function matrix S0RLDS can then be considered as the RLDS containing
the linear shaker transfer function S0 and the transfer function of the coherent
(linear) part of the nonlinear distortions. The noise sources NS[i] , i = 1, . . . , Ni
are considered as the SNS containing the non-coherent (and thus stochastic) part
of the nonlinear distortions introduced by the shaker-amplifier system.
Hence the following relation holds
[m]
F0
[m]
[m] + N
=R
SNS
(2.54)
[m] = S RLDS R[m] and stochastic noise source

where the related linear output R
[m] [m]H 0
[m]
[m]H
N
NSNS are uncorrelated, i.e. E R
= 0 as well E R[m] NSNS = 0. Based
SNS
on this model, the effect of nonlinear distortions, introduced by the excitation
devices, on the asymptotical properties of the Hiv estimator can now be studied.
Asymptotic Properties
Referring to 2.5.3, it is proven that the consistency is not influenced by the transfer function matrix S0 and thus non-linear distortions do not affect the consistency
of the Hiv estimator.
Considering the EV model (2.5), assuming that the noise sources NF , NX are
uncorrelated with NS , the data covariance matrix is also given by
CEZ
= E EZ EZH =
N
N + G
G
F
S
N N
N + G
H0 G
X F
S
N N
N H H + G
G
0
F
X
S
N
N H H + G
H0 G
0
X
S
(2.55)
Substitution of (2.55) in
XF
Cov(H
iv )
H
1 FRT T FR
XR
FR1 CF H
1FRH H
XR
H1
GR H1
CX + H
1 H1
1
M
XRH
FRH H
FR1 Cov(F, X) + Cov(X, F )H
XR H
(2.56)
+H
1
1
1
1
yields
XF
Cov(H
(2.57)
iv )
1 FRT T FR
XRH
FRH H
N )H
N + G
FR1 (G
XR H
=
H
GR H1
H1
1
1
1
1
F
S
M
!
1
H
XR
FR
H
N N )
1 H
1
N H0 + G
N
N H0 + G
2herm H
(G
H0 G
S
2.7. Validation
33
N is an Hermitian matrix and considering the consistency property

Given that G
S
finally results in
XF )
(2.58)
Cov(H
iv
1
1 FRT T FR
N N )
N H H 2herm(H0 G
N + H0 G
G
=
H1 GR H1
0
F
X
F
X
M
However, when nonlinear distortions are introduced by the excitation devices
and arbitrary excitation signals are used, the covariance matrix (2.58) will not be
equal to the Cramer Rao lower bound, meaning that some efficiency will be lost.
This can be explained by considering the model proposed by Bendat and the EV
model (2.5) from which it follows that
[m]
F [m] = F0
[m]
+ NF
[m]
[m] + N
=R
S
[m]
+ NF
(2.59)
[m] and the stochastic

Consequently, assuming the related linear output signal R
noise source signal NSNS to be complex normally distributed (which follows from
theorem 4.4.1 in [21]). Then, based on Eq. (2.59), the Cramer Rao lower bound
(2.12) is given by
1
1
N
CCR =
C
(2.60)
GR + G
SNS
M
Hence,
R
FR = G
T H
FRT G
N
R + G
H
G
R 1
1
SNS
(2.61)
the Cramer Rao lower bound will be lower than the covariance matrix (2.58)
and some efficiency will be lost when using the IV approach in the presence of
nonlinear distortions introduced by the excitation devices. This is explained by
[m]
and
the fact that the stochastic noise source NSNS is not coherent with the R
thus, when applying the IV estimator, the contributions of the SNS are going to
zero for the number of measurements going to infinity (i.e. are averaged out when
FR ). It should be noticed that, in the case of periodic excitation,
computing H
1
the Hiv is equivalent to the Hev estimator, and no loss of asymptotical efficiency
appears.
2.7
Validation
In order to validate the results of this chapters the 7 DOF system, shown in Figure 2.5, was used in order to simulate a modal testing experiment. The simulated
data is used to compare the different nonparametric FRF estimators. Monte carlo
simulations are used to study the validity of the expressions for the FRF covariance matrices. The influence of different possible noise sources on this matrix is
studied as well.
34

exciter
me = 1.5kg
ce = 75.62N s/m
ke = 3000N/m
system
ms = 10kg
cs = 6N s/m
ks = 150000N/m
Table 2.1: Mass, damping and stiffness characteristics of 7 DOF system and exciter.
The system is excited by 2 electrodynamic shakers with their body grounded.

The exciter coil has mass me and is supported on a flexure of stiffness ke and
damping ce . The forces acting on the exciter coil, due to the electromagnetic
effect are fe1 and fe7 , while the forces actually applied to the system, i.e f1 or f7 ,
are measured on the mass fixed to the coil (rigid stinger). The structure consists of
7 subsystems with an equal effective mass ms , damping cs and stiffness ks except
for the fourth (middle) subsystem, that has an effective mass that equals 3m s .
The characteristics of this system and the electrodynamic shakers are given in
Table 2.1.
The electrodynamic shakers are considered in order to have a good approximation of a typical experimental modal test setup. After all, the excitors will
affect the dynamical behaviour of the structure in just the same way as a force or
response transducer. Obviously the larger the excitation device, the greater the
influence on the system under test. In practice, by mounting the force transducer
as close as possible to the system the primary effects of the shakers can be avoided
in the measured forces. However, responses of the shaker on its suspension will
still be observed in the Auto Power Spectra of the force measurements as so-called
force drop-outs, which can be reduced by the use of a constant current amplifier
ensuring the application of a constant force.
Given the effective mass, damping and stiffness matrices of the system, the
exact frequency response functions can be computed using Newtons equation of
motion:
Mx
(t) + C x(t)
+ Kx(t) = f (t)
f1
f e1
fp
ke
f7
ks
me
ms
ce
x1
ks
cs
f e7
ks
...
ms
cs
(2.62)
ke
ms
cs
me
ce
x7
Figure 2.5: Simulated 7 DOF system excited by 2 electrodynamic shakers.
2.7. Validation
case
a
b
c
d
35
1E-4
1E-2
1E-4
1E-2
F
(40dB
(20dB
(40dB
(20dB
SNR)
SNR)
SNR)
SNR)
1E-12
1E-12
1E-12
1E-10
X
(40dB
(40dB
(40dB
(20dB
SNR)
SNR)
SNR)
SNR)
P
1E-2
1E-2
1E-1
1E-1
Table 2.2: Different noise cases considered for comparison of nonparametric FRF estimators.
with M , C and K respectively the mass, damping and stiffness matrices, and f (t)
and x(t) the applied force and structural response vectors. Using the Laplace
transform, the frequency-domain equivalent is given as
G(s)X(s) = F (s)
(2.63)
with G(s) = M s2 + Cs + K the dynamic stiffness, where the transfer function

matrix H(s), is found as
1
H0 (s) = G(s)
(2.64)
The applied forces fe are assumed noise-free and equal to the electrical signals
(sent to the power amplifier) up to a certain force/current factor. Two forces f 1 and
f7 and seven responses (displacements) x1 , . . . , x7 were considered. Both channel
and process noise are are taken into account during the simulations. The shaker
noise sources NS (f ) (cf. Eq. 2.5) are not considered during the simulations.
Based on the exactly known FRFs, all signals could be simulated directly in the
frequency-domain, with the advantage that no leakage phenomena are present
in the data and as a result, the application of time-windows, such as Hanning,
was not needed. Throughout this chapter it is implicitly assumed that leakage is
negligible, while Chapter 3 will address the problem of leakage.
Example 1: A first simulation experiment was done to study the performance of
following FRF estimators: H1 , Hv and Hiv using a random noise excitation and Hev
using a multisine excitation. For a random excitation (with normal distribution
N (0, 1)) and no measurement noise added on the data, Figure 2.6 shows the Auto
Power Spectrum of the force f1 and response x1 obtained during 50 simulated
experiments (M = 50) as well as the exact FRF between these DOFs with 6
modes in the frequency band 540Hz.
The effect of measurement noise on the nonparametric FRF estimators is studied by considering both channel and process noise. Gaussian distributed random
noise is added on the force (N (0, F )) and response (N (0, X )) Fourier data to
simulate the effect of channel noise. The process noise is simulated by applying an additional but unmeasured random noise (N (0, P )) excitation on mass
2. Table 2.2 summarizes the different noise cases used in this example, with an
indication of the overall SNR on the force and response signals.
36
20
|G | (N )
10
0
10
20
10
15
20
25
Freq (Hz)
30
35
40
10
15
20
25
Freq (Hz)
30
35
40
10
15
20
25
Freq (Hz)
30
35
40
140
|G | (m2)
160
180
200
220
240
260
|H| (m/N)
50
100
150
Figure 2.6: Example of the Auto Power Spectra of force f1 (top) and response x1 (middle) measured during 50 experiments and the exact FRF between these DOFs (bottom).
The force drop-outs are present due to the effect of the excitators.
For these 4 noise cases, the FRFs estimated using the H1 , Hv , Hiv and Hev ,
are compared with the true FRF and respectively shown in Figures 2.7, 2.8, 2.9
and 2.10 for the FRF(1,1), i.e. an impedance measurement at the mass 1 during
50 experiments. Since for modal analysis, one is mainly interested in the modes,
a zoom around the resonance peaks is shown.
The H1 yields biased estimates for all 4 cases. The systematic errors are largest
in the resonances and the bias is related to the signal-to-noise ratios of the input
signals (cf. Eq. 2.25). Comparing cases a and b, denoted as (a b), and (a d)
learns that the bias effect is mainly due to the input channel noise, which is not
taken into account by the H1 noise model. The comparison (a c) illustrates the
effect of the process noise, since due to the feedback through the shaker-system
interaction, a part of this noise is found in the input measurements and is not
taken into account. Notice however that, given the smaller bias error, this noise
contribution is smaller compared to the increase in channel noise (a b). On
2.7. Validation
37
the other hand, the part of the process noise (present on the outputs) as well
as the output channel noise are included in the noise model and hence have no
contribution to the bias of the H1 estimates.
The results obtained by means of the Hv approach, shown in Figure 2.8, are
very similar to those of the H1 estimator. This can be explained by the important difference in signal amplitudes between the forces and responses (i.e.
rms(F )
rms(X) + 10E4). This causes the Hv estimator, which is bounded by the H1
and H2 estimator, to over-emphasize the input signals, and so the input-noise,
leading to a result close to the H1 estimates. By scaling the amplitude of the
signals with for instance their rms value prior to the Hv calculations, the Hv tends
in general to be more consistent as illustrated for noise case d in Figure 2.11.
As shown in Figure 2.9, the Hiv estimator yields fairly accurate FRF estimates
for all the considered noise cases. This shows that the use of an errors-in-variables
noise model will improve the FRF measurements by taking also the noise on the
input measurements into account. Since the process noise, present on the force
and response measurements, is uncorrelated with the (noise-free) source signals
fe , the effect of this noise is eliminated by averaging over a sufficient number
of experiments. This illustrates that the Hiv estimator, is certainly preferred to
the H1 , especially in the case of low channel SNRs and a significant presence of
(unmeasured) environmental noise. An important example are forced vibration
tests on a bridge where the wind load can introduce significant process noise.
In practice, the Hiv estimator only requires the additional storage of the source
(electrical) signals that are send to the shaker amplifiers. Since the number of
inputs is typically not more than 5, this amount of data is still small compared to
the number of response DOFs.
Comparison of the Hiv and Hev learns that both estimators perform well. Nevertheless, the use of a multisine excitation still has the advantage of better SNRs,
since a constant amplitude is applied at each spectral line of the observed frequency band (in the case that no interaction exists with shaker). As a result, for
the same number of experiments, the uncertainty on the measurements will be
lower in the case of the multisine. Although the SNR for random noise signals
increases rapidly for a small number of experiments, because dips in the averaged
input power spectrum (due to the random behaviour) disappear, it is shown in
[109] that four experiments are needed to guarantee that 95% of the measurement points have a SNR corresponding to that of a well-designed, deterministic
excitation after one period. A rule of thumb for practical applications is that 5
experiments with random noise are needed to have a comparable SNR to 1 experiment with a deterministic excitation (such as a multisine). This is certainly
an important factor with respect to the overall testing time, although multiple
experiments are required by the averaging process. On the other hand, random
noise excitation is still often used in modal analysis since possible nonlinear distortions are reduced by averaging a sufficient number of records, say 10 or more.
38
50
60
55
65
60
|H| (m/N)
|H| (m/N)
70
65
75
70
80
75
80
8.4
8.6
8.8
Freq (Hz)
85
13.4
13.6
13.8
Freq (Hz)
14
95
80
100
|H| (m/N)
|H| (m/N)
90
100
105
110
110
120
130
24
25
26
Freq (Hz)
27
115
34.5
35
35.5
Freq (Hz)
36
36.5
Figure 2.7: Effects of channel and process noise on the H1 FRF estimator. Comparison
of estimated FRF(1,1) with true FRF(1,1) ( ) for noise cases a (), b ( ), c ( )
and d ( ). Both the process and channel noise on the input result in an important
systematic error.
The same averaging effect is obtained by means of random multisine excitation,

which, however, is not commonly available in modal testing equipment.
Using Eq. (2.22), the covariance matrix can be computed for the H1 estimate
of the FRF matrix, from which the uncertainty bounds on the FRF estimates are
found. In a similar way the covariance matrix is computed for the Hiv and the Hev
(Hiv using a periodic excitation) based on Eq. (2.39). Figure 2.12 shows, for noise
2.7. Validation
39
50
60
55
65
|H| (m/N)
|H| (m/N)
60
65
70
70
75
80
75
80
8.4
8.6
8.8
Freq (Hz)
85
13.4
70
13.6
13.8
Freq (Hz)
14
95
80
100
|H| (m/N)
|H| (m/N)
90
100
105
110
110
120
130
24
25
26
Freq (Hz)
27
115
34.5
35
35.5
Freq (Hz)
36
36.5
Figure 2.8: Effects of channel and process noise on the Hv FRF estimator. Comparison
and d ( ). Both the process and channel noise on the input result in an important
systematic error due to an important difference in signal amplitude.
case d, the amplitude difference between the true FRF(1,1) and respectively the
H1 (top), Hiv (middle) and Hev (bottom) estimate as well as the corresponding
99% confidence interval. This result illustrates the systematic errors in the H 1
estimate around the resonance frequencies since the amplitude difference is clearly
larger than the 99% uncertainty bound. On the other hand, both the Hiv and
Hev estimates fall within the 99% uncertainty bounds, and hence the amplitude
40
50
60
55
65
60
|H| (m/N)
|H| (m/N)
70
65
75
70
80
75
80
8.4
8.6
8.8
Freq (Hz)
85
13.4
13.6
13.8
Freq (Hz)
14
95
80
100
|H| (m/N)
|H| (m/N)
90
100
105
110
110
120
130
24
25
26
Freq (Hz)
27
115
34.5
35
35.5
Freq (Hz)
36
36.5
Figure 2.9: Effects of channel and process noise on the Hiv FRF estimator. Comparison
and d ( ).
difference are solely the result of the uncertainty on the measurement data. This
confirms the results presented in Figures 2.7, 2.9 and 2.10 and confirms that the
H1 is a biased estimator, while the Hiv and Hev are unbiased. Furthermore, Figure
2.12 shows a lower uncertainty on the FRF estimates for the Hev , which is explained
by the better SNRs during the multisine measurements.
After all, the bias effect of the H1 estimator will result in errors on the modal
parameters, and especially the damping and mode shapes. This is illustrated in
2.7. Validation
41
50
60
55
65
60
|H| (m/N)
|H| (m/N)
70
65
75
70
80
75
80
8.4
8.6
8.8
Freq (Hz)
85
13.4
13.6
13.8
Freq (Hz)
14
95
80
100
|H| (m/N)
|H| (m/N)
90
100
105
110
110
120
130
24
25
26
Freq (Hz)
27
115
34.5
35
35.5
Freq (Hz)
36
36.5
Figure 2.10: Effects of channel and process noise on the Hev FRF estimator. Comparison of estimated FRF(1,1) with true FRF(1,1) ( ) for noise cases a (), b ( ), c (
) and d ( ).
Figure 2.13 by comparing the operational deflection shapes (i.e. amplitude of the
FRF at each response location at a resonance frequency) for the different estimators (noise case d). The difference between the true mode shape and the different
estimates is clearly larger for the H1 due to the bias in the FRF estimates, while
the Hiv and Hev are in good agreement. Again, the benefit of using a multisine is
observed since the Hev has a smaller difference compared to the Hiv .
42
75
80
85
|H| (m/N)
90
95
100
105
110
115
120
23.5
24
24.5
25
25.5
Freq (Hz)
26
26.5
27
27.5
24
24.5
25
25.5
Freq (Hz)
26
26.5
27
27.5
75
80
85
|H| (m/N)
90
95
100
105
110
115
120
23.5
Figure 2.11: Scaled Hv () (top) and unscaled Hv () (bottom) compared to H1 ( )

and Hiv ( ) FRF(1,1) estimate.
Example 2: In order to assess the validity of the methodology for obtaining the
covariance matrix of the Hiv FRF estimates based on Eq. (2.39), Monte carlo
simulations were performed and compared to the true covariance matrix. Using
the expressions (2.29) and (2.39), 100 experiments were simulated in the frequency
band of 628Hz using 20 averages (M = 20) and the noise settings of case c (cf.
Table 2.2), from which a Monte Carlo estimate of the FRF covariance matrix was
obtained. The true covariance matrix can be computed based on the true FRF
matrix H0 and shaker dynamical stiffness matrix G0 as well as the knowledge of
the added noise and the expressions for C(Hiv ).
Figure 2.14 illustrates the validity of these expressions, given the good agreement between the true and the Monte Carlo results. The differences, for some of
the covariances are only due to the number of 20 averages used to compute the
FRF matrix estinates in each Monte Carlo run. By increasing M the uncertainty
2.7. Validation
43
50
10
|H| (m/N)
|H| (m/N)
20
10
20
8
10
12
Freq (Hz)
14
50
10
|H| (m/N)
|H| (m/N)
24
25
26
Freq (Hz)
27
24
25
26
Freq (Hz)
27
24
25
26
Freq (Hz)
27
50
20
10
20
8
10
12
Freq (Hz)
14
50
50
20
10
|H| (m/N)
|H| (m/N)
10
20
8
10
12
Freq (Hz)
14
50
Figure 2.12: Amplitude difference () between the true FRF(1,1) and respectively the
H1 (top), Hiv (middle) and Hev (bottom) estimate and the 99% confidence interval ( ).
on the estimates will further decrease to the true values. It can be seen that the
effects of the process and input noise become important in the resonance peaks resulting in the higher uncertainties. This can be explained by the shaker-structure
interaction, which introduces mutually correlated noise on all the signals. The
feedback to the input-side is largest in the resonances, since then this interaction
becomes most important, which was also indicated by the important force-drops
in Figure 2.6, also resulting in lower SNRs on the inputs.
Example 3: As discussed in the previous example, the true covariance matrix
can be computed based on the true FRF matrix H0 and shaker dynamical stiffness
44
x 10
|H| (m/N)
2
1
0
1
2
3
4
DOF
4
DOF
x 10
|H| (m/N)
0.5
0
0.5
1
Figure 2.13: Comparison of the operational deflection shapes of the first (top) and
second mode (bottom) obtained from the true (+ ) FRFs and H1 ( ), Hiv ( )
and Hev ( ) FRF estimates (noise case d).
matrix G0 as well as the knowledge of the added noise. By means of this, the effects
of the different possible noise sources on the correlation between the FRFs can be
studied.
Using again the 7 DOF system (cf. Figure 2.5), the correlation matrix of
C(Hiv ) is computed for three cases by considering respectively only output, input
and process noise. Figure 2.16, shows the results at a resonance frequency of
the mode at 8.74Hz (left) and a near frequency at 9.23Hz (right) which is still
near this mode as shown in Figure 2.6. The effect of output channel noise on
the FRF correlations can be seen in the top figures. Although this channel noise
is mutually uncorrelated, an important interaction between the shakers and the
structure, also results in a correlation between the FRFs measured between the
two input and same output locations. Referring to Eqs. (2.37), (2.39) and (2.39),
R
learns that, in the case that no input or process noise exists and CX and G
2.7. Validation
45
60
Var(H11) [(m/N) ]
60
80
80
100
100
120
120
Var(H12) [(m/N) ]
60
Cov(H11,H12) [(m/N) ]
80
100
120
140
10
15
20
25
140
11
21
10
15
20
25
160
Cov(H ,H ) [(m/N) ]
60
140
11
60
31
80
80
80
100
100
120
120
120
140
140
140
160
10
15
20
25
72
60
62
15
20
25
10
Cov(H ,H ) [(m/N)2]
72
60
52
80
100
100
100
120
120
120
140
140
140
160
160
160
180
180
25
41
10
15
20
Freq (Hz)
25
15
72
60
80
15
20
Freq (Hz)
25
20
25
Cov(H ,H ) [(m/N)2]
80
10
20
160
10
Cov(H ,H ) [(m/N)2]
15
11
60
100
160
10
Cov(H ,H ) [(m/N)2]
Cov(H ,H ) [(m/N) ]
180
10
42
15
20
Freq (Hz)
25
Figure 2.14: Comparison between the true () and estimated (Monte Carlo) ( ) FRF
covariance matrix CHiv .
are diagonal, correlations can still be introduced by the non-diagonal elements

FR , which can become large due to an interaction between
of the FRF matrix H
1
the exciters and the structure (cf. Figure 2.15). This figure gives again the nondiagonal elements of the same FRF matrix when using smaller exciters (i.e. ke, ce
and me 10 times smaller than values in Table 2.1). Since the input locations are
not a node for the considered mode, this interaction is large near the resonance
frequency resulting in the additional correlations as shown in Figure 2.16. In order
to illustrate the effects of this interaction, Figure 2.17, shows similar results for
the smaller exciters, where no correlation is introduced.
46
0
10
|H| (m/N)
20
30
40
50
60
6
10
12
14
16
18
Freq (Hz)
20
22
24
26
28
FR
Figure 2.15: The non-diagonal elements of the FRF matrix H
for the output noise
1
only case when using large () and small ( ) exciters.
For the case that only input channel noise is considered, the results are shown
R are diagonal,
in Figures 2.16 (middle) and 2.17 (middle). Although CF and G
correlations are again introduced between the FRFs corresponding to the multiple
FR
input (and same output) locations by the shaker-structure interaction (i.e. H
1
non-diagonal). The correlation between the FRFs corresponding to the same input
XR in the matrix
location results from the fact that near a resonance the matrix H
1
H
FR
product MCov(H1 )M has a rank 1. Consequently the FRFs corresponding to

the same input location will be totally correlated. In the case of the first bending
the mode, DOF 4 is a node (cf. Figure 2.13) and hence the elements for that DOF
in the transfer function matrix M will be zero. This difference of correlation for
2 shaker setups can be noticed by comparing Figures 2.16 and 2.17.
Based on the EV relationship, process or environment noise introduces mutually correlated noise on all the outputs. At the same time, the shaker-structure
interaction introduces a feedback to the inputs as well. Hence, since all elements
of CEZ generally differ from zero, and since only one source of process noise is
considered in the simulations, the total correlation of all the FRFs corresponding
to the same input location and partial correlation between FRFs with different
input is explained again by Eq. (2.39).
As can be expected from the results shown Figure 2.14, the measurement noise
can certainly introduce important correlations between the different FRFs. It
is however important to remember that a high correlation between the FRFs,
does not necessarily has an important influence in terms of the actual FRF noise
covariances depending on the process and measurement noise as well as on the
2.7. Validation
47
2
0.8
0.6
FRF
FRF
6
8
0.8
0.4
10
0.6
0.4
10
0.2
12
14
2
8
FRF
10
12
14
0.2
12
14
8
FRF
10
12
14
2
0.8
0.6
0.8
FRF
FRF
0.4
10
0.6
0.4
10
0.2
12
14
2
8
FRF
10
12
14
0.2
12
14
8
FRF
10
12
14
2
0.8
4
6
0.6
0.4
10
0.8
FRF
FRF
0.6
0.4
10
0.2
12
14
2
8
FRF
10
12
14
0.2
12
14
2
8
FRF
10
12
14
Figure 2.16: Effects of the measurement noise on the FRF correlations at the frequency
corresponding to the first bending mode at 8.74Hz (left) and at a near frequency at 9.23Hz
(right). The correlation matrix shows the effect of respectively only output noise (top),
input noise (middle) and process noise (bottom).
48
2
0.8
0.6
FRF
FRF
6
8
0.8
0.4
10
0.6
0.4
10
0.2
12
14
2
8
FRF
10
12
14
0.2
12
14
8
FRF
10
12
14
2
0.8
0.6
0.8
FRF
FRF
0.4
10
0.6
0.4
10
0.2
12
14
2
8
FRF
10
12
14
0.2
12
14
8
FRF
10
12
14
2
0.8
4
6
0.6
0.4
10
0.8
FRF
FRF
0.6
0.4
10
0.2
12
14
2
8
FRF
10
12
14
0.2
12
14
2
8
FRF
10
12
14
Figure 2.17: Similar results as in Figure 2.16, but using smaller excitors, reducing the
shaker-structure interaction significantly.
2.8. Conclusions
49
dimensioning of the excitors. Moreover, the output correlations can be eliminated

by using a Scanning Laser Doppler Vibrometer setup (cf. 2.4).
2.8
Conclusions
Classical modal analysis starts from the frequency response functions that are
estimated using FRF estimators. The data is typically acquired using a random
excitation. However, in general, too simplified noise models assumed by these
estimators, result in biased FRF estimates. Furthermore, the uncertainty on the
FRF estimates is yet never considered in modal analysis during the parametric
identification process.
The application of parametric stochastic estimators, such as for instance the
frequency-domain ML estimator, in the domain of modal analysis requires that the
noise information for the FRF data is known. Moreover, since the accuracy of the
parametric results is related to the quality of the measured data, the maximum
likelihood approach is only meaningful in the case that consistent FRF estimates
can be obtained.
This chapter has revised the nonparametric FRF identification in a maximum
likelihood framework based on an errors-in-variables stochastic noise model. This
model provides an accurate representation of the measurement setup taking the
effects of channel, shaker and process noise on all the inputs and outputs into
account. Table 2.3 summarizes the results of the comparison for the different
FRF estimators. From this, it can be concluded that the estimators currently
used in practice, i.e. H1 , H2 and Hv are generally not suited for accurate FRF
identification since no general errors-in-variables noise model is considered. If the
noise covariance matrix would be a priori known, the Hv can be generalized to
the Hgtls , which belongs to the class of maximum likelihood estimators. However,
since this noise covariance matrix is never a priori known in practice, this H gtls
estimator is not applicable.
EV noise model
ML properties
No a priori noise info
MISO
General applicability
H1
++
++
++
H2
++
++
Hv
++
++
Hgtls
++
++
Hiv
++
+
++
++
+
Hev
++
++
++
++
Table 2.3: Comparison of different nonparametric FRF estimators.
For this reason, the so-called nonparametric instrumental variables or H iv estimator was studied for FRF identification for arbitrary excitation signals. It has
50
been shown in this chapter that this estimator has maximum likelihood properties,
without requiring any a priori information about the measurement noise. Nevertheless, this approach uses additional information under the form of the generator
signals, which in practice only requires the additional acquisition of a very limited number of signals. Multivariable expressions were derived for both the FRF
and noise covariance matrix from the same measurement data. When nonlinear
distortions are introduced by the excitation devices, the consistency property remains valid, only resulting in some loss of the asymptotical efficiency. Moreover,
the Hiv remains also consistent when a MISO approach is used to process the
extensive amount of data, which is an important property for the application of
modal analysis. The methodology is validated and compared with the classical
nonparametric approach for a simulated vibration experiment that considers the
interaction between the shaker and the structure. It is shown that both the noise
on the measurements and these interactions can result in important correlations
between the FRFs.
For the case of periodic signals, the errors-in-variables FRF estimator Hev is
a special case of this instrumental variables Hiv estimator. The Hev remains its
ML properties when nonlinear distortions are introduced by the excitation devices.
Although the Hev does not require the generator signals, its practical applicability
is constraint by the class of excitation signals, since periodic signals are still not
often used for modal analysis practices.
2.9. Appendix A The GTLS Estimator per Output
2.9
51
Appendix A The GTLS Estimator per Output
H diagonal, the cost function (2.8) can

E B
Assuming the (No No ) matrix BC
Z
also be written as
H ) 1 B
E B
G
Z B
H
`(H)
= M trace diag(BC
Z
1 [G]
= M trace diag[C]
(2.65)
with
diag[C]
[G]
c11
0
...
0
m11
m12
...
m1No
0
m21
c
.
.
.
0
m
.
.
.
m2No
22
22
.
.
.
.
.
..
.
.
..
..
.. ..
..
..
..
.
0
0
. . . c No No
mNo 1 mNo 2 . . . m No No
c11 m11
c11 m12
...
c11 m1No
c22 m21
c22 m22
...
c22 m2No
..
..
..
.
.
.
.
.
.
c No No mNo 1 c No No mNo 2 . . . c No No mNo No
proving that the cost function boils down to the sum of cost functions obtained
= [H,
IN ] C No Ni +No separately
by estimating each row of B
o
= M (c11 m11 + c22 m22 + . . . + cN N mN N )
`(H)
o o
o o
(2.66)
G
Z B
H is diThe same result is found for the case that the (No No ) matrix B
G
Z B
H is asymptotically
Z = GZ +CE , with GZ the true data, B
agonal. Since G
0
Z
0
H is diagonal.
E B
(M ) a diagonal matrix if BC
Z
From this it is also trivial that Equation (2.65) is equivalent with
= M trace diag(BC
H ) 1 diag(B
G
Z B
H )
E B
`(H)
Z
2.10
(2.67)
E
Appendix B Unbiased Estimate of G
X
For simplicity, the scaling factor occurring in Eq. (2.24) will be proven assuming
there is only one output. The relation between the exact (noise-free) input Fourier
vectors and the output Fourier coefficient at spectral line f can be written as
[m]
X [m] = HF [m] + EX
(2.68)
52
Consider the following normalized cost function

2
M E [m]
X
X
`=
GEX0
m=1
(2.69)
E
[m]
with GEX0 the true variance of the output noise. Assuming GXo
EX0
distributed
as N (0, 1) then the cost function ` is 2 (0, 1) (Chi-quadratic) distributed and its
expected value equals
E{`} = M Ni
(2.70)
where M Ni corresponds to the number of degrees of freedom of the Chi-quadratic
distribution. M is the number of equations and Ni number of constraints when
minimizing the cost function (2.69), i.e.
`
=0
H1
...
`
=0
HNi
(2.71)
An estimate of the output noise variance is given by (2.23) as

M
X
[m] 2
E = 1
E
G
X
M m=1 X
(2.72)
which expected value also equals, given the above derivation
GEX0
(M Ni )
E
E G
=
E{`} =
GEX0
X
M
M
(2.73)
Chapter 3
FRF Identification in the

Presence of Leakage
In practice, the estimation of frequency response functions (FRFs) is often complicated by the influence of noise on the measured data as well as by spectral leakage
in the case of random excitation. Chapter 2 discussed accurate FRF identification
in the presence of measurement noise by considering FRF estimators developed in
an errors-in-variables framework. This chapter deals with the problems related to
leakage, due to the finite measurement time, when arbitrary excitation signals are
used. In modal analysis, an exponential time-window is often applied, for reducing
both the effects of leakage and measurement noise, in the case of a time-limited
excitation such as for pulse or burst random signals. The possibility of applying an
exponential time-window in combination with random noise excitation is studied
by revising the estimation of Auto and Cross Power Spectra in Section 3.3, while
the same is done starting directly from FRFs in Section 3.4. A comparison of
the different approaches in Section 3.7, including other methods suggested in the
literature, learns more about their effectiveness to cope with leakage phenomena.
Finally, an experimental case study for a subframe of a car studies the applicability
in practice of the most important methods derived in both chapters 2 and 3.
53
54
3.1
Chapter 3. FRF Identification in the Presence of Leakage
Introduction
Chapter 2 discussed accurate FRF identification in the presence of measurement

noise by considering FRF estimators with ML properties that were developed in
an errors-in-variables framework. However, another important source of errors for
modal analysis applications is spectral leakage due to a finite measurement time
of non-periodic signals. Errors-in-variables FRF identification in the presence of
leakage can still be seriously hampered since the damping of modes is usually
overestimated corresponding to a bias (systematic) error.
Periodic (broadband) excitation [39], certainly avoids leakage problems and
can improve the signal-to-noise ratio compared to arbitrary signals (cf. 2.7).
Arbitrary signals, such as random noise, have the advantage of reducing possible nonlinear distortions by averaging over a sufficient number of measurements.
As a compromise between random and deterministic excitation, periodic random
signals can be used, however with the important drawback of an increased measurement time. Each new random sequence should be applied during a certain
time before starting the actual measurement in order to avoid the transient effects
(such as leakage) appearing each time a new sequence is applied. This transient
behavior, however, depends on the damping of the system as well as on the desired
frequency resolution. For low damped systems, the length of each period should
be sufficiently long in order to have enough information (i.e. frequency resolution)
in the small resonance peaks. In that case, the additional waiting time generally
is not more than one period (i.e. typically 20 80% of the measurement period).
The same is true for pseudo random (i.e random multisine signals), however with
the advantage that the amplitude spectrum is constant (while the phase is still
random), resulting in better SNRs compared to periodic random.
However, arbitrary signals, such as random noise, are still most often used
for modal testing because of their general availability as well as the reduction of
non-linear distortions. Furthermore, in general, arbitrary signals are a better approximation for the operational conditions of a wide class of mechanical structures
compared to deterministic signals. A common approach to reduce the leakage effects is the application of a time-window (e.g. a Hanning window) that certainly
helps to reduce the effects of leakage. However the problem will not be completely
eliminated, especially for the case of lightly damped structures.
For that reason, the use of a time-limited burst random excitation (combined
with the implicit rectangular finite time-window) has gained popularity in the
modal analysis community since it allows the signals to decay to zero before the
end of the time window. Analogously to impact testing, an exponential window
can also be applied in order to amplify this decay by artificially increasing damping
by means of an exponential window. The use of an exponential window has the
important advantage that the effect of this window can be totally compensated for
on the modal parameters. However, disadvantages of burst random, and transient-
3.2. Simulated Vibration Experiment
55
30
20
|m/s /N| (dB)
10
0
10
20
30
40
50
10
15
20
25
Freq. (Hz)
30
35
40
Figure 3.1: The exact FRF(1,1) between the acceleration and force at DOF 1.
like excitation in general, are a worse signal-to-noise ratio and a higher crest factor
compared to random noise excitation.
Therefore, this chapter studies the possibility for reducing both the effects of
leakage and measurement noise by combining the application of a random noise
excitation and an exponential time-window. A comparison of several approaches is
made with existing techniques available in literature. The different approaches will
be studied by means of the simulated vibration experiment that is first discussed
in the next section. At the end of this chapter, an experimental case study for a
subframe of a car illustrates the applicability in practice of the most important
methods derived in both chapters 2 and 3.
3.2
Simulated Vibration Experiment
The same 7 DOF system as presented in 2.7 is used again in order to simulate a
modal testing experiment. For simplicity reasons, only a single excitor is considered at DOF 1, without any loss of generality of the presented results. In order
the study the problem of leakage, time-domain signals were now generated using
a discrete-time state space formulation based on the shaker and system characteristics given in Table 2.1. Using this model, the exact FRFs (cf. Figure 3.1)
and modal parameters for the 6 modes in the band of 540Hz (cf. Table 3.1) were
computed.
56

mode
1
2
3
4
5
6
fd (Hz)
8.6711
13.6263
24.2989
26.4022
35.1201
35.6307
(%)
0.0939
0.1300
0.2270
0.2645
0.4047
0.4185
Table 3.1: Damped natural frequency and damping ratio for the 7 DOF system.
Normal distributed noise was used to excite the system in a band of 050Hz.
The force fe , applied to the shaker coil, is assumed noise-free and equal to the
electrical signal (sent to the power amplifier) up to a certain force/current factor,
while the actual force acting on the system and the 7 measured responses are
respectively denoted as f1 and x1 , . . . , x7 . Given the low damping of the structure, time histories with a total number of 128K (1K = 1024) time samples were
measured (corresponding to a total measurement time of about 1300s) given the
sample rate of Ts = 0.01s.
Based on these time histories, the frequency response function (FRF) data is
derived using averaging-based nonparametric processing techniques. As discussed
in chapter 2, when using random excitation, this FRF identification, is classically
done using the noise-on-output H1 estimator (for angular frequency f )
1
XF
H
1 = GXF GF
(3.1)
or in an EV framework by means of the Hiv estimator

XR G
1 = G
XF = G
XR G
1 G
RG
1 = H
FR1
XR H
H
iv
1
1
FR
R
FR
(3.2)
However, given the finite total amount of data samples acquired, the user
is always faced with a tradeoff in an effort to maximize the number of spectral
averages (i.e. minimizing the uncertainty) while keeping a sufficient frequency
resolution (i.e. maximizing the number of samples within the record). When
the number of data samples within one record becomes small, leakage errors will
certainly be introduced in the FRF estimates. Since this chapter focusses on the
effects of leakage on the nonparametric processing, measurement noise was now
not considered in the simulation experiment.
3.3
Power Spectra Estimators
Basically, two classical approaches exist to estimate the Auto and Cross Power
Spectra in Eqs. (3.2) and (3.1) [89]. The Periodogram estimator operates directly
3.3. Power Spectra Estimators
57
on the Fourier data to yield a Power Spectra (PS) estimate. On the other hand,
the Correlogram estimator first estimates correlation functions, and then obtains
the PS estimate by means of a DFT. These approaches are now revised for the
possibility of applying exponential windowing techniques, in order to reduce the
effects of leakage when using random noise excitation.
3.3.1
Periodogram PS Estimator
In practice, Auto and Cross Power Spectra are generally estimated using the Welch
Periodogram estimator [149, 89] where this approach was also implicitly used in
2.7, for the computation of the nonparametric FRF estimates. This PS estimator
introduces time-windows to reduce the bias error caused by leakage and optionally
segment overlapping to increase the number of records for the (asynchronous)
averaging process.
Given the time histories of the reference (i.e. force on shaker coil), force and
response signals with a total number of Ntot samples measured at a sampling
rate Ts . Dividing these histories in M adjacent records results in the time records
r[m] (tn ), force f [m] (tn ) and response x[m] (tn ) as defined by the EV model Eq. (2.1),
where m = 1, . . . , M and tn = nTs , n = 0, 1, . . . , Ns 1 with Ns the number of
samples within each record. In general, given the finite total number of samples
acquired Ntot = M Ns , there is a trade-off between a high spectral resolution
(Ns as large as possible) and minimizing the estimation variance (M as large
as possible) since the variance decreases as O(M 1 ) while the bias error due to
leakage increases for decreasing Ns .
In order to increase the number of segments for averaging, and therefore to
decrease the variance of the PSD estimate, segment overlapping can be introduced
by shifting over S (S M ) samples between adjacent records. With a total
number of time samples Ntot = M Ns the maximum number of segments available
for averaging is then given by the integer part of (Ntot Ns )/S + 1. However,
caution is needed since overlap also introduces correlations between the adjacent
records [106].
Compared to the rectangular time-window (corresponding to the finite time of
each record), applying for example a Hanning or Hamming window to the records
prior to the DFT transform, reduces the effect of the sidelobes and so the leakage
error at the price of only a slight decrease in resolution. Such a time-window
forces the random signals to be periodic by reducing the discontinuities at the
boundaries of the finite record. However, this also reduces the total energy in the
measured signal resulting in an attenuation of the amplitude, which is typically
a bias error. Although, the amount of attenuation solely depends on the window
shape and so this error can be compensated. However, since the estimate of the
FRF matrix (3.1) or (3.2) is found as the ratio of Power Spectra, a compensation
58
for the attenuation of the time-window is cancelled out and hence is not considered
here in the derivation of the expressions for the Power Spectra.
Applying a time-window to the records z [m] (tn ) R Ni +No 1 and r [m] (tn )
T
R
, where z [m] (tn ) = f [m]T (tn ), x[m]T (tn ) results in the following samples
Ni 1
[m]
zw
[n]
= w[n]z [m] [n]
[m]
[n]
rw
= w[n]r [m] [n]
(3.3)
for 0 n Ns 1 and 1 m M , where the windowed segment Periodogram

matrix is computed
[m] (f ) = Z [m] (f )R[m] (f )H
(3.4)
G
ZR
over the angular frequency range f = 2f /Ns Ts , f = 0, 1, . . . , Ns /2 1, where
Z [m] (f ) and R[m] (f ) are the DFT (cf. Eq. 2.2) of the records (3.3). The Welch
Cross Power Spectra estimator is then given by the linear averaging of the segment
Periodograms
M
X
[m] (f )
P (f ) = 1
G
(3.5)
G
ZR
M m=1 ZR
P (f ). Although a wide
where a similar expression is found for the Auto PS G
R
range of possible time-windows w(tn ) can be considered for this approach (Hanning, Hamming, Nuttall, Gaussian, Chebyshev, ...), an exponential window is not
applicable for this approach. However, since the modal parameters (i.e. damping)
can be compensated for an exponential window, a possible alternative approach
for the calculation of the PS estimates is studied in the next section.
As discussed in Chapter 2, the FRF covariance matrix is found using the expressions (2.22) and (2.39) respectively for the case of the H1 and Hiv estimator.
Remark 1 : Alternatively, Power Spectral Density functions, which are scaled to
the unity frequency resolution and compensated for the energy attenuation due to
the time-window, can be computed as
[m]
SZR (f ) =
1
Z [m] (f )R[m] (f )H
U N s Ts
(3.6)
over the angular frequency range f = 2f /Ns Ts , f = 0, 1, . . . , Ns /2 1, where

U is the discrete-time window energy
U = Ts
NX
s 1
w2 (tn )
(3.7)
n=0
This factor U removes the effect of the window energy bias.

Remark 2 : Depending on the considered FRF estimator, H1 or Hiv , the notations used for the signals differ. For the case of the H1 estimator (3.1), the force
59
25
20
M = 1,2,4,8,16
M = 1,2,4,8,16
18
16
20
|m/s N| (dB)
|m/s N| (dB)
14
15
12
10
8
10
6
8.64
8.66
8.68
Freq. (Hz)
8.7
13.6
13.65
Freq. (Hz)
13.7
Figure 3.2: The H1 estimator using Periodogram approach for increasing number of
averages M = 1, 2, 4, 8, 16 and decreasing blocksize Ns = 128, 64, 32, 16, 8K (Ntot = 128K
samples) together with the true FRF (bold line).
signals f [m] (tn ) are used instead of the generator signal r [m] (tn ) and hence the
XF (f ) and Auto PS G
F (f ) are only computed. This convention is used
Cross G
throughout this chapter.
Example: The effect of leakage is now illustrated for the H1 estimator using
the Periodogram estimator to compute the FRF estimates for the 7DOF system,
introduced in 3.2. Figure 3.2 shows one of the estimated FRFs, obtained by
decreasing the blocksize Ns and increasing number of averages M , while the total
number of time samples was fixed to Ntot = 128K (1K=1024). A Hanning window
was applied to reduce leakage. Zooming in on the first 2 modes clearly illustrates
the effects of leakage that are present in the data, resulting in an important bias
error. This error can be reduced by using only a small number of averages, however
resulting in noise FRF estimates. The low damping of the considered system,
explains why the leakage errors become important even in the case of the large
blocksizes.
3.3.2
Correlogram PS Estimator
Given again M adjacent records of Ns of the generator r [m] (tn ), force and response
z(tn ) obtained by the division of the time histories. The (Ni + No ) Ni unbiased
cross correlation matrix estimate between the two signal vectors z(tn ) and r(tn ) ,
60
is computed as
(
gzr [j] =
PNs j1
1
Ns j
1
Ns |j|
z[n + j]r T [n]

Pn=0
Ns |j|1
z[n]r T [n + |j|]
n=0
for
for
0 j Ns 1
(Ns 1) j < 0
(3.8)
The biased correlation estimate uses 1/Ns rather than 1/(Ns j).
Next, an estimate of the windowed segment Correlogram matrix is obtained
by taking the DFT of the finite sequence of the cross correlation estimate (3.8)
C (f ) =
G
ZR
L
X
w[l]
gzr [l]e(if lTs )
(3.9)
l=L
The maximum lag index L is typically less than the number of data samples in
order to avoid the greater statistical variance associated with the higher lags of
the correlation estimates. Although a maximum of L Ns /10 was suggested
by Blackman and Tukey [18], for mechanical systems this cut-off length strongly
depends on the damping of the considered system. Furthermore, the correlogram
PS estimator is asymptotically unbiased as the number of lags increases [18]. The
FFT algorithm can be used to perform a fast computation of the correlogram
estimate by the convolution method [35].
An odd-length (2L+1)-samples lag time window w[l] (e.g. Hanning, Hamming,
Chebyshev, ...) over the interval L l L and symmetric about the origin can
be applied in (3.9). This window reduces the effect of leakage due to the large
sidelobes of the implicit rectangular window and therefore the bias error in the
estimate. This is illustrated, for the considered 7 DOF system, in Figure 3.3, where
the cross-correlation function gx1 f (3.8) with Ns =8192 samples is shown. At the
higher lags, the greater statistical variance can be noticed. Applying for example
a Hanning window of length (2L + 1)-samples with L = Ns results in Figure 3.3,
forcing the correlation function to zero at the higher lags to reduce leakage effects.
However, when applying a time-window, the poles (especially the damping)
will be affected. In the case that an exponential window is used, the poles can be
exactly compensated for the added damping, which is not the case for other time
windows such as e.g. Hanning and Hamming. Hence, for the purpose of modal
parameter estimation (i.e. the poles and residues), the correlogram estimator
is best combined with an exponential window. The exponential window adds
artificial damping shifting the stable poles more to the left an the unstable poles
more to the right in the Nyquist plane.
C (f ), a double-sided exponenGiven the form of the correlogram estimate G
ZR
tial time-window w[l]
w[l] = e|l|Ts
with
Ll L
(3.10)

gxf[l]
0.1
61
0.05
0.08
0.04
0.06
0.03
0.04
0.02
0.02
0.01
0.02
0.01
0.04
0.02
0.06
0.03
0.08
0.04
0.1
5000
0
lag number
5000
0.05
w[l]gxf[l]
5000
Hanning
0
lag number
5000
Figure 3.3: The cross-correlation function gx1 f with Ns =8192 (left) and the effect from
applying a Hanning window (right).
w[l]g [l]
xf
0.05
Expon 1%
w[l]g [l]
0.04
0.04
0.03
0.03
0.02
0.02
0.01
0.01
0.01
0.01
0.02
0.02
0.03
0.03
0.04
0.04
0.05
5000
0
lag number
xf
0.05
5000
0.05
5000
Expon 0.1%
0
lag number
5000
Figure 3.4: The effect of applying an exponential window to the cross-correlation function gx1 f with different factor (left: 1% at end of window) and (right: 0.1% at end of
window).
can be applied in order to reduce the effect of leakage (due to the implicit rectangular window) and noise (greater statistical variance) associated with the higher
lags of the correlation estimates. Typically, for a time-limited signal, such as a
pulse or burst random excitation, the factor is chosen such that amplitude at
the last time lag of the exponential window is 1% of its initial amplitude [62]. Fig-
62
ure 3.4 shows, for the cross-correlation function gx1 f , the effect of an exponential
window with different values for . As can be seen in this case, 0.1% at the end of
the exponential window seems a more appropriate choice since the higher damping
only reduces the greater variance at the higher lags, while the system dynamics
(around the central lag) are not much affected.
Moreover, an exponential time-window offers the advantage that the extracted
poles of the modal model can be corrected for the artificial damping introduced
by the window that is a priori known by the choice of the -factor
pcorrected
= pestimated
+
r
r
(3.11)
Remark : In the case that the instrumental variables approach is used in combination with a random noise excitation, the generator signals generally are white
noise sequences, in which case this approach can be further simplified. Given the
relation in the continuous-time domain between the measured time signals z(t)
and the generator signals r(t), with h(t) R (No +Ni )Ni the impulse response
function matrix
z(t) = hzr (t) r(t)
(3.12)
Convolution with r(t) yields
gzr (t) = hzr (t) grr (t)
(3.13)
Given that the generator signals r(t) are normalized white noise signals, the autocorrelation function matrix is then given by grr (t) = INi (t), with (t) a Dirac
pulse function. Hence, Eq. (3.13) indicates that the correlation functions g zr (t)
are impulse response functions since for causal systems the negative lags are zero.
This implies that it is sufficient to consider only the positive lags of the correlogram
estimate (3.9) since they contain all the information needed to identify the modal
parameters
L
X
C (f ) = Ts
w[l]
gzr [l]e(if lTs )
(3.14)
G
ZR
l=0
Figure 3.5 shows the both cross-correlations gf r and gx1 r . As can be seen from this
figure, these correlation functions have the form of an impulse response functions
and hence only the positive lags contain the dynamics of the system.
Given the generator r(tn ), force f (tn ) and response x(tn ) signals for the M
C [m] (f ) (3.9) can
subsequent records of Ns samples, a correlogram PS estimate G
ZR
be computed for each record (m = 1, . . . , M ). Assuming the segment correlograms
statistically independent, then the elements of the matrix
M
X
C [m] (f )
C (f ) = 1
G
G
ZR
M m=1 ZR
represent the sample mean of a set of M independent variables.
(3.15)

w[l]g [l]
fr
0.9
Expon 0.1%
w[l]g [l]
xr
0.08
0.8
Expon 0.1%
0.06
0.7
0.04
0.6
0.02
0.5
0.4
0.3
0.02
0.2
0.04
0.1
0.06
0
0.1
63
5000
0
lag number
5000
0.08
5000
0
lag number
5000
Figure 3.5: The cross-correlations gf r (left) and gx1 r (right) after applying an exponential window (0.1% at end of window).
An estimate of the sample variance of the correlogram estimates is given as

C (f )) = 1
var vec(G
ZR
M
!
M
2
X
[m]
1
C (f ))
C (f )) vec(G
vec(G
ZR
ZR
M 1 m=1
(3.16)
As for the Periodogram estimator there is a trade-off between a high spectral

resolution (Ns as large as possible) and minimizing the estimation variance (M as
large as possible). However, since in this case an exponential window can be used
to reduce the effects of leakage, the user could opt for more records of a smaller
blocksize.
Example: The H1 FRF estimates using the Correlogram approach estimated
for an increasing number of averages M = 2, 4, 8 and decreasing blocksize Ns =
64, 32, 16K (Ntot = 128K samples) are shown in Figure 3.6. An exponential window was applied such that the window was 0.1% at the end of the correlation
function (i.e. at lags Ns 1 and Ns + 1). Similar results are found for other
values of . Although an exponential window is applied to the correlation function in order to reduce leakage, the FRF estimates found by using the correlogram
approach clearly still suffer from leakage, at least in this case of the considered low
damped system. Moreover, taking the ratio of 2 Correlogram PS estimates that
already include the effects of an exponential window seems less robust than the
classical H1 using the ratio of 2 Periodogram estimates.
64
25
3
20
4
15
|m/s N| (dB)
|m/s N| (dB)
5
10
9
8.5
8.6
8.7
Freq. (Hz)
8.8
8.9
35
35.2
35.4
Freq. (Hz)
35.6
35.8
Figure 3.6: The H1 FRF using the Correlogram PS estimates for an increasing number
of averages M = 2, 4, 8 and decreasing blocksize Ns = 64, 32, 16K (Ntot = 128K samples)
and true FRF. The exponential window was 0.1% at the end of the correlation function.
3.4
Exponential Windowing of FRF Estimates
Instead of reducing the effects of leakage by acting on the calculation of the Power
Spectra, an exponential window can also be applied to the inverse DFT of the
FRF estimate given by (3.1) or (3.2). In this section it is shown how exponential
windowing can be used in combination with these FRF estimates in order to cope
with the trade-off between minimal variance and bias for the final FRF estimates.
3.4.1
Hexp FRF Estimator
An obvious way to minimize the effects of leakage is to measure long enough by

choosing a sufficient number of samples Ns within a record. Hence, the basic
idea of this approach consists of computing the FRF matrix estimate in Eq. (3.1)
or (3.2) by means of the Periodogram PS estimator (3.5), where Ns is sufficiently
large to ensure that both the effects of leakage and the time-window (e.g. Hanning)
are sufficiently small.
The damping behavior of the system determines the significant length of the
corresponding impulse response function (IRF). An idea of the optimal value for
Ns can be found for instance from the inverse Fourier transform of the Empirical
Transfer Function Estimate. This ETFE is computed as the ratio of the DFT of
the full-length random input and output signals, which is basically an H 1 estimator
using the Periodogram with no averages (M = 1). However, this often results in
3.4. Exponential Windowing of FRF Estimates
65
very noisy FRF estimates as can be seen in Figure 3.2.

A schematic overview of this approach is shown in Figure 3.7. In practice, given
the fixed total number of time samples Ntot , only a small number of averages, say
typically M = 14 can be used to obtain the FRF estimate in order to avoid effects
of leakage and the Hanning window. In this case, however, taking the inverse FFT
of the FRF estimate H(), results in an impulse response function h(t). In order to
improve the quality of the IRF, by reducing the noise for the higher time samples,
an exponential window can then be applied resulting hexp (t) and subsequently
Hexp () by computing again the FFT.
This is equivalent to a convolution with a low pass filter LP() in the frequencydomain
exp
HXF (f ) = HXF (f ) LP(f )
(3.17)
where LP(f ) is the DFT of an exponential window in the time-domain
LP(f ) =
NX
s 1
w[n]e(if nTs )
with
w[n] = e|n|Ts
(3.18)
n=0
The FRF matrix estimate HXF is computed using the Periodogram Power Spectra
estimator according the H1 (cf. Eq. 3.1) or Hiv (cf. Eq. 3.2) approach.
Since, the effect of the exponential window can be compensated for using
Eq. (3.11), this can be done without introducing a bias error on the final parameter
estimates. In the case of the classical approach, the Hanning window, applied to
smaller blocks in order to have a sufficient number of averages, introduces a bias
error that cannot be compensated and increases for a decreasing blocksize. Also
notice that data compression of the initial FRF data can be obtained by cutting
the IRF at a lower time sample (L Ns ) and next applying an exponential lag
window w[n] (0 n L) on the shorter IRF. Another possible approach for data
compression can be obtained by first applying an exponential on the full IRF and
next using cyclic averaging (cf. 3.6) to reduce the frequency resolution.
Remark : As for any of the nonparametric methods considered in this chapter,
this method requires a sufficient total number of time samples such that no leakage
is present in the empirical transfer function estimate.
The covariance matrix of HXF
exp
(f ) is found as
exp
Cov HXF (f ) = Cov HXF (f ) |LP(f )|2
(3.19)
exp
where Cov HXF (f ) is given by Eqs. (2.22) and (2.39) for respectively the case
of the H1 and Hiv FRF estimators. In [145], the validity of this expression has
been demonstrated by means of Monte Carlo simulations.
66
H( )
60
50
50
40
40
30
30
20
20
10
10
0
-10
Hexp ( )
60
0
0
500
1000
1500
2000
-10
2500
500
1000
IFFT
w(t) = e t
0.9
0.8
0.7
2.5
2
1.5
1
0.6
0.5
0
-0.5
-1
-1.5
0.5
0.4
1000
2000
3000
4000
5000
0.5
0
0.3
-0.5
0.2
-1
0.1
2500
hexp (t)
1.5
-2
2000
FFT
h(t)
2.5
1500
-1.5
0
1000
2000
3000
4000
5000
-2
1000
2000
3000
4000
5000
Figure 3.7: Schematic overview of Hexp estimator.
Example: Based on the ETFE of the FRF (i.e H1 with M = 1), the effect of
using an exponential window was studied for an increasing value of in Eq. (3.18).
Choosing = 1 , 21 , 41 , 81 , with 1 = 3.5135E-3, resulted in an exponential
window with a value of 1% at respectively Ns , Ns /2, Ns /4, Ns /8 and Ns /16,
thus increasingly cutting off. The length of the IRF (and thus the frequency
resolution) was kept the same (Ns = 128K) in all cases. In the frequency-domain,
this results in an increased smoothing effect of the low pass filter, which can be
seen in Figure 3.8 by zooming on the first mode and the modes 5 and 6 that
have a higher damping. It can be seen that the noise on the ETFE is smoothed,
while damping is artificially added, however, with an a priori known factor .
For the case that the exponential window is defined by increasing this factor 16
times, the added damping is approximately equal to the real part of the pole of
the first mode, meaning that the pole is located twice as far from the imaginary
axis of the Nyquist plane. In 3.7, the parametric results will be compared after
compensating the real part of the pole for the added damping.
3.4. Exponential Windowing of FRF Estimates
67
25
1 , 2, 4, 8, 16
4
20
|m/s N| (dB)
|m/s2N| (dB)
15
8
10
9
8.63
8.64
8.65
8.66 8.67 8.68

Freq. (Hz)
8.69
8.7
8.71
35
35.2
35.4
Freq. (Hz)
35.6
35.8
Figure 3.8: The true FRF (bold line), ETFE (dashed line) and the Hexp
estimator for
1
increasing value of , while length of the IRF (and hence the frequency resolution) was
kept the same.
3.4.2
Nonparametric Least Squares IRF Estimator
Another way for obtaining the IRF from the data is found by considering (for
simplicity) the (SISO) relation in the discrete-time domain between a measured
input and output signal
x[m] (tn ) = hxf (tn ) f [m] (tn )
(3.20)
where the input f [m] (tn ) and input x[m] (tn ) were divided over M adjacent records.
The IRF hxf (tn ) = hxf (nTs ) for n = 0, . . . , Ns 1 can be computed directly in
the time-domain by solving the following over-determined set of equations as a
classical least squares problem [f ]{hxf } = {x} or explicitly given by Eq. (3.21).
The submatrices of [f ] are non-symmetric (Ns Ns ) with a Toeplitz structure.
Solving this LS problem (3.21) yields an estimate of hxf (tn ). Notice however,
that the first submatrix of [f ] contains the samples of the generator signal at the
negative time instants, which in practice correspond to the initial conditions and
are generally not known.
Remark : The generalization for an errors-in-variables noise model using an instrumental variables measurement setup is straightforward, by considering the relation
(3.20) between the generator and respectively the force and response signals and
substituting the final FRF estimates in Eq. (3.2).
68
f0
f1
..
.
fN 1
s
f Ns
f
Ns +1
..
f2Ns 1
f2Ns
..
.
fM Ns 1
f1
f0
..
.
f2
f1
..
.
fNs 2

fNs 1
fNs +1
..
.
f2Ns 2

f2Ns 1
..
.
fM Ns 2
fNs 3

fNs 2
f Ns
..
.
f2Ns 3

f2Ns 2
..
.
fM Ns 3
..
.
..
.
..
.
f(Ns 1)
f(Ns 2)
..
.
f0

f1
f2
..
.
f Ns

fNs +1
..
.
f(M 1)Ns
hxf
0
hxf
hxf
..
xf
hNs 1
x0
x1
x
2
x
N
1
xN s
=
x
Ns +1
..
x
2Ns 1
x2Ns
..
xM Ns 1
(3.21)
In order to reduce the effect of leakage as well as noise, a one-sided exponential

window can be applied again to this IRF
hxf
LSexp
LS
(tn ) = hxf (tn ).e|n|Ts
with 0 n Ns 1 and finally H
LSexp
(3.22)
() is given by the FFT of (3.22).
An advantage of this approach, compared to the approach in 3.4.1, is that

the system IRF can be computed without having to transform to the frequencydomain. However, in order to solve the LS problem in a time-efficient manner,
Fast Toeplitz solvers need to be used to reduce the computation time when
Ns is typically 2048 or higher. This can be considered as a disadvantage of this
approach since these solvers can be numerically bad conditioned and assume that
[f ] is symmetric and positive definite. For Ns smaller than 2048 a simple matrix
inversion of [f ] can be used.
Instead of using all data samples to formulate the large LS problem (3.21),
M 1 compact LS problems can be solved from the M 1 last blocks of Ns
equations. This yields M 1 different estimates of the IRF hLSexp[l] (l=2,. . . ,M)
from which a sample mean and sample variance can be computed. Assuming the
M 1 IRFs statistically independent observations, the sample mean is given by
H
LSexp
(f ) =
1 X LSexp[l]
H
(f )
M 1
(3.23)
l=2
while an estimate of the sample variance is computed as
!
M
2
1
1 X LSexp[l]
LSexp
LSexp
()} =
(f )
var{H
(f ) H
H
M 1 M 2
l=2
(3.24)
3.5. Iterative Nonparametric FRF Estimator
3.5
69
Iterative Nonparametric FRF Estimator
A different approach for leakage reduction was proposed in [126]. The basic idea
is to remove the effect of leakage on the FRF by iterative improvement. Although
the approach was first presented for the H1 estimator (for SISO systems), it is
straightforward to generalize the results to the multivariable Hiv estimator. The
proposed algorithm comprises of the following steps:
ZR (f ) , f = 0, 1, . . . , Ns /2 1 using Eq. (3.1) and the peri1. Calculate H
1
odogram PSD estimator (cf. 3.3.1). A Hanning window is used to reduce
the leakage error.
ZR
= IFFT H
ZR : h[n]
(f ) .
2. Calculate the inverse FFT of H
1
1
3. Eliminate the offset errors in h[n]
with
h[n] = h[n]
NX
s 1
1
h[n]
Ns n
(3.25)
n=N2
The
reason for this is that the offset error, which is a systematic error, is
O( Ns ) if the input signal contains no DC component, else the error becomes
O(Ns ) [126]. A typical choice is N2 = 0.9Ns , since the algorithm requires
that h[n] can be set to zero for n > Ns .
4. Determine the optimal length N1 of the IRF, by looking for the minimum
of the approximation error s(N1 )
s(N1 ) =
N3
X
e2N1 [n]
(3.26)
n=N1 +1
PN 1
h[l]r[n l] and N3 a user-selected
with eN1 [n] = z[n] zN1 [n], zN1 [n] = l=0
number. Since the optimum is typically rather flat, it is sufficient to check
only a small number of different filter lengths (e.g. Ns /4, Ns /2, 3Ns /4, Ns .
If the optimal length is Ns , then it is strongly advised to increase the length
of the subrecords Ns .
5. Calculate the new (updated) FRF estimate
P 1
P
ZR
H
1 (f ) = HN1 + GeN1 R (f )GRR (f )
(3.27)
with HN1 the transfer function of the finite impulse response (FIR) filter
P 1
P
with length N1 . The error terms G
eN1 R (f )GRR (f ) can be interpreted as
a transient error as shown in [111] starting from the observation that what
is called a leakage error is essentially a transient problem.
70
25
20
M = 2, 4, 8, 16
M = 2, 4, 8, 16
18
16
|m/s N| (dB)
20
15
14
12
10
8
10
6
8.63
8.64
8.65
8.66
8.67
8.68
8.69
8.7
8.71
13.56 13.58
13.6
13.62 13.64 13.66 13.68

Freq. (Hz)
13.7
Figure 3.9: The iterative (H1 ) FRF estimator for increasing number of averages M =
2, 4, 8, 16 and decreasing blocksize Ns = 64, 32, 16, 8K (Ntot = 128K samples).
6. Repeat the procedure from step 3 with
ZR
= IFFT H
1 (f )
h[n]
(3.28)
The iteration process is controlled by monitoring the minimum of s(N1 ).

Once this error starts to increase, the iteration process is stopped.
Example: Using this iterative approach, the FRFs of the 7 DOF system were
estimated subsequently using M = 2, 4, 8, 16. The results, given in Figure 3.9,
show that this approach improves the FRF estimates compared to the classical
H1 . As cited in [126], this approach requires a large number of data samples in the
case of low damped systems as is typical for mechanical structures. Reducing N s
in order to reduce the variance, results in a bias error, which is however smaller
than for the H1 for comparable blocksizes (cf. Figure 3.2).
3.6
Cyclic Averaging Approach
Recently, the use of cyclic averaging was proposed as a method for reducing the
leakage error in the estimation process of frequency response functions for modal
analysis applications [6, 103, 104]. Cycling averaging, historically, has been used
in rotating machinery data analysis to enhance characteristics in the data that
are functions of the rotating speed. The same averaging technique has been used
for estimating FRFs to enhance the frequency characteristics in the data that
are integer functions of the observation window while minimizing the nonperiodic
3.6. Cyclic Averaging Approach
71
characteristics and so reducing the leakage. The main idea is to use the technique
of cyclic averaging prior to the FRF computation (e.g. H1 , Hiv ) by means of the
Periodogram estimator (cf. 3.3.1).
Consider again M adjacent records of the generator r [m] (tn ), force f [m] (tn ) and
response x[m] (tn ) signals with Ns samples in each a record and the total number of
time samples Ntot = M Ns . To illustrate the cyclic averaging technique, consider
for example a record of one of the measured response signals denoted as x o (tn ).
The Discrete Fourier Transform (DFT) of this record (cf. Eq. (2.2)) can also be
written as
NX
s 1
i2nf P
x(tn )e( Ns )
Xo (f P ) =
(3.29)
n=0
with f P = 2f P/Ns Ts and P an integer number such that f P = 0, P, . . .

Ns 1. Choosing for example P = 2, Eq. (3.29) for f = 1 becomes
Xo (2 )
NX
s 1
x(tn )e(
i4n
Ns )
(3.30)
n=0
Ns /21
i2n
x(tn )e( Ns /2 ) +
n=0
NX
s 1
i2n
x(tn+Ns /2 )e( Ns /2 )
n=Ns /2
where the Fourier transform is computed by describing the unity circle twice for
n = 0, 1, . . . , Ns . Using the time shift theorem of the Fourier transform this can
also be written
Xo (2 ) =
Ns /21 h
n=0
i i2n
x(tn ) + x(tn+Ns /2 ) e( Ns /2 )
(3.31)
and in general Eq. (3.29) becomes

Ns /P 1
Xo (f P ) =
n=0
" P 1
X
l=0
x(tn+lNs /P ) e(
i2nf
Ns /P
(3.32)
From this it follows that the cyclic averaging process is tuned by choosing P since
this divides a record of Ns samples into P blocks that are averaged by describing
the unity circle P times. This results in a Fourier spectrum that has Ns /P spectral
lines and thus with a frequency resolution that is P times smaller. Before cyclic
averaging, a Hanning window of length Ns is first applied to the records r [m] (tn ),
f [m] (tn ) and x[m] (tn ). The FRF computation is done by means of the Periodogram
approach (cf. 3.3.1) starting from the M cyclic averaged Fourier spectra.
Figure 3.10, shows the results for the classical Periodogram H1 estimator
(with M = 4) combined with cyclic averaging. The data was first cyclic averaged using P = 1, 2, 4, 8 blocks with a corresponding decrease of the blocksize
72
25
20
exact
M=4,P=1
M=4,P=2
M=4,P=4
M=4,P=8
18
16
|m/s N| (dB)
20
|m/s2N| (dB)
exact
M=4,P=1
M=4,P=2
M=4,P=4
M=4,P=8
15
14
12
10
8
10
6
8.63
8.64
8.65
8.66 8.67 8.68

Freq. (Hz)
8.69
8.7
8.71
13.56 13.58
13.6
13.62 13.64 13.66 13.68

Freq. (Hz)
13.7
Figure 3.10: The Periodogram H1 estimator combined with cyclic averaging with M = 4
and P = 1, 2, 4, 8 resulting in a decreasing blocksize Ns = 32, 16, 8, 4K samples (Ntot =
128K samples).
(Ns = 32, 16, 8, 4K) for a total of 128K samples. A Hanning window was applied
to reduce leakage.
Although cited in literature [6, 104] as an approach for the reduction of leakage,
it follows from these results that cyclic averaging is merely a signal processing
technique to reduce the computation time for deriving the FRFs. The amount
of memory required to store this data is also reduced since it contains a factor P
less frequency lines compared to the large blocksize case (i.e. Ns /2). However,
no gain in terms of SNR can be obtained while the measurement (testing) time
itself cannot be reduced. Moreover, the lower frequency resolution can also result
in a less accurate reading of resonance frequencies as can be seen in Figure 3.10
and closely spaced modes may be less accurate estimated during the parametric
identification. Furthermore, in this case of a low damped system, this approach
suffers from leakage in the same way as the classical H1 FRF estimators (for a
same number of averages M ).
3.7
Comparison of Methods
Since the decrease in frequency resolution and the additional damping (bias) introduced by the exponential window make a nonparametric comparison a difficult
task, the comparison is based on the modal parameters. These were estimated using a frequency-domain maximum likelihood estimator (cf. Chapter 5). Since the
bias error mainly affects the damping of the modes, the parametric results were
3.8. Experimental Case Study
73
compared by means of the relative bias error for the damping defined as
Re(pr p0 )
Re(p0 ) .
The results for the classical H1 , Hexp

and the iterative Hiter
approach are given
1
1
in Table 3.2 for the damping ratio of first mode. The Correlogram approach was
not considered, given the poor nonparametric results.
Ns (K)
H1
128
64
32
16
8
-0.0068
-0.0170
0.0683
0.0886
0.2369
Hexp
1
(M = 1)
0.0006
0.0008
0.0015
0.0018
0.0020
Hexp
1
(M = 2)
/
-0.0168
-0.0167
-0.0168
-0.0207
Hiter
1
/
-0.0172
0.0217
0.0552
0.0964
Table 3.2: Relative bias error for the damping ratio (%) of the first mode for different
approaches.
When decreasing the blocksize Ns , the H1 estimator (cf. 3.3.1) clearly suffers
from leakage, resulting in a systematic error up to 24% for the case of Ns = 8K,
while the newly proposed FRF estimator Hexp
(cf. 3.4) using the ETFE has
1
only a very small bias error of 0.2%. Notice, that even in the case that the added
damping is as large as the damping of the mode (i.e. Ns = 8K where = 161 ),
a proper compensation for the exponential window (cf. Eq. 3.18) can still be done
on the parametric results. When applying the Hexp
estimator with M = 2, it
1
can already be seen that a small systematic error exists on the initial H 1 estimate,
before applying the exponential window (3.17), since the relative bias error remains
similar for an increasing Ns . The iterative approach Hiter
clearly performs better
1
that the classical H1 , although this method also introduces an important bias error
once the blocksize becomes too small as was also explained in 3.5.
3.8
Experimental Case Study
This section will discuss the application of the most important methods derived
in Chapters 2 and 3 for the nonparametric processing of the data obtained from
a subframe of a car. Figure 3.11 shows the multi-input modal test setup using
two electrodynamic shakers. The subframe is supported by elastic chords in order
to approximated the free-free conditions. Time histories of 256K samples are
measured for the two applied forces and 23 responses (accelerations in vertical or
+Z direction) using a random noise excitation in a frequency band 01024Hz with
a sampling rate of 2048Hz. Since this subframe structure is characterized by a low
damping ( 0.5 1.5%), caution is needed with respect to leakage.
The Hiv estimator was used to estimate the FRF (2.29) and corresponding
74
noise covariance matrix (2.39). Hereto, the time histories were divided into 32
blocks (M = 32) of 8K time samples (Ns = 8K) on which a Hanning window
was applied to reduce possible effects of leakage. Figure 3.12 shows the FRF(1,1)
as well as its variance obtained by the Hiv estimator, from which it can be seen
that the uncertainty on the data is highest at the resonance frequencies. This
is explained by considering the signal-to-noise (SNR) ratios for the corresponding
force, due to force-drops in the input at resonance frequencies. The response signal
has lower SNRs at the anti-resonance frequencies. Furthermore, the harmonics of
50Hz can be noticed in both the force and response. These are introduced by the
power amplifiers, and so can be considered as a part of the excitation explaining
why these harmonics are not present in the FRF estimates.
Next, different FRF estimators are compared for this practical case in Figure 3.13 (top) by zooming in on the first mode at 55.5Hz. The signal processing
parameters were the same as for the Hiv , i.e. M = 32, Ns = 8K and a Hanning
window. As can be seen, the Hiv is situated between the H1 and H2 estimate, while
the Hv almost coincides with the H2 . Computing the bias for each frequency using Eq. (2.25) and Eq. (2.28) for respectively the H1 and H2 estimator, results in
Figure 3.13 (bottom), indicating that both estimators yield a biased result. Comparing the difference between the H1 and Hiv and the same for the H2 with the
theoretically computed bias, clearly indicates that Hiv is a consistent FRF estimate since an errors-in-variables noise model is considered. On the contrary, the
bias error is largest for the H2 explained by the poor SNR of the response signal
in the lower frequency range.
Based on the knowledge of the full noise covariance matrix, it was also possible
to compute the degree of correlation between each of the Hiv FRF estimates,
as shown for 3 close frequencies around the first damped resonance frequency at
55.5Hz in Figure 3.14 (a). As can be seen at the resonance frequency of 55.5Hz,
the FRFs corresponding to the same input have a high correlation, which can be
XR in Eq. (2.39) has a
explained by the fact that near a resonance the matrix H
1
rank 1. As a result, in practice, the FRFs corresponding to the same input have
a high correlation, except for the DOFs corresponding to nodal points of that
mode. Furthermore, some correlation is introduced by the interaction between the
shakers and the structure, explaining the medium correlation between FRFs for
the same response DOF and the two input locations. From this experimental case
study, it also follows that higher correlations do appear at resonance frequencies,
while being small even only a few DFT lines away from the resonance. Moreover,
in this case the shaker-structure interaction is small, avoiding high correlations
between FRFs corresponding to different input and same response location.
When increasing the number of blocks M to 64 and 128 and hence decreasing
the blocksize Ns to 4K and 2K, leakage certainly poses a problem given the low
damping of the subframe. Although the use of a Hanning window reduces the
leakage effects resulting in Figure 3.15 (top), an attenuation (bias error) is present.
The effect of leakage is clear from Figure 3.15 (bottom) where a rectangular window
3.9. Conclusions
75
is used resulting in important errors in the H1 FRF estimate as could be expected.

On the contrary, the results obtained by the Hiv are still acceptable even for the
short blocks of 2K samples (M = 128). This indicates that the Hiv is clearly
more robust for leakage, compared to the H1 , which can be explained by the fact
that leakage mainly translates into stochastic errors that are non-coherent with the
generator signals (or instruments) and hence are averaged out by the Hiv approach.
Finally, the subframe data was processed with the classical H1 , Hiter
and Hexp
1
1
estimators (cf. Chapter 3), for which the damping ratios estimated by the parametric ML estimator (cf. Chapter 5) are summarized in Figure 3.16. For the H1
k
and Hiter
1 , the number of blocks M was increased as M = 2 for k = 1, 2, . . . , 7.
From the parametric results given in Figure 3.16 based on the H1 (top) and Hiter
1
(middle) it clearly follows that an overall improvement is obtain by means of the
nonparametric iterative estimator, especially for the case that important leakage
effects are present in the data (i.e. M > 5). However, applying the Hexp
esti1
mator, for an increase of the -factor (and thus increasing cut-off) of the applied
exponential window as = k1 for k = 1, 2, . . . , 7, while Ns is fixed to 128K samples, results in Figure 3.16 (bottom). This clearly illustrates the robustness for
leakage of this last estimator since the initial FRF estimate (cf. Figure 3.7) is computed based on a small number of large (and leakage free) blocks. Also notice that
the additional damping introduced by the exponential window can be correctly
compensated on the modal damping even for the higher cut-off () values.
3.9
Conclusions
The problem of leakage, encountered when using arbitrary excitation for modal
testing, has been studied in this chapter. Leakage effects result from using arbitrary signals, such as random noise, within a finite measurement time. It is
shown that, in the case of the classical approach for FRF estimation, using a Periodogram estimator, the severity of these errors depends on the trade-off between
maximizing the number of samples within a time record (and hence the frequency
resolution) to reduce the effects of leakage, and maximizing the number of records
available for averaging to reduce the statistical variations on the FRF estimates.
In modal analysis, the use of an exponential window is often applied to reduce
leakage in the case of a time-limited excitation such as for an impact or burst
random excitation. In this chapter the use of an exponential window has been
generalized for random noise excitation. Several possible approaches for leakage
reduction on the estimated FRFs have been studied and compared with techniques
proposed in the literature. The most important characteristics of the different
methods are summarized in Table 3.3.
The robustness for both leakage and measurement noise is poor for the classical
FRF estimators Hper and Hcycl given the trade-off conflict. As a result, the number
76
Robustness leakage / noise

Window Compensation
Computation efficiency
Hper
++
Hcycl
++
Hcorr
+
+
Hexp
++
++
++
HLS
++
++
Hiter
+
+
+
Table 3.3: Comparison of different estimators for nonparametric FRF identification in

the presence of leakage.
of records M needed to have an acceptable variance will introduce problems of

leakage, especially in the case of low damped mechanical systems. For the same
reason, the effect of a Hanning window, which cannot be completely compensated
for, becomes more important for a decreasing number of samples within a record.
On the other hand, these approaches are very fast, which is a requirement for
processing the large data sets in the case of modal analysis applications.
Although, the use of the Correlogram approach for FRF estimation, makes the
application an exponential window in combination with random noise excitation
possible, the ratio of the correlogram Power Spectra that already contain the effects
of the window, appears not robust in practice, again resulting in important errors.
On the contrary, the Hexp FRF estimator, based on the use of an exponential
window in combination with the classical FRF estimators (such as e.g. H 1 and
Hiv ), turns out to be a robust method with respect to both leakage and measurement noise. By using large blocksizes, the effects leakage on the FRF estimates
are minimized, while the exponential window is actually used to reduce the statistical variance on the FRF estimate. Although this results in biased nonparametric
results, the additional damping is known and its effect on the modal parameters
can be compensated, which is not the case for e.g. a Hanning window. Furthermore, the noise covariances for the windowed FRF estimates Hexp can still be
determined. The robustness for leakage of this method was clearly demonstrated
compared to the classical Hper and the iterative Hiter FRF estimators. Furthermore, this method is simple to apply in practice, since commercial measurement
equipment for modal analysis often only stores the measured FRF data. This
method requires a sufficient total number of time samples such that no leakage
is present in the empirical transfer function estimate, which however is a general
requirement for each of the discussed nonparametric FRF estimators.
An alternative method, based on the same idea, starts from the impulse response functions that are found by solving a nonparametric Least Squares problem
directly in the time-domain. This HLS FRF estimator has similar performance
with respect to accuracy for leakage and measurement noise as the Hexp , but requires too much computation time when used in practice for typical modal analysis
problems.
Finally, the nonparametric iterative FRF estimator Hiter , that reduces some
3.9. Conclusions
77
of the leakage effects by iterative improvement, clearly performs better than the
classical approach. Nevertheless, this method is also confronted with the trade-off
problem, requiring sufficient records in order to yield an acceptable stochastic error
on the FRF estimates. As a result, the effects of leakage and the Hanning window
become important in the case that the blocksize becomes too small. Nevertheless,
further optimization of this method by means of exponential windowing, can also
result in a robust method for FRF identification in the presence of leakage, besides
the Hexp approach.
78
Figure 3.11: Multi-input modal test setup using two electrodynamic shakers. The responses are measured in 23 DOFs using accelerometers.
3.9. Conclusions
60
79
H
iv
var(Hiv)
40
Ampl. (dB)
20
20
40
60
80
100
200
300
400
Freq (Hz)
500
600
100
200
300
400
Freq (Hz)
500
600
700
50
Ampl. (dB)
40
30
20
10
SNR F
0
700
50
Ampl. (dB)
40
30
20
10
SNR X
0
100
200
300
400
Freq (Hz)
500
600
700
Figure 3.12: FRF(1,1) and variance obtained by the Hiv estimator (top) and SNRs for
corresponding force (middle) and response signal (bottom).
80
20
H1
H2
Hv
Hiv
15
Ampl. (dB)
10
10
54.5
55
55.5
Freq (Hz)
56
56.5
H
2
H
1
4
3
Bias Error (dB)
2
1
0
1
2
3
4
5
54.5
55
55.5
Freq (Hz)
56
56.5
Figure 3.13: Comparison of different FRF estimators for subframe data (top) and theoretical bias for H1 and H2 result (bottom).
3.9. Conclusions
81
1
5
0.9
10
0.8
15
0.7
0.6
20
FRF
Ampl. (dB)
10
0.5
25
0.4
30
0.3
35
15
0.2
40
0.1
20
45
54
56
58
60
62
10
20
Freq (Hz)
(a)
FRF
30
40
(b)
0.9
0.9
10
0.8
10
0.8
15
0.7
15
0.7
0.6
0.5
25
0.4
30
0.6
20
FRF
20
FRF
0.5
25
0.4
30
0.3
35
0.3
35
0.2
40
0.1
45
10
20
FRF
(c)
30
40
0.2
40
0.1
45
10
20
FRF
30
40
(d)
Figure 3.14: Degree of correlation between each of the Hiv FRF estimates for 3 close
frequencies around the first damped resonance frequency at 55.5Hz shown in (a), i.e. at
53.5Hz (b), 55.5Hz (c) and 57.5Hz (d).
82
60
55
50
Ampl. (dB)
45
40
35
30
25
20
15
10
H
1
H
iv
540
550
560
570
580
590
600
610
620
630
640
Freq (Hz)
60
55
50
Ampl. (dB)
45
40
35
30
25
20
15
10
H
1
H
iv
540
550
560
570
580
590
Freq (Hz)
600
610
620
630
640
Figure 3.15: Comparison of the H1 and Hiv for increasing M to 64 (solid and dashed
line) and 128 ( and ) when using a Hanning (top) and rectangular (bottom) window.
3.9. Conclusions
83
1.8
mode 1
mode 2
mode 3
1.6
1.4
Damping ratio (%)
1.2
1
0.8
0.6
0.4
0.2
0
1.8
4
k (M = 2k)
mode 1
mode 2
mode 3
1.6
Damping ratio (%)
1.4
1.2
1
0.8
0.6
0.4
0.2
0
k (M = 2 )
0.2
mode 1
mode 2
mode 3
0.18
0.16
Damping ratio (%)
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
4
beta = k.beta1
Figure 3.16: Comparison of the damping ratios (%) estimated by the parametric ML
based on the H1 , Hiter
FRFs for a decreasing blocksize (M = 2k for k = 1, 2, . . . , 7) and
1
exp
the H1 for an increasing -factor ( = k1 for k = 1, 2, . . . , 7).
84
Chapter 4
Introduction to Modal
Parameter Estimation
This chapter introduces the concepts of parametric system identification for modal
parameter estimation. First, parametric models that can be used to describe the
vibration behaviour of mechanical systems are discussed in Section 4.2. The different models are interconnected by means of a matrix polynomial description. Next
Section 4.3 gives an extensive overview of parametric identification methods originating from control theory and different engineering areas, with their potentials
and drawbacks from the point of view for MPE.
85
86
4.1
Chapter 4. Introduction to Modal Parameter Estimation
Introduction
The next section discusses several models that can be used to represent the information about the vibration behaviour of a structure, present in the measurements, with a limited number of parameters. Although the modal model is most
appropriate for a physical understanding of this vibration behaviour, most identification techniques do not directly identify this model since it is nonlinear-inthe-parameters. Instead, models such as ARMAVX and State Space and Matrix
Fraction Descriptions originating from electrical engineering and control theory
are often used. The modal parameters can be derived once a model is identified
from the data, given the linear relation between the different models.
An overview of identification methods that can be used to extract the parameters of these models is given in Section 4.3. Based on the properties of the identified
model and the algorithm implementation, a better insight in the potential for extracting the model parameters from large modal data sets is obtained. A major
grouping is usually done based on the domain in which the data is treated numerically resulting in time-domain and frequency-domain methods. This overview also
gives a good introduction and motivation for the parametric approach that will be
used in Chapters 5 and 6.
4.2
Parametric Models
Matrix Fraction Descriptions

A first model to describe the vibration behaviour of a mechanical structure with
Nm degrees of freedom is given by Newtons equation of motion:
Mx
(t) + C x(t)
+ Kx(t) = f (t)
(4.1)
with M , C and K respectively the mass, damping and stiffness matrices, and f (t)
and x(t) the applied force and structural response vectors. Using the Laplace
transform, the frequency-domain equivalent is given as
G(s)X(s) = F (s)
(4.2)
with G(s) = M s2 + Cs + K the dynamic stiffness. A matrix inversion yields

F (s)H(s) = X(s)
(4.3)
with H(s) the transfer function matrix, which can be expressed as
1
Gadj (s)
H(s) = G(s)
=
|G(s)|
(4.4)
4.2. Parametric Models
87
While the numerator Gadj (s) (the adjoint matrix) is an (Nm Nm ) matrix containing polynomials in s of order 2(Nm 1), the denominator is a polynomial in
s of order 2Nm . Therefore, Eq. (4.4) can also be written as
B1,1 (s) . . . B1,Nm (s)
..
..
..
.
.
.
H(s) =
B(s)
=
A(s)
BNm ,1 (s)
. . . BNm ,Nm (s)

A(s)
(4.5)
which is called a common denominator model or also scalar matrix fraction description.
In fact, this expression can also be considered as a special case of multivariable
transfer function models described using a Matrix Fraction Description (MFD),
i.e. the ratio of two matrix polynomials [74, 51]. Two kinds of MFDs are possible,
i.e. a left MFD (LMFD)
1
H(s) = AL (s)
BL (s)
(4.6)
where the matrix coefficients ALr are (No No ) matrices and BLr are (No Ni )
matrices or a right MFD (RMFD)
1
H(s) = BR (s) AR (s)
(4.7)
where the matrix coefficients ALr are (Ni Ni ) matrices and BLr are (No Ni )
matrices. When the frequency response function matrix H(f ) is used as primary
data, both descriptions can be linearized in the matrix coefficients (parameters),
while for input/output Fourier data only the LMFD results in a linear relationship.
From the LMFD, Eq. (4.4) is found for the specific case that the number of output observations No equals the number of degrees of freedom N since AL (s) = G(s)
and BL (s) a constant (No Ni ) matrix. Notice that AL (s) now is a second (low)
order matrix polynomial since the matrix coefficients are (Nm Nm ) matrices.
Since the dynamic stiffness G(s) is a matrix polynomial of second order with the
coefficients (Nm Nm ) matrices, the characteristic equation, given by the determinant of G(s), is of order 2Nm . However, in the case that the number of
response locations No is smaller than the number of degrees of freedom (modes)
Nm , this model has to be expanded to a higher order (2n) matrix polynomial with
smaller matrix coefficients in order to identify Nm modes (roots, poles or modal
frequencies)
2n s2n + 2n1 s2n1 + 2n2 s2n2 + . . . + 0 = 0
(4.8)
of which the characteristic equation and hence the roots are the same (up to a
scale factor) as for the second order matrix polynomial equation (4.2). The limit
of this expansion process results in a polynomial of high order 2Nm with scalar
coefficients
2Nm s2Nm + 2Nm 1 s2Nm 1 + 2Nm 2 s2Nm 2 + . . . + 0 = 0
(4.9)
88
and likewise, the roots of this scalar characteristic equation are the same as for
the second order matrix polynomial equation.
Therefore, the number of characteristic values that can be determined depends
upon both the size of the (Nc Nc ) matrix coefficients and the order of the matrix
polynomial (nj ) given by the following relation Nc .nj = 2Nm .
For the case of modal analysis, the number of responses No is typically much
larger than the number of applied forces Ni . Hence, for a LMFD (4.6), AL (s)
is a low order matrix polynomial with large matrix coefficients (No No ), while
for the RMFD (4.7), AR (s) is a high order matrix polynomial with small matrix
coefficients (Ni Ni ). Based on this relation, a so-called Unified Matrix Polynomial
Approach (UMPA) was proposed in [5], for the comparison of different estimation
algorithms using a common mathematical structure.
Another class of models often used for time-domain system identification that
also uses matrix fraction descriptions, are the so-called ARMAVX models [85, 131].
A commonly-used formulation for the input-output relationship is described as
x(t) + A1 x(t 1) + . . . + Ana x(t na )
= B1 f (t 1) + . . . + Bnb f (t nb ) + e(t)
(4.10)
where the Ar are (No No ) matrices and the Br are (No Ni ) matrices. Defining
the backward shift operator as q 1 f (t) = f (t1), the following matrix polynomials
in q 1 can be introduced
A(q)
1 + A1 q 1 + . . . + Ana q na
B(q)
B1 q 1 + . . . + Bnb q nb
(4.11)
and equation (4.10) becomes

x(t) = H(q)f (t) + G(q)e(t)
(4.12)
with
H(q) = A1 (q)B(q)
G(q) = A1 (q)
(4.13)
where the factorization of H(q) in terms of two matrix polynomials is also called a
LMFD. This model is called the ARX model, where AR refers to the autoregressive
part A(q)x(t) and X to the input or exogeneous part B(q)f (t). In the special case
where na = 0, x(t) is modelled as a finite impulse response (FIR). By introducing
more matrix polynomials in the equation error model structure other variants are
derived, such as ARMA, ARMAX,... [85].
Pole-Residue Model
Often used in modal analysis is the Partial-Fraction Description (PFD), also called
the pole-residue parameterization, which is found through a partial fraction ex-
4.2. Parametric Models
89
pansion of Eq. (4.4)

Nm
X
H(s) =
r=1
m
X
Rr
Rr
+
s pr r=1 s pr
(4.14)
with pr and Rr respectively the poles and (Nm Nm ) residue matrices as model
parameters corresponding to the rth mode of the structure. Although a PFD is
numerically better conditioned than high-order polynomial models (MFDs), the
PFD is nonlinear-in-the-parameters. Equivalently, the inverse Laplace transform
yields this model in the time-domain
h(t) =
Nm
X
R r ep r t +
Nm
X
Rr epr t
(4.15)
r=1
r=1
For modal analysis purposes, the residue matrices are further decomposed as [62]
Rr = Qr r Tr = r LTr
(4.16)
with r , Lr respectively the mode shape and modal participation factor vector
and qr the modal scale factor for mode r. Modal participation factors are a result
of multiple reference (input) modal parameter estimation and define how each
mode is excited from each of the input locations.
State Space Formulation
Reformulating Eq. (4.1) as

0
x(t)
=
M 1 K
x
(t)
I
M 1 C
x(t)
x(t)
0
M 1
f (t)
(4.17)
and defining respectively the state vector y(t) and the output and direct inputoutput transmission matrices C and D as
x(t)
and C = [I 0] , D = [0]
(4.18)
y(t) =
x(t)
a state-space description of the system is given by

y(t)
Ay(t) + Bf (t) + w(t)
x(t)
Cy(t) + Df (t) + v(t)
(4.19)
with the matrices A, B respectively the (dynamical) system and input matrix,
which together with C and D represent the systems realization describing the
systems response at any time to a known input signal. The eigensolutions of
the (2Nm 2Nm ) system matrix A yield the system poles, while the mode shape
and modal participation factor vectors are respectively found from the matrices
90
C and B. The zero-mean white noise vectors w(t) and v(t) can be added to
model respectively the process noise and output measurement noise. By using an
auxiliary state vector, Eq. (4.1) has become a first order differential (or difference)
expression, i.e. the state space model. The frequency-domain version of the state
space formulation (4.19) (with the noise vectors w(t) = 0 and v(t) = 0) yields a
general expression for the transfer function matrix
1
H(s) = C sI A B + D
4.3
(4.20)
Overview of MPE Algorithms
During the last three decades, extensive effort has been devoted to the development of methods that aim to produce a reliable identification of the dynamical
behaviour of a structures by means of the modal parameters. As a result of the
introduction of the Fast Fourier Transform (FFT) and the development of multichannel acquisition and powerful storage and computation capacity, the methods
have evolved from very simple single degree of freedom (SDOF) techniques to
methods that analyse data from multiple-input excitation and multiple-output
responses simultaneously in a multiple degree of freedom (MDOF) approach.
SDOF frequency-domain methods
The earliest methods work in the frequency-domain and belong to the class of
SDOF methods. These methods assume that the modes are real and the dynamical behaviour within a small frequency band can be described by a single mode.
The Peak Amplitude method [17] simply takes the natural frequencies from the
observation of the peaks on the graphs of the magnitude of the frequency response
function. The damping ratios are derived from the sharpness of the peaks and the
mode shapes are calculated from the ratios of the peak amplitudes at various response locations. Like other SDOF methods, as for instance Quadrature Response
[101] or Circle Fitting method [76], the Peak Amplitude method will yield erroneous results, especially for the damping, when the modes are closely spaced and
coupled, because only one mode can be estimated at the same time. Moreover,
the SDOF methods are very sensitive to measurement noise and require extensive interaction from the user. Nevertheless, these methods are still used even in
commercial analysis software, mainly because of the limited computation time,
although during the recent years the multiple degree of freedom methods have
completely taken over when an accurate MPE is wanted.
MDOF time-domain methods
In the class of MDOF methods, a major grouping is usually done based on
the domain in which the data are treated numerically resulting in time-domain
and frequency-domain methods. One of the first MDOF time-domain algorithms
4.3. Overview of MPE Algorithms
91
specifically designed for MIMO MPE was the Ibrahim Time-Domain (ITD) method
[66]. By means of the Modal Confidence Factor (MCF) [65], this method has been
adapted for an automated model order selection in order to minimize the required
user-interaction. The main disadvantage of the ITD is the computational complexity that origins from solving the large eigenvalue problems that result from
reformulation the time-domain equivalent of the modal model (4.15) as an eigenvalue problem. Since this method starts from the free-response time data, the
eigenvectors (mode shapes) cannot be mass-normalized. Taking sparse matrix
structures into account it was possible to reduce the computational load [67], although this faster version is still too slow to be used for current MPE practices.
Another well-known time-domain algorithm, which starts from the state space
formulation (4.19) (with the noise vectors w(t) = {0} and v(t) = {0}) to derive
the minimum realization of the system matrices, is the Eigensystem Realization
Algorithm (ERA) [71, 70]. Although this algorithm has been successfully applied in
many modal analysis applications, it is computationally involved since it requires
the singular value decomposition of the so-called block Hankel matrix, which is
constructed from the impulse response function (IRF) data and in general a large
matrix (No Ns Ni Ns ) with Ns the number of time samples. An investigation on
the effects of noise on the estimated parameters using ERA is discussed in [72],
while the faster ERA/DC variant, which uses autocorrelation functions as data, is
introduced in [73].
During the recent years the state space formulation (4.19) has also been the
basis for subspace identification methods [141]. Several combined deterministicstochastic algorithms, such as for instance Balanced Realization (BR), Multivariable Output-Error State sPace (MOESP) and Canonical Variate Algorithm (CVA)
were proposed to identify the state matrices as well as the stochastic variables in
(4.19). Subspace identification algorithms are non-iterative, and thus always yield
a solution without the risk for convergence problems. However, since the measured force and response time sequences are used without any preprocessing, the
data is in general characterized by high levels measurement noise. Also no reduction of the amount of data is obtained through an averaging process as is the
case for IRFs, making the computational burden one of the drawbacks. So far in
the domain of modal analysis, subspace identification has been mainly applied for
MPE from output-only data (forces are not measured) acquired from civil [100]
and mechanical structures [59, 60, 10].
Another class of time-domain system identification methods, having its roots
in the domain of control theory and applications, are the PEM-based methods,
such as ARMA, ARMAX,..., using a variant of the ARMAVX model structure.
However, since this model is a LMFD, the matrix polynomials have a low order
with large matrix coefficients, where the product of the model order nj and number
of responses has to be greater than 2Nm in order to find 2Nm system poles. The
PEM approach is less suited for MPE, since an increase of the model order by
one results in large increase of the number of system poles, making it difficult to
92
determine the optimal model order by means of a so-called stabilization chart. In

[59], the ARV method has been used for MPE from output-only data.
A much faster time-domain method is the Least Squares Complex Exponential
(LSCE) algorithm first introduced by [23] as an extension of Pronys method [112],
which states that the roots of an underdamped system always occur in complex
conjugate pairs and therefore can be used to form a characteristic polynomial with
real coefficients. From the IRF data at subsequent discrete time instants, these coefficients can be computed and the roots then result in the system poles. A MIMO
version of the LSCE is referred to as the Polyreference LSCE [148]. The LSCE
method is fast and does not require initial estimates for the modal parameters.
Moreover, based on auto- and cross-correlations between output measurements,
the LSCE algorithm has also been successfully applied for MPE from output-only
measurements as demonstrated in [59]. The only unknown is the number of modes
that must be considered in the analysis for which an overspecified number of modes
is usually used. However, as for most parametric estimators used for modal analysis, this requires the distinction between the physical and computational modes for
which a so-called stabilization chart was proposed. Displaying the poles (on the
frequency axis) for an increasing model order, indicates the physical poles since
they in general tend to stabilize for an increasing model order, while the mathematical poles scatter around. For low noise levels the PLSCE in combination with
the stabilization chart is suitable for an accurate MPE. However, for high noise
levels the estimation process becomes susceptible to errors and results can differ
significantly depending on the users experience. The LSCE only yields the poles
and modal participation factors, while the mode shapes have to be estimated in
a second step, for which the Least Squares Frequency-Domain (LSFD) algorithm
[94] is usually applied. The LSFD is based on the fact that, once the poles (and
modal participation factors) are known, equation (4.14) becomes a linear-in-theparameters problem that can be solved in least squares sense. Even for a high
number of outputs (e.g. No = 500 or more) the LSCE-LSFD combination can still
be executed within a reasonable computation efforts, explaining its frequent use
for modal analysis applications.
The LSCE and ITD methods can be considered as special cases of the Timedomain Direct Parameter Identification (TDPI) algorithm [79]. This MDOF
method yields global estimates for the system poles, mode shapes and modal
participation factors (suitable for multiple input data) and directly estimates the
coefficients of a (low order) ARX model (4.10) (in modal analysis literature often
referred to as ARMA model) in a least squares sense starting from the measured
force and response sequences. If the force is an impulse (Dirac pulse), the LSCE
is derived.
Within the domain of modal analysis it is a commonly accepted rule to prefer
time-domain estimators for systems with a low damping (damping ratio < 1%),
while frequency-domain estimators are better for highly damped systems ( >
2.5%). This is often motivated by the fact that FRF data from low damped systems
93
contains most of its energy in a very limited number of frequency lines, while the
equivalent IRF data has a large magnitude for a large number of time samples.
In practice, however, an IRF is computed from a FRF by means of the inverse
FFT (IFFT) and since this is an orthogonal transformation exactly the same
information will be available in both the frequency and time-domain. As a result,
the validity of the previously mentioned rule in practice can be seriously questioned
and moreover, the computation of the IRF using the IFFT of the FRF results in
aliasing effects, which become important in the case of low frequency resolution
or small damping [8]. The argument that, when using arbitrary excitation signals,
time-domain is preferred because of spectral leakage phenomena is also invalid
since the IRFs, used for time-domain identification, are computed from the FRFs
containing this leakage. It should be noticed that, transient effects in general (e.g.
leakage) can be modelled in the frequency domain in an analogous fashion as is the
case in the time-domain [111, 56]. Moreover, the use of residual terms to model the
effect of modes outside the analysis frequency band as well as the possibility to use
an non-equidistant frequency grid and frequency band selection are also advantages
of frequency-domain identification. Problems of numerical conditioning related to
a high model order can be considered as a drawback, although the latter greatly
depends on the choice of parameterization, as will be discussed in Chapter 5.
MDOF frequency-domain methods
A first MDOF frequency-domain method already mentioned is the LSFD estimator, which uses the modal model (4.14). In the case that the poles (and modal
participation factors) are not estimated in a first step, this method is nonlinearin-the-parameters resulting in a nonlinear least squares problem that has to be
solved using an optimization algorithm. Good starting values are required to reduce the number of iterations. This method as such never became popular, except
in combination with the LSCE, where it reduces to a linear Least Squares problem
yielding global estimates for the mode shapes (i.e. independent of the reference
degree of freedom).
As in the time-domain, a number of frequency-domain methods use a state
space parameterization. A frequency-domain version of the ERA method (ERAFD), is presented in [73], which formulation is closely related to the time-domain
algorithm. The primary data for the ERA-FD are the FRFs forming a complex block matrix from which, through an SVD, a state model is derived and
transformed to the modal space. As a part of this ERA-FD algorithm, accuracy
indicators, namely the output modal amplitude coherence and modal spectrum
coherence are developed to distinguish between the physical and mathematical
modes. A number of important features in frequency-domain analysis, such as
overlap averaging and data windowing and since a frequency band selection is
possible, considerably reduce computation and storage requirements.
As for the ERA method, frequency-domain versions of time-domain subspace
identification schemes having a reduced computation and storage requirements
94
were proposed. Starting from MIMO frequency response function measurements,

frequency-domain algorithms based on discrete-time deterministic subspace identification [31, 141] have been described in [84, 90]. These methods already use the
ideas of frequency weighting presented in [12], in order to avoid a distortion of
the Markov parameter sequences by time aliasing effects when these are computed
from FRFs using the IFFT. A continuous-time deterministic subspace variant is
discussed in [140]. The latter approach requires the introduction of orthogonal
(Forsythe) polynomials to avoid the problem of badly conditioned data-matrices
as for other frequency-domain identification schemes as well [117].
Another method that first identifies the frequency-domain low order complete
direct (state space) model (4.20) from MIMO FRF measurements is the Frequencydomain Direct Parameter Identification (FDPI) method [77]. The poles, mode
shapes and participation factors are derived from the system matrices of a state
model. In practice, a data reduction based on a Principal Component Analysis
(PCA) of the frequency response function matrix is necessary for computation
speed and memory requirements as well as to construct a proper stabilization diagram for proper model order detection . The algorithm assumes that the number
of modes within the analysis band is less than or equal to the number of outputs.
The Complex Mode Indicator Function (CMIF) [102] that uses a SVD of the
frequency response function matrix, has gained some popularity because of its
simple use. The CMIF is a plot of the log-magnitude of the singular values of the
FRF matrix as a function of the frequency. Peaks in the CMIF plot indicate the
damped natural frequencies. Based on the assumption that near a resonance the
FRF matrix is dominated by a single term in equation (4.14), it follows that the
mode shape and modal participation factor vectors are found from the left and
right singular vectors associated with the largest singular value. Since the CMIF
is a multiple input algorithm, it can detect multiple roots resulting in several
significant singular values peaking at a specific frequency. It should be noticed
that, for the case of multiple poles, the singular vectors corresponding to the
singular values significantly differing from zero are not the actual mode shapes.
Instead, these are given by a (unknown) linear combination of the singular vectors.
A large number of frequency-domain estimators are based on a matrix fraction
description model and are closely related to the identification techniques discussed
in the next two chapters. Similar to the PEM approach [85] in the time-domain,
these estimators are based on the minimization of an equation error between
the measured and the modelled FRF matrix. Considering for example the scalar
matrix fraction description (4.5), this equation error is defined as
(f ) = H(f )
B(f , )
A(f , )
(4.21)
where f is the angular frequency (f = 1, . . . , Nf ) and the polynomial coefficients

the parameters to estimate. The quadratic-like cost function is then defined as the
sum of the Frobenius norm of the squared error function for each angular frequency
95
PN f
(f )2 . The most straightforward method to minimize `()
f , i.e `() = f =1
F
with respect to the parameters using a Least Squares approach. However, since
this cost function is nonlinear-in-the-parameters, it requires an iterative algorithm
based on for instance a Gauss Newton approach [11]. Alternatively, a two step
approach can be used to obtain a suboptimal solution in analogy with the LSCELSFD procedure [92]. Due to the iterative character of estimators based on the
nonlinear cost function, starting values are required and to achieve this, Levi [81]
proposed a way to linearize the cost function resulting from the multiplication of
the equation error (4.21) with A(f , )
`()LS =
Nf
X
A(f , )H(f ) B(f , )2
(4.22)
f =1
Based on this idea, a large number of linear least squares estimators have been
describe in literature during the last two decades. The so-called Rational Fraction (Orthogonal) Polynomial (RFOP) method was presented in 1982 [113] for
SISO systems. Since the formulation of the LS problem in the continuous-time
frequency-domain resulted in ill-conditioned matrices for high-order systems, the
Scalar MFD model was expressed in terms of Forsythe polynomials. This method
was extended to a global method (GRFOP) for SIMO identification [114], while
Van der Auweraer [135] formulated the multiple reference version (OPOL).
According to the two major classes, introduced in [105, 51], all the previously
discussed estimators belong to the class of so-called deterministic algorithms, which
in essence are curve fitting techniques. Using however a stochastic approach based
on an errors-in-variables model, as defined in Chapter 2, yields a more realistic
description of real-life experiments. By taking knowledge about the noise on the
measured data into account in the cost function, it is possible to derive estimators
with significant higher accuracy compared to the ones developed in the deterministic framework. Most of the described stochastic estimators start from the Fourier
transform of the input (force) and response time sequences measured using a periodic excitation signal derived in an errors-in-variables framework, from which a
sample mean and sample variance can be derived [106] (cf. 2.5.4). Based on this
information, the model parameters can be derived using the so-called frequencydomain Maximum Likelihood Estimator (MLE) developed by Schoukens and Pintelon [122] and extended to multivariable systems by Guillaume [41]. In the same
stochastic framework, Weighted Total Least Squares algorithms with nearly maximum likelihood properties were developed for both SISO [137, 46, 107] and MIMO
[52, 108] system identification. Although these stochastic algorithms feature nice
statistical properties, the computational burden related to the iterative character
and the initial formulation of the algorithms has been a bottleneck to apply these
algorithms for MPE. Modal data is typically characterized by a large number of
spectral lines and response locations as well as a high modal density. Furthermore,
the initial implementations are based on Fourier data acquired under periodic excitation, which is also different from the FRF data that is most often acquired
96
using an arbitrary excitation signal. As a result until a few years ago, these algorithms were not well-known in the modal analysis community except for a few
specific applications in the domain of modal analysis, such as for instance flutter
testing [110, 47, 134]. Flutter analysis requires a stochastic approach since the
measured data is characterized by significant measurement noise. In [55] a dedicated multivariable implementation for frequency-domain estimators, based on a
common denominator transfer function model, has been given. Typical for modal
analysis is the large number of output measurements. This extensive amount of
data requires optimized algorithms that balance between accuracy and memory
and computation efficiency. This work has been the basis for the further development of frequency-domain system identification for modal analysis. Recently,
the ML approach has also been adapted for modal parameter estimation from
output-only measurements by using the spectral density matrix between outputs
instead of FRF matrix [45, 58]. Furthermore, the accurate identification of the
frequency components (sinusoids) present in the vibration behaviour of rotating
machinery, based on a pole/residue model, is possible as well by means of the
frequency-domain MLE [44].
4.4
Conclusions
This chapter has given an introduction of the concepts of parametric system identification for modal parameter estimation. Different parametric models to describe
the vibration behaviour of mechanical systems have been discussed, where three
major groups of parameterizations can be considered: matrix fraction description,
pole-residue or state space.
The interconnection of the different models is illustrated by means of a matrix
polynomial approach. This provided a proper basis for understanding the relation
between many of the parametric identification methods that originate from control
theory and different engineering areas. An extensive overview of the parametric
identification methods with their potential and drawbacks from the point of view
for modal parameter has been given. Besides the methods specifically developed
for modal parameter estimation, such as for instance the LSCE, ITD and ERA,
recent research in the field of system identification has resulted in more advanced
subspace and frequency-domain maximum likelihood methods.
This overview also introduced the so-called stochastic estimators, such as the
frequency-domain Maximum Likelihood or Weighted Total Least Squares methods
that take the noise information of the data into account. This thesis focusses on
the applicability of these estimators for modal analysis, where this is motivated by
several reasons. New applications in the domain of modal analysis, require a high
accuracy of the estimated parameters, with important examples such as damage
detection and structural health monitoring. Other applications, such as flight
4.4. Conclusions
97
flutter analysis, typically result in very noisy data from which accurate damping
estimates should be derived. Since the stochastic estimators do consider the noise
on the data, accurate models can be estimated and confidence bounds for the
parameters can be obtained.
A methodology for determining the noise on the FRF data in an EV framework
was already developed in Chapter 2. The next two chapters will study the numerical efficiency of the parametric estimators using FRF or Input/Output Fourier
data. The initial formulation of the algorithms will be adapted in order to process
large data sets characterized by a high modal density and dynamical range.
98
Chapter 5
Frequency-domain MPE
from FRF Data
This chapter presents the optimization of the (stochastic) frequency-domain methods, such as the Maximum Likelihood estimator, for modal parameter estimation.
Large modal data sets require optimized algorithms that balance between accuracy
and computation efficiency. A multivariable implementation, based on a commondenominator transfer function model, is given. First, improvements of the classical
implementation of the frequency-domain linear Least Squares (LS) approach will
be presented in Section 5.3. Aspects, such as robustness for high modal density
and numerical performance (computation time, memory usage) of the different algorithm implementations and parameterizations are studied. The final result is a
fast LS solver, i.e. the so-called Least Squares Complex Frequency (LSCF) method.
This estimator offers the user, in analogy with the well-known LSCE method, a
stabilization chart. The fast implementation is also extended to the class of the
stochastic estimators based on a Total Least Squares (TLS) and the Maximum
Likelihood (ML) framework in Sections 5.4 and 5.5.
99
100
5.1
Chapter 5. Frequency-domain MPE from FRF Data
Introduction
Since modal analysis usually starts from FRF measurements, the optimization of
the stochastic frequency-domain methods for modal parameter estimation from
FRF data is one of the key issues of this thesis. The case for I/O will be considered in Chapter 6 for specific applications where only short data records can be
measured.
Typical for modal analysis is the large number of output measurements. An
extensive amount of data requires optimized algorithms that balance between accuracy and memory/computation efficiency. A multivariable implementation for
frequency-domain estimators, based on a common-denominator transfer function
model, will be derived.
Improvements of the classical implementation of the frequency-domain linear
Least Squares (LS) approach will be presented for extracting the modal parameters
from typical modal FRF data. Aspects, such as robustness for high modal density
and numerical performance (computation time, memory usage) of the different
algorithm implementations and parameterizations are studied and validated by
means of experimental modal data obtained from a safety critical aircraft component. The final result is a fast LS solver, i.e. the so-called Least Squares Complex
Frequency (LSCF) method. In analogy with the well-known LSCE method, a fast
construction of a stabilization chart is studied as well, where different representations are possible. A second step LS solver is proposed for the estimation of the
residues based on the poles selected by the user from this diagram.
Furthermore, it is shown in this chapter how the fast implementation can also
be extended to the class of stochastic Total Least Squares (TLS) estimators and the
Maximum Likelihood (ML) method. The FRF data and required noise information
can be obtained by means of the Hiv FRF approach presented in Chapter 2. Using
experimental data, the robustness of the algorithms is demonstrated for model
orders up to 200.
5.2
Parametric model
Given the global character of the poles of a mechanical system, a scalar matrixfraction description (cf. 4.2) better known as a common-denominator model
(CDM) will be used for the development of the frequency-domain estimators.
The measured Frequency Response Function (FRF) at DFT frequency f , between
output o and input i, is then modelled as
k (f , ) = Bk (f , )
H
A(f , )
(5.1)
5.2. Parametric model
101
for k = 1, 2, . . . , No Ni with
Bk (f , ) =
n
X
bkj jf
(5.2)
j=0
the numerator polynomial between the output/input DOF combination k and

A(f , ) =
n
X
aj jf
(5.3)
j=0
the common-denominator polynomial. The coefficients aj and bkj are the unknown
parameters to be estimated. In general, the order of the denominator polynomial
and numerator polynomials can differ.
For a discrete-time domain model, the generalized transform variable f , evaluated at DFT frequency f , is given by f = e(if Ts ) (Z-domain) with Ts the
sampling period. For continuous-time domain models, other choices for f are
possible such as for instance
(if )
lumped continuous-time systems (Laplace domain)
orthogonal polynomials, e.g. Forsythe, Chebyshev

p
f =
( (if ))
diffusion phenomena
(tanh R (if )) microwaves (Richardson domain) [116]
To obtain an identifiable parameterization (5.1), it is needed to impose a

(scalar) constraint. This is readily verified by considering the following expression
k (f , , ) = Bk (f , ) = Bk (f , )
H
(5.4)
A(f , )
A(f , )
Clearly, for every non-zero scalar another equivalent scalar matrix-fraction description is obtained. The parameter redundancy can be removed by fixing one
coefficient of the denominator, such as for instance the highest order coefficient of
the denominator, i.e. an = 1 or by imposing a norm-1 constraint, i.e. H = 1
with containing all polynomial coefficients aj and bkj , the parameters to be esH
A = 1 with A containing
timated. Other constraints are possible such as e.g. A
only the denominator coefficients.
A linear Least Squares approach requires model equations that are linear-inthe-parameters. An often used approximation, first presented by Levi [81] for
k in (5.1) by the measured FRF
SISO systems, consists of replacing the model H
Hk and multiplying with the denominator polynomial
n
n
X
X
Wk (f )
bkj jf
(5.5)
aj jf Hk (f ) 0
j=0
j=0
102
for k = 1, . . . , No Ni and f = 1, . . . , Nf the number of spectral lines. By introducing an adequate weighting function Wk (f ) in the equations (5.5), the quality of
the LS estimate can often be improved as discussed in 5.6.
5.3
(Weighted) Linear Least Squares
The first part of this chapter presents a linear Least Squares solver for the minimization of the LS cost function (4.22), which using the model (5.1) becomes
`()LS =
Nf
N
o Ni X
X
k=1 f =1
Bk (f , ) A(f , )Hk (f )2
(5.6)
Starting from the LS formulation based on both Jacobian and normal equations
several algorithms are discussed and optimized for deriving the modal parameters
from complex high order systems in a time-efficient way.
5.3.1
LS Formulation based on Jacobian Matrix
Since Eqs. (5.5) are linear-in-the-parameters and because a common-denominator

model is used, they can be reformulated as
B1
1 0
0
1
B2
0 2
0
2
..
0
(5.7)
..
.
.
.
..
..
B No Ni
0
0 No Ni No Ni
A
with
Bk
and
where
bk0
bk1
=
..
bkn
k (1 )
k (2 )
k =
..
k (Nf )
k (f ) = Wk (f ) 0f , 1f , . . . , nf
a0
a1
A =
..
an
k (1 )
k (2 )
k =
..
k (Nf )
,
(5.8)
k (f ) = k (f )Hk (f )
(5.9)
5.3. (Weighted) Linear Least Squares
103
The Jacobian matrix J of the least-squares problem
1 0
0
1
0 2
0
2
J = .
..
.
..
..
.
0
0 No Ni No Ni
(5.10)
has Nf No Ni rows and (n + 1)(No Ni + 1) columns (with Nf n, where n is the

order of the polynomials). Because every equation in (5.5) has been weighted with
Wk (f ), the matrix entries k in (5.10) generally differ.
Contrary to the k matrices, the k matrices occurring in (5.10) do not contain
measured data and thus are not subject to errors. Because of this, it is possible
to apply the mixed LS-TLSalgorithm presented in [139].
The matrix J in (5.10) can be partitioned as J = [, ] with the matrix
exactly known and subject to errors, where is Nf No Ni (n + 1)(No Ni ) and
is Nf No Ni (n + 1). is the parameter vector with (n + 1)(No Ni + 1) elements.
The mixed LS-TLS algorithm determines the denominator coefficients A such
that [, ] = 0 with an perturbation matrix with a minimal Frobenius
T
T T
norm and = [B
, A
] with B a vector of length (n + 1)(No Ni ) and A a vector
with n + 1 elements. The algorithm consists of the following steps
Mixed LS-TLS algorithm using full QR decomposition
step 1 : Compute the QR-decomposition of J
R11
[, ] = [Q1 , Q2 ]
0
R12
R22
(5.11)
with R22 a (n + 1 n + 1) upper triangular matrix.

step 2 : Compute the SVD R22 = U SV H with S = diag(1 , . . . , n+1 ) and
V = [1 , . . . , n+1 ] an orthonormal matrix.
step 3 : The denominator coefficients are found as the eigenvector corresponding to the smallest singular value n+1 , i.e. A = n+1 .
The prove of this algorithm is given in [139] (pp.9295). The number of single
floating point operations (flops) gives a first insight in the computation speed
of the algorithm by counting each sum or product as one flop [40]. The total
number of flops involved in steps 13 is approximately O(2(No Ni )3 n2 Nf ). Based
on the knowledge of the denominator coefficients A , this algorithm also allows to
compute the numerator coefficients B by means of a back-substitution approach,
O((No Ni )2 n) [40].
104
In the first step, a QR-decomposition of the large matrix J must be computed. This can be done in a numerical efficient way by exploiting the block
structure of this matrix. Bayard proposed an algorithm for a so-called sparse QRdecomposition [11]. However, this algorithm assumes that the submatrices k in
J are all equal for k = 1, . . . , No Ni . In general, this is not the case for the LS
problem (5.7) since the weighting functions Wk (f ) can differ for each k. Therefore, the Sparse QR algorithm, first presented in [11], was generalized solving the
LS problem using the following algorithm.
Mixed LS-TLS algorithm using sparse QR decomposition

step 1 : For each k, with k = 1, . . . , No Ni , compute the QR-decomposition
of the submatrices k yielding k = Qk Rk .
step 2 : Compute the orthogonal projection of the submatrices k on the
space defined by the orthonormal column vectors of the matrix Qk by means
H
H
of the projection operator defined as Q
k = I Qk Qk (with Qk Qk = I) in
order to form the following matrix (with = No Ni )

Q1 1
Q
2 2
V =
(5.12)
..
.
Q

step 3 : Perform the QR decomposition of V yielding V = QV RV .
step 4 : The matrix R22 (cf. step 2 of LS-TLS algorithm) is given by RV .
Proof : Given the QR-decomposition of the sparse matrix J (5.11) (cf. step 1 of
LS-TLS algorithm)
R1 0
0 R1
Q1 0
0 Q1
0 R2
0 R2
0 Q2
0 Q2
..
..
.
.
(5.13)
J = .
..
.
..
.
.
..
.
.
0
0 R R
0
0 Q Q
0
0
0
R
with = + 1 and where R = R22 . For the computation of the denominator

coefficients A using the mixed LS-TLS algorithm, it is sufficient to know the small
(n + 1 n + 1) upper triangular matrix R . Indeed, from (5.13) it follows that
k
Q k Rk
Qk Rk + Qk R
105
The orthogonal projection of the submatrices k on the space defined by the

orthonormal column vectors of the matrix Qk is obtained by multiplying with the
H
projection operator Q
k = I Q k Qk
Q
k k = Qk Qk Rk + Qk Qk R = 0 + Qk Qk R = Qk R
(5.14)
Since Qk and Qk are orthogonal matrices, [I Qk QH

k ]Qk = Qk + 0. Consequently, the matrix V (5.12) can be written as
Q1
Q2
V = . R = Q V RV
(5.15)
..
Q
. H H
H .
where [QH
1 , Q2 , ., Q ] is an (Nf No Ni n) orthogonal matrix and R an (nn)
upper triangular matrix, representing the QR-decomposition of the large matrix
V . Hence, the required matrix R22 (cf. step 2 of the mixed LS-TLS algorithm),
is given by R = RV .
Similar as for the mixed LS-TLS algorithm [139], it is possible, based on the
quantities computed in steps 13 of the sparse matrix QR algorithm, to compute
also the numerator coefficients based on a back-substitution approach as discussed
in [11].
The most important drawback of Bayards sparse matrix QR algorithm, requiring approximately Nf 2 (5nNo Ni +Nf ) flops, stems from the orthogonal projections
3
Q
k k derived by the explicit matrix multiplications, O(N f ). This also explains
the assumption of equal matrices k in order to form the projection matrix only
once to limit the computation time, since in the case of differing weights, a total of
No Ni Nf 2 (5n + Nf ) flops is required. Another drawback is the QR-decomposition
of the large matrix V . However, this algorithm can be further improved by computing these orthogonal projections by means of QR decompositions (a technique
which is commonly used in subspace identification algorithms [141]). This improvement was necessary to handle the problem with different matrix entries k
efficiently.
Mixed LS-TLS algorithm using sparse QR via orthogonal projections
The orthogonal projections Q
k k (k = 1, . . . , No Ni ) forming the matrix V , can be
computed avoiding the explicit matrix products as well as the QR decomposition
of the large matrix V , as follows
step 1 : Compute the QR-decomposition
[k , k ] = [(Qk )a , (Qk )b ]
(Rk )a
0
(Rk )ab
(Rk )b
(5.16)
106

for k = 1, . . . , No Ni
step 2 : Compute the QR decomposition of the matrix Rb
H
H H
Rb = (R1 )H
= Q V RV
b (R2 )b . . . (R )b
(5.17)
with RV the matrix R22 in the mixed LS-TLS algorithm.

Proof : The QR-decomposition of [k , k ] yields
k
(Qk )a (Rk )a
(Qk )a (Rk )ab + (Qk )b (Rk )b
As the QR-decomposition of k is unique, Qk = (Qk )a and Rk = (Rk )a . Leftmultiplication of k with the orthogonal complement of (Qk )a , i.e. (Qk )
a =
[I (Qk )a (Qk )Ta ], results in
(Qk )
a k
(Qk )
a (Qk )a (Rk )ab + (Qk )a (Qk )b (Rk )b
[I (Qk )a (Qk )Ta ](Qk )a (Rk )ab + [I (Qk )a (Qk )Ta ](Qk )b (Rk )b
=
=
(Qk )a (Rk )ab (Qk )a (Rk )ab + (Qk )b (Rk )b 0

(Qk )b (Rk )b
since (Qk )Ta (Qk )a = I and (Qk )Ta (Qk )b = 0. As a result, Q

k k = (Qk )a k =
(Qk )b (Rk )b . This proves that the (Nf n) submatrices Qk k of the matrix V
(5.12) are found by means of a (thin) QR-decomposition of [k , k ] instead of
explicitly computing the matrix products Q
k k .
2
The explicit computation of matrix product (Qk )
a k requires Nf (n + 1)
flops without taking the formulation of the matrix (Qk )a into account. The QRdecomposition only requires Nf (2n + 2)2 flops, with Nf (2n + 2).
Using the orthogonal projection algorithm

V then results in

(Q1 )b (R1 )b
(Q1 )b
0
(Q2 )b (R2 )b 0
(Q
2 )b

V =
=
..
..
..

.
.
.
(Q )b (R )b
for a fast formulation of the matrix
0
0
..
.
(Q )b
(R1 )b
(R2 )b
..
.
(R )b
(5.18)
where the block diagonal matrix with submatrices (Qk )b is orthogonal again (by
construction).
Finally, the QR-decomposition of the (n n) matrix Rb , defined as Rb =
H
H H
(R1 )H
, is given by Rb = QR RR and hence this corresponds to
b (R2 )b . . . (R )b
107
the QR-decomposition of the large matrix V = QV RV (5.15), where
(Q1 )b
0
0
(Q2 )b
0
RV = RR and QV =
QR
..
..
..
.
.
.
0
0
(Q )b
(5.19)
In practice, only the matrix RR is used. Hence the total number of flops required
for steps 12 is now approximately O(8No Ni Nf n2 ). As a result, by optimizing
the Sparse QR algorithm, an important reduction of the computation time as well
as memory usage can be obtained. Nevertheless, since the number of measured
frequencies Nf is typically large for modal testing, the LS formulation based on
the so-called normal equations is commonly-used by many of the modal parameter
estimators discussed in Chapter 4.
5.3.2
LS Formulation based on Normal Matrix
Instead of deriving the parameter estimates directly from the Jacobian matrix,
many estimators used in modal analysis form the normal equations explicitly by
computing
H
0
H
1 1
1 1
H
H
2 2
0
2 2
(5.20)
JHJ =
..
..
..
..
.
.
.
.
P
No Ni H
H
H
1 1 2 2
k=1 k k
with the entries of the submatrices given by
[H
k k ]rs
Nf
X
f =1
[H
k k ]rs
Nf
X
f =1
[H
k k ]rs
|Wk (f )|2 r1
s1
f
f
H
|Wk (f )Hk (f )|2 fr1 s1

f
Nf
X
f =1
|Wk (f )|2 Hk (f )fr1 s1

f
(5.21)
In 5.3.3 it is shown that the submatrices of the normal matrix are structured
matrices. By exploiting the specific matrix structure, an important reduction of
the computation time and memory requirements is obtained.
Defining the following submatrices
Rk = H
k k
Sk = H
k k
T k = H
k k
(5.22)
108
the normal equations are rewritten as
R1
0
..
.
0
R2
..
.
..
.
S1H
S2H
B1
B2
..
PN o N i
B No Ni
k=1 Tk
A
S1
S2
..
.
(5.23)
Since the submatrices Rk = H

k k in Eq. (5.23) do not contain any measurement
data (i.e. they are not subjected to errors), it is possible to apply again the mixed
LS-TLS approach presented in [139] as discussed in 5.3.1.
Although the number of rows of the normal matrix (5.20) is much smaller than
the number of rows of the Jacobian matrix (5.10), its size (i.e. (n + 1)(No Ni + 1))
is often still of importance for the computation time in the case of typical modal
test data (No > 100) since solving the Eqs. (5.23) for still requires (No Ni n)3
flops.
Under the condition that, the parameter constraint only applies to the denominator coefficients A , the numerator coefficients can be eliminated from the normal
equations
Bk = Rk1 .Sk .A
(5.24)
by substitution in the last (n + 1) equations of (5.23)

N
o Ni
X
k=1
SkH Bk +
N
o Ni
X
k=1
Tk A 0
yielding a very compact problem

#
"N N
o i
X
H 1
Tk Sk Rk Sk A = DA 0
(5.25)
(5.26)
k=1
The square matrix D has a size (n + 1) and thus is much smaller than the original
normal matrix (5.20) with size (n + 1)(No Ni + 1). Notice, that in the case of
the Jacobian matrix the mixed LS-TLS algorithm uses QR decompositions for the
elimination of the numerator coefficients Bk .
A LS solution of A is found by fixing for instance the highest order coefficient
an to 1
[D(1 : n, 1 : n)]1 {D(1 : n, n + 1)}

(5.27)
ALS =
1
The mixed LS-TLS solution is given by the eigenvector ve corresponding to the

smallest eigenvalue e found by solving an eigenvalue problem Dve = e ve .
The LS or mixed LS-TLS solutions for A , obtained by solving the compact
linear LS problem (5.26) is the same as obtained by solving the full LS problem
109
(5.23) (with the same constraint). It is proven in [74] (matrix inversion lemma)
that the inverse of the normal matrix in Eq. (5.20) is as well a positive definite
Hermitian symmetric matrix given as
R

SH
1
E
S
=
FH
| T
F

| G
(5.28)
with the submatrices E = (R ST 1 S H )1 , F = R1 S(T S H R1 S)1 and

G = (T S H R1 S)1 where E and G are both Hermitian matrices. As shown in
[30], the jth (j = 1 , . . . , n + 1) column of (T S H R1 S)1 gives the LS solution
for the denominator coefficients A under the constraint aj = 1. Notice that, in
the case that none of the entries is subject to errors (i.e. no noise is present on the
data), the matrix is not of full rank and so the inverse does not exist. However, in
that case, (5.23) and (5.26) are exactly equal to zero and as a result the solution
A is uniquely defined.
Once the A coefficients are known, back-substitution based on (5.24) can be
used to derive the numerator coefficients B . The total number of flops required
for the elimination of the numerator coefficients, solving D for A and the backsubstitution is O(2No Ni n3 ). This approach is more time efficient than solving
(5.23) directly, i.e approximately No 2 .Ni 2 times faster.
In Appendix A it is proven that this LS algorithm based on the compact
normal equation yields the same solution as the sparse QR algorithm presented in
5.3.1 (under the assumption that both problems are numerically well conditioned
cf. 5.3.3). Nevertheless, the LS algorithm, based on the compact normal equations
has the important advantage to be faster.
5.3.3
Optimal LS Implementation for MPE
Based on the results of the previous sections, different implementations of a

frequency-domain LS estimator can be derived by varying following algorithm
characteristics:
Jacobian or Normal matrix based LS formulation. (The Jacobian implementation is based the fast sparse QR decomposition via projections)
real or complex valued coefficients
continuous or discrete-time model, defined by the generalized transform variable
parameter constraint, i.e. a LS or mixed LS-TLS constraint
110
SLAT TRACK
LEADING
EDGE
OF WING
SLAT
Figure 5.1: A Slat track (top) mounted in the wing of an Airbus320 aircraft (bottom).
In order to determine an optimal implementation for the problem of modal parameter estimation, the performance of the various implementations is assessed
using modal test data obtained from on a slattrack of an Airbus A320 commercial
aircraft.
Slat tracks are located at the leading edge of an aircraft wing and make part
of a gliding mechanism that is used to enlarge the wing surface (see Figure 5.1).
The enlargement of the wing surface is needed in order to increase the lift force at
reduced velocity during landing and take off. The A320 airplane has 5 slats per
wing. The first slat (i.e. the inboard slat between fuselage and engine) contains 4
slat tracks. The other 4 slats have 2 slat tracks each.
Using a Scanning Laser Doppler Vibrometer setup, shown in Figure 5.2, the
response (velocities) to a single input excitation was measured in 352 scan-points.
An electrodynamic shaker was used to apply a random noise excitation in a frequency band of 0-4kHz with a resolution of 1.25Hz. Frequency response functions
111
SLAT TRACK
FORCE SENSOR
STINGER
SH AK ER
SHAKER
S C ANN ING L ASE R

DO PP LER V IBR OM ET ER
Figure 5.2: Force (input) measurement with shaker, stinger and force sensor attached
to the slat track (left). Velocity (output) measurement with scanning laser Doppler
vibrometer (right).
were estimated using the H1 estimator with 5 averages. Using the coherence functions, the variances of the FRFs were obtained as well. This data set is a good
representation of a typical modal data set.
Unless otherwise stated, a common-denominator model of 50 modes and a
frequency band of 10003725Hz is used for the parametric identification by means
of the various LS implementations in this section. Given the variances, obtained by
means of the knowledge of the coherence functions based on Eq. (2.22), a Weighted
LS formulation was used for the various implementations. This was done by means
of the nonparametric weighting function (5.76) that will discussed later in 5.6.
Real or Complex Valued Coefficients

The equations derived in the previous two sections implicitly assume that the
coefficients are complex-valued. Hence, since a common-denominator model
(5.1) is used, the denominator polynomial A(f , ) has scalar coefficients and an
order n equal to the number of modes to be estimated Nm (cf. 4.2).
To obtain real-valued coefficients, the Jacobian matrix has to be transformed
into a real valued matrix. This can be done by reformulating the Jacobian matrix
as follows
re
1
0
0
re
1
0 re
0
re
2
2
(5.29)
JRE = .
.
..
..
..
.
0
re
No Ni
re
No Ni
112
with
re
k (f )
Re(k )
Im(k )
and
re
k (f )
Re(k )
Im(k )
The normal matrix is transformed to a real-valued matrix simply by taking the

T
real part of J H J since JRE
JRE = Re(J H J).
In the case that real-valued coefficients are estimated, the order of the denominator polynomial A(f , ) has to be doubled in order to identify Nm modes again.
This, however, is not in favor of the numerical conditioning of the equation matrices, especially when the normal equations are formulated in the Laplace domain
(f = if ) as is will be shown now.
Continuous-time domain Laplace variable
The classical implementation of the frequency-domain linear LS estimator uses
a continuous-time model with real-valued coefficients. A linear, time-invariant
(LTI) continuous-time system of order n is modelled by the common-denominator
model (5.1) by taking the generalized transform variable = i. As a result, the
submatrices K and K of the Jacobian matrix of the LS problem (5.7) contain
entries of the form of a power basis
k (f ) = Wk (f ) (if )0 , (if )1 , (if )2 , . . . , (if )n

k (f )
k (f )Hk (f )
(5.30)
It can be noticed that these submatrices have a graded structurestrongly related

to a so-called structured Vandermonde matrix. The formulation of the Jacobian
matrix (cf. Eq. 5.7) is approximately O(12nNo Ni Nf ) for the case of complex
coefficients.
H
H
From this, it follows that the submatrices H
k k , k k and k k appearing
in the normal equations (5.23) convert into structured matrices as well
0
i 1
2
i(c1) n
i 1
2
i 3
ic (n+1)
(c+1)
2
3
4
(n+2)
(c+2)
3
4
5
(n+3)
..
..
..
..
..
.
.
.
.
.
i(r1) n
ir (n+1)
i(r+1) (n+2)
(i)(r1) i(c1) 2n
with r, c = 1, . . . , n + 1 and where the operator n is defined as

n
Nf
X
f =1
G(f )fn
(5.31)
113
0
50
Ampl. (dB)
100
150
200
250
300
1000
1500
2000
2500
Freq. (Hz)
3000
3500
1500
2000
2500
Freq. (Hz)
3000
3500
20
0
Ampl. (dB)
20
40
60
80
100
120
1000
Figure 5.3: Jacobian WLS using a frequency-scaled continuous-time model with real
(top) and complex (bottom) coefficients. Measurements (dotted line) and estimated
transfer function model (solid line).
2
with G(f ) = |Wk (f )|2 for the matrices H
k k , G(f ) = |Wk (f )Hk (f )| for
2
H
H
the matrices k k and G(f ) = |Wk (f )| Hk (f ) for the matrices k k . Based
H
on the Hermitian symmetry of H
k k and k k , it is sufficient to compute and
store the elements of the first row and last column (i.e. 2n + 1 instead of (n + 1) 2
summations) in order to reconstructed the complete matrix. The same is true
for the submatrices H
k k , which however are not Hermitian symmetric but only
for the reason of the sign of the elements in the lower triangular part. Taking the
114
matrix structure into account, the normal matrix formulation requires 48nN o Ni Nf
flops, i.e. only 4 times the number of flops for the Jacobian formulation, whereas
the explicit product J H J is O((No Ni )3 n2 Nf ).
However, for a continuous-time model, Eqs. (5.23) become numerically illconditioned, especially for high-order systems. This is certainly the case for complex mechanical systems with a high modal density in the studied frequency band.
Normalization (scaling) of the frequency axis by a factor s = (Nf 1 )/2 can
improve the numerical conditioning to a certain extend. Nevertheless, in practice,
a model order of n = 20 (i.e. 10 modes in the case of real coefficients) is often
appears to be a the limit for preserving an acceptable numerical conditioning.
Figure 5.3 shows a synthesized FRF for the Jacobian LS implementation using
a frequency-scaled continuous-time model with real (top) and complex (bottom)
coefficients. As can be seen a bad numerical conditioning, i.e. = 6E75 for real
and = 6E45 for complex case (cf. Table 5.2), results in important estimation
errors. It can be seen that the higher frequencies are over-emphasized, typical for
this type of model. Nevertheless, the model obtained is completely wrong for both
cases. For the case of the slattrack, reasonable results could only be obtained with
this implementation by analyzing small frequency bands with Nm not higher than
10.
Continuous-time domain Orthogonal Polynomials
The numerical conditioning of the jacobian or normal matrix can be significantly
improved by decomposing the numerator and denominator polynomials of each
k (f , ) for k = 1, . . . , No Ni into a well-chosen basis of
transfer function model H
orthogonal polynomials
Pn
j=0 ukj pkj (if )
(5.32)
Hk (f , ) = Pn
j=0 vj qj (if )
where pkj (if ) and qj (if ) are the sets of orthogonal polynomials evaluated at the
angular frequency f (f = 1, . . . , Nf ) and where are the new (real) coefficients
to be estimated
= [u10 , u11 , u12 , . . . , u1n , u20 , . . . , uNo Ni n , v0 , . . . , vn ]
(5.33)
In the domain of modal analysis, Richardson and Formenti [113, 114] used this approach in order to improve the numerical conditioning of their so-called Global Rational Fraction Polynomial (GRFP) method which was also extended to a MIMO
method (OPOL) by Van der Auweraer [133]. To obtain a better conditioning of
the normal matrix in Eq. (5.23), Forsythe polynomials can be used. Orthogonal
Forsythe polynomials can be generated through a recursive scheme presented in
[38, 113]. The OPOL estimator, for example, uses the Forsythe polynomials to
115
10
20
30
Ampl. (dB)
40
50
60
70
80
90
100
110
1000
1500
2000
2500
Freq. (Hz)
3000
3500
Figure 5.4: Normal-based WLS implementation for continuous-time model with real
coefficients using Forsythe polynomials.
obtain unity matrices for the block matrices on the diagonal of the normal matrix
improving the numerical conditioning of this matrix as well as allowing a reduction
of the computation time and required memory.
For the practical implementation of the LS approach using Forsythe polynomials, the same polynomial basis was used for each transfer function model using
real coefficients, where the polynomials were orthogonal with respect to the weight
T (f ) = 1
Nf
X
T (f )pi (f )pj (f ) = ij
(5.34)
f =1
and the result for the normal-based implementation is shown in Figure 5.4. Given
the good numerical conditioning ( = 2E3) of the normal matrix, the results for
the Jacobian and normal equations yield a very similar result.
However, an important drawback of this parameter estimation approach is the
necessity to transform the estimated coefficients from the orthogonal basis back
to the original power polynomial basis in order to solve for the modal parameters
(cf. 5.3.4). Even if frequency scaling is applied, this again represents a numerically ill-conditioned problem. A solution to this problem is discussed in [117],
i.e. the extension of the use of orthogonal polynomials to the complete identification process. This includes the extraction of the roots of each numerator and the
denominator polynomial and hence the modal parameters by rewriting each characteristic equation into a low order state space model and solving the eigenvalue
116
10
20
30
Ampl. (dB)
40
50
60
70
80
90
100
110
1000
1500
2000
1500
2000
2500
Freq. (Hz)
3000
3500
3000
3500
10
20
30
Ampl. (dB)
40
50
60
70
80
90
100
110
1000
2500
Freq. (Hz)
Figure 5.5: Jacobian-based (top) and Normal-based (bottom) WLS implementation for
a continuous-time model with complex coefficients and using Chebyshev polynomials.
problem of the so-called orthogonal companion matrix of this model. This is a

numerically well-conditioned matrix since it has been formulated in the orthogonal basis as well. By doing so, it is possible to identify the modal parameters of
high order systems in a numerically stable way as demonstrated for SISO systems
in [117]. Nevertheless, in the case of MIMO systems, a different set of Forsythe
k expolynomials pkj (if ) has to be generated for each transfer function model H
plaining the high computational involvement of this approach. This is a major
drawback for using such estimation schemes for modal parameter estimation.
117
The use of Chebyshev polynomials instead of Forsythe polynomials can reduce

the computation time and memory usage significantly. Using a similar approach as
discussed in [1] for SISO systems, the construction of the square (n+1) submatrices
Rk , Sk and Tk (cf. Eq. 5.21) boils down to the computation of 2n + 1 entries (i.e.
the entries of the first row and last column). Indeed, the other entries can be
computed as a sum from two of these 2n + 1 entries, since the product of two
Chebyshev polynomials Ci and Cj is given as Ci Cj = 21 (Ci+j + Cij ).
The result for slattrack data obtained by this approach is shown in Figure 5.5
for both a Jacobian and Normal based implementation. While results for a Jacobian Chebyshev implementation are good, it can be seen that, since a continuoustime model is used and Chebyshev polynomials are only approximately orthogonal, the numerical conditioning ( = 9E8) becomes a problem for Nm = 50 in
the case of the normal equations. Comparison with Figure 5.4, shows that the
results obtained by the normal-based implementation using Forsythe polynomials
is certainly better in the band of 15003000Hz, however at the price of a higher
computation time.
Discrete-time domain Z variable
Considering the common-denominator model (5.1) in the discrete-time domain, the
generalized transform variable f is defined as f = e(if Ts ) . Since these complex
polynomial basis functions are implicitly orthogonal with respect to the unity
circle, a well-conditioned Jacobian matrix J is usually obtained, which also justifies
the explicit calculation of the normal equations. It turns out from past experience
in modal analysis with discrete-time domain estimators such as the LSCE that the
numerical conditioning of the normal matrix (or so-called covariance matrix) is not
a major problem. The computation of the poles from the estimated denominator
coefficients is also well-conditioned (cf. 5.3.4).
H
H
The submatrices H
k k , k k and k k appearing in
(5.23) are structured matrices of the following form
z 0
z 1
z 2
z 3
0
1
2
z 1
z
z
z
1
0
1
z 2
z
z
z
2
1
0
z 3
z
z
z
..
..
..
..
.
.
.
.
.
.
.
z n
z (n1)
z (n2)
z (n3)
where the z n operator is defined as

n
z =
Nf
X
f =1
G(f )eif Ts n
the normal equations

z n
z (n1)
z (n2)
z (n3)
..
.
z 0
(5.35)
118
10
20
30
Ampl. (dB)
40
50
60
70
80
90
100
110
1000
1500
2000
2500
Freq. (Hz)
3000
3500
1500
2000
2500
Freq. (Hz)
3000
3500
10
20
30
Ampl. (dB)
40
50
60
70
80
90
100
110
1000
Figure 5.6: Jacobian-based (top) and Normal-based (bottom) WLS implementation for
a discrete-time model with complex coefficients.
2
with G(f ) = |Wk (f )|2 for the matrices H
k k , G(f ) = |Wk (f )Hk (f )| for
H
2
H
the matrices k k and G(f ) = |Wk (f )| Hk (f ) for the matrices k k . Having
entries that are constant along each diagonal, this matrix has a so-called Toeplitz
structure. Toeplitz matrices belong to the larger class of persymmetric matrices
and the inverse of a nonsingular Toeplitz matrix is persymmetric as well. Based
H
on the Hermitian symmetric character of the structured matrices H
k k and k k
n
n
(since z
= (z ) if G(f ) is real), it is sufficient to compute and store the
elements of the first row (i.e. n + 1 instead of (n + 1)2 summations) in order to
119
reconstructed the complete matrix. For the matrices H

k k also the first column
is required, since these are not Hermitian symmetric and hence 2n + 1 elements
must be stored. Taking the matrix structure into account, the normal matrix
formulation now requires approximately 32nNo Ni Nf flops, i.e. only 2.66 times
the number of flops for the Jacobian formulation.
In Appendix B it is shown that, if the frequencies are uniformly distributed
(e.g., f = f. with = N2
), a fast computation of the matrix entries (5.21)
s Ts
can be done using the Fast Fourier Transform [53, 125], in this case requiring
15No Ni Nf log2 (Nf ) flops for the normal matrix formulation. This results in a
further reduction of the computation time if 15 log 2 (Nf ) < 32n, thus depending
on the model order n and the number of DFT frequencies Nf . In practice, this
typically results in a further reduction of a factor 210.
Figure 5.6 shows the comparison for the Jacobian and normal-matrix based
WLS implementation for a discrete-time model of 50 modes with complex coefficients. As can be seen both results are the same, as is theoretically expected,
since the normal matrix still has a good conditioning ( = 7E4) for this high number of modes, indicating the important advantage of using a discrete-time model.
The benefit from using complex coefficients can be seen by comparing the normalbased implementations using a discrete model with complex and real coefficients,
shown in Figure 5.7, where in the case of the real-valued coefficients, the numerical
conditioning ( = 6.5E5) has some affect on the accuracy of the model for the 3
closely-spaced modes around 1700Hz (since the order of the polynomial mode is
doubled).
As discussed in [123] the assumption made with respect to the discrete character of the measured data and the model used to represent the LTI system should
be the same. The two most commonly-used assumptions are Zero-Order Hold
(ZOH) and Band Limited (BL). In practice, since anti-aliasing filters are applied
during the measurements, the measured signals have a limited bandwidth (BL).
A discrete-time model, implicitly assumes that the measured signals remain constant between two consecutive samples (ZOH). Mixing the use of both assumptions
introduces modelling errors, which, however, by sufficient over-modelling remain
small in practice.
Using a bilinear transformation [91], it would be possible to model continuoustime systems exactly by means of a discrete-time model wherefore a pre-distortion
of the frequency-axis DT = T2s arctan (CT ) is required. However, due to this
distortion, the complex exponential functions f = e(if Ts ) are no longer orthogonal resulting again in a worse numerical conditioning. Moreover, the FFT
algorithm (cf. Appendix B) is not applicable anymore since this requires a uniform
frequency grid, resulting in a increase of the computation time.
120
0
20
Ampl. (dB)
40
60
80
100
120
1000
1500
2000
2500
Freq. (Hz)
3000
3500
4000
Figure 5.7: Normal-based WLS implementation for a discrete-time model with real
coefficients.
20
Ampl. (dB)
40
60
80
100
120
1000
1500
2000
2500
Freq. (Hz)
3000
3500
Figure 5.8: Normal-based Weighted mixed LS-TLS implementation for a discrete-time

model with complex coefficients.
LS or mixed LS-TLS Approach

By choosing another constraint for the parameters a different solution is found.
All the previous LS solutions were found by fixing the highest order denominator
121
coefficient an to 1. The mixed LS-TLS solution is found by using a norm-1 conH

straint on the denominator coefficients, i.e. A
A = 1, resulting in Figure 5.8. It
is clear that the choice of the constraint has an important effect on the quality of
the estimated model since in the case of the norm-1 constraint more mathematical
poles appear in the transfer function estimate compared to the LS estimate in Figure 5.6 . This will be illustrated as well in 5.3.5, where a stabilization diagram
is constructed by varying the LS constraint by fixing the lowest to highest order
denominator coefficient.
Comparison of Numerical Properties

Concerning the numerical properties for the different possible implementations and
parameterizations two important numerical aspects were considered:
number of flops: gives a first indication of the computational speed of the
different possible algorithm implementations when using the Jacobian or
Normal equations. The number of flops is found by counting each sum or
product as one single floating point operation (flop) according to the new
definition of flops given in [40].
matrix condition number and rank : gives an indication for the numerical
robustness of the implementation for a high model order. Depending both
on the choice of parameterization and type of LS equations, the accuracy of
the transfer function estimate will deteriorate, once the condition number
of the Jacobian or Normal matrix becomes too high. At the same time,
the rank of this matrix indicates the maximum number of modes that can
be identified by the number of linear independent rows or columns in the
matrix.
These quantities are summarized in Tables 5.1 and 5.2 for the practical case of the
slat track data, where No = 352, Ni = 1, Nf = 1075 and Nm = 50. In terms of
computation speed, the important gain of the optimized algorithms can be seen
from the normalized gain. For the case of the slat track example, the Jacobianbased implementation using the mixed LS-TLS algorithm based on the sparse QR
via projections is approximately 10000 times faster (in terms of flops) than solving
the problem by a full QR decomposition. Compared to the sparse QR algorithm
proposed by Bayard, using different weights for the Weighted LS version, there is
still a gain of approximately a factor 100. However, the Normal-based implementation based on the use of the Toeplitz structure, FFT and elimination approach
achieves an additional gain of approximately a factor 100 compared to the fast QR
approach for the Jacobian approach. Furthermore, it can be seen from these results that the number of flops required for the formulation of the Normal equations
by exploiting the Toeplitz structure is comparable to the Jacobian matrix.
122

Flops
Slat track
example
Normalized
gain ()
Form. J
12nNo Ni Nf
2.3E8
full QR(J)
8(No Ni Nf )3 n2 Nf
9.4E14
sparse QR(J) = W
8Nf 2 (nNo Ni + Nf )
1.7E11
5.5E3
sparse QR(J) 6= W
12No Ni Nf 2 (2n + Nf )
5.7E12
1.6E2
sparse QR(J) via proj.
32No Ni Nf n2
3.0E10
3.1E4
Jacobian-based methods
Normal-based methods
Form. Nsum
32nNo Ni Nf
6.1E8
Form. NFFT
15No Ni Nf log2 (Nf )
5.7E7
full EIG(Nsum )
4(No Ni n)3
2.2E13
4.3E1
compact EIG(Dsum )
16No Ni n(n2 + 2Nf )
1.3E9
7.2E5
compact EIG(DFFT )
No Ni (16n3 + 15Nf log2 (Nf ))
7.6E8
2.1E6
Table 5.1: Comparison of number of flops for different mixed LS-TLS algorithms using a discrete-time model with complex coefficients, including the flops count for the
formulation of the Jacobian or Normal matrix.
Table 5.2 compares the condition number and rank of the Jacobian matrix R v
(cf. Eq. 5.17) and the compact Normal matrix D (cf. Eq. 5.26) for the different
studied parameterizations, where the last row and column are omitted for the
LS problem with an = 1. These results obviously indicate the numerical problems
related to the use of the Laplace variable. Besides the very high condition numbers,
the rank of the matrices is not higher than 10, which explains that only a single or
few modes are more or less found for the high frequencies in Figure 5.3. The use
of Forsythe polynomials clearly improves the numerical conditioning, where the
Jacobian and Normal matrices have a comparable condition number and are both
of full rank. On the contrary, the normal-based implementation using Chebyshev
polynomials already suffers from the fact that these polynomials are not perfectly
orthogonal resulting in a poor overall synthesis, although most of the physical
modes are still identified since the normal matrix is still of full rank. Finally, the
use of a discrete-time model also yields a good numerical behaviour. Although,
the condition number for the normal matrix using real coefficients is somehow
higher, one can conclude that an implementation based on discrete-time model
is characterized by good numerical properties, combining both the benefits of
accuracy and computation speed.
Conclusions of Comparative Study

The conclusions drawn from this comparative study of various possible parameterizations for the frequency-domain least squares implementations for experimental
modal data can be summarized as follows:
123
Cond
Rank
Rv (1 : n, 1 : n) R
5.94E75
10
Rv (1 : n, 1 : n) C
1.67E45
D(1 : n, 1 : n) R
2.30E143
D(1 : n, 1 : n) C
3.10E75
Continuous-time Laplace
Continuous-time Forsythe
Rv (1 : n, 1 : n) R
1.04E3
100
D(1 : n, 1 : n) R
1.92E3
100
Continuous-time Chebyshev
Rv (1 : n, 1 : n) C
3.55E3
50
D(1 : n, 1 : n) C
9.27E8
50
Rv (1 : n, 1 : n) R
8.65E2
100
Rv (1 : n, 1 : n) C
8.58E2
50
D(1 : n, 1 : n) R
6.50E5
100
D(1 : n, 1 : n) C
7.35E4
50
Discrete-time
Table 5.2: Comparison of condition number and rank of Jacobian matrix Rv and Normal
matrix D for different parameterizations and a model with 50 modes (Nm = 50). The
last row and column are omitted for the LS problem with an = 1, where n = 2Nm or
n = Nm for respectively real and complex coefficients.
The use of the normal equations has the advantage, compared to the Jacobianbased LS formulation, that these equations can be constructed in a computational efficient way, while the size of the normal matrix is also smaller. This
normal matrix has a block structure with the submatrices having a predefined Vandermonde-like or Toeplitz structure. When using a discrete-time
domain model, these submatrices can be computed using the FFT algorithm,
while a good numerical conditioning of the normal matrix is preserved. A
Jacobian-based implementation is preferred in the case of continuous-time
model, although numerical conditioning is often a major problem. Both
approaches allow the elimination of the numerator coefficients in order to
first identify the system poles. This results in an important gain of computation efficiency. The numerator coefficients can be found by means of
back-substitution.
Although classically real-valued coefficients are estimated, the use of complex
coefficients is preferable with respect to both the numerical conditioning and
the computational performance. For complex coefficients the order of the
denominator polynomial equals the number of modes that can be presented
by the common-denominator model, while this model order is doubled in the
case of real-valued coefficients. For modal analysis applications, this number
of modes can be high (a typical number is 50 modes).
124

The choice of the generalized transform variable is an important factor
with respect to the numerical properties of the LS implementation. Using a
discrete-time model with complex-valued coefficients, models with 100 modes
or more can be estimated without any numerical problems (cf. 5.7). Although, modelling errors will be introduced by this type of model, a sufficient
over-modelling allows to minimize these errors. Given the robustness for high
model orders, this over-modelling does not introduce numerical problems. In
the case that the modelling errors become too large or when a very accurate
model is required, a continuous-time model might be preferred. As discussed
before, the use of Chebyshev polynomials offers a possible alternative for the
fast formulation of the structured matrices and hence it is still possible to
implement a fairly fast version of the linear Least Squares estimator based
on the normal equations. Although, since the Chebyshev polynomials are
only orthogonal in approximation, a normal-based formulation tends to suffer again from a deteriorating numerical condition, once the model orders
become high. On the contrary, Forsythe polynomials yield a very robust
formulation for a continuous-time model identification, however at the price
of an important increase of the computation time. Hence, a MIMO implementation is considered too slow for practical use in modal analysis.
The parameter constraint has an important effect on the quality of the estimated model as well. Fixing the highest order coefficient to 1, has the tendency to yield better results than the mixed LS-TLS (norm-1 constraint),
although some caution is needed as will be shown in 5.3.5.
5.3.4
Transformation to Modal Model
Since the goal is to determine the structural dynamics by means of the modal
model, the modal frequencies, damping ratios and modal residues have to be derived from the estimates of the polynomial coefficients . This is done by transforming the common-denominator model (5.1) into a pole-residue parameterization
(4.14), which is done as follows
Poles: the poles pr (r = 1, . . . , Nm ) are found as the roots of the common
denominator polynomial A(, A ) with coefficients A . From the poles the
modal frequency fdr and damping ratio r are readily obtained as
f dr =
Im(pr )
2
and
r =
Re(pr )
|pr |
(5.36)
Residues: the residue matrices Rr (No Ni ) can be calculated from the

coefficients as follows (k = 1, . . . , No Ni )
Rkr =
k (, )( pr )
lim H
pr
(5.37)
125
If a discrete-time pole-residue model (Z-domain) is used, the poles pr and residues

Rr have to be transformed to the Laplace domain by means of the impulseinvariant transformation (z = esTs ), where the damped natural frequency and
damping ratio are subsequently obtained from the poles as (5.36).
5.3.5
Constructing Stabilization Charts
The presence of noise (measurement noise, computation noise, ...) and possible
modelling errors (discrete-time domain model) require the model order n to be
chosen high enough in order to find all physical modes. Over-modelling, however,
introduces many computational poles, which complicates the modal parameter estimation process. In order to assist the user in distinguishing the physical (structural) from the computational poles, a so-called stabilization chart was proposed.
By displaying the poles (on the frequency axis) for an increasing model order (i.e.
number of modes in the model), indicates the physical poles since in general they
tend to stabilize for an increasing model order, while the computational poles
scatter around. First presented in the same period as the LSCE estimator, it has
become a common tool in modal analysis today. As a result a fast construction of
the stabilization chart is one of the requirements of a MPE algorithm.
To construct a stabilization chart, the poles have to be estimated for increasing
model orders. Based on the knowledge of the square matrix D with size (n + 1)
D=
"N N
o i
X
k=1
Tk
SkH Rk1 Sk
(5.38)
this can be done in a time efficient way, by solving the eigenvalue problem of
submatrices of D for an increasing size. By doing so, a set of LS or mixed LS-TLS
solutions (and thus the poles) are obtained for a varying order of the denominator
polynomial, while the order of the numerator polynomial is kept constant and
equal to the maximum specified order. Notice that the matrix D plays a similar
role as the so-called covariance matrix of the LSCE estimator.
Figure 5.9 shows the relation between the compact and full normal LS equations for a varying order of the denominator polynomial, while the order of the
numerator polynomials is fixed. As can be seen from Eq. (5.38), omitting for example n j rows and columns of the compact matrix D is equivalent to omitting
nj columns in the submatrices Sk and nj rows and columns in the submatrices
Tk of the full problem. The submatrices Rk remain the same and consequently,
the order of the numerator polynomials is fixed to the maximum order n. As a
result, according Eq. (5.38), the solution found by solving the compact eigenvalue
problem, where D is now a square (n j + 1) matrix, is the same as found from
solving the full problem, since the submatrices that are used for both formulations
are still the same.
126

#
!"
0
0
a0
a1
an
R1
S1

b10
b11
0
0
b1n
S2
R2
No Ni
S1H
S2H
Tk
b20
b21
b2 n
a0
a1
an

0
0
0
0
0
Figure 5.9: Relation between compact and full normal LS equations for varying order
of denominator polynomial, while order of numerator polynomials is fixed.
Remark : A similar approach is possible for the Jacobian-based implementation.

Based on the knowledge of the square matrix RV returned by the first step of the
mixed LS-TLS algorithm (cf. 5.3.1) for a given maximum model order, the SVD
of submatrices of RV can be computed for an increasing size of this matrix. Given
the equivalency between the compact and full normal LS problem as well as the
results of Appendix A proving that the Jacobian and Normal equations yield the
same solution, this also implies that same equivalency exists for the results found
by omitting the last n j rows and columns in Rv and last n j columns the full
Jacobian J.
Figure 5.10 (top) shows stabilization diagram for the case of the slattrack,
127
40
35
Number of modes
30
25
20
15
10
5
0
1000
1500
Freq. (Hz)
2000
2500
2000
2500
40
35
Number of modes
30
25
20
15
10
5
0
1000
1500
Freq. (Hz)
Figure 5.10: Stabilization diagram for normal discrete LS estimator with complex coefficients with the highest order coefficient an = 1 (top) and a norm-1 constraint (bottom)
with (+ = stable), ( = unstable) (i.e. positive real part) and (solid line = averaged sum
FRFs).
when using the normal-based discrete estimator with complex coefficients (with
the highest order coefficient an = 1). The poles are plotted for an increasing model
order, with (+) indicating stable poles, () the unstable poles (i.e. positive real
part) and the line is the averaged sum of all FRF measurements.
The effect of the parameter constraint is clearly noticed by comparing with
Figure 5.10 (bottom), where a norm-1 constraint is used to yield the mixed LS-
128
45
40
Denominator coefficient
35
30
25
20
15
10
5
0
1000
1500
Freq. (Hz)
2000
2500
Figure 5.11: Stabilization diagram for normal discrete mixed LS estimator with complex
coefficients for a constant model order and variable constraints.
TLS solution. However as can be noticed from both results, the LS implementation
indicates the pole at 1870Hz as unstable, while a fairly consistent behaviour of
this solution can be seen for the increasing model order. The mixed LS-TLS
diagram, shows this pole as a stable solution. This learns us that, although the
LS implementation leads to clear stabilization diagram, it can sometimes happen
that a physical pole is estimated as unstable, however with a real part very close
to zero.
Instead of computing the LS or mixed LS-TLS solutions for an increasing
model order, a variant for the classical stabilization chart can be constructed by
considering the n + 1 LS solutions (for the maximum model order), derived by
fixing one by one all coefficients of the denominator to 1. In practice, this can
be done in an efficient way by computing the inverse matrix of D. As shown in
[32], the jth (j = 1, . . . , n + 1) column of D 1 = (T S H R1 S)1 gives the LS
solution for the denominator coefficients A under the constraint aj = 1. Plotting
the n poles obtained for each fixed coefficient, results in Figure 5.11. The physical
and computational poles can be clearly distinguished since the physical poles are
found as stable poles for each parameter constraint, while the computational ones
are more scattered and evolve from stable to unstable poles. This result also
demonstrates that the constraint applied on the denominator coefficients has an
important influence on the clarity of the stabilization chart since the computational
poles tend to shift into the unstable part of the Nyquist plane when fixing the
higher order coefficients to 1. This approach can also be applied for the Jacobianbased LS.
129
40
35
Number of modes
30
25
20
15
10
5
0
1000
1500
Freq (Hz)
2000
2500
Figure 5.12: Stabilization diagram for normal discrete mixed LS estimator with complex
coefficients with the highest order coefficient an = 1, where relative criteria of 1% for the
frequency and 5% for the damping are used. The markers stand for (? = stable pole),
(+ = stable frequency), (x = stable damping), (o = unstable pole), where stable here
means within the criterium.
Remark : It should be noticed, however, that in practice the evaluation of the

poles in order to determine their stabilization behavior is often based on the use
of relative criteria, e.g frequency variation over 2 model orders 1% and damping 5%, as shown in Figure 5.12. However, such criteria should be used with
caution. Especially in the case of lightly damped modes at lower frequencies (0
100Hz) for which the damping ratios are small, the percentage variation will be
large, making that relative criteria are more severe for poles that are close to the
origin of the Nyquist plane. Moreover, such criteria do not take into account the
uncertainty on the measurement data, which will introduce variations on the subsequent estimates. For this reason, other mode validation criteria are presented in
Chapter 7 in order to automate the physical poles selection process.
5.3.6
LSCF Estimator
The final practical implementation of the frequency-domain Linear Least Squares

estimator optimized for modal parameter estimation is called the Least Squares
Complex Frequency (LSCF) algorithm. The choice of the most important algorithm characteristics was based on the results obtained in 5.3.3 and can be
summarized as:
130

Formulation: the normal equations (5.26) are constructed for the common
denominator discrete-time model in the Z-domain. Consequently, by looping
over the outputs and inputs, the submatrices Rk , Sk and Tk are formulated
through the use of the FFT algorithm as Toeplitz structured (n + 1) square
matrices. Using complex coefficients, the FRF data within the frequency
band of interest (FRF-zoom) is projected in the Z-domain in the interval
of [0, 2[ in order to improve the numerical conditioning. (In the case that
real coefficients are used, the data is projected in the interval of [0, [.) The
effects of the discontinuities at the edges of the interval, can be reduced by
projecting on an interval that does not completely describes the unity circle,
say [0, 2[ where is typically 0.90.95. Deliberately over-modelling is
best applied to cope with the discontinuities. This is justified by the use of a
discrete-time model in the Z-domain, which is much more robust for a high
order of the transfer function polynomials.
Solver : the normal equations can be solved for the denominator coefficients
A by computing the LS or mixed LS-TLS solution. The inverse of the
square matrix D for the LS solution is computed by means of a pseudo
inverse operation for reasons of numerically stability, while the mixed LSTLS solution is computed using an SVD.
Stabilization Chart: The stabilization chart is computed in a time-efficient
way as explained in 5.3.5. As discussed in 5.3.5, fixing the highest order
coefficient (an ) generally results in the most clear stabilization charts.
Residue Estimation: Depending on the purpose of the LSCF estimator, the
numerator coefficients and residues are derived differently. In the case that
the LSCF is used as a starting value generator for the ML estimator, the
numerator coefficients are computed by back-substitution through Eq. (5.24).
On the other hand, if only the physical poles are selected by means of the
stabilization chart, the residues are estimated by a second-step LS solver as
presented in the next section.
5.3.7
LS Residue Estimator
An alternative for deriving the estimates of the residues consists of using a 2-step
approach analogously to the traditional LSCE-LSFD approach. As discussed in
5.3.5, the LSCF estimator can be used to estimate the poles for an increasing
model order in a time-efficient manner where the results are summarized in a
stabilization chart. In a second step, the selection of the physical poles is passed to
a second-step solver, which estimates both the mode shape and modal participation
values at the same time. For this purpose, the so-called Least Squares Frequencydomain Residue estimator was developed. This estimator uses the frequencydomain formulation of the pole-residue parameterization which is given (for the
131
case that displacements and forces are considered) as

H(i) =
Nm
X
j=1
N
X
Rj
Rj
LR
+
+ UR 2
i pj j=1 i pj
(5.39)
where pj and Rj are respectively the poles and (No Ni ) residue matrices. U R and
LR are the (No Ni ) upper and lower residual terms that approximate the effects
of modes below and above the frequency band of interest. Since the poles p j are
known from the first step, the model (5.39) becomes linear in the parameters, i.e.
the residues Rj , U R and LR, and so a linear least squares problem has to be solved.
The least squares estimate of the complete residue matrices Rj , j = 1, . . . , Nm ,
can be further decomposed in a mode shape j and modal participation factor Lj
vector by means of the (thin) SVD algorithm [40]
Rj = U V T
(5.40)
Assuming that rank Rj = 1 (i.e. only one singular different from zero), the first
column of U represents the mode shape vector and the first row of V T represents
the modal participation factor vector LTj , while the singular value 1 can be used
to scale these vectors.
Similar to the CMIF method, for the case of multiple poles, the singular vectors
corresponding to the singular values significantly differing from zero are not the
actual mode shapes. Instead, these are given by a (unknown) linear combination
of the left singular vectors.
As an example, Figure 5.13, shows the mode shapes of 2 modes in the lower part
of the considered frequency band estimated by means of the LS Residue estimator.
Remark : It should be noticed that, when using a discrete-time model, it is not
possible to take the effects of modes below and above the frequency band of interest
into account by means of the so-called upper and lower residual terms, as is the
case for the continuous-time modal model (5.39). Nevertheless, over-modelling for
the effects of the discontinuities, generally results in a good FRF synthesis, even
at the edges of the analysis band. Moreover, since the LSCF estimator allows to
determine both the poles and residues of the physical modes within the band in
a single step, the unknown U R and LR can still be derived from a linear least
squares problem based on
Nm
N
X
X
Rj
R
j
H(i)
= +U R LR
(5.41)
p
i
p
2
j
j
j=1
j=1
5.3.8
Comparison of LSCF with LSCE
A comparison of the LSCF estimator with the well-known and commonly-used

LSCE estimator learns more about their similarities and differences
132
(a)
(b)
Figure 5.13: Mode shapes of slattrack for modes at 1156.6Hz (a) and 1348.6Hz (b)
estimated by the LS Residue estimator.
Data: both estimators start from FRF data for which a frequency band of
interest can be selected. However, since the LSCE is a time-domain method,
the selected data is transformed into IRFs by means of an IFFT. Consequently, possible leakage errors in the FRFs are also present in the IRF
data. Furthermore, the practical implementation of the LSCE only uses the
first, say 100 samples of the IRFs (containing the most information about
the system) in order to reduce the computation time.
Model : the LSCF uses a discrete-time scalar matrix fraction description
model (SMFD) in the Z domain with complex polynomial coefficients. The
poles are determined as the roots of the common denominator polynomial.
On the other hand, in the algorithm of the Polyreference LSCE, the poles
and participation factors are derived as the eigenvalue solutions of a problem
that is formulated using an AR model with real (Ni Ni ) matrix coefficients
[62], where this model corresponds to a right matrix fraction description
(RMFD) starting from the transpose of the IRF matrix.
Solver : in both cases, the least squares solution is found based on the formulation of the normal equations. In the case of the LSCE, the normal matrix
is called the covariance matrix and plays the same role as the normal matrix
D in (5.26).
Stabilization Chart: As discussed in 5.3.5, both estimators can construct
the stabilization chart in a time-efficient manner by solving the LS problem
for an increasing size of the normal matrix. However, since the PLSCE uses
5.4. Weighted Generalized Total Least Squares
133
an AR model with real matrix (Ni Ni ) coefficients, the relation between

the number of modes Nm that can be described by the model and the order
of the matrix polynomial n is given as nNi 2Nm . Hence the increment of
the model order in the stabilization chart is not necessarily equal to one. In
the case of the LSCF estimator this increment is always equal to one.
Modal Parameters: since the LSCF estimator uses a SMFD model, both
the poles and residues, and hence the mode shape and modal participation
factor vectors can be estimated in one step. Alternatively, similar to the
LSCELSFD approach, the LSCF can be used to yield only the poles in
a first step from which the user can make a selection using a stabilization
chart. Next, the residues are then obtained by means of the LS Residue
estimator. On the other hand, an AR model is used by the LSCE implying
that besides the poles, also the modal participation factors can be estimated
simultaneously during a first step, while the LSFD estimator yields the mode
shape vectors in a second step. In the case of inconsistent data, the LSCE in
generally performs worse [136], which partially can be explained by the fact
that both the poles and the participation factors are inconsistent hampering
the estimation process.
Figure 5.14, shows the poles obtained by means of the LSCE estimator (N m =
50). Comparing this result with Figure 5.10 (top) illustrates that the LSCE estimator does not properly identify the closely-spaced poles around 1650Hz. Although, the LSCE estimates real-valued coefficients, the use of a discrete-time
domain model avoids a bad numerical conditioning. Nevertheless, since the LSCE
is a time-domain method, no frequency weighting could be applied, which also
explains the better results obtained by the frequency-domain LS.
5.4
By taking also the uncertainty on the estimated FRF data into account during
the parametric identification, more accurate results can be obtained compared to
the LS approach. This can be done by the so-called (Weighted) Generalized Total
Least Squares (WGTLS) and Maximum Likelihood estimators. These approaches
are discussed for modal parameter estimation from FRF data in this and the
following section.
5.4.1
Cost Function Formulation
Given the linearized LS problem formulated as (cf. Eq. 5.7) and shortly noted as
J 0, the (Weighted) Generalized Total Least Squares (WGTLS) solution for
134
40
35
Model order
30
25
20
15
10
5
0
1000
1500
Freq (Hz)
2000
2500
Figure 5.14: Stabilization diagram for the LSCE estimator using real coefficients with
the highest order coefficient a2n = 1 (Ni = 1).
this estimation problem is given as [139]

1 k2 subject to J
= 0 and H = I
arg mink(J J)C
F
J
(5.42)
J,
A frequency-dependent weighting can be introduced by means of the scalar weighting Wk (f ) in the linear LS equations (5.5). The weighting matrix CJ is the square
root of the covariance matrix of J, i.e. C = CJH CJ = E J H J , with J = J J

a measure for the noise contribution on the Jacobian matrix. These disturbances
are due to the uncertainty on the FRF data.
In [107] it is proven that the elimination of J in (5.42) results in the equivalent
cost function minimized by the WGTLS estimator
arg min
H J H J
subject to H = 1
H CJH CJ
(5.43)
where this cost function is equivalent to the Rayleigh quotient of the parameter
vector . In practice, the solution is found by computing the Generalized Singular
Value Decomposition (GSVD) of the matrices J and CJ .
With the matrix J H J equivalent to the normal matrix (5.20), denoted as M
and C = CJH CJ the covariance matrix of this normal matrix, it implicitly follows
from (5.43) that the equivalent cost function minimized by the WGTLS estimator
135
for the estimation problem (5.23) is given as

arg min
H M
subject to H = 1
H C
(5.44)
In practice, the solution is now found by means of an Eigenvalue Decomposition (EVD) for the generalized eigenvalue problem that directly follows from this
equivalent cost function
M = C
(5.45)
where the solution WGTLS is given by eigenvector corresponding the smallest
eigenvalue .
Based on the equivalency between (5.43) and (5.44) the WGTLS solution for
this estimation problem is also found as
)C 1 kF subject to M
= 0 and H = 1
arg mink(M M
(5.46)
,
M
where the WGTLS solutions given by (5.42) and (5.46) are the same (cf. Appendix C).
5.4.2
Generalized Eigenvalue Problem Formulation
Referring to the generalized eigenvalue problem (5.45)

J H J = C
(5.47)
the covariance matrix C contains a measure for the errors on the normal matrix
M and hence has the same block structure
C S1
CR1
C S2
C R2
C = E J H J = ..
(5.48)
..
..
..
.
.
.
.
P
No Ni
CSH1 CSH2
k=1 CTk
In the case of FRF data, only the submatrices k contain this data and hence
only the last n+1 columns of the Jacobian matrix
to errors on
(5.10) are subjected
the measurements, i.e. J = J0 + J with J = [0] . . . [0] | . Consequently, the

covariance matrix C = E{J H J} is then given as (5.48), where the submatrices
CRk = CSk = [0] for k = 1, . . . , No Ni , while only the square (n + 1) matrix
P No Ni
CT = k=1
CTk is different from zero.
Notice that, by using SMFD model, the FRFs Hk (f ) are implicitly assumed
to be uncorrelated. Referring to Chapter 2, the extent of correlations between
136
the FRFs strongly depends on the modal test setup used as well on the noise
sources interfering with the setup. In the case that electrodynamic shakers are used
to excite the structure, important correlations can be introduced near resonance
frequencies. However, in order to obtain a fast implementation of the frequencydomain algorithms for MPE possible correlations have to be omitted. Nevertheless,
the results obtained by taking only the variances into account are in general more
accurate that the traditional or so-called deterministic LS algorithm discussed in
5.3. Furthermore, by using a SMFD, the WGTLS is consistent without taking
any possible FRF correlations into consideration.
With the entries of the submatrices Tk given by
[Tk ]rs =
[H
k k ]rs
Nf
X
f =1
|Wk (f )Hk (f )|2 fr1 s1

f
(5.49)
the entries of the submatrices CTk are computed as

[CTk ]rs = E
H
k k rs
Nf
X
f =1
H
|Wk (f )|2 var Hk (f ) fr1 s1
f
(5.50)
with var Hk (f ) = E HkH (f )Hk (f ) the variance of the measured FRF.

When a discrete-time model in the Z-domain is used with uniformly distributed
frequencies, the entries (5.50) can be computed again in a time-efficient manner
by means of the FFT algorithm (cf. Appendix B) with the matrices CTk having a
Toeplitz structure.
From this it follows that the mixed LS-TLS solution corresponds to the GTLS
solution of the Generalized Total Least Squares problem (5.45) under the condiPN i N o
tions that k=1
CTk [I]n+1 (i.e. proportional to identity matrix) and that
CRk , CSk are all zero for k = 1, . . . , No Ni . Since the submatrices k in Eq. (5.10)
do not contain any measurement data, CRk = CSk = [0] is always satisfied. The
mixed LS-TLS approach has two advantages compared to the classical TLS approach:
the assumptions with respect to the measurement noise are more realistic
since the classical TLS assumes CRk [I]n+1 with the same scaling factor
PN i N o
as for k=1
CTk [I]n+1
the elimination of the numerator coefficients (5.24) is only possible when

CRk = CSk = [0], which is the case for the the mixed LS-TLS approach.
Based on this, the generalized total least squares problem (5.45) is explicitly writ-
5.5. Maximum Likelihood
137
ten as
R1
0
..
.
S1H
0
R2
0
S2H
0
..
.
S1
S2
..
.
B1
B2
..
B
N
o Ni
0
0
..
.
0
0
..
.
..
.
B1
B2
0
..
..
.
.
B
N
o Ni
CT
(5.51)
From this it can be easily seen that, by using a common denominator model,
the WGTLS estimator is consistent without requiring any possible correlations
between the FRFs to be taken into account. As for the LS approach, the numerator
coefficients B can be eliminated from (5.51) by
Bk = Rk1 .Sk .A
(5.52)
while substitution in the last n + 1 equations of (5.51) now yields a compact GTLS
problem
"N N
#
o i
X
H 1
(5.53)
Tk Sk Rk Sk A = CT A
k=1
where the compact normal matrix is the same as the matrix D found for the LS
estimator (cf. Eq 5.26).
Again, once the A coefficients are known, back-substitution based on (5.52)

can be used to derive all B coefficients. This approach is more time efficient than
solving (5.51) directly, i.e approximately No 2 .Ni 2 times faster.
For the practical implementation of the fast WGTLS estimator, a similar approach as for the LSCF estimator (cf. 5.3.6) was followed. Again a fast construction of the stabilization chart is possible by means of computing the EVD for
an increasing order n, i.e. an increasing size of the D and CT matrices.
5.5
5.5.1
Maximum Likelihood
ML Equations
Under the assumptions that the FRFs are complex normally distributed and the
measured FRFs are uncorrelated, the maximum likelihood cost function reduces
to
Nf
N
o Ni X
X
k (f , ) Hk (f )|2
|H
`ML () =
(5.54)
var Hk (f )
k=1 f =1
k (f , ) is the FRF for k = 1, . . . , No Ni , modelled using the common
where H
denominator model (5.1). Omitting the FRF covariances in the cost function
138
results in a loss of statistical efficiency of the estimates, while the estimator is still
consistent (cf. 5.8).
The ML estimate ML of the polynomial coefficients is obtained by minimizing
the cost function (5.54) with respect to the parameters , where (5.54) is nonlinearin-the-parameters. This can be done for instance by means of a Gauss-Newton
optimization algorithm, which takes advantage of the quadratic form of the cost
function. The Gauss-Newton iterations are given by
1. solve the normal equations
H
H
Jm
Jm m = Jm
em
for
(5.55)
2. compute an update of the previous solution m

m+1 = m + m
(5.56)
with em = e(m ) the equation error and Jm = e()/ |m the Jacobian matrix
where the equation error or so-called residual vector is given as
1 (1 ,)H1 (1 )
H
var(H1 (1 ))
H
(
,)H
(
)
1
N
1
N
f
f
var(H1 (Nf ))
(5.57)
e() =
2 (1 ,)H2 (f )
H
var(H2 (f ))
..
N N (N ,)HN N (N )
H
o i
o i
f
f
var(HNo Ni (Nf ))
The explicit computation of the Jacobian matrix Jm yields a matrix with the same
block structure as the Jacobian matrix (5.10) where the submatrices k and k
are now given by
0 (1 )
n (1 )
var(Hk (1 ))A(1 ,)
var(Hk (1 ))A(1 ,)
..
..
k =
(5.58)
.
0 (Nf )
n (Nf )

var(Hk (Nf ))A(Nf ,)
var(Hk (Nf ))A(Nf ,)
and
k =
0 (1 )Bk (1 ,)
var(Hk (1 ))|A(1 ,)|2
..
.
0 (Nf )Bk (Nf ,)
var(H
k (Nf
))|A(Nf ,)|2
n (1 )Bk (1 ,)
var(Hk (1 ))|A(1 ,)|2
..
.
n (Nf )Bk (Nf ,)
var(H
k (Nf
))|A(Nf ,)|2
(5.59)
139
Given the block structure of the jacobian matrix Jm , the equations (5.55) are
explicitly given (for iteration m) as
R1
0
..
.
S1H
0
R2
0
S2H
0
..
.
B1
S1
B2
S2
..
..
.
PN o N i
B
No Ni
k=1 Tk
A
H
1 e1
H
2 e2
..
.
N N e N o N i
PNoo Nii H
k=1 k ek
(5.60)
H
H
with the submatrices now defined as Rk = H
k k , Sk = k k and Tk = k k
using (5.58) and (5.59). Given the block structure of the Eq. (5.60), a time-efficient
H
formulation of the square matrix Jm
Jm with (No Ni + 1)(n + 1) rows and columns
H
and the vector Jm em with (No Ni +1)(n+1) rows is possible in an analog manner as
presented in 5.3.2. In the case that a discrete-time model is used, these square
(n + 1) submatrices again have a Toeplitz structure. In the case of uniformly
distributed frequencies, the entries can be computed using the FFT algorithm.
Furthermore, because of the block structure, the first No Ni 1 blocks of can

be rewritten as
Bk = Rk1 H
(5.61)
k ek + Sk A
where Bk and A are the perturbations on the numerator and denominator
coefficients computed in Eq. (5.55) and ek the part of the residual vector (cf.
Eq. 5.57) corresponding to the output/input location k. The numerator coefficients
can be eliminated from the last n + 1 equations of (5.55) by means of (5.61),
resulting in
N
o Ni
X
k=1
N
o Ni
X
H 1 H
H
Tk SkH Rk1 Sk A =
k S k R k k e k
(5.62)
k=1
This elimination approach decreases the required storage space by a factor N o Ni

and the computational operations by a factor No 2 Ni 2 . As a result, the computation time for one iteration is comparable with the LSCF estimator. A time-efficient
implementation of the frequency-domain maximum likelihood algorithm is solely
possible when only the variances of the FRFs are considered in the ML cost function (5.54). Hence this implementation is merely a compromise between accuracy
and computation time. As a result, it is significantly faster than the exact ML
algorithm, while the results are in general more accurate that the traditional or socalled deterministic algorithms for modal parameter estimation. The covariances
can optionally be considered for a more accurate computation of the uncertainty
bounds for the parameter estimates, as shown in the next section.
140
5.5.2
Parameter Uncertainty Bounds
Given the uncertainty on the measurements as the covariance matrix of the FRF
matrix (cf. Ch. 2), the covariance matrix of the ML estimate ML can be computed
according the following approximation [109] (p.287)
H
Cov(ML ) [Jm
Jm ]1
(5.63)
with Jm the Jacobian matrix evaluated in the last iteration step of the GaussNewton algorithm. As one is mainly interested in the uncertainty on the modal
frequencies and damping ratios, only the covariance matrix of the denominator
coefficients is required. Starting from (5.62), the covariance matrix of the denominator coefficients A can be computed without loss of generality but an important
reduction of the computation time as
Cov(A )
o
n
H
E A A
( N N
o i
X
Tk SkH Rk1 Sk
k=1
N N
o i
X
l=1
eH
l
l Rl1 Sl
!1 N N
o i
X
k=1
! N N
o i
Tl
H 1 H
H
k Sk R k k e k
SlH Rl1 Sl
l=1
!H )
!
(5.64)
where A are perturbations of the estimated coefficients of the denominator polynomial, for which E{A } is taken equal to zero, which is justified by Eq. (5.62)
if the estimate of A equals the exact denominator coefficient vector. Under the
noise assumption that no correlation
exists between any

of the FRFs, the same
H
holds for the residuals, i.e. E eH
e
=
I
and
E
e
e
= 0Nf for k 6= l, and
k
N
l
f
k
k
hence Eq. (5.64) reduces to
Cov(A )
"N N
o i
X
k=1
Tk
SkH Rk1 Sk
!#1
(5.65)
Hence it is not necessary to invert the full matrix defined by (5.63). Moreover,
this covariance matrix is the inverse of the matrix that is also computed to solve
the normal equations (5.62). From (5.65), the uncertainty on the poles (modal
frequencies and damping ratios) can be computed using the approach given in
[54].
Remark : In the case that the FRFs are correlated, it is necessary to compute
the covariance matrix according the full expression (5.64). In [143], the influence
of correlations between FRFs measured in near points is studied by means of
monte carlo simulations demonstrating a systematic overestimation of the bounds
(typically up to a factor 2) when only the variances are taken into account.
141
40
35
Iteration number
30
25
20
15
10
5
0
1000
1200
1400
1600
1800
2000
Freq. (Hz)
2200
2400
2600
Figure 5.15: Stabilization diagram for the ML estimator.
5.5.3
ML Stabilization Chart
Given the poles derived from each iteration step, a variant of the stabilization chart
(cf. 5.3.5) can be constructed by plotting the poles for an increasing number
of iterations as a function of the frequency. Since the model order is fixed, the
same number of solutions (poles) is found for each iteration step. However, it is
observed that the physical poles will have a more consistent behaviour over the
iterations resulting in clear stable lines, as shown in Figure 5.15 for the case of the
slattrack data.
5.5.4
Logarithmic ML
The logarithmic equation error is given by
k (f ) Hk (f )
LOG
(f ) = log H
k
(5.66)
The variance of this equation error is given by

var
LOG
(f )
k
var Hk (f )
= var log(Hk (f ))
Hk (f )2
(5.67)
i.e. a measure for the noise-to-signal

(power)
ratio of the measured FRFs. Hence, if
a nonparametric estimate of var Hk (f ) is available (cf. Chapter 2), the following
142
weighted complex logarithmic cost function is obtained

`LOG () =
Nf
N
o Ni X
X
k=1
k (f ) Hk (f ) 2
log H
Hk (f )2
f =1 var Hk (f )
(5.68)
This estimator is in strict sense ML-likewith an asymptotic efficiency close to

the ML estimator while being practically consistent [49]. The logarithmic ML
estimator is more appropriate for the analysis of FRF measurements with a very
high dynamical range (i.e. 100dB and more) and is more robust to outliers in the
data.
The same fast algorithm as used for the minimization of (5.54) can be applied
to (5.68) to derive the parameter estimates with their uncertainty bounds. The
entries of the Jacobian matrix Jm (cf. Eq. 5.55) are now found to be
and
k =
k =
0 (1 )|Hk (1 )|
var(Hk (1 ))Bk (1 ,)
..
.
0 (Nf )|Hk (Nf )|

var(Hk (Nf ))Bk (Nf ,)
0 (1 )|Hk (1 )|
var(Hk (1 ))A(1 ,)
..
.
0 (Nf )|Hk (Nf )|
var(H
k (Nf
))A(Nf ,)
n (1 )|Hk (1 )|
var(Hk (1 ))Bk (1 ,)
..
.
n (Nf )|Hk (Nf )|

var(Hk (Nf ))Bk (Nf ,)
(1 )|Hk (1 )|
var(Hk (1 ))A(1 ,)
..
.
n (Nf )|Hk (Nf )|
var(H
k (Nf
))A(Nf ,)
(5.69)
(5.70)
Remark : The idea of fitting the logarithm (of the amplitude) of the FRF data, was
first proposed by [128] for the nonlinear least squares estimator. The logarithmic
LS cost function tends to be smoother than the traditional nonlinear least squares
cost function resulting in a larger convergence region [9]. A generalization for
MIMO systems has been given in [68]. In [49], the asymptotical properties are
studied in an EV framework.
5.6
Choice of Frequency Weighting
Depending on the solver used, i.e. a WLS (cf. 5.3) or WGTLS (cf. 5.4) and the
choice of the weighting Wk (f ) a number of other frequency-domain estimators
can be derived using the same fast implementation.
5.6. Choice of Frequency Weighting
143
Weighted Least Squares Solver

A drawback of the linearized equation error k (f , ) = A(f , )H(f )Bk (f , ),
in the case of a continuous-time model (a polynomial in f = if ), is the overemphasizing of high-frequency measurements in the cost function. This may result
in a poor modelling at low frequencies and ill-conditioned normal equations in the
case of data with a high modal density and a large dynamic frequency
By inPNo Nrange.
i
troducing a frequency-dependent weighting in the cost function k=1
|k (f , )|2 ,
this problem can be reduced. In the case that a common-denominator model is
used, this is done by weighting the linearized equation error k (f , ) by means of
the scalar weight Wk (f )
w
k (f , ) = Wk (f ) A(f , )Hk (f ) Bk (f , )
(5.71)
for k = 1, . . . , No Ni the output/input DOFs and f = 1, . . . , Nf the number of

spectral lines.
The basic idea was introduced by Sanathanan and Koerner [119], where the
following (deterministic) weighting function evaluated using the denominator coefficients known from a previous estimation
1
Wk2 (f ) =
A(f , m1 )2
(5.72)
results in an iteratively weighted linear least squares problem minimizing the following cost function
`IWLS (m ) =
Nf
N
o Ni X
X
k=1
A(f , m )Hk (f ) Bk (f , m )2
A(f , m1 )2
f =1
(5.73)
As a result, a lower weight is given to the data at the higher frequencies. Several
other weighting functions (Stahl, Strobel, Gy
urki, t Mannetje, Whitfield, etc) are
summarized in [105]. This iterative approach, however, requires adequate starting
values [145].
The optimal weighting function equals the maximum likelihood weighting in
the cost function (5.54) which, for the linearized least squares equations, can be
introduced by the following (stochastic) weighting
1
Wk2 (f ) =
2
A(f , m1 ) var Hk (f )
(5.74)
evaluated using the denominator coefficients known from a previous estimation.

This results in an iteratively weighted estimator with nearly ML properties, called
the Iterative Quadratic Maximum Likelihood (IQML) estimator minimizing the
144
0
20
Ampl. (dB)
40
60
80
100
120
1000
1500
2000
1500
2000
2500
Freq. (Hz)
3000
3500
3000
3500
10
20
30
Ampl. (dB)
40
50
60
70
80
90
100
110
1000
2500
Freq. (Hz)
Figure 5.16: Unweighted (top) and weighted (bottom) normal-based LS implementation

for a discrete-time model with complex coefficients.
following cost function

`IQML (m ) =
Nf
N
o Ni X
X
k=1
A(f , m )Hk (f ) Bk (f , m )2
A(f , m1 )2 var Hk (f )
f =1
(5.75)
In the case of a discrete-time model, the problem over-emphasizing the highfrequency measurements in the cost function is not experienced, which is one of
the advantages of using this type of model for modelling modal data with a high
145
modal density and a large dynamic frequency range. Nevertheless, the idea of
using a frequency-dependent stochastic weighting can still offer advantages since
the importance of each measurement in the LS estimation process is weighted
according the quality of each measurement. Analogously to the nonparametric
weighting function as proposed in [107] for Input/Output Fourier data, the following frequency-dependent weighting function can be used for FRF data in the
case of a discrete-time model, where possible correlations between the FRFs are
neglected,
Wk2 (f )
Hk (f )2
=
var Hk (f )
(5.76)
Given the nonparametric character of the weighting function an iterative process is

not needed. In order to illustrate the effect of this frequency weighting, Figure 5.16
shows the results obtained for the slattrack data using an unweighted (top) and
weighted (bottom) version of the Normal-based LS for a discrete-time model with
complex coefficients (Nm = 50). As can be noticed, the 3 closely-spaced modes
around 1650Hz are not properly identified by the unweighted LS, while a correct
model is found for the analyzed frequency band in the case of the weighted LS.
Weighted Generalized Total Least Squares Solver

The WGTLS cost function (5.44) in terms of the linearized weighted equation
error (5.71) is given as
2
P N o N i PN f
A(f , )Hk (f ) Bk (f , )2
k=1
f =1 Wk (f )
`WGTLS () =
2
PN o N i P N f
A(f , )2 var Hk (f )
k=1
f =1 Wk (f )
(5.77)
By choosing the weighting function Wk (f ) as the nonparametric weighting

(5.76), the accuracy can be improved since the importance of each measurement is
weighted according the quality of each measurement without the need for starting
values.
By using Wk (f ) the ML-like weighting function (5.74) evaluated using the
denominator coefficients known from a previous estimation, the asymptotical efficiency of the GTLS estimator can be improved. This results in an iteratively
WGTLS estimator with nearly maximum likelihood properties, also called the Bootstrapped Total Least Squares (BTLS) estimator.
146
20
30
40
Ampl. (dB)
50
60
70
80
90
100
110
120
130
1000
2000
3000
4000
5000
6000
7000
5000
6000
7000
Freq. (Hz)
20
30
40
50
Ampl. (dB)
60
70
80
90
100
110
120
130
1000
2000
3000
4000
Freq. (Hz)
Figure 5.17: Results for WLSCF (top) WGTLS (bottom) estimators for Nm = 80.
147
20
30
40
Ampl. (dB)
50
60
70
80
90
100
110
120
130
1000
2000
3000
1000
2000
3000
4000
Freq. (Hz)
5000
6000
7000
4000
5000
6000
7000
20
30
40
50
Ampl. (dB)
60
70
80
90
100
110
120
130
Freq. (Hz)
Figure 5.18: Results for ML (top) and BTLS (bottom) estimator for Nm = 80.
148
5.7
Validation of Stochastic Estimators for Experimental Data
In order to validate the accuracy of the stochastic algorithms for a high model
order, the slattrack data was analyzed by estimating a model of 80 modes in
frequency band of 2007800Hz using the WLS, WGTLS shown in Figures 5.17
and ML (50 iterations) and BTLS (50 iterations) estimator shown in Figure 5.18.
As can be seen, the and WLSCF and WGTLS yield very similar results, although the WLSCF tends to estimate a more accurate model for the very high
frequencies, which can be explained by the difference in parameter constraint. The
ML estimator clearly yields the most accurate model, where all the modes in the
considered frequency band are correctly identified. Interesting modes for comparison are the closely-spaced and highly damped modes at 2600Hz and the closely
spaced mode close to the dominant resonance peak at 4000Hz.
5.8
Asymptotical Properties of Stochastic Estimators
Based on the cost function formulation it is possible to evaluate the asymptotical properties of the different frequency-domain estimators. However, since the
possible correlations between the FRFs are neglected in the covariance matrix of
the FRFs, none of the presented estimators will be asymptotically efficient. The
asymptotical consistency of the estimators can be examined using the quick tools
proposed in [109] (p. 189190).
A necessary condition for consistency is the minimization of the expected value
of the cost function by the exact modal parameters (assuming that the true model
belongs to the model set)
e = arg min E{`()}
(5.78)
Remark : However, in order to prove the consistency of an estimator (i.e. a necessary and sufficient condition), one should prove that the cost function `() converges in means square sense to exact cost function for the number of frequency
spectral lines Nf [109].
Least Squares
Given that the measured FRF can be written as Hk (f ) = Hk0 (f ) + NHk (f ),
where E{Hk0 (f )NH

(f )} = 0. The expected value of the cost function of the
k
5.8. Asymptotical Properties of Stochastic Estimators
149
LS estimator (5.6) is computed as

E{`WLS ()}
=
(5.79)
Nf
o Ni X
n NX
o
A(f , )H(f ) Bk (f , )2
E
k=1 f =1
Nf
o Ni X
n NX
o
(A(f , )Hk0 (f ) Bk (f , ) + A(f , )NH (f )2
E
k
k=1 f =1
Nf
N
o Ni X
X
k=1 f =1
A(f , )Hk0 (f ) Bk (f , )2 + A(f , )2 var NH (f )

k
(5.80)
2
Hence, in the exact parameters, E{`WLS (e )} = 0 + A(f , Ae ) var NHk (f ) ,

is not minimal since this still depends on the parameter and the data.
Generalized Total Least Squares
A similar result as for the LS is found for the Total Least Squares estimator. However, referring to the cost function of the GTLS (5.77), where the measured FRF
Hk (f ) = Hk0 (f ) + NHk (f ), with E{Hk0 (f )NH

(f )} = 0 and the variances
k
var Hk (f ) are assumed to be exactly known, the expected value is given as
E{`GTLS ()}
(5.81)
nP
o
P
Nf
No Ni
A(f , )Hk0 (f ) Bk (f , )2 + A(f , )2 NH (f )
E
k
k=1
f =1
=
P N o N i PN f
k=1
f =1
PN o N i P N f
A(f , )Hk0 (f ) Bk (f , )2
k=1
f =1
=
+1
2
PN o N i P N f
k=1
f =1 A(f , ) var Hk (f )
From this it follows that the expected value of the GTLS cost function is minimal
in the exact parameters, i.e the necessary condition for consistency.
Maximum Likelihood
Under the assumption
that the FRFs are complex normally distributed and the

variances var Hk (f ) are exactly known, computing the expected value of the
cost function (5.54) based on the linearized equation error results in
Nf
N
o Ni X
A(f , )Hk0 (f ) Bk (f , )2
X
+1
(5.82)
E{`ML ()} =
k=1 f =1
150
where the measured FRF Hk (f ) = Hk0 + NHk , with E{Hk0 NH

} = 0. From this
k
it follows that the expected value of the ML cost function is minimal in the exact
parameters, i.e the necessary condition for consistency.
5.9
Conclusions
In this chapter, the LS problem formulation using the Jacobian or normal equations
has been numerically optimized in terms of the required memory and computation
time by exploiting the block structure resulting from using a common-denominator
model. For the Jacobian-based implementation, this was achieved by developing
a fast sparse QR algorithm for the elimination of the numerator coefficients, while
this elimination follows directly from the block structure of the normal equations,
where this latter approach is still faster than using the Jacobian formulation.
A study of different types of parameterization for experimental modal data
has demonstrated that the use of the normal-based formulation is justified when
using a discrete-time model with complex coefficients, resulting in a very robust
approach in terms of numerical conditioning and accuracy, where models of 100
modes or more could be estimated without any numerical problems. The memory
usage and computation time were further optimized by exploiting the Toeplitz
structure of the block matrices and the use of the FFT algorithm in the case of
the discrete-time model.
For a continuous-time model, the use of orthogonal polynomials is required in
order to preserve a good numerical conditioning, as shown by the use of Forsythe
or Chebyshev polynomials. In the case of Chebyshev polynomials, a time-efficient
implementation could be implemented. However, a model order of typically 50 or
more still results in numerical problems for the normal matrix since the Chebyshev
polynomials are only orthogonal in approximation.
Furthermore, it is shown that the parameter constraint has an important effect
on the obtained solution and a method for the fast construction of the stabilization
diagram has been derived. Likewise the well-known LSCE estimator, the final
implementation of the LS algorithm (LSCF) can be used in a 2 step approach in
combination with the proposed LS Residue estimator.
The fast normal-based implementation was extended to the class stochastic
(frequency-domain) estimators such as the Maximum Likelihood and the Weighted
Generalized Total Least Squares. Notice that the time-efficient implementation is
a balance between accuracy and memory/computation efficiency since it does not
allow to take possible correlations between the FRFs into account. Although this
implies that the stochastic estimators are not asymptotically efficient anymore, the
consistency is preserved for the ML and WGTLS estimator. It has been shown that
a fast implementation of the WLS and WGTLS variants, such as IQML and BTLS,
5.9. Conclusions
151
can be readily obtained. The same features, such as Toeplitz structured matrices,
use of FFT algorithm and a fast construction of the stabilization diagram were
applicable. The robustness and accuracy of the stochastic estimators has been
illustrated for experimental data obtained from a safety critical component of an
aircraft.
152
5.10
Appendix A Jacobian and Normal Matrix

based LS Algorithms
Given the very compact LS problem, based on the (n + 1) last equations of (5.23),
#
"N N
o i
X
H 1
Tk S k R k Sk A 0
(5.83)
k=1
Rewriting this as a function of the submatrices k and k of the Jacobian matrix

H
(5.10), based on the definition of the submatrices Rk = H
k k , Sk = k k ,
(cf.
Eq
5.23),
yields
Tk = H
k
k
#
"N N
o i
X
H
H
H
1 H
(5.84)
k k k (k k ) k k A 0
k=1
or
N
o Ni
X
k=1
+
H
k I k k k A 0
(5.85)
+
with +
k the pseudo-inverse of the matrix k . Furthermore, the matrix I k k
is idempotent since
+
+
+
+
(5.86)
I k +
I k +
k = I 2k k + k k k k = I k k
k
with +
k k = I. The QR-decomposition of the (Nf n + 1) matrix k = Qk Rk ,
from which it follows that
+ +
H
k +
k = Q k R k R k Qk = Q k Qk
(5.87)
H 1 H
+
Qk = IQH
since
Q+
k (Qk is orthogonal). As a result, I k k =
k = Qk Qk
I Q k QH
k , which is the (Hermitian) orthogonal projection operator Q k as defined in step 2 of the Sparse matrix QR algorithm (cf. 5.3.1).
As a result, given Eq. (5.86), Eq. (5.85) can be written as
Q
1 1
H H
2 2
Q
1 Q1 , 2 Q2 , . . . , H
A 0
.
No Ni No Ni
..
(5.88)
Q
No Ni No Ni
which based on the definition of the matrix V in step 3 of the mixed LS-TLS
algorithm using Sparse QR decomposition boils down to
H
V H V A = RVH QH
V QV R V A = R V R V A 0
(5.89)
proving that normal equations based on the matrix RV will yield the same solution
for A as the normal equations (5.26).
5.11. Appendix B Fast Calculation of Normal Matrix Entries
5.11
153
Appendix B Fast Calculation of Normal

Matrix Entries
If a discrete time-domain model is used, i.e. jf = e(if Ts ) and if the frequencies

), then the summations
are uniformly distributed (e.g., f = f. with = N2
s Ts
(5.21) can be rewritten as
[H
k k ]rs
Nf
X
f =1
[H
k k ]rs
Nf
X
f =1
[H
k k ]rs
|Wk (f )|2 ei2(rs)f /Ns

|Wk (f )Hk (f )|2 ei2(rs)f /Ns
Nf
X
f =1
|Wk (f )|2 Hk (f )ei2(rs)f /Ns
(5.90)
Defining the sequences

k (n)
k (n)
k (n)
|Wk (f )|2
n = 1, . . . , Nf
0
n = 0, Nf + 1, . . . , Ns 1
|Wk (f )Hk (f )|2

n = 1, . . . , Nf
0
n = 0, Nf + 1, . . . , Ns 1
|Wk (f )|2 Hk
n = 1, . . . , Nf
0
n = 0, Nf + 1, . . . , Ns 1
(5.91)
(5.92)
(5.93)
it is observed that
[H
k k ]rs
[H
k k ]rs
[H
k k ]rs
=
=
F k (r s)
F k (r s)
F k (r s)
(5.94)
with
F k (m)
NX
s 1
k (n) expi2mn/Ns
n=0
F k (m)
NX
s 1
k (n) expi2mn/Ns
n=0
F k (m)
NX
s 1
k (n) expi2mn/Ns
n=0
the discrete Fourier transform (DFT) of the above defined sequences. Using an
FFT to form the normal equations [53, 125] results in a further reduction of the
computation time (typically a factor 210).
154
5.12
Appendix C Equivalency of WGTLS Solutions
It can be assumed that C = I without any loss of generality. Given the singular
value decomposition of the Jacobian matrix J
J = U V H
(5.95)
the Frobenius norm of the matrix J is then given as the sum of the singular values
(n is the rank of the matrix J)
v
u n
q
uX
j2
(5.96)
kJkF = trace J H J = t
j=1
The Frobenius norm of the matrix J H J results in

v
uX
u n 4
H
kJ JkF = t
j
(5.97)
j=1
which, in general, is not equal to kJkF . However, the TLS solution is only found
in the case that the rank of the perturbed matrix J is equal to n 1 in order that
2 0. The SVD of the matrix J of rank n 1 is given as
k(J J)k
F
J =
n1
X
uj j vjH
(5.98)
j=1
Hence, based on equations (5.96) and (5.97) it follows that

2F
k(J J)k
F
k(J H J JH J)k
=
=
n2
p
n4 = n2
(5.99)
where n is the smallest singular value. This proves that both the (WG)TLS
problems based on the Jacobian (5.42) and normal matrix (5.46) yield the same
result.
Chapter 6
Frequency-domain MPE
from Input/Ouput Data
For the case of an arbitrary excitation and that only a very limited amount of data
samples can be acquired, the nonparametric FRF identification based on averaging
techniques becomes difficult. Sufficient time samples must be available to have
an acceptable frequency resolution to detect closely spaced modes. An alternative
approach then consists of estimating the modal parameters based on the Input
and Output (I/O) Fourier sequences. Based on the results of Chapter 5, a fast
implementation of the frequency-domain (Total) LS and ML methods is derived for
I/O data in Sections 6.4, 6.5 and 6.7. Moreover, when using the I/O data instead
of the FRFs, effects of leakage (or transients in general) can be compensated by
estimating the system and transient parameters simultaneously as will be discussed
in Section 6.8. This approach can also be generalized to estimate the parameters
from auto and cross power spectra, which can be preferable when longer time records
are available and a parametric compensation for possible leakage effects is desired
(cf. 6.9).
155
156
Chapter 6. Frequency-domain MPE from Input/Ouput Data
6.1
Introduction
As explained before, the common approach for experimental modal identification

starts from frequency response functions (FRF), which are derived using a nonparametric FRF estimator such as H1 or Hv . However, FRF-based identification
can be seriously complicated by one or more of the following problems
First of all, too restrictive assumptions about the noise model result in inconsistent FRF estimators. As discussed in Chapter 2, accurate FRF estimation
is possible based on errors-in-variables nonparametric estimators, such as the
Hiv (arbitrary excitation) or Hev (periodic excitation).
However, since FRF estimators are based on averaging techniques, sufficient
time samples must be available in order to have an acceptable frequency
resolution to detect all the dynamical behaviour, including closely-spaced
modes. When only short data sequences are acquired under a random noise
excitation, it is preferable to estimate the parameters based on the raw (I/O)
Fourier series.
Another problem relates to the use of arbitrary excitation signals (e.g random noise) which introduces leakage effects, especially in the case of short
time sequences obtained from lightly damped structures. Although, periodic
signals, such as multisine or periodic chirp signals, avoid effects of leakage,
random noise excitation is still most often used in modal testing requiring
special attention as was also discussed in Chapter 3.
Several applications are confronted with one or more of these problems. In the
case of flutter testing, the time to perform measurements is very limited due to
safety and cost reasons. Only very short time sequences can be measured at each
flutter speed and the measured data typically contains high noise levels [47, 134].
When using a SLDV setup to measure in a high number of response DOFs and the
overall testing time is limited, the measurements can also be restricted to short
data sequences.
In [110, 47, 42] a frequency-domain Maximum Likelihood algorithm and [108,
42] frequency-domain Total Least Squares-based algorithms are presented to estimate modal parameters for both SISO (Single Input Single Output) and MIMO
(Multiple Input Multiple Output) models from Input/Output data. However the
numerical (computation time, memory usage and numerical conditioning) performance was not optimized to handle extensive data-sets characterized by a high
modal density and corrupted by leakage.
In this chapter, the frequency-domain estimators for FRF-based modal parameter estimation presented in Chapter 5, are adapted for the case of I/O data. As
discussed in the beginning of this section, the use of I/O data is certainly prefered when only short data sequences corrupted with leakage errors are available.
6.2. Parametric Model
157
The formulation of the I/O estimators is based on an errors-in-variables set-up

(cf. 6.3), where both the noise on the input and output signals are considered.
However, as for the FRF-based estimators where possible correlations between the
FRFs are ommited, the fast implementation of the I/O estimators requires that
only the output variances are considered. Nevertheless, in the case of a SLDV
measurement setup, this noise model is still exact.
In 6.4, both the Jacobian and Normal-based formulations of the LS estimator
are derived for the I/O data. Using the same approach as in 5.3.2, a fast implementation of the LS estimator is possible. However, as will be shown in 6.5,
the fast formulation of the WGTLS I/O estimator was less straightforward since
two approximations are required. Besides ommiting possible output correlations,
the derivation the WGTLS requires a linear approximation in the formulation of
a compact WGTLS problem, which has been validated by means of Monte Carlo
simulations. The I/O-based ML estimator is discussed in 6.7.
By estimating simultaneously the initial conditions and the system model parameters, it is possible to deal with arbitrary signals in the frequency-domain
without any approximation and under the same assumptions as in the time-domain
[111, 56]. Using this approach, the proposed I/O estimators can be made robust
for errors due to leakage effects (cf. 6.8).
In 6.9, discusses the generalization of the I/O estimators to the case of Auto
and Cross Power Spectral Density functions. The use of PSD functions has the
advantage, compared to using the raw I/O data, that the noise on the data is
reduced by means of the averaging process, however, this requires sufficient data.
On the other hand, compared to using FRF data, the effects of leakage can now
be compensated for as shown in 6.8.
6.2
Parametric Model
Analogous to Chapter 5, a common denominator model is used to model the

Frequency Response Function between output o and input i (for o = 1, 2, . . . , No
and i = 1, 2, . . . , Ni ) at angular frequency f with f = 1, . . . , Nf (Nf the number
of spectral lines)
oi (f ) = Boi (f , )
H
A(f , )
(6.1)
Pn
Pn
j
j
with Boi (f , ) =
j=0 aj f the
j=0 boij f the numerator and A(f , ) =
common-denominator polynomial, where n is the order of the polynomials and f
the generalized transform variable (cf. 5.2). The polynomial coefficients a j and
boij are the parameters to be estimated.
158
6.3
Errors-in-Variables Noise Model
In practice, the measured responses (velocities and/or accelerations) as well as

forces are affected by errors. As discussed in 2.2, the measured I/O Fourier data
Z = [F H , X H ]H can be represented using a errors-in-variables stochastic noise
model
Z(f ) = Z0 (f ) + EZ (f )
f = 1, . . . , Nf
(6.2)
[H0 (f ), INo ] Z0 (f ) = 0
with Z0 = [F0H , X0H ]H the true input-output (I/O) Fourier data and EZ =
H H
[EFH , EX
] some random perturbations and Z(f ), Z0 (f ), EZ (f ) C (Ni +No 1) .
The errors EZ (f ) are assumed to be complex normally distributed with the

following a-priori known covariance matrix
CEF EX (f )
CEF (f )
H
CEZ (f ) = E{EZ (f )EZ (f ) } =
(6.3)
CEX (f )
CEX EF (f )
where the non-diagonal elements of the output noise covariance matrix C EX are
not considered throughout this chapter. This follows from the practical issues
related to the derivation of a full rank covariance matrix CEZ (f ) for a large
number of responses (No ) by means of measurements (cf. 2.4). Furthermore, a
time-efficient implementation of the I/O estimators requires that CEX is diagonal
as will turn out in this chapter.
For the particular case of scanning laser Doppler vibrometer (SLDV) measurements, the errors on the (subsequent) output measurement setup are uncorrelated
since each response location is measured separately resulting in zero non-diagonal
elements of CEX (zk ). On the other hand, since the number of inputs is typically
small (Ni 5), input correlations as well as input-output correlations can be determined, resulting in CEF (f ) and CEX EF (f ) being full matrices for each measured
output. Taking, besides the noise variances, also the cross-correlations of the errors on the I/O Fourier data into account, can further decrease the uncertainty of
the estimates.
6.4
(Weighted) Linear Least Squares
Using the model (6.1) and given the input and output Fourier data (2.1) as defined
in 6.3, the linearized (Levi) weighted equation error for output o and spectral
line f is then written as
Wo (f ) bBo (f , )c {F (f )} A(f , )Xo (f ) 0

(6.4)
where {F (f )} represents the input (Ni 1) Fourier vector and Xo the response
(output) Fourier coefficient measured at DOF o for spectral line f as a result of
159
the Ni inputs. The (1 Ni ) row vector bBo (f , )c is the oth row of the (No
Ni ) numerator polynomial matrix B(f , ). An adequate (frequency-dependent)
weighting function Wo (f ) in the equations (6.4), can improve the quality of the
parameter estimates. Notice that in the specific case of SLDV measurements, a
different input Fourier vector {Fo (f )} is measured for each output DOF.
Since the equations (6.4) are linear in the parameters
they can be formulated as J 0
B1
1 0
0
B2
0 2 0
..
..
..
.
.
.
.
.
.
.
.
.
.
.
0
0 No No
B No
A
and in the Fourier data,
(6.5)
with J the Jacobian matrix now having Nf No Ni rows and (n + 1)(No Ni + 1)

columns. A similar formulation can be obtained for real coefficients for which the
number of equations and coefficients is doubled. The entries of the submatrices
o = [o1 , . . . , oNi ] (Nf Ni .(n + 1)) and o (Nf (n + 1)) are given by
oi (f )
Wo (f )[0f , 1f , . . . , nf ]Fi (f )
o (f )
Wo (f )[0f , 1f , . . . , nf ]Xo (f )
(6.6)
while the parameter vector entries contain the (unknown) coefficients

Bo
A
=
=
[bo10 , bo11 , . . . , boNi n ]T

[a0 , a1 , . . . , an ]
(6.7)
(6.8)
Since an analogous formulation is obtained as in 5.3.1, the Least Squares

solution can again be found by means of the sparse QR algorithm. However, for
the same reason as in 5.3.2, (i.e. Nf n), the normal equations J H J 0
results in a more compact formulation of the identification problem (J H J having
(n + 1)(No .Ni + 1) rows and columns)
B1
R1
0
S1
B2
R
0
S
2
2
..
(6.9)
..
.
..
. 0
..
.
.
0
PN o
B No
S1H S2H
o=1 To
A
with Ro having (n + 1).Ni rows and columns,
Ro11 Ro1Ni
Ro21 Ro2Ni
Ro = .
..
..
..
.
.
RoNi 1 RoNi Ni
160
So having (n + 1).Ni rows and (n + 1) columns,

So = [So1 , So2 . . . , SoNi ]T
and To having (n + 1) rows and columns. The parameter vector entries containing
the (unknown) coefficients are given by
bo10
a0
bo11
a1
Bo =
, A =
..
..
.
.
boNi n
an
The entries of the submatrices are given by
[Roi1 ,i2 ]rs
Nf
X
f =1
[To ]rs
Nf
X
f =1
[Soi1 ]rs
|Wk (f )|2 Fi1 (f )Fi2 (f )fr1 s1

f
H
s1
|Wk (f )Xo (f )|2 r1
f
f
Nf
X
f =1
s1
|Wk (f )|2 Fi1 (f )Xo (f )r1
f
f
(6.10)
The same time efficient approach as presented in 5.3.2 can be used to derive an
estimate of the coefficients . The elimination of the numerator coefficients Bo
Bo = Ro1 So A
o = 1, . . . , No
(6.11)
yields
"N
o
X
o=1
To SoH Ro1 So A 0
(6.12)
Once the denominator coefficients A are known, a back-substitution based on

Eq. (6.11) can be used to find B . Based on (6.12), a stabilization chart can be
constructed again in a time efficient manner as was explained in 5.3.5.
From Eqs. (6.9) and (6.10) it clearly follows that the submatrices of the normal
matrix have a predefined structure. By exploiting the specific matrix structure
in a similar way as discussed in 5.3.2 and 5.3.3, a further reduction of the
computation time and memory requirements is obtained. More specific, in the
case that a discrete-time model is used (i.e. f = e(if Ts ) ) the submatrices Ro ,
So and To have a Toeplitz structure while the entries can be computed by means
of the FFT algorithm if the frequencies are uniformly distributed. Using a discrete
time-domain model leads to both a well-conditioned Jacobian and normal matrix.
In the case of a continuous time-domain model, orthogonal polynomials have to
be used to significantly improve the numerical conditioning (cf. 5.3.3).
6.5
161
6.5.1
Cost Function Formulation
The same general formulation of the Weigthed Generalized Total Least Squares,
as derived in 5.4.1 for FRF-based identification, is applicable for the I/O formulation of this estimator. With the Jacobian matrix J and normal matrix M now
formulated as in Eqs. (6.5) and (6.9), the WGTLS solutions are then respectively
given by Eqs. (5.42) and (5.46).
Hence, given the normal equations (6.9), the equivalent cost function minimized
by the WGTLS estimator for this estimation problem is given as (where M = J H J)
arg min
H M
subject to H = 1
H C
(6.13)
In practice, the solution is found by means of an Eigenvalue Decomposition (EVD)

for the generalized eigenvalue problem that directly follows from this equivalent
cost function
M = C
(6.14)
where the solution WGTLS is given by the eigenvector corresponding to the smallest eigenvalue . The covariance matrix C is a measure for the errors on the
matrix M resulting from the measurement noise on
the I/O
Fourier data. This
covariance matrix C is defined as C = CJH CJ = E J H J , with J = J J a
measure for the noise contribution on the Jacobian matrix (6.5).
6.5.2
Generalized Eigenvalue Problem Formulation
Considering the errors-in-variables model with the noise assumptions made in 6.3,
the generalized eigenvalue problem (6.14) is explicitly given as
R1
0
..
.
0
R2
..
.
S1H
S2H
..
.
B1
B2
..
A
T
S1
S2
..
.
CR1
0
..
.
0
C R2
..
.
..
.
C S1
C S2
..
.
CSH1
CSH2
CT
B1
B2
..
(6.15)
162
PN o
PN o
where T = o=1
To and CT = o=1
CTo . Given the Jacobian entries (6.6), the
entries of the covariance matrix C are computed as (i1 , i2 = 1, . . . , Ni )
N
[CRi1 i2 ]rs
o
f
X
H
E H
s1
|Wo (f )|2 covar Fi2 (f ), Fi1 (f ) r1
oi1 oi2 rs =
f
f
f =1
[CS i1 ]rs
o
f
X
H
|Wk (f )|2 covar Xo (f ), Fi1 (f ) r1
s1
E H
oi1 o rs =
f
f
f =1
Nf
[CTo ]rs
H
|Wk (f )|2 var Xo (f ) fr1 s1
E H
o o rs =
f
(6.16)
f =1
where var Xo (f ) are the diagonal

elementsof the (diagonal)
output noise
covari
ance matrix CEX and covar Fi2 (f ), Fi1 (f ) and covar Xo (f ), Fi1 (f ) , respectively the elements of the input noise CEF (f ) and input/ouput noise covariance
matrices CEF EX (f ) (cf. Eq. 6.3). When a discrete-time model in the Z-domain
is used with uniformly distributed frequencies, the entries (6.16) can be computed
in a time-efficient manner by means of the FFT algorithm with the matrices C Ro ,
CSk and CTk having a Toeplitz structure.
As for the I/O WLS estimator (cf. 6.4), the computation time can be significantly reduced by eliminating the numerator coefficients Bo , since only the
denominator coefficients A are required to compute the system poles. In the case
of I/O Fourier data, elimination of the numerator coefficients Bo in (6.15) yields
(o = 1, . . . , No )
1 1
(6.17)
Ro CSo Ro1 So A
Bo = I Ro1 CRo
from which it can be seen that additional terms in (with unknown) appear in
the case that also the input noise sources are taken into account (i.e. C Ro 6= 0
and CSo 6= 0). Notice that for FRF-based GTLS identification this problem does
not occur since errors are only present on the FRF data appearing only in the
last (n + 1) columns/rows of the normal equations (cf. 5.4.2). Consequently,
a straightforward formulation of a compact generalized eigenvalue problem in the
denominator coefficients A is not possible by eliminating Bo from the last n
equations of (6.15), i.e
No
X
o=1
SoH Bo + T A =
No
X
CSHo Bo + CT A
(6.18)
o=1
and as a result, the exact solution of the GTLS identification problem (6.14)
is only given by solving the generalized eigenvalue problem (6.15). However, the
computational load for solving this eigenvalue problem, in the case of typical modal
analysis applications (i.e. No 500 , Ni = 3 , n 50), is not acceptable in
practice.
163
One obvious way to benefit from the elimination approach is by considering

only errors on the measured output sequences (i.e. H1 noise-model assumptions
cf. 2.5.2), with all matrices CRo and CSo equal to zero in (6.15) and (6.17). This
results, after eliminating the numerator coefficients, in the following generalized
eigenvalue problem in A
#
"
No
X
H 1
(6.19)
T
So Ro So A = CT A
o=1
Once the A coefficients are known, Eq. (6.17) (with CRo = CSo = 0) can be used
to derive all numerator coefficients. However, ignoring the input noise sources has
the important disadvantage of loosing the errors-in-variables noise-model characteristics together with the asymptotical consistency of the GTLS approach.
Therefore, another option was considered in order to combine the time-efficient
elimination process with the advantages of EV (GTLS) identification. Using a
1
in Eq. (6.17) can be apTaylor expansion, the inverse matrix I .Ro1 .CRo
proximated as
(6.20)
I Ro1 CRo
= I + Ro1 CRo + (Ro1 CRo )2 + . . .
Introducing this approximation for the first order in (6.17) and substitution in
(6.18) results in, after retaining the first order terms in
Bo
= Ro1 So A Ro1 CRo Ro1 So A + Ro1 CSo A
(6.21)
which now again yields a compact generalized eigenvalue problem in A

No
X
o=1
To SoH Ro1 So A
No
X
o=1
"
CTo +
(6.22)
SoH Ro1 CRo Ro1 So
2herm
N
o
X
o=1
SoH Ro1 CSo
!#
Compared to (6.19), two additional terms in appear after linearizing, containing

the noise characteristics of the input and between the input and output Fourier
data, i.e. CRo and CSo . Notice that, the matrix in the right part of (6.22) is the
covariance matrix of the compact normal matrix
#
"
No
X
H 1
(6.23)
D= T
So R o So
o=1
Indeed, the covariance matrix CD = E{D} where

D =
No
X
o=1
To
No
X
o=1
SoH Ro1 So SoH Ro1 Ro Ro1 So + SoH Ro1 So
(6.24)
164

mode
1
2
3
4
5
6
pole
-1.5234E-3 + 0.7500E+1i
-1.3441E-3 + 1.2167E+1i
-0.7553E-3 + 2.2164E+1i
-1.2464E-3 + 2.3837E+1i
-0.5506E-3 + 2.5336E+1i
-1.1167E-3 + 4.4167E+1i
residue
3.9282E-8 + 2.5015E-3i
1.2874E-7 + 1.4997E-3i
-4.9174E-6 + 1.1966E-3i
1.0541E-5 + 1.3087E-3i
-7.9402E-7 + 1.7064E-3i
1.2672E-9 + 1.9000E-3i
Table 6.1: Poles and residues of theoretical SISO transfer function.
and hence taking the expected value results in

!#
N
"N
No
o
o
X
X
X
1
H 1
H 1
CD =
S o R o C So
So Ro CRo Ro So 2herm
CTo +
o=1
o=1
(6.25)
o=1
since CTo = E{To }, CRo = E{Ro } and CSo = E{So }.

Instead of solving (6.15) directly, this elimination approach results again in an
important reduction of the computation time since the GSVD algorithm can be
applied to much smaller matrices (dimensions reduced by No .Ni ). From equation
(6.20) it follows that the linear approximation will be valid in the case that the
eigenvalues of (Ro1 CRo ) , where Ro1 CRo is a measure for the signal-to-noise
ratio (SNR) of the input Fourier data.
Similar to the approach used in 5.3.5, a time efficient construction of the
stabilization chart is possible again. Since the model equations in (6.12) and
(6.19) are formulated for a given maximum model order, the generalized eigenvalue
problem can be re-solved for all smaller model orders by simply taking sub-matrices
of the full matrices. By doing so, a set of solutions is obtained for an increasing
order of the denominator polynomial, while the order of the numerator polynomial
is kept constant and equal to the maximum specified (polynomial) model order.
6.5.3
Validation of Linear Approximation
The validity of this linear approximation is now studied by means of monte carlo
simulations. A Single Input Single Output system having 6 modes with poles
and residues, given in Table 6.1 and shown in Figure 6.1, was used to generate
an I/O data set of 1200 equally distributed frequencies in the band 0.67.9Hz
i.e. F0 (f ) = 1 , X0 (f ) = H0 (f ) (f = 1, . . . , 1200).
Independent zeromean Gaussian noise was added to both the input and output data. In order to
evaluate the validity of approximation (6.20), the variance of the input noise was
increased over 21 logarithmically distributed steps in the range of [1E-8,1E+0]
while the output noise level remained constant with a variance of 1E-6. For each
I/O noise combination, 1000 disturbed data sets were generated and for each
10
180
160
165
140
10
120
Phase (degrees)
Amplitude (dB)
20
30
100
80
40
60
50
40
60
70
20
4
f (Hz)
4
f (Hz)
Figure 6.1: Theoretical transfer function used for SISO Monte Carlo simulations.
set the modal parameters of 20 modes (poles, residues) were estimated using the
estimators summarized in Table 6.2. The frequency weighting function W o (f )
was equal to 1 for all estimators. The accuracy of the different estimators can be
abbreviation
GTLS
FGTLS
FGTLSX
TLS
FTLS
eqn.
eqn.
eqn.
eqn.
eqn.
description
(6.14)
(6.22)
(6.19) with CR = 0
(6.22) with C = diag(1, 1)
(6.19) with CT = 1
Table 6.2: Different estimators studied during Monte Carlo simulations. (CS = 0 for
all considered cases.)
compared with respect to their squared bias, total variance and mean squared error
(MSE) for both the estimated poles and residues. These quantities are related as
follows
n
o
MSE(P ) = E (P P0 )H (P P0 )
n
o
= E (P E{P })H (P E{P }) + (E{P } P0 )H (E{P } P0 )
=
total variance(P ) + squared bias(P )
(6.26)
with P and P0 the estimated and theoretical values of the poles (residues). The
Monte Carlo simulation results can be compared for the different estimators as
166
10
10
GTLS
FGTLS
FGTLSX
TLS
FTLS
10
10
10
MSE
MSE
10
10
10
10
10
10
10
10
10
noisevar input
10
10
10
10
10
10
10
noisevar input
10
10
noisevar input
10
noisevar input
10
10
10
10
10
10
10
10
10
10
10
10
10
11
11
10
10
10
noisevar input
10
10
10
10
10
10
10
GTLS
FGTLS
FGTLSX
TLS
FTLS
10
GTLS
FGTLS
FGTLSX
TLS
FTLS
10
total variance
10
10
10
10
10
10
10
10
GTLS
FGTLS
FGTLSX
TLS
FTLS
10
squared bias
10
squared bias
10
GTLS
FGTLS
FGTLSX
TLS
FTLS
10
10
10
10
total variance
GTLS
FGTLS
FGTLSX
TLS
FTLS
10
10
10
noisevar input
10
10
10
10
10
Figure 6.2: Monte Carlo simulation results: mean squared error (MSE), squared bias
and total variance for poles (left) and residues (right).
6.6. Validation of (WG)TLS for Experimental Data
167
shown in Figure 6.2. The important errors in case of both TLS estimators are due
to the high bias errors, originating from the inconsistency properties of the TLS
approach, since no noise covariance information is taken into account. Considering
only output noise results in the FGTLSX, improving the accuracy as long as the
input variance is small. However, once the effect of the input noise outweighs the
output noise, the errors for the FGTLSX increase significantly. Since input noise
is disregarded, the FGTLSX is still inconsistent, explaining the increase in bias
and the agreement with the TLS results for high input noise variances. The gain
in accuracy by taking input noise into account is clearly illustrated by comparing
the FGTLSX and the GTLS where the MSE is 5dB to 10dB smaller for the GTLS.
The good agreement of the errors for the GTLS and FGTLS indicates the validity
of the approximation (6.20). Only for very high variances of the input noise (SNR
around 0dB), the errors are slightly higher for the FGTLS. This result proves
that no loss in accuracy is encountered when using the approximation (within
practical noise levels), making the fast implementation of the fast algorithm of the
(W)GTLS suitable for an accurate analysis of large modal data sets as shown in
the next section.
6.6
Validation of (WG)TLS for Experimental Data
Consider again the experiments performed on the slat track presented in Chapter 5,
using the scanning Laser Doppler Vibrometer setup as shown in Figure 5.2. Notice
that this setup completely satisfies the errors-in-variables noise model defined in
section 6.3 since no correlations exist between the output signals. Since a periodic
chirp excitation was used, the noise information could be obtained by computing
the sample variances of the measured signals during 5 experiments, while leakage
was avoided.
Using the fast implementations of the I/O TLS and GTLS estimator, a SMFD
model with 22 modes was estimated in a frequency band of 2501500Hz containing
1000 spectral lines. No frequency weighting was used in this case. Figure 6.3 shows
the ratio of the FFT of a measured velocity and the corresponding force (dots) as
well as the estimated model (solid line). The FGTLS has a better performance,
explained by the use of noise covariance matrix information during the estimation
process, making the FGTLS consistent and less sensitive to measurement noise.
Comparing the stabilization diagram for both estimators, shown in Figure 6.4,
it is clear that the use of a proper noise (EV) model results in an improved stabilization behavior of the physical modes. In addition, the computational poles are
estimated as unstable in the case of the FGTLS, improving the user-friendliness
of the chart, since in general, only the stable poles are plotted.
Another important part of the modal model is the spatial information present
in the estimated mode shapes. Figure 6.5 shows the mode shape for the mode
168
Ampl.(dB)
50
Phase (degrees)
50
400
600
800
1000
1200
1400
200
100
f (Hz)
100
200
400
600
800
1000
Freq. (Hz)
1200
1400
400
600
800
1000
1200
1400
800
1000
Freq. (Hz)
1200
1400
Ampl. (dB)
50
50
Phase (degrees)
200
100
f (Hz)
100
200
400
600
Figure 6.3: Parametric results for the slat track using the FTLS (top) and FGTLS
(bottom). Amplitude and phase of the measurements (dots) and the estimated transfer
function model (solid line).
6.6. Validation of (WG)TLS for Experimental Data
169
22
20
18
Number of modes
16
14
12
10
8
6
4
2
0
400
600
800
1000
f (Hz)
1200
1400
400
600
800
1000
f (Hz)
1200
1400
22
20
18
Number of modes
16
14
12
10
8
6
4
2
0
Figure 6.4: Stabilization chart with the stable (+) and unstable () poles (i.e. positive
real part) for the FTLS (top) and FGTLS (bottom) method.
170
Figure 6.5: Mode shape for mode of 1234Hz, obtained by FTLS (left) and FGTLS
(right).
at 1234Hz. Again the results obtained by the FGTLS are better since an accurate noise model is taken into account. The mode shape for the FTLS is clearly
corrupted by estimation errors due to the measurement noise.
13
12
10
10
12
10
11
10
11
10
10
10
Flops
Flops
10
10
10
10
10
10
10
10
20
40
60
Outputs
80
100
10
20
40
60
80
model order
100
Figure 6.6: Flops count for GTLS (dashed line) and FGTLS (solid line) for varying
number of outputs and Nm = 20 (left) and varying model order and No = 10 (right).
6.7. Maximum Likelihood Estimator
171
Finally, the numerical efficiency is illustrated in Figure 6.6 by comparing the

GTLS and FGTLS algorithms in terms of (Matlab) flops count for a varying
number of outputs No (with n=20, Ni =1) and a varying model order n (with
No =10, Ni =1). Considering the output dimension No , the results indicate that
the GTLS problem is solved in O(No3 ) flops, whereas O(No ) flops for the FGTLS,
resulting in a gain of O(No )2 flops, explained by the elimination approach. On the
other hand, for an increasing model order, the number of flops used by the GTLS
and FGTLS is similar, i.e. O(n3 ).
6.7
Maximum Likelihood Estimator
Based on Eq.(6.4), the following equation error can be defined

o (f , ) = bBo (f , )c {F (f )} A(f , )Xo (f )
(6.27)
The maximum likelihood (ML) cost function based on the weighted equation error
1/2
so (f , ), i.e. o (f , )var o (f , )
, is given by
`ML ()
Nf
No X
X
o=1 f =1
|so (f , )s
o (f , )|
Nf
No X
2
X
|bBo (f , )c {Fo (f )} A(f , )Xo (f )|
var o (f , )
o=1 f =1
(6.28)
where the variance of the (scalar) equation error (6.27)
var o (f , )
bBo (f , )c CF (f ) bBo (f , )cH + |A(f , )|2 var Xo (f )
(6.29)
2herm bBo (f , )c CF Xo (f )A(f , )
with CF (f ) = CEF (f ) and CF Xo (f ) = CEF EX [:,o] (f ) (cf. Eq. 6.3). The ML

T
T T
T
] are obtained
, A
estimates of the polynomial coefficients ML = [B
, . . . , B
1
No
by minimizing (6.28) with respect to these parameters .
This can be done by means of a Gauss-Newton optimization algorithm, which
takes advantage of the quadratic form of the cost function (6.28). The GaussNewton iterations are now given by
1. solve the normal equations
H
H
Jm
Jm m = Jm
em
for
(6.30)
2. compute an update of the previous solution m

m+1 = m + m
(6.31)
172
with em = e(m ), Jm = e()/ |m and e() = [s1 , . . . , sNo ]T .

However, in the case that complex-valued coefficients are used, the equation
error so (f , ) is not analytical in the (complex) parameters and hence, the
entries of the Jacobian matrix Jm can only be computed by considering the real
and imaginary part of the coefficients as
m = [Re(B1 )T , Im(B1 )T , . . . , Re(A )T , Im(A )T ]T
JRem
JImm
e(Re,Im )/Re |Rem ,Imm
e(Re,Im )/Im |Rem ,Imm
(6.32)
(6.33)
The entries (6.33) of the Jacobian matrix Jm , for the weighted equation error
at spectral line f , are found to be given as
bJRem ,Imm co =
1
o
o
var(o )
(6.34)
3
var(o ) Re,Im Rem ,Imm 2 var(o ) 2 Re,Im Rem ,Imm
1
1
2
where o (, Rem ,Imm ) is denoted as o .
Since the correlations between the outputs are ommited and a SMFD model
is used, the Jacobian matrix Jm has a similar block structure as the Jacobian
matrix of the Least Squares formulation (6.5). As a result, it is again possible to
H
H
em ) in a time-efficient way using
Jm and Jm
form the normal equations (i.e., Jm
the Toeplitz structure and FFT algorithm in the case that a discrete-time model
is used. Furthermore, given the block structure of Jm , the first No blocks can be
eliminated from the normal equations analogously to the MLE algorithm presented
in 5.5.
The practical implementation of the ML algorithm for MPE from I/O data has
yet been done only for the case that the variances of the outputs are taken into
account (i.e. CF = 0 and CF Xo = 0 in Eq. 6.29). While the cost function (6.28)
is uniquely defined, this is not the case for the equation error so (f , ) and hence
the equation error o (f , ) can as well be formulated as
j
k
o (f , ) {F (f )} Xo (f )
o (f , ) = H
j
k
o (f , ) =
with H
bBo (f ,)c
A(f ,)
(6.35)
and the following expression for the variance
var o (f , ) = var Xo (f )
(6.36)
1/2
Weighting the equation error (6.35) by var o (f , )
results in an expression
that is analytical in the complex polynomial coefficients . The submatrices o
6.8. MPE in the Presence of Leakage Phenomena
173
and o of the Jacobian matrix Jm , have the same structure as the Jacobian matrix
in (6.5)
n
0
(1 )Fi (1 )
(1 )Fi (1 )
var(Xo (1 ))A(1 ,)
var(Xo (1 ))A(1 ,)
..
..
(6.37)
o =
.
0
n
(Nf )Fi (Nf )

(Nf )Fi (Nf )

var(Xo (Nf ))A(Nf ,)
var(Xo (Nf ))A(Nf ,)
and
o =
(1 )bBo (1 ,)c{F (1 )}
2
var(Xo (1 ))|A(1 ,)|
..
.
0 (Nf )bBo (Nf ,)c{F (Nf )}
var(X
o (Nf
))|A(Nf
,)|2
(1 )bBo (1 ,)c{F (1 2)}

var(Xo (1 ))|A(1 ,)|
..
.
n (Nf )bBo (Nf ,)c{F (Nf )}

var(X
( ))|A( ,)|2
o
Nf
Nf
(6.38)
The uncertainy bounds on the estimated parameters ML can be derived using the
same approach as discussed in 5.5.2.
6.8
MPE in the Presence of Leakage Phenomena
A typical problem for frequency-domain estimators in general arises from the presence of transient phenomena in the data, of which spectral leakage is an important
example. Until recently, frequency-domain system identification assumed that the
input and output signals are periodic or time-limited within the observation window as in the case of e.g. an impulse or burst random excitation. In [111] the
applicability of frequency-domain system identification techniques has been generalized to arbitrary signals and these results can be applied for modal parameter
estimation from I/O data.
When the modal parameters are derived directly from the I/O data there is in
general a difference between time and frequency-domain methods. This difference
is mainly due to the manner how the initial conditions of the system are taken
into account. This can be understood by considering the Laplace transform of the
j-th derivative of the force time signal f (t)
j
j1
r
X
d f (t)
j
j1r d f (t)
=
s
F
(s)
s
(6.39)
L
dtj
dtr
r=0
t=0
where the sum is a polynomial in the Laplace variable of order j 1 which only
exists if the initial conditions differ from zero. Hence, when taking the initial
conditions into account, the Laplace transformed SISO input-output differential
equation is given as
A(s)X(s) = B(s)F (s) + T (s)
(6.40)
174
where the order of T (s) equals the maximum order of B(s) and A(s) minus 1 with
unknown coefficients T = [c0 , c1 , . . . , cn1 ]T .
From this it follows that a transient polynomial T (f , ) can be used to model
the initial conditions (e.g. the effects of leakage) present in the I/O data. For
example in the case of the ML estimator presented in 6.7, adding the polynomial
T (f , ) to the equation error o results in the following generalized cost function
with
j
2
k
Nf B
No X
o (f , ) {F (f )} A(f , )Xo (f )
X
`ML () =
var o (f , )
o=1 f =1
j
{F (f )}
o (f , )
B
(6.41)
bBo (f , ), T (f , )c
(
)
{F (f )}
1
The (unknown) coefficients of T (f , ) are included in and are estimated simultaneously with the coefficients Bo (f , ) and A(f , ). This result indicates
that taking a transient polynomial into account is equivalent with having one
additional input with a constant Fourier spectrum equal to 1, and as a result,
the same I/O algorithms can still be applied. This result shows that parametric
frequency-domain estimators can be made robust for leakage by estimating the
initial (and final) conditions of the system together with the system parameters.
The use of a transient polynomial is now illustrated for the case of the slattrack
structure. Using a similar setup as discussed in 5.3.3, the time histories of the
applied force and response (acceleration) were now measured using an impedance
head. In this case the structure was excited using both a multisine and a random
noise excitation in the frequency band of 0-4000Hz with 16384 frequency lines.
Using the time records obtained by the multisine, containing 32768 samples,
the empirical transfer function estimate (ETFE) was computed. Based on this
ETFE, a SMFD discrete-time model was estimated using the FRF ML estimator
(discussed in 5.5). Since no leakage effects appear when using multisine excitation, this model will be used as a reference for the comparison of the different
possible approaches for analyzing the random noise time histories.
As a first case, the complete random noise time records (Ns = 32768 = 32K)
were considered. Figure 6.7 (top) compares the model obtained by the I/O ML
estimator using the FFT of the time records, when using a transient polynomial
or not. The dashed-dotted line corresponds to the reference model obtained from
the multisine data. The ETFE for the random noise data (dots) is a very noise
function as can be expected. From this data, the I/O ML estimator derived the
model without (dashed line) and with using a transient polynomial (solid line).
6.9. MPE from Auto and Cross Power Spectra
175
A model with 30 modes was estimated for a frequency band of 200-1500Hz. By

zooming in, the effects of leakage can already be noticed in the case that no
transient polynomial is used, as for instance for the mode at 1156Hz. In this case
the zoomed frequency band has 921 spectral lines (Nf = 921).
Decreasing the size of the records clearly introduces more leakage effects in the
case of random noise data, as is shown in Figure 6.7 for Ns = 16K (Nf = 461)
(middle) and Ns = 8K (Nf = 231) (bottom) and Figure 6.8 for the cases Ns = 4K
(Nf = 116 in zoom) (top), Ns = 2K (Nf = 59) (middle) and Ns = 1K (Nf = 30)
(bottom). As can be seen for this experimental case, the approach using the
I/O ML with transient polynomial is very robust for short data records. Even
in the case of 1K time samples corresponding to a frequency resolution of 8Hz,
this approach can still estimate a fairly accurate model compared to the approach
without using this transient polynomial.
The I/O ML approach can also be compared with the FRF-based ML. The
FRF data can be obtained by either computing a H1 estimate, which however
requires the averaging of a number of subrecords, or by computing the ETFE also
using a Hanning window. Since, in the general case of short data records, the first
possibility is not applicable, a comparison was done using the ETFE. Figure 6.9,
shows a zoom on the results found when using the FRF ML for the same analysis
parameters as for the I/O estimation for a decreasing size of the records. As can be
seen, the estimated model is acceptable in the case of the 32K samples, however,
due to imporant leakage errors the results detoriate fast for a decreasing blocksize
when compared to the I/O ML approach that also estimates the transients. The
use of a Hanning window is not sufficient once the records become short, in this
case Ns = 2K or less.
6.9
MPE from Auto and Cross Power Spectra
In the case that sufficient data is measured to perform an average based nonparametric processing, the presented I/O estimators can also be used for MPE
from the Auto and Cross Power Spectra functions. Indeed, assuming that the
input F [m] (f ) and output X [m] (f ) Fourier sequences, as defined in 6.3, are
observed during M records (m = 1, . . . , M ), the Auto and Cross Power Spectra
can be estimated. Using for instance the Periodogram estimator (cf. 3.3), the
Power Spectra matrix is obtained as
M
X
Z (f ) = 1
Z [m] (f )Z [m]H (f )
G
M m=1
(6.42)
H
where Z [m] (f ) = F [m] (f )H , X [m] (f )H
C (Ni +No )1 is the Fourier data
observed during the mth measurement.
176

70
60
50
Ampl. (dB)
40
30
20
10
0
10
1000
1050
1100
Freq (Hz)
1150
1200
1050
1100
Freq (Hz)
1150
1200
1050
1100
Freq (Hz)
1150
1200
70
60
50
Ampl. (dB)
40
30
20
10
0
10
1000
70
60
Ampl. (dB)
50
40
30
20
10
0
10
1000
Figure 6.7: Comparison of I/O ML results without (dashed) and with (solid) transient polynomial with Reference model (dashed-dotted) and the ETFE for random noise
Fourier data (dots). Cases: Ns = 32K (Nf = 921 in zoom) (top), Ns = 16K (Nf = 461)
(middle) and Ns = 8K (Nf = 231) (bottom).
6.9. MPE from Auto and Cross Power Spectra
177
70
60
Ampl. (dB)
50
40
30
20
10
0
10
1000
1050
1100
Freq (Hz)
1150
1200
70
60
Ampl. (dB)
50
40
30
20
10
0
10
1000
1050
1100
Freq (Hz)
1050
1100
Freq (Hz)
1150
1200
70
60
Ampl. (dB)
50
40
30
20
10
0
10
1000
1150
1200
Figure 6.8: Comparison of I/O ML results without (dashed) and with (solid) transient polynomial with Reference model (dashed-dotted) and the ETFE for random noise
Fourier data (dots). Cases: Ns = 4K (Nf = 116 in zoom) (top), Ns = 2K (Nf = 59)
(middle) and Ns = 1K (Nf = 30) (bottom).
178
70
60
Ampl. (dB)
50
40
30
20
10
0
10
1000
1050
1100
Freq (Hz)
1150
1200
Figure 6.9: Comparison of FRF (ETFE with Hanning window) ML results with Reference model (dashed-dotted) for the cases: Ns = 32K (Nf = 921 in zoom) (solid),
Ns = 8K (Nf = 231) (dashed) and Ns = 2K (Nf = 59) (dotted).
[m]
Referring to Eq. (6.9), the normal equations J H J
can be computed for
[m]
each set of the measured records Z(f ) , which can be averaged over the number
of experiments M . The entries of the submatrices of the obtained normal equations
are then given as
[Roi1 ,i2 ]rs
[To ]rs
[Soi1 ]rs
Nf
X
!
M
H
1 X [m]
[m]
=
|Wk (f )|
Fi1 (f )Fi2 (f ) fr1 s1
f
M m=1
f =1
!
Nf
M
X
H
1 X [m]
[m]
2
s1
X (f )Xo (f ) r1
|Wk (f )|
=
f
f
M m=1 o
f =1
!
Nf
M
X
H
1 X [m]
2
[m]
=
|Wk (f )|
Fi1 (f )Xo (f ) r1
s1
f
f
M m=1
2
f =1
where the expressions between the parentheses represent the Auto and Cross Power
Spectra as defined in (6.42). This shows that the generalization of the I/O estimators, as developed in this chapter, is straightforward. The use of the Power
Spectra has the advantage, compared to using the raw I/O data, that the noise
on the data is reduced by means of the averaging process. This, however, requires
that sufficient data samples could be measured. On the other hand, compared to
using FRF data, the effects of leakage can still be compensated for as shown in
6.8.
6.10. Conclusions
6.10
179
Conclusions
The use of an I/O estimator is preferred in the case of short data sequences, since
then the nonparametric processing for FRF estimation is not possible. This can
be the case for flight flutter testing and structural health monitoring applications.
The formulation of the I/O estimators is based on an errors-in-variables set-up,
where both the noise on the input and output signals are considered. Similar to
the FRF-based estimators, where possible correlations between the FRFs are neglected, the fast implementation of the I/O estimators omits possible correlations
among the noise on the output measurements, balancing between computational
efficiency and accuracy. Moreover, in the case of a SLDV measurement setup, this
simplified noise model is still exact.
Based on the results of Chapter 5, a fast implementation of the frequencydomain LS method has been derived for the case of I/O data, where both the
Jacobian and Normal equations-based formulations can be used, while again a
fast construction of a stabilization diagram is possible. It has been shown that
in the case of the WGTLS estimator, a reduction to a compact formulation was
not straightforward in the case that input noise is considered. Nevertheless, the
input noise can be taken into account at the price of using a linear approximation
for which, however, the validity has been demonstrated for a wide range of input
signal-to-noise ratios by means of Monte Carlo simulations. Furthermore, neglecting the output correlations only results in a loss of asymptotical efficiency while
the WGTLS remains consistent. Next, it has been shown that a fast implementation of the I/O ML estimator is possible for the case that also the input noise is
considered.
Furthermore, by estimating simultaneously the initial conditions and the system model parameters, it is possible to deal with arbitrary signals in the frequencydomain without any approximation and under the same assumptions as in the
time-domain. Using this approach, the proposed I/O estimators can be made robust for errors due to leakage effects, as is demonstrated for experimental data
obtained from a safety critical component of an airplane.
The fast I/O estimators could also be generalized to estimate the parameters
starting from auto and cross power spectra, which is preferable when longer time
records are available. Compared to using the raw I/O data, the noise on the data
is reduced by means of the averaging process. On the other hand, compared to
using FRF data, the effects of leakage can now be compensated for during the
parametric identification.
180
Chapter 7
Automated Modal
Parameter Estimation and
Tracking
In this chapter it is demonstrated that, by using a frequency-domain Maximum
Likelihood estimator, features such as high accuracy and confidence bounds for the
estimated parameters and robustness for high measurement noise levels can significantly contribute to the automation of the modal identification process. For the
validation of the model, criteria based on a statistical approach were developed in
addition to well-known criteria such as modal phase colinearity and mode complexity. Two different approaches for the final mode selection are based on a weighted
validation criterium or a fuzzy clustering algorithm. In addition, an accurate
mode-tracking algorithm is presented. It is shown that the problem of coinciding
poles, due to shifting or crossing modes under changing structural dynamics, can
hamper the mode tracking and a robust solution is proposed. The possibilities and
limitations of the proposed automated modal parameter estimation approaches are
investigated and tested for two cases. The first application concerns the automated
analysis of data obtained from a slat track of an Airbus A320 commercial airplane. Serving as a preliminary study for future monitoring practices during life
cycle tests on slat tracks, the method for mode tracking is tested on this same data.
The second application focusses on the automated processing of data from a bridge.
181
182
Chapter 7. Automated Modal Parameter Estimation and Tracking
7.1
Overview of State-of-the-Art
During the last few years a lot of effort has been put in the development of vibration based damage detection techniques [34]. Since the dynamic behaviour
of structures is influenced by damage, one is able to detect the occurrence and
evolution of damage by monitoring the modal parameters (such as resonance frequencies) [61, 132]. However, varying environmental and operational conditions
can seriously complicate in situ structural health monitoring affecting the modal
parameters as well [3, 100]. Nevertheless, an important element for the applicability of damage detection techniques as a part of monitoring practices is an
automated identification and tracking procedure since traditional modal identification still requires extensive interaction from an experienced user. In the case of
modal analysis, the model order is typically over-specified in order to capture all
physical modes present in the frequency range of interest. However, this requires
the distinguishing between the physical and mathematical modes, which is more
involved than merely determining a proper model order.
Most of the classical model order selection tools (Akaike, F-test, MDL,...) that
are used in time-domain identification were developed in the framework of control
theory [2, 129, 115]. The shortcomings of these methods have been extensively
studied in [78]. They only allow to verify if the model order used is appropriate
or not (detection of under-modelling or over-modelling).
Order selection techniques, used for frequency-domain identification, have been
proposed in the literature, based on a statistical analysis of the residual errors
(such as a whiteness test) [121]. This approach can also detect the presence of
non-linear distortions. A more recent model order selection approach is based
on a statistical analysis of the global minimum of a weighted non-linear leastsquares cost function [118]. Another possible order selection criterion consists in
determining the rank of the Jacobian matrix or the sampled covariance matrix
[118, 29]. These techniques can be used to verify if the model order is appropriate
or not, or to see if there are non-linear distortions present in the measurements,
but they do not allow to separate physical from computational modes. Moreover,
the model orders considered in modal analysis are usually much higher and also
the amount of data to be processed is much larger than for control applications
(typically systems with a very limited number of inputs and outputs).
Today, in modal analysis, a stabilization diagram is still an important tool to
assist the user in separating physical from mathematical modes. Since this is done
manually, selecting the physical poles can be very difficult and time-consuming
depending on the type and quality of the measurement data, the performance of
the estimator and the experience of the user. Moreover, this manual approach
is certainly not suitable for monitoring purposes.
Recently, several approaches for automated modal identification for diagnostic
7.2. Autonomous Modal Parameter Identification
183
purposes were proposed, most of them based on the use of Neural Networks [83],
Fuzzy Logic [150] or Genetic Algorithms [25]. Unfortunately the computational
load for these methods can be quite high. Classical time-domain identification
algorithms (e.g. ERA/DC) were used, not taking any information about the measurement noise into account. The main application focus of the research conducted
by the NASA [96, 97, 150] and Zimmerman and James [25, 69] is to reduce the
user-time to perform a detailed ground vibration test/analysis of the Space Shuttle
Orbiter after each flight. Essentially, they follow a rule-based approach to build
up a system model pole table by consecutively adding or updating 1 pole. The
followed approaches are very heuristic, and as a result dedicated for their specific
application.
In [146, 144] automated modal identification approaches are presented based
on the use of a frequency-domain Maximum Likelihood estimator (MLE) and
stochastic validation criteria, yielding promising results. In particular the availability of accurate parameter estimates with their uncertainty information contributes greatly to a more simplified validation and selection of the physical poles.
This statistical information is furthermore very important to make a distinction
between changes in the modal parameter caused by damage and changes caused
by measurement noise. While, the computational load of the MLE algorithm has
been significantly reduced such that one iteration of the algorithm is as fast as
the well-known LSCE algorithm (cf. Chapter 5), the use of this estimator is still
currently not applicable for the on-line identification of fast time-varying systems. Examples are a rocket consuming fuel or dangerous and expensive flight
flutter tests, requiring fast online data processing techniques for which some possible approaches are discussed in [28, 19, 63]. On the other hand, numerous other
vibration-based condition monitoring applications are conducted on a totally different time scale, resulting in sufficient time for online data analysis using more
accurate algorithms such as the MLE. Examples are the monitoring of in-operation
large structures such as bridges [100, 34] or offshore platforms [22, 34], or ground
inspections of airplanes and the space shuttle orbiter [64, 34]. Another application
within the same category is the condition monitoring of structures during life cycle
tests, which will be studied in this chapter as well.
7.2
Autonomous Modal Parameter Identification
The approach used to automate the modal parameter identification process consists of 3 main steps, discussed in detail in the following paragraphs:
1. Data Analysis: Identification of a model with sufficiently high order using a
frequency-domain Maximum Likelihood estimator
2. Mode Validation: Computation of the mode validation criteria
184
3. Physical Mode Selection: Based on these criteria, the structural (physical)

modes can be pinpointed by means of a mode quality index or by classification
using a fuzzy clustering approach
7.2.1
Data Analysis
The FRF-based frequency-domain Maximum Likelihood estimator (MLE), as presented in 5.5, is used as data analysis engine. In this case, the MPE algorithm
starts from the LSCF estimators to obtain starting values. The MLE algorithm
minimizes the log-likelihood cost function based on the FRF and corresponding
variance data. Important algorithm parameters are the model order, the number
of iterations and the frequency band of interest. A maximum of 50 iterations is
chosen since, for structural dynamics analysis, this typically leads to convergence.
For each iteration, the modal parameters and poles and zeros are calculated.
In order to capture all physical modes in the model, a model with a high order
is estimated. Due to this over-modelling, a large number of computational modes
occur. A first straightforward elimination of such modes is done by removing all
modes having poles with a frequency within 1% of the edges of the analysis band
or a damping ratio smaller than 0% (unstable pole) or larger than say 10% (application dependent). The frequency cutoff is applied to eliminate the computational
poles on the edges of the analysis bandwidth that result from the fact that the
MLE uses a discrete-time model. This first mode elimination step is called the
physical thresholding step.
7.2.2
Mode Validation Criteria
Based on the knowledge of the uncertainty on the measured data, it is possible,

besides the classical deterministic validation criteria, to define stochastic criteria
that give more information regarding the physicalness of a mode.
Stochastic Mode Validation Criteria

Given the scalar matrix fraction description used as model for the MLE implementation (cf. 5.5), the polynomial coefficients are related to a poles/zeros model and
as discussed in [130, 118] for SISO systems, the computational or mathematical
modes are characterized by so-called cancelling pole-zero pairs. This is explained
by the fact that the estimator tries to minimize the contribution of a computational pole in the model by placing a cancelling zero very close to that pole. Such
pole-zero pairs will occur for two possible reasons:
185
Computational mode: a computational mode should not affect the transfer

function and as a result the pole should be located far from the imaginary
axis (high damping) or close to a (cancelling) zero.
Poorly excited/measured physical mode: when a mode is poorly excited or
measured, e.g. in a nodal point, it has only a small contribution to the
transfer function, and hence the physical pole will be cancelled by a zero.
Based on this property and given the uncertainty information for the poles and
zero, obtained from the MLE algorithm, the following stochastic mode validation
criteria can be defined:
Pole-zero pairs: a large number of zeros within a uncertainty circle around
pole pr indicates a computational mode. The radius of the circle is calculated
as
p
(7.1)
R = 2log(1 pb)pr
where pr = std(pr ) the uncertainty on pole pr , for r = 1, . . . , Nm the

number of modes and with pb the probability for the event that the true
pole falls inside the circle. As discussed in [146], the uncertainty levels of
computational poles are often higher (i.e. typically order 10 to 100) than for
the physical poles.
Pole-zero correlation: the correlation between poles and zeros gives more
insight for the detection of cancelling pole-zero pairs, since these are always
strongly mutually correlated. In the case of a high modal density or a high
uncertainty on the measurement data, it is possible that other zeros fall
within the uncertainty circle and as a result the correlation matrix is then
required to determine which pole and zeros are cancelling pairs. Based on
the uncertainty on the measurement data, the MLE can calculate the covariance matrix of the ML estimates (cf. Eq. 5.63), from which the correlation
coefficient between each pole and zero is derived. In order to limit the computational load, only the nearest zero is used in practice resulting in the
following correlation coefficient for pole pr
corr(pr , z) =
covar(pr , zc )
std(pr ) std(zc )
(7.2)
where zc is the nearest cancelling zero of each estimated transfer function.

However, in the case of weakly excited modes, it is possible that both stochastic
mode validation criteria falsely indicate physical modes as computational. The
autonomous procedure can be made robust for this type of failure by performing
in addition an analysis for each FRF separately (SISO analysis) using again the
MLE or for example the frequency-domain Least Squares (LSCF) estimator (cf.
5.3.6) in order to reduce analysis time. In the case of a weakly excited mode, the
186
pole should be identified for most of the FRFs, while computational poles do vary
for the different FRF analyses. The probability of having poorly excited modes
can be reduced by using a multi-input excitation setup. However, in the case of
single-input excitation, an optimal experimental set-up can tackle the problem of
weakly excited modes for monitoring practices, where only the modes that vary
most during the deterioration of the structure must be tracked over time and
should be well excited. On the other hand, when a robust autonomous estimation
process is required for analyzing data without a priori knowledge of the structure,
it is best to incorporate this check for weakly excited poles at the expense of
additional analysis time.
Deterministic Mode Validation Criteria

Besides the validation approach based on cancelling pole-zero pairs, well-known
validation criteria specifically developed for modal analysis purposes can be considered as well. Most of them give an indication based on the mode shape properties,
such as the Modal Phase Collinearity (MPC), Mean Phase Deviation (MPD) and
the Mode Overcomplexity (MOV)
The Modal Phase Collinearityas defined in [71] is a measure for the degree
of complexity of a mode shape by evaluating the functional linear relationship between the real and imaginary parts of the mode shape coefficients.
For proportionally or lightly damped structures, physical modes behave as
normal (real) modes and the MPC index approaches unity. A mode with
a low index is rather complex, indicating a computational or noisy mode.
Caution is needed in the case of complex physical modes, which are typical
for highly damped systems.
The Mode Overcomplexityas defined in [62], is based on the mode complexity check, which evaluates the sensitivity of the damped natural frequency
for a mass change at each response degree of freedom (DOF). The percentage
of the response DOFs for which a mass addition yields a negative frequency
sensitivity is the MOV for a mode. Physical modes yield a MOV close to
unity. The MOV criterion requires the availability of a driving point measurement, which, in practice, is not always available.
One of the major drawbacks of both criteria is their deterministic character (no use
of uncertainty information) and as a result they can be expected to be less robust
for poor quality data. Especially in the case of medium values (30 70%) they can
indicate a computational or a noisy mode, where the latter can still be a structural
mode. Nevertheless, given the limited computational load, their integration in this
autonomous process can increase the robustness for mode evaluation.
7.2.3
187
Physical Mode Selection
Based on the results of each validation criterium, the modes remaining after the
data analysis step (cf. 7.2.1) have to be grouped in 2 classes, i.e. the physical
and computational modes. Two possible approaches are now proposed.
Fuzzy Clustering Approach

An iterative Fuzzy C-means clustering algorithm [15], proven to be useful for the
automation of the modal mode extraction [144], is used in this case. The algorithm
is implemented in the MATLAB Fuzzy Logic Toolbox [151]. The input variables for
clustering are based on the validation criteria discussed above. An input variable
Xk consists of Nm objects, with Nm the total number of modes (physical and
mathematical) after the data analysis step. The following input variables were
considered:
Variable X1 : standard deviation of the estimated pole pr (r = 1, . . . , Nm ).
Variable X2 : the number of transfer functions having a zero within e.g. the
99% uncertainty circle (pb=99%) of the estimated pole pr divided by the
total number of transfer functions (o = 1, . . . , No number of responses, i =
1, . . . , Ni number of forces)
n
o
# |pr zc | < R
(7.3)
pairs (pr ) =
No Ni
Variable X3 : the number of cases where the correlation between pole pr and
the nearest zero zc in each transfer function model is higher than e.g. 90%
divided by the total number of transfer functions
n
o
# corr(pr , zc ) > 90
corr (pr ) =
(7.4)
No Ni
Variable X4 : the inverse of the modal phase colinearity (MPC).
Variable X5 : the inverse of the RMS magnitude of the mode shape r .
Variable X6 : the number of SISO analyses that could find a pole within
a certain interval around the global pole estimate pr divided by the total
number of transfer functions.
# pr pk pr
(7.5)
siso (pr ) =
No Ni
188

where pk is a pole estimate for each SISO analysis (k = 1, . . . , No Ni ). The
interval can be user-defined or related on the uncertainty of global pole
estimate. Thus, X6 =1 when all No Ni SISO estimations return a pole close
to the global pole.
An interval scaling of the input variables transforms them into the same interval
(e.g. [0;1]). In a first step of the interval scaling procedure, the variables with a
large range (i.e. variable X1 ) are transformed using a logarithmic scale. Next, each
variable is subtracted by its minimum and divided by its range (i.e. the maximum
minus the minimum). Finally, all variables, except X6 , are subtracted from unity.
Values larger than 0.5 indicate physical modes. The output of the algorithm is a
membership function for the 2 classes, i.e. physical and computational modes. If
the membership function for an object with respect to a class is larger than 0.5
(or 50%) the object belongs to that class.
Weighted Mode Quality Index

Another straightforward method to synthesize the results of the mode validation criteria is to compute a (weighted) index, defined as the Mode Quality Index (MQI), which is a measure for the physical behavior of a mode. Using the
variables X1 X6 , scaled between [0;1], this MQI can be defined as a weighted
ratio
M QIr
std (1 X1r ) + pairs (1 X2r ) + corr (1 X3r )

+mpc (1 X4r ) + rms (1 X5r ) + siso X6r
(7.6)
P
with
i = 1. A weighted index has the advantage that, since the stochastic
criteria (pole-zero pairs and correlation) are more robust for noisy measurement
data, they can have a larger contribution to the overall MQI than the deterministic
criteria (MPC and MOV). A high MQI, typically above 70%, indicates a structural
mode, while lower values are found for computational. The final set of modes is
selected from the results of the last iteration, by identifying those modes having a
Mode Quality Index above or equal to the threshold of physicalness with a typical
value of 5070%. The mode quality index is certainly appropriate for automated
monitoring practices were the proper weights are chosen during the setup of the
monitoring system.
7.3. Case Study I: Airbus A320 Slat Track
7.3
7.3.1
189
Case Study I: Airbus A320 Slat Track

Test Structure and Experiments
The autonomous approach is illustrated for the application of structural health

monitoring of a slat track, which was already discussed in 5.3.3. Safety critical
components such as slat tracks are rigorously tested by means of rig endurance
tests to prove their ability to withstand all safety regulations. It is commonly
accepted that the track should outlive the plane by five times, i.e. 240000 flights.
Until today, commonly used damage detection techniques in the aircraft industry
are mostly visual and/or local (ultrasound techniques, analysis of magnetic fields,
radiology, thermal methods). All these techniques require that the sensors be close
to the damaged area of the investigated part. Vibration-based damage detection
however allows a more global analysis of the structure with the potential of being
done in situ. Besides the development of shortened endurance tests without losing
their ability to represent actual in flight loads, the availability of vibration-based
condition monitoring system can result in an improved damage assessment of the
crack initiation and propagation in the slat track throughout the life cycle tests.
In this section, the results of a preliminary study of an approach for autonomous
modal parameter identification and tracking, applied to the slat track monitoring
problem, are presented. Structural changes were introduced in a controlled manner and forced-vibration tests resulted in frequency response function data. This
allowed simulating the effect of crossing modes, which can significantly hamper
the mode tracking process as well as the damage assessment, which is discussed
in [147, 99]. Nevertheless, it is important to notice that the approach presented
can be applied to a large number of other applications, including in-operation
condition monitoring of bridges, and (off-line) flight flutter analysis.
Given the scope of the research presented, i.e. a feasibility study of an automated identification and tracking process, the measurements were performed
under controlled free-free conditions using a shaker and accelerometers. As shown
in Figure 7.1, seven accelerometers were attached to the slat track on the boarder
of the inner-side of the surface. This part of the slat track was also measured by
means of a scanning laser Doppler vibrometer in 500 points, serving as a reference model for optimal sensor location and a better understanding of the dynamic
behaviour.
A broadband periodic excitation was applied by means of a multisine signal
[57] in the frequency band 04096Hz with 16384 spectral lines. The frequency
response and coherence functions were calculated using the H1 FRF estimator
with 5 averages. Figure 7.2 shows some of the seven FRFs indicating a high modal
density, while Figure 7.3 shows 9 of the mode shapes obtained from data measured
by the scanning laser vibrometer. In order to simulate varying structural modes
in a controlled manner, an increasing mass was added (by means of magnets) in
190
SLAT
TRACK
S1
FORCE
SENSOR
S2
STINGER
S3
S4
S5
SHAKER
S6
S7
Figure 7.1: Measurements using accelerometers. Excitation with a shaker and force
measurement by means of a force sensor. Response measurements by means of 7 accelerometers.
60
40
Ampl. (m/s2/N dB)
20
20
40
60
400
600
800
1000
Freq. (Hz)
1200
1400
Figure 7.2: FRFs measured between the force transducer and the accelerometers 1
(dots), 2 (solid) and 5 (dashed).
between accelerometers 2 and 3, during 14 steps with an increment of 22.5 10-3kg.

The total mass added was 30.5 10-3kg, i.e. 0.77% of the total weight of the slat
track (3.96kg). This approach was preferred to a difficult controllable initiation
and propagation of cracks by means of saw cuts.
191
578.8Hz
614.7.Hz
622.5Hz
764.3Hz
926.6Hz
936.6Hz
1136.9Hz
1203.0Hz
1310.6Hz
Figure 7.3: Mode shapes obtained by SLDV measurements.
192
7.3.2
Autonomous Estimation Results for Slat Track
Data Analysis
A model order of 50 modes, 50 iterations and an analysis frequency band of 300
1375Hz were used during the ML identification. Analysis of the reference data set
(i.e. no mass added) yields 17 modes after removing the modes with unstable and
highly damped poles from the model (cf. 7.2.1).
Mode Validation Criteria

Based on the stochastic mode validation criteria presented in 7.2.2 the following
results were derived for the slat track:
Pole-zero pairs: Figure 7.4, shows the real and imaginary part of the poles
(+) and zeros (o) as well as the corresponding (dashed) 99% uncertainty
circle in a band 550630Hz (both the axes are scaled by a factor 2). One can
clearly recognize the clustering pattern of zeros around the computational
poles within their uncertainty circles, while no or only a few zeros are present
in the small uncertainty circles around the physical poles. (The horizontal
dashed lines indicate the physical poles.)
Pole-zero correlation: Figure 7.5 gives the correlation coefficient between the
pole and nearest zero in each transfer function for some of the modes in the
band 550630Hz. High correlation coefficients for the modes at 554.3Hz and
564.4Hz indicate a computational character for these modes.
Notice however that, in the case of the weakly excited mode at 554.3Hz, both
stochastic mode validation criteria falsely indicate a physical mode as computational. This is easily understood by the fact that, since a weakly excited mode
barely appears in the measured data (Figure 7.2), a zero in each FRF approximately cancels the physical pole. The autonomous algorithm is made robust by
performing in addition a SISO analysis for each FRF separately using e.g the
LSCF estimator. Figure 7.6 (left) shows the absolute value of the difference between the global pole at 554.3Hz and the closest poles found by each separate
FRF analysis. The dashed line indicates the product of the absolute value of the
global pole and its uncertainty. The same was done for the computational pole at
564.4Hz in Figure 7.6 (right). This result demonstrates that the weakly excited
physical mode can be identified for most of the FRFs
193
630
620
610
imag/2/pi
600
590
580
570
560
550
50
25
0
real/2/pi
25
50
Figure 7.4: Nyquist plot of the real and imaginary part (scaled by a factor 2) of poles
(+) and zeros (o) after physical thresholding. The computational poles are characterized
by a typically large uncertainty circle (dashed line) containing a cluster of zeros. The
horizontal dashed lines indicate the physical poles.
7.3.3
Fuzzy Clustering
The various input variables that can be used for the fuzzy clustering are shown
in Figure 7.7. Considering the threshold of 0.5%, it can be noticed that the
deterministic criteria (X4 and X5 ) have a significant higher error ratio than the
stochastic criteria. Indeed, variable X4 indicates all 4 computational poles as
physical, while 2 physical poles are below 0.5%. Variable X5 indicates the 2
computational poles as physical. On the other hand, the stochastic criteria X 1
and X2 only indicate the pole of the weakly excited mode falsely as computational,
while X3 and X6 give no erroneous indications.
In order to study the influence of the choice of input variables on the classification result, Figure 7.8 shows the results for 4 possible combinations of variables.
Considering the threshold of 0.5, where all physical poles should be above in order
to be correctly classified, where the best result is obtained by using the variables
X2 , X3 and X6 (at least in the case of the slattrack data). Figure 7.8, also demon-
194

554.3 Hz
Correlation coeff.
0.8
Correlation coeff.
0.8
0.6
0.6
0.4
0.4
0.2
0
0.2
2 3 4 5 6 7
Transfer function
578.8 Hz
0.5
2 3 4 5 6 7
Transfer function
2 3 4 5 6 7
Transfer function
614.7 Hz
2 3 4 5 6 7
Transfer function
1
Correlation coeff.
Correlation coeff.
0.5
564.4 Hz
0.8
0.6
0.4
0.2
0
Figure 7.5: Correlation coefficients between the pole and nearest zero in each transfer
function indicating computational modes at 554.3Hz and 564.4Hz.
strates that considering all the possible input variables does not necessarily yield
a better clustering result. This is explained by the use of the deterministic criteria (i.e. (MPC)1 and (rms(|r |)1 ), which have a higher error ratio than the
other (stochastic) criteria. It is clear that the variable X6 makes the automated
procedure more robust when weakly excited modes are present. In the analyzed
frequency band, 13 poles were correctly classified as physical. The natural frequency and damping ratio of the classified poles are summarized in Table 7.1 for
the case of the reference model.
It is important to notice that for structural health monitoring applications only
a few or even a single sensor will be used. As a result, care should be taken for
some of the proposed input variables. In the case of a single sensor, the modal
phase colinearity has no physical meaning, while the phase difference between the
force and response (acceleration) should still be close to 90 degrees for physical
modes. Similarly, the variable X2 should be used with care when the sensor is in
a node of a mode while variable X6 is not applicable anymore. For this reason, it
is important to first select the modes to be tracked and then choose an optimal
sensor position, i.e. not in any node of the selected set of modes. The other
variables remain applicable without any constraints.

554.3Hz
60
50
40
30
20
10
564.4Hz
70
|(p)|
|(p)|
195
3 4 5 6
FRF number
3 4 5 6
FRF number
Figure 7.6: The absolute value of the difference between the global pole and the closest
poles found by each separate FRF analysis. The dashed line indicates the product of the
absolute value of the global pole and its uncertainty.
mode
1
2
3
4
5
6
7
8
9
10
11
12
13
fMLE (Hz)
362.24
554.19
578.84
603.91
614.66
622.48
667.89
764.27
926.63
936.56
1136.88
1203.02
1310.62
MLE (%)
0.103
0.439
0.657
1.032
0.611
0.192
0.109
0.174
0.218
0.127
0.177
0.099
0.186
fLSCE (Hz)
362.17
554.08
578.77
603.90
614.91
623.81
668.03
764.72
926.56
936.55
1136.47
1203.79
1311.74
LSCE (%)
0.102
0.428
0.567
0.696
0.662
0.241
0.093
0.174
0.213
0.125
0.172
0.098
0.191
Table 7.1: Autonomous identification results for the slat track (reference data set):
natural frequency and damping ratio and comparison with LSCE results.
Mode Quality Index

Alternatively, the Mode Quality Index can be used to classify the physical and
mathematical modes. Based on the results in Figure 7.7, proper weights can be
196

Variable 1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
400
600
800
1000
1200
Variable 3
0.8
0.6
0.6
0.4
0.4
0.2
0.2
400
600
800
1000
1200
Variable 5
0.8
0.6
0.6
0.4
0.4
0.2
0.2
400
600
800
1000
600
1200
800
1000
1200
Variable 4
400
600
800
1000
1200
Variable 6
0.8
400
0.8
Variable 2
400
600
800
1000
1200
Figure 7.7: Input variables X1 X6 that can be used for classifying the poles for the
reference data. The X-axis represents the frequency in Hz. The dotted lines indicate the
physical poles (in the upper part > 50%).
chosen for the MQI (7.6). Figure 7.9 shows the result for the reference data and
last ML iteration, for two different choices:
case 1: std = 0.20, pairs = 0.20, corr = 0.30, siso = 0.30
case 2: pairs = 0.30, corr = 0.35, siso = 0.35
197
X X X
1
X X X X X X
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.5
0.4
0.2
0.1
0.1
800
1000
1200
400
600
X X X X
1
0.9
0.8
0.8
0.7
0.7
0.6
0.5
0.4
1200
0.2
0.1
0.1
1200
0.4
0.3
800
1000
Freq (Hz)
1000
0.5
0.2
600
0.6
0.3
400
800
0.9
X X X
Membership fct.
Membership fct.
0.4
0.2
600
0.5
0.3
400
0.6
0.3
Membership fct.
Membership fct.
400
600
800
1000
Freq (Hz)
1200
Figure 7.8: Classification result for the slat track (reference data set) by using several
combinations of the input variables X1 X6 for fuzzy clustering. The X-axis represents
the frequency in Hz. The horizontal dashed line indicates a threshold of 0.5, where
all physical poles should be above in order to be correctly classified. The dotted lines
indicate the physical poles.
As can be expected, these results are in line with those of the fuzzy clustering
approach. Nevertheless, the use of the clustering approach has the advantage that
it avoids the user-defined choice of the weights for the different input variables,
making this approach more automated and in general resulting in an improved
separation of the physical and computational poles.
100
100
90
90
80
80
70
70
60
60
MQI (%)
MQI (%)
198
50
50
40
40
30
30
20
20
10
10
10
mode
15
20
10
mode
15
20
Figure 7.9: Mode Quality Index for slat track (reference data set) for the 2 cases of the
weights, case 1 (left), case 2 (right). This result is in line with the results in the bottom
figures of Figure 7.8
Validation
In order to validate the final model extracted by the autonomous identification

process, a conventional modal analysis on the reference data set was performed
as well. The results of the LSCE estimator (in LMS CADA-X software) [86] are
summarized in the stabilization diagram presented in Figure 7.10 for the band
490805Hz. Analyzing in several sub-bands with extensive over-modelling was
necessary in order to identify all physical modes. Based on the pole indicators,
(s, f and o), an experienced user should be able to select all physical poles, even
the weakly excited ones at 554.3Hz and 620Hz, although a large number of mathematical poles complicate this selection. All physical modes identified by the
LSCE algorithm where also returned by the autonomous process. However, it is
in general, even for an experienced user, a difficult task to select all the physical
modes using the classical approach. The use of the MLE algorithm as well as the
mode validation tools is an important step towards an automated estimation process. Table 7.1 demonstrates a good agreement between the natural frequencies
and damping ratios of the selected poles using the LSCE and those returned by
the autonomous process. Around 600Hz and 925Hz two groups of closely spaced
modes can be noticed, which will be considered to illustrate the robustness of the
proposed tracking algorithm in 7.5.
7.4. Case Study II: I40 Bridge
199
Figure 7.10: LSCE Stabilization diagram for the reference data set.
7.4
7.4.1
Case Study II: I40 Bridge

Test Structure and Experiments
A prototype of the proposed autonomous approach has been successfully applied

for a number of experimental test cases [146] under laboratory conditions resulting in data-sets with relatively good signal-to-noise ratios. However, data-sets
corrupted with high noise levels are not uncommon in practice. Typical examples
are modal test data obtained from large structures such as civil structures (bridges
and buildings), automotive structures (fully assembled cars) and aerospace structures (aircrafts, rockets). For this reason, the automated approach is also validated
by means of forced vibration test data obtained on a highway bridge.
The bridge was located along Interstate Highway 40 across the Rio Grande
River in Albuquerque, New Mexico. A series of modal tests was performed on this
bridge after it had been closed for traffic prior to demolition in 1993. The section
of the bridge that was instrumented for this series of modal tests consisted of 3
spans with a combined length of about 130 m (see Figure 7.11).
200
Figure 7.11: Geometry of the portion of the I40 bridge (sideview) that was tested.
Figure 7.12: Measurement point geometry of the I-40 highway bridge. The measurement points, numbered 126, are indicated by . The shaker position is indicated by
SE.
Furthermore 13 accelerometers were placed on each of the 2 plate girders along

the length of the 3 spans, resulting in a total of 26 DOFs and the structure was
excited in DOF 3 (Figure 7.12). A typically measured FRF together with its coherence function is depicted in Figure 7.13. The coherence function indicates the poor
quality of the measurements, mainly due to the resistance of the large structure to
the forced excitation, especially at resonance frequencies [37] and lower signal to
noise ratios around the anti-resonances. More details on the bridge and performed
vibration experiments can be found in [37]. The data can be downloaded at the Los
Alamos National Laboratory website (http://ext.lanl.gov/projects/damage id/).
7.4.2
Model Order Reduction
In general, the number of structural modes, within the frequency band of analysis,
is a priori not known. Since the optimal model order depends on the type of
structure and frequency band studied, an optional step can be considered to reduce
the initial arbitrary model order to a more appropriate size before the actual modal
parameter estimation based on the MLE algorithm is performed. Two possible
approaches for model order reduction are illustrated for the case of the I40 bridge.
201
dB(FRF) (m/N/s2)
80
100
120
140
160
6
Freq (Hz)
10
11
6
Freq (Hz)
10
11
Coherence
0.8
0.6
0.4
0.2
0
Figure 7.13: Typical FRF (top) and corresponding coherence function (bottom) obtained on the I-40 Highway Bridge.
Pole Uncertainty Analysis

Running the Weighted LSCF estimator with an arbitrary chosen (high) model
order, e.g. 50 modes, results in a first set of poles. Due to the extensive overmodelling, a large number of computational modes are present. A first approach
to reduce this initial model order, is the analysis of the uncertainties on the estimated poles. Availability of the variances on the FRFs allows to do a sensitivity
analysis of the parameters for noise on the measured data resulting in an uncertainty level on the estimated poles. Figure 7.14 shows the uncertainties of the
poles in case of the bridge data. Ranking the uncertainty (standard deviation) of
all estimated poles from low to high and calculating the derivative of this function results in Figure 7.15. Since the uncertainties of physical poles are generally
significantly smaller compared to the computational ones, the maximum of this
derivative function should appear close to the number of physical poles. As shown
in Figure 7.14 for the lower frequency part (25Hz) of the analysis band, a clear
gap of 1015dB separates the uncertainties for the physical (indicated by dashed
lines) and computational poles. However, between 59Hz the measurement data is
characterized by poor signal-to-noise ratios and as a result the uncertainties of the
physical poles increase, making a clear distinction between physical and computational poles impossible. Although pole uncertainty analysis can be successfully
applied in numerous cases [146], this illustrates clearly the shortcomings of this
approach for measurement data of poor quality.
202

30
25
20
dB
15
10
10
6
Freq (Hz)
Figure 7.14: Uncertainties (standard deviation) for poles estimated by WLSCF for 50
modes (dotted lines: physical poles).
Pole-Zero Pairs Elimination

The second approach tested for model order reduction, is based on the detection
of cancelling pole-zero pairs as presented in 7.2.2, a typical characteristic for
computational modes. Eliminating modes corresponding to these pole-zero pairs,
often results already in a fairly good indication for the number of structural modes.
Figure 7.16 shows most of the poles () and zeros () estimated by the weighted
LSCF estimator for an initial model order of 50 modes. The resonance frequencies
are indicated on the scaled imaginary axis. Again, it can be noticed that a large
number of poles are surrounded by dense clusters of zeros, which is typical an
indication of cancelling pole-zero pairs. Furthermore, most of these poles are
unstable (in right part of Nyquist plane).
The probability parameter in Eq. (7.1) is chosen as pb = 68% in order to
identify only the most obvious cancelling pole-zero pairs since the goal is only to
reduce the model to a more appropriate order without the risk of excluding any
of the physical poles. As shown in figure Figure 7.16 by the ellipses (due to scaled
imaginary axis) all computational poles clearly have dense clusters of zeros fallen
within the threshold circle. By eliminating these poles, the model order can be
203
f(stdp)=sort(db(stdp))
40
Ampl.
30
20
10
0
10
10
15
20
25
30
35
40
45
50
20
25
30
number of modes
35
40
45
50
diff(f(stdp))
6
5
Ampl.
4
3
2
1
0
10
15
Figure 7.15: Analysis of pole uncertainties. Ranking of the uncertainty (standard deviation) of all estimated poles from low to high (top) and calculating the derivative function
(bottom).
reduced and this is done in an recursive way until no more pole-zero pairscan
be detected for the given threshold parameters. In the case of the bridge data, the
initial model order was reduced from 50 to 28 modes during 5 WLSCF runs.
7.4.3
Autonomous Estimation Results for I40 Bridge
For the analysis of the I40 bridge data, a model order of 28 modes, 50 iterations
and an analysis frequency band of 29Hz were used. The averaged sum of all
FRFs and the ML poles over 50 iterations for a fixed model order of 28 are given
by the pole-diagram in Figure 7.17. Next, the physical thresholding step consists
of a first elimination of unfeasible modes generated by the ML estimator. After
applying the heuristic criteria discussed in 7.2.1, the poles in Figure 7.17 were
obtained.
Although the physical thresholding eliminated a large number of non-physical
modes, this is often not sufficient for eliminating all computational modes. This
is also the case for the bridge data. As can be seen in Figure 7.17, several closely
204

9
IM/2/PI
0.5 0.4 0.3 0.2 0.1
0
RE
0.1
0.2
0.3
0.4
0.5
Figure 7.16: Poles () and zeros () in the Nyquist plane. The ellipses (circles re-scaled
by 2) indicate the 2 dimensional 68% uncertainty bounds around each pole.
spaced lines appear, whereas in the frequency range 57Hz no structural modes
are expected [37].
7.4.4
Although, all lines exhibit a stable behavior, it is most likely that not all of
these lines correspond to structural modes and a manual selection of the physical
poles based in this diagram is likely to yield erroneous results. For this reason,
the physical pole selection was done again by using a fuzzy clustering approach as
presented in 7.2.2.
The input variables X1 X4 were available for the Fuzzy clustering approach,
in the case of the bridge data, and are shown in Figure 7.18. Again in this case,
the Modal Phase Colinearity has a higher error ratio compared to the stochastic
criteria (X2 and X3 ). From the classification result shown in Figure 7.19, it follows
that 10 modes are classified as physical, where there is a very clear distinction
between the physical and computational poles in the case that only the stochastic
7.5. Mode Tracking
205
50
45
40
Iteration Number
35
30
25
20
15
10
Freq (Hz)
Figure 7.17: Pole-diagram: poles () remaining after physical thresholding.
criteria (i.e. X1 X3 ) are considered. The first 6 physical modes indicated in

Figure 7.19 are in good correspondence with those reported in [37], while the
other 3 modes were out of the analysis band considered in that report.
7.5
Mode Tracking
Since the natural frequencies and damping ratios of the modes change due to
changes in the structure, it is not always straightforward to determine which modes
from two subsequent models are corresponding. Furthermore, it can happen that
two modes cross each other in terms of natural frequency or damping ratio or even
that the estimation algorithm returns a single mode at the moment of crossing,
masking the other mode and seriously hampering the tracking process. In this
section, a robust mode tracking approach is presented for the case of the slat
track.
206

Variable 1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
Variable 3
0.8
0.6
0.6
0.4
0.4
0.2
0.2
6
Freq. (Hz)
Variable 4
0.8
Variable 2
6
Freq. (Hz)
Figure 7.18: Input variables X1 X4 that can be used for classifying the poles. The
X-axis represents the frequency in Hz. The dotted lines indicate the physical poles.
7.5.1
Tracking Approach
One obvious way for mode tracking is to monitor the changes of the estimated
natural frequencies. However in the case of closely spaced or crossing modes more
information is needed to determine which modes are related over the subsequent
monitoring instances. The damping ratio could help in some cases to separate
close modes. In [33] the use of mode shape information in the form of the Modal
Assurance Criterion (MAC) is successfully applied for mode tracking for flight flutter test data. For tracking purposes, the MAC evaluates the degree of correlation
[t1]
[t]
between two mode shapes over subsequent instants, k
and l , resulting in a
7.5. Mode Tracking
207
X1 X2 X3 X4
Membership fct.
0.8
0.6
0.4
0.2
0
X1 X2 X3
Membership fct.
0.8
0.6
0.4
0.2
0
6
Freq. (Hz)
Figure 7.19: Classification result for the bridge data by using all (top) and only the
stochastic input variables X1 X3 (bottom) for fuzzy clustering. The X-axis represents
the frequency in Hz. The horizontal dashed line indicates a threshold of 0.5, where
all physical poles should be above in order to be correctly classified. The dotted lines
indicate the physical poles.
value close to one for corresponding modes

MAC
[t1]
[t]
k , l
[t1]H [t] 2
.
= [t1]H k [t1] l [t]H [t]
k
.k
l .l
(7.7)
with k = 1, . . . , Nm t1 the number of modes at instant t 1, l = 1, . . . , Nm t the

number of modes at instant t.
However, structural health monitoring in normal operating conditions is often
done by means of a very limited number of sensors, and as a result, the MAC
criterion is then not always applicable. This can be overcome by using the distance
[t1]
(or resemblance) between two residue matrices over subsequent instants, R k
[t]
and Rl (with Rr = Qr r Tr = r LTr , with Lr the row vector of participation
factors). The Frobenius norm of their difference, results in a value close to zero
208
for corresponding modes

[t1] [t]
(R
, R l )F =
k
trace
[t1]
Rk
[t] H
Rl
[t1]
Rk
[t]
Rl
(7.8)
In the case of a single response sensor, the use of the distance measure will only
fail when two closely spaced modes have a similar residue values at the sensor
location. The use of high spatial resolution (laser vibrometer) measurements can
then contribute to an optimal selection for the sensor location for the modes to be
tracked.
Notice that in general the distance criterion cannot be used in the case of
operational data (output-only data), since in that case the participation factors
cannot be estimated. An example is the monitoring of a bridge by means of road
traffic excitation, where the level of excitation varies. However, bridge inspection
systems integrate sufficient sensors such that the MAC criterion can then be used
to perform the in-operation mode tracking. This is different for monitoring applications having only a limited number of sensors, such as in the case of routine
airplane ground inspections or life cycle tests. In these cases however, output-only
measurements over the subsequent instances are then performed in similar conditions by means of a patch actuator or acoustic excitation. Although the MAC
criterion is not applicable (not enough sensors), the distance criterion can then
be used since the source of excitation is identical during the periodic inspections
and the scaling factor remains similar. However, in the case of general operational
conditions, the input forces vary as a function of the amplitude, frequency coloring
and location where the forces are applied.
7.5.2
Tracking Algorithm Problem of Missing Modes
Initially, the modes of interest for the tracking are selected from the reference
model obtained from an analysis of the first data set. Mode tracking over the
consecutive instants is done based on a model comparison between the new and
previous modal model. Considering the poles in the Nyquist plane, the distance
[t1]
[t1]
between pole pk
of mode mk
(k = 1, . . . , Nmt1 number of modes at t 1),
and all poles of the modes at the next instant t can be determined. Depending on
the amount of a structural change and the sensitivity of the mode for that change,
the influence on a certain pole (frequency and/or damping) can be large and as a
[t1]
result the closest pole at instant t does not necessarily correspond to the pole p k .
For this reason the y closest poles at instant t are considered (y being an algorithm
parameter which is typically chosen between 5Nm t ). Next the residue distance
measure (7.8) is used to determine which of the y modes at instant t corresponds
[t1]
to the mode mk . After having determined the closest mode at instant t for
each mode at instant t 1, an additional step is necessary to check if each mode is
uniquely matched over the subsequent instants. It can happen that a mode from
7.5. Mode Tracking
209
the model at instant t is used more than once for different modes at t 1, due to
the problem of missing modes at instant t. This problem can appear in the case
of crossing modes with very close or coinciding poles, with the possibility that the
broadband MLE identification only returns a single mode. In this case, a model reestimation by means of the autonomous identification process in a small frequency
band is initiated. Given the performance of the MLE approach, this quasi always
results again in a set of very close but different modes. Using the same model
comparison approach based on the distance measure (7.8) for the small band, the
different modes can now still be tracked over the subsequent time instants. In the
case that the re-estimation process is not successful, the modes still missing on
instant t are ultimately replaced by the modes at the previous instant ensuring the
continuity of tracking all modes once they are estimated again for the following
instants.
7.5.3
Tracking Results for Slat Track
To illustrate the robustness of the tracking approach for missing modes, the slat
track modes at 926.6Hz and 936.6Hz in the reference model are followed over the
subsequent measurements for an increase of the mass between sensors 2 and 3
(see Figure 7.1). Applying the tracking algorithm without checking for missing
modes results in Figure 7.20 (top), where it is clearly shown that for instant 7
only a single mode is estimated during the broadband MLE analysis, used twice
for two, very closely spaced but different modes at that instant. As a result, for
the subsequent instants only the initial mode of 936.6Hz is further tracked, missing
the initial mode of 926.6Hz. The type of structural changes, i.e. a mass change
between sensors 2 and 3 allowed for simulating the crossing effect of the modes,
since this mass is located in a node for the mode of 926.6Hz while in a point of
high amplitude for the mode of 936.6Hz (see Figs. 3 and 5). Re-estimation of
the model for instant 7 in a small band around 926Hz (25Hz) results in two
closely spaced but different modes at 926.16Hz and 926.41Hz, with a MAC value
of 33.36%. The re-estimation step makes the tracking algorithm robust to missing
or coinciding modes as is also shown in Figure 7.20 (bottom).
In order to illustrate that the tracking algorithm also works for larger changes
in the structure, different amounts of mass changes were considered. Besides the
increment of about 2 10-3kg over 14 steps, tracking results were obtained for
an increment of 10 10-3kg over 3 steps and 30 10-3kg over one single step. As
presented by Figure 7.21, the tracking algorithm is perfectly capable of tracking
the crossing modes at 926.6Hz and 936.6Hz even if the structural changes become
larges. For example, the natural frequency of the mode at 936.6Hz shifted by
2.63% for a mass increment of 30 10-3kg.
The robustness of the algorithm for closely spaced modes is also illustrated in
Figure 7.22 for the modes around 610Hz, for which tracking, based on the distance
210
940
50
45
freq. (Hz)
930
40
920
30
910
0
25
3
20
15
10
5
0
910
10
15
10
15
10
15
10
15
damp. ratio (%)
Iteration number
35
915
920
925
Freq (Hz)
930
x 10
2.5
2
1.5
1
0.5
0
935
instant
940
30
freq. (Hz)
930
25
Iteration number
920
20
910
0
15
3
damp. ratio (%)
10
0
910
5
3
915
920
925
Freq (Hz)
930
935
x 10
2.5
2
1.5
1
0.5
0
5
instant
Figure 7.20: Mode tracking results (modal frequency and damping ratio) for the modes
at 926.6Hz (*) and 936.6Hz (o) using the tracking algorithm without (top) and with
(bottom) checking for missing modes.
criterion, is successful again. The mass change between sensors 2 and 3 clearly
affects some of the modes while others remain unchanged since they have a node
at that location as can also be seen in Figure 7.3. A crossing of the damping ratios
can be noticed now.
Figure 7.23 compares the use of the MAC (7.7) and the distance (7.8) criterion
for 3 different measurement setups. In the case that the 7 response sensors are
used (dash-dotted line), there is no difference in the tracking results obtained
by using the MAC or the distance criterion. However, in the case that a single
7.5. Mode Tracking
211
freq. (Hz)
940
930
920
910
0
10
12
14
16
10
12
14
16
damp. ratio (%)
x 10
2.5
2
1.5
1
0.5
0
instant
Figure 7.21: Mode tracking results (modal frequency and damping ratio) for the modes
at 926.6Hz and 936.6Hz using different mass changes for the subsequent instants. Tracking of mode at 926.6Hz over 14 steps (dotted *), 3 steps (o) and 1 step (). Tracking of
mode at 936.6Hz over 14 steps (dotted +), 3 steps (2) and 1 step ().
sensor is used, the MAC criterion always returns one, and thus is not applicable
for tracking purposes. On the other hand, the distance criterion still gives good
results in the case that a single response sensor is located in position 1 (dashed line)
(see Figure 7.1 for sensor positions) since it uses the amplitudes of the difference
in residues which is still be meaningful in the case of a single sensor. Only in
the case that the residues of two closely spaced modes have similar amplitudes
for that specific sensor position, also this criterion fails. This is the case when
the single sensor is located in position 7 (dotted line). This can also be seen in
Figure 7.3, where the bottom part of the slat track (i.e. sensor 7) has similar
vibration behaviour for both modes at 926.6Hz and 936.6Hz. Prior knowledge
of the dynamic behaviour by means of a scanning laser vibrometer allows for an
optimal sensor positioning depending on the modes selected for tracking.
The availability of the confidence levels on the estimated parameters gives
additional information to determine if a small change of a modal parameter is due
to a structural change or due to the uncertainty on the measured data. Figure 7.24
shows the natural frequency and damping ratio for the mode of 926.6Hz with the
95% confidence interval. Since the mass changes are applied in a quasi-nodal
point for this mode, the change of model parameters remains small. Nevertheless,
212
freq. (Hz)
630
620
610
600
0
10
12
14
16
8
instant
10
12
14
16
damp. ratio (%)
0.015
0.01
0.005
0
0
Figure 7.22: Mode tracking results for the closely spaced modes around 610Hz.
the changes are significantly larger than the sum of the confidence levels over two
subsequent instants clearly still indicating that in the structural changes appeared.
Only for instant 7, the uncertainty levels are higher explained by the difficult
estimation problem of the two coinciding modes.
7.5.4
Damage Assessment during Tracking
Mode tracking can serve autonomous monitoring systems by detecting structural

changes, and threshold levels for different modal parameters can be set for notifying
an operator. However, this is not sufficient in the case that one wants to know
the location and severity of the damage. As a result a damage assessment has
to be done based on the results obtained from the autonomous identification and
tracking process. Recently, a novel sensitivity-based damage assessment technique
was proposed in [98]. However, since a driving point is required, the applicability
of this method was restricted to forced vibration tests only, while most monitoring
applications only have access to output-only measurement data. The applicability
of the sensitivity-based approach has been extended to output-only measurements
and illustrated based on the results discussed in this chapter [99].
7.5. Mode Tracking
213
freq. (Hz)
940
930
920
910
0
10
12
14
16
10
12
14
16
10
12
14
16
10
12
14
16
damp. ratio (%)
x 10
2.5
2
1.5
1
0.5
0
instant
freq. (Hz)
940
930
920
910
0
damp. ratio (%)
x 10
2.5
2
1.5
1
0.5
0
instant
Figure 7.23: Comparison of modal frequency and damping ratio for the two modes
at 926.6Hz (*) and 936.6Hz (o) obtained from the tracking algorithm using the MAC
(top) and distance criterion (bottom) for different measurement setups using 7 sensors
(dash-dotted), one sensor at position 1 (dashed), one sensor at position 7 (dotted).
214
freq.(Hz)
927
926.5
926
925.5
0
10
12
14
16
8
instant
10
12
14
16
damp. ratio (%)
2.8
x 10
2.6
2.4
2.2
2
0
Figure 7.24: MLE uncertainty information during tracking process. Horizontal bars
indicate the 95% confidence intervals for the natural frequency and damping ratio at
each instant.
7.6
Conclusions
This chapter presented a novel approach for an autonomous modal parameter

identification and mode tracking illustrated by means of two practical case studies.
The results are very promising since the approach is capable of identifying and
tracking all physical modes of interest even in difficult cases of closely spaced or
even crossing modes due to changes (damage) in the structure. The proposed
approach is applicable for both forced-vibration and output-only measurements
and provides a tool for structural health monitoring.
It is shown that accurate parameter estimates with confidence intervals are estimated by means of the frequency-domain Maximum Likelihood algorithm, while
stochastic mode validation criteria, make the validation process more robust to
measurement noise. The final model is selected using a mode quality index or a
fuzzy clustering approach, which separate physical from computational modes.
The mode tracking process is based on model comparison between subsequent
measurements. Using the minimal distance between mode residue matrices, corresponding modes can be successfully determined even without the need of extensive
7.6. Conclusions
215
spatial information, which is important for the implementation of many monitoring

systems. The tracking approach provides a small band re-estimation in the case
of missing or wrongly estimated modes due to closely spaced or crossing modes
not captured by the model during the broadband analysis. Eventually, it has been
demonstrated that the proposed tracking approach still works for larger structural
changes.
216
Chapter 8
Conclusions
8.1
Thesis Summary
In this thesis, the applicability of frequency-domain identification methods in the

domain of modal analysis has been studied. Special attention has been paid to the
problems and requirements related the different steps in the overall identification
process. It is now worthwhile to review the most important results achieved in the
different chapters.
Chapter 2 is devoted to the first step of modal analysis. The nonparametric
identification of FRFs has been revised in a maximum likelihood framework based
on an errors-in-variables stochastic noise model. A nonparametric approach for
the determination of both the FRF matrix and its covariance matrix from the same
measurement data set has been presented for arbitrary excitations. It is shown
that the so-called nonparametric instrumental variables (Hiv ) estimator has quasi
maximum likelihood properties, without requiring any a priori information about
the measurement noise. A study of the applicability of the proposed estimator for
the specific case of modal analysis has shown that the nonparametric H iv estimator
remains consistent when a MISO approach is used to process the extensive amount
of data, while the practical limitations with respect to the determination of the
noise information have been pinpointed.
The problem of leakage, encountered when using arbitrary excitation for modal
testing, has been addressed in Chapter 3. Several approaches for reducing the
leakage on the estimated FRFs by means of exponential windowing have been
studied and compared with techniques proposed in literature. This has resulted
in a method that is based on the use of an exponential window in combination
with the classical FRF estimators. Although this results in biased nonparametric
217
218
Chapter 8. Conclusions
results, the additional damping is known and its effect on the modal parameters
can be compensated, which is not the case for e.g. a Hanning window. Moreover,
the proposed approach is simple to apply in practice, using commercial measurement equipment for modal analysis that often only stores the measured FRF data.
Furthermore, the noise information can still be determined when using this approach. Based on the results of Chapter 2, this method has been generalized to
an errors-in-variables framework.
An introduction of the concepts of parametric system identification for modal
parameter estimation is given in Chapter 4. Different parametric models to describe the vibration behaviour of mechanical systems have been discussed, and an
overview of parametric identification methods with their potentials and drawbacks
from the point of view of modal parameter estimation has been given. This chapter motivated the application of so-called stochastic frequency-domain estimators,
such as the Maximum Likelihood estimator.
Chapter 5 has been devoted to the optimization of the (stochastic) frequencydomain modal parameter estimators using FRF measurements. An extensive
amount of data, typical for modal analysis, requires optimized algorithms that
balance between accuracy and memory/computation efficiency. A multivariable
implementation of frequency-domain estimators, based on a common denominator
transfer function model, was derived omitting however possible correlations between the FRFs. A study of different types of parameterization for experimental
modal data has demonstrated that the use of normal equations is justified when
using a discrete-time transfer function model with complex coefficients. The final
LS implementation resulted in the so-called Least Squares Complex Frequency
(LSCF) method. This estimator offers the user, in analogy with the well-known
LSCE method, a stabilization chart. The fast implementation could also be extended to the class of stochastic Total Least Squares (mixed LS-TLS) estimators
and the Maximum Likelihood (ML) methods, making them applicable for modal
analysis. Using experimental data, the robustness of the algorithms was demonstrated for model orders up to 200.
Based on the results of Chapter 5, a fast implementation of the frequencydomain methods has been derived for I/O data. The use of these estimators
is preferred in the case of short data sequences, since then the nonparametric
processing for FRF estimation is not possible. This can be the case for flight
flutter testing and structural health monitoring applications. The formulation
of the I/O estimators is based on an errors-in-variables set-up, where both the
noise on the input and output signals are considered. Similar to the FRF-based
estimators, where possible correlations between the FRFs are neglected, the fast
implementation of the I/O estimators omits possible correlations among the noise
on the output measurements. It has been shown that a compact formulation for
the WGTLS estimator was only possible by using a linear approximation for the
contributions of input noise. Furthermore, by estimating simultaneously the initial
conditions and the system model parameters, the proposed I/O estimators could
8.2. Ideas for Further Research
219
be made robust for errors due to leakage effects. Next, a generalization has been
given of the fast I/O estimators to estimate the parameters starting from auto and
cross power spectral density functions. This is preferable when longer time records
are available while a parametric compensation of the leakage is still possible.
Finally, Chapter 7 presents an approach for an autonomous modal parameter
identification and mode tracking . The approach is capable of identifying and
tracking all physical modes of interest even in difficult cases of closely spaced or
even crossing modes due to changes (damage) in the structure. It is shown that
accurate parameter estimates with confidence intervals are estimated by means
of the frequency-domain Maximum Likelihood algorithm, while stochastic mode
validation criteria make the validation process more robust to measurement noise.
The mode tracking process is based on model comparison between subsequent measurements. It was demonstrated that modes could be successfully tracked without
the need of extensive spatial information. This is important for the implementation of many monitoring systems. The tracking approach also checks for missing
or wrongly estimated modes due to closely spaced or crossing modes. The tracking
approach remains applicable even for larger structural changes. The approach has
been validated for two experimental cases from the domain of mechanical and civil
engineering.
8.2
Ideas for Further Research
In chapters 5 and 6, it has been shown that the use of a common-denominator

model results in structured matrices, which has been exploited to improve the
speed and memory usage of the algorithms. Nevertheless, further optimization
can be considered in order to broaden the possible applications that can benefit
from the accurate stochastic WGTLS and ML estimators. Possible ideas for further
improvement in terms of computation speed and memory usage are related to:
the exploitation of the structured Toeplitz matrices for internal operations
such as e.g. the matrix product, inversion and factorizations (QR, SVD, ...).
the implementation of the algorithms using parallel processing. The compact
normal matrix
#
"N N
o i
X
H 1
(8.1)
Tk S k R k Sk
D=
k=1
is perfectly suitable for a parallel implementation that divides the computation of the submatrices Rk , Sk and Tk over multiple processors, since the
matrix D can be composed as the sum of multiple submatrices corresponding to different batches of input/output DOFs. This will allow to further
optimize the stochastic estimators for applications such as online monitoring
of the flutter behaviour of an aircraft during flight testing.
220
Chapter 8. Conclusions
In chapter 2, the noise covariance matrix for the FRFs or Input/Output signals
can be computed using the Hiv methodology. Since in practice M < No , this
matrix is not of full rank and hence cannot be inverted. During the parametric
estimation, the correlations between the FRFs or output signals are not taken
into account, which results in some loss of asymptotically efficiency while the
consistency is remained. Nevertheless, the uncertainty bounds obtained for the
parameters using the approach discussed in 5.5.2 will not be correct, which can
be a problem for applications such as e.g. damage detection and structural health
monitoring and hence further research should be devoted to this problem. A
first idea consists of computing the uncertainty on the modal parameters directly
from the nonparametric noise covariance matrix as C = S H CHiv S, where S is a
sensitivity matrix. This is possible since no inversion of the matrix C Hiv is required,
although the computational burden is too high to be applied in practice requiring
a numerical optimization of this approach. Another idea for deriving uncertainty
bounds for the parameters obtained by the LSCF estimator is as follows. Based on
the stabilization chart using variable LS constraints, clusters of poles are found for
a specific model order for which a statistical analysis yields the mean and variance
of each pole. How are these variances related to the stochastic variances (or
confidence bounds) introduced by the noise on the measurements is the question
to be assessed. Comparing this with Monte Carlo simulations will indicate if this
approach can yield acceptable confidence bounds for the LSCF estimator.
Other types of matrix fraction description models can be considered. The use
of a RMFD will lead to a polyreference implementation of the LSCF, where the
poles and modal participation factors can be estimated together in a first step.
The computation time will be limited since the RMFD model still results in structured matrices by uncoupling the output DOFs. At first sight, this polyreference
approach would avoid the use of the SVD of the residue matrices, where information is lost due to the rank reduction. Nevertheless, the performance in terms
of accuracy and stabilization chart should be compared by the SMFD implementation, especially in the case of inconsistent data due to the effects of e.g. mass
loading.
Furthermore, extending the applicability of the algorithms to operational modal
analysis is an important topic for research. This requires first the revision of the
nonparametric processing in order to obtain suitable data, i.e. Auto and Cross
Power Spectra between the output signals as well as the uncertainty on this data.
Since the operational conditions often correspond to random noise excitation, the
applicability of windowing techniques should be studied in order to reduce effects
of leakage. Moreover, if the excitation consists of a part that is measured and a
part that is unknown due to the presence of process noise, the possibility of using a
combined approach forced/operational is an interesting topic for further research.
Bibliography
[1] J.L. Adcock. Curve fitter for pole-zero analysis. Hewlett-Packard Journal, pages 3336,
January 1987.
[2] H. Akaike. A new look at the statistical model identification. IEEE Transactions on
Automatic Control, AC-19(6):716723, 1974.
[3] S. Alampalli. Influence of in-service environment on modal parameters. In Proceedings of
16th International Modal Analysis Conference, pages 111116, Santa Barbara, CA, USA,
1998. SEM.
[4] R. Allemang. Modal analysis - where do we go from here? In Proceedings of 11th International Modal Analysis Conference, page keynote paper, Florida, USA, 1993. SEM.
[5] R. Allemang, D. Brown, and W. Fladung. Modal parameter estimation: a unified matrix
polynomial approach. In Proceedings of 12th International Modal Analysis Conference,
Honolulu, Hawaii, USA, 1994. SEM.
[6] R. Allemang and A. Phillips. Cyclic averaging for frequency response function estimation.
In Proceedings of 14th International Modal Analysis Conference, Michigan, USA, 1996.
SEM.
[7] B. D. O. Anderson. Identification of scalar errors-in-variables models with dynamics.
Automatica, 21(6):709716, 1985.
[8] J. R. F. Arruda. Transient truncation and periodization in modal testing. Journal of
Sound and Vibration, 149(2):327329, 1991.
[9] J. R. F. Arruda. Objective functions for nonlinear curve fit of frequency response functions.
AIAA Journal, 30(3):855857, 1992.
[10] 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. Transactions of the ASME: Journal of Dynamic Systems, Measurement, and
Control, 123(4):668676, 2001.
[11] D. S. Bayard. High-order multivariable transer function curve fitting: algorithms, sparse
matrix methods and experimental results. Automatica, 30(9):14391444, 1994.
[12] D.S. Bayard. An algorithm for state space frequency domain identification without windowing distortions. In Proceedings of the 31st IEEE Conference on Decision and Control,
pages 17071712, Tucson, Arizona, USA, 1992.
[13] J. S. Bendat. Nonlinear system analysis and identification from random data. John Wiley
& Sons, New York, 1990.
[14] J. S. Bendat and A. G. Piersol. Engineering applications of correlation and spectral analysis. John Wiley & Sons, New York, 1980.
[15] J. C. Bezdek. Pattern recognition with fuzzy objective function algorithms. Plenum Press,
New York, 1981.
221
222
Bibliography
[16] P. Billingsley. Probability and Measure. John Wiley & Sons, New York, 1986.
[17] R. Bishop and G. Gladwell. An investigation into the theory of resonance testing. Philosophycal Transactions of the Royal Society of London, 255 A 1055:241280, 1963.
[18] R. B. Blackman and J. W. Tukey. The measurement of power spectra from the point of
view of communication engineering. Dover Publications, Inc., New York, 1958.
[19] A. Bosse, F. Tasker, and S. Fisher. Real-time modal parameter estimation using subspace
methods: applications. Mechanical Systems and Signal Processing, 12(6):809823, 1998.
[20] J. W. Brewer. Kronecker products and matrix calculus in system theory. IEEE Transactions on Circuits and Systems, 25:772781, 1978.
[21] D. R. Brillinger. Time series : data analysis and theory. McGraw-Hill, New York, expanded edn. edition, 1981.
[22] R. Brincker, P. H. Kirkegaar, P. Anderson, and M.E. Martinez. Damage detection in an
offshore structure. In Proceedings of the 13th International Modal Analysis Conference,
pages 661667, 1995.
[23] D. Brown, R. Allemang, R. Zimmerman, and M. Mergeay. Parameter estimation techniques
for modal analysis. In SAE paper 790221, pages 828846. SAE, 1979.
[24] B. Cauberghe, P. Guillaume, B. Dierckx, and P. Verboven. Identification of modal parameters from inconsistent data. In Proceedings of the 20th International Modal Analysis
Conference, pages 809815, Los Angeles, CA, USA, 2002.
[25] K. S. Chhipwadia, D. C. Zimmerman, and III G. H. James. Evolving autonomous modal
parameter estimation. In Proceedings of the 17th International Modal Analysis Conference,
pages 819825, Kissimmee, FL, USA, 1999.
[26] R. E. Cobb and L. D. Mitchell. Estimates of variance and confidence bands for the threechannel frequency response function estimator. International Journal of Analytical and
Experimental Modal Analysis, 5(3):185194, 1990.
[27] R. D. Cook. Concepts and applications of finite element analysis. John Wiley and Sons,
1981.
[28] J.E. Cooper and K. Worden. On-line physical parameter estimation with adaptive forgetting factors. Mechanical Systems and Signal Processing, 14(5):705730, 2000.
[29] C.E. Davila and H. Chiang. An algorithm for pole-zero system model order estimation.
IEEE Transactions on Signal Processing, 43(4):10131017, April 1995.
[30] B. De Moor, M. Gevers, and G.C. Goodwin. l2 -overbiased, l2 -underbiased and l2 -unbiased
estimation of transfer functions. Automatica, 30(5):893898, 1994.
[31] B. De Moor and J. Vandewalle. A geometrical strategy for the identification of state space
models of linear multivariable systems with singular value decomposition. In Proceedings
of 3rd Symposium on Applications of Multivariable Systems Techniques, pages 5969,
Plymouth, UK, 1987.
[32] B. De Moor and J. Vandewalle. A unifying theorem for linear and total linear least squares.
IEEE Transactions on Automatic Control, 35(5):563566, May 1990.
[33] M.J. Desforges, J.E. Cooper, and J.R. Wright. Mode tracking during flight flutter testing
using the modal assurance criterion. Proceedings of the Institution of Mechanical Engineers, 210:2737, 1996.
[34] 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 vibration
characteristics: a literature review. Technical Report LA-13070-MS, Los Alamos National
Laboratory, 1996.
[35] DSP Committee. Programs for digital signal processing. Technical report, Digital Signal
Processing Committee, New York, USA, 1979.
Bibliography
223
[36] D. J. Ewins. Modal testing : theory and practice. Research Studies Press, 2000. Second
edition.
[37] C. R. Farrar, W. E. Baker, T. M. Bell, K. M. Cone, T. W. Darling, T. A. Duffey, A. Eklund,
and A. Migliori. Dynamic characterization and damage detection in the I-40 bridge over
the rio grande. Technical Report LA-12767-MS, Los Alamos National Laboratory, 1996.
[38] G. E. Forsythe. Generation and use of orthogonal polynomials for data-fitting with a digital
computer. Journal Soc. Industr. Appl. Math., 5(2):7488, 1957.
[39] K. Godfrey, editor. Perturbation signals for system identification. Prentice Hall International series in Acoustics, speech and signal processing. Prentice Hall International, 1993.
[40] G. H. Golub and C. F. Van Loan. Matrix computations. The John Hopkins University
Press, Baltimore, third edition, 1996.
[41] P. Guillaume. Identification of multi-input multi-output systems using frequency-domain
models. PhD thesis, dept. ELEC, VUB, Brussels (Belgium), june 1992.
[42] P. Guillaume. Errors-in-variables identification techniques applied to modal analysis. In
Proceedings of Design Engineering Technical Conferences (DETC), pages DETC97/VIB
4253, Sacramento, CA, USA, 1997. ASME.
[43] P. Guillaume. Frequency response measurements of multivariable systems using nonlinear averaging techniques. IEEE Transactions on Instrumentation and Measurement,
47(3):796800, 1998.
[44] P. Guillaume, K. Badredin, and M. Van Overmeire. Frequency domain maximum likelihood identification of simusoids applied to rotating machinary. In Proceedings of the 18th
International Modal Analysis Conference, pages 411417, San Antonio, TX, USA, 2000.
SEM.
[45] 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, Florida, USA, 1999. SEM.
[46] P. Guillaume and R. Pintelon. A gauss-newton-like optimization algorithm for weighted
non-linear least-squares problems. IEEE Transactions on Signal Processing, 44(9):2222
2228, 1996.
[47] P. Guillaume, R. Pintelon, and J. Schoukens. Description of a parametric MLE in the frequency domain for MIMO systems and its application to flight flutter analysis. Mechanical
Systems and Signal Processing, 4(5):405416, 1990.
[48] P. Guillaume, R. Pintelon, and J. Schoukens. Nonparametric frequency response function
estimators based on nonlinear averaging techniques. IEEE Transactions on Instr. Meas.,
41(6):739746, 1992.
[49] P. Guillaume, R. Pintelon, and J. Schoukens. Robust parametric transfer function estimation using complex logarithmic frequency response data. IEEE Transactions on Automatic
Control, 40(7):11801190, 1995.
[50] P. Guillaume, R. Pintelon, and J. Schoukens. Accurate estimation of multivariable frequency response functions. In Proceedings of the 13th Triennial World Congress, pages
423428, San Francisco, USA, 1996.
[51] 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.
[52] P. Guillaume, R. Pintelon, and J. Schoukens. A weighted total least squares estimator for
multivariable systems with nearly maximum likelihood properties. IEEE Transactions on
Instrumentation and Measurement, 47(4):818822, 1998.
[53] 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.
224
Bibliography
[54] P. Guillaume, J. Schoukens, and R. Pintelon. Sensitivity of roots to errors in the coefficient
of polynomials obtained by frequency-domain estimation method. IEEE Transactions on
Instrumentation and Measurement, 38(6):10501056, 1989.
[55] P. Guillaume, P. Verboven, and S. Vanlanduit. Frequency-domain maximum likelihood
identification of modal parameters with confidence intervals. In Proceedings of the 23rd
International Seminar on Modal Analysis, pages 359366, Leuven, Belgium, 1998.
[56] P. Guillaume, P. Verboven, S. Vanlanduit, and E. Parloo. Accurate damping estimates
using adapted frequency-domain identification techniques. In Proceedings of the 18th International Modal Analysis Conference, pages 411417, San Antonio, TX, USA, 2000.
SEM.
[57] 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, pages 15431549, Kissimmee, FL, USA, 2001. SEM.
[58] L. Hermans, H. Van der Auweraer, and P. Guillaume. A frequency-domain maximum
likelihood approach for the extraction of modal parameters from output-only data. In
Proceedings of 23rd International Seminar on Modal Analysis, pages 367376, Leuven,
Belgium, 1998.
[59] L. Hermans and H. Van der Auweraer. Modal testing and analysis of structures under
operational conditions: Industrial applications. Mechanical Systems and Signal Processing,
13(2):193216, 1999.
[60] 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, pages 300306, Santa Barbara, CA, USA,
1998. SEM.
[61] L. Hermans, H. Van der Auweraer, and L. Mevel. Health monitoring and detection of
a fatigue problem of a sports car. In Proceedings of 17th International Modal Analysis
Conference, pages 4248, Kissimmee, FL, USA, 1999.
[62] W. Heylen, S. Lammens, and P. Sas. Modal analysis theory and testing. PMA, KU Leuven,
Leuven, Belgium, 1998.
[63] T. Hiyama and N. Suzuki. Real-time modal analysis of power system oscillations. In IEEE
International Symposium on Circuits and Systems, pages IV225IV228, 2000.
[64] D. L. Hunt, S.P. Weiss, W.M. West, T.A. Dunlap, and S.R. Freesmeyer. Development and
implementation of a shuttle modal inspection system. Journal of Sound and Vibration,
24(9):3442, 1990.
[65] S. Ibrahim. Modal confidence factor in vibration testing. The Shock and Vibration Bulletin,
48(1):6575, 1977.
[66] S. Ibrahim and E. Mikulcik. A method from the direct identification of vibration parameters
from the free response. The Shock and Vibration Bulletin, 47(4):183198, 1977.
[67] S. R. Ibrahim. An upper hessenberg sparse matrix algorithm for modal identification of
mini-computers. In Proceedings of the 10th International Seminar on Modal Analysis,
Leuven, Belgium, 1985.
[68] R. N. Jacques and D. W. Miller. Multivariable model identification from frequency response
data. In Proceedings of the 32nd Conference on Decision and Control, pages 30463051,
San Antonio, USA, 1993.
[69] III G. H. James, D. C. Zimmerman, and K. S. Chhipwadia. Application of autonomous
modal identification to traditional and ambient data sets. In Proceedings of the 17th
International Modal Analysis Conference, pages 840845, Kissimmee, FL, USA, 1999.
[70] J. J. Juang. Applied system identification. Prentice-Hall, Inc., New Jersey, USA, 1994.
[71] J.-N. 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.
Bibliography
225
[72] J.-N. Juang and R. S. Pappa. Effects of noise on modal parameters identified by the
eigensystem realization algorithm. Journal of Guidance, Control and Dynamics, 9(3):294
303, 1986.
[73] J. N. Juang and H. Suzuki. An eigensystem realization algorithm in frequency domain
for modal analysis identification. ASME Transactions on Vibrations, Acoustics, Stress,
110(24):2429, January 1988.
[74] T. Kailath. Linear systems. Prentice-Hall, New Jersey, 1980.
[75] M. Kendall and A. Stuart. Inference and relationship, volume 2 of the advanced Theory
of Statistics. Charles Griffin & Company Limited, London, UK, 4 edition, 1979.
[76] C. C. Kennedy and C. D. P. Pancu. Use of vectors in vibration measurement and analysis.
Journal of the Aeronautical Sciences, 14(11):603625, 1947.
[77] F. Lembregts. Frequency domain identification techniques for experimental multiple input
modal analysis. PhD thesis, dept. PMA, KULeuven, Leuven (Belgium), december 1988.
[78] I. J. Leontaritis and S. A. Billings. Model selection and validation methods for non-linear
systems. Int. Journal on Control, 45(1):311341, 1987.
[79] J. Leuridan. Some direct parameter model identification methods applicable for multiple
input modal analysis. PhD thesis, dept. of Mech. and Ind. Eng., University of Cincinnati,
Ohio, USA, 1984.
[80] J. Leuridan, D. De Vis, H. Van der Auweraer, and F. Lembregts. A comparison of some frequency response function measurement techniques. In Proceedings of the 4th International
Modal Analysis Conference, pages 908918. SEM, 1986.
[81] E. C. Levi. Complex curve fitting. IEEE Transactions on Automatic Control, 4(1):3744,
1959.
[82] S. Li, D. L. Brown, and H. Vold. A scaled total least squares method for modal parameter estimation using the unified matrix polynomial approach. In Proceedings of the 12th
International Modal Analysis Conference, Honolulu, Hawaii, 1994.
[83] T. Lim, R. Cabell, and R. Silcox. On-line identification of modal parameters using artificial
neural networks. Journal of Vibration and Acoustics, 118(4):649656, 1996.
[84] K. Liu, R. N. Jacques, and D. W. Miller. Frequency domain structural system identification by observability range space extraction. In Proceedings of the American Control
Conference, pages 107111, Baltimore, Maryland, 1994.
[85] L. Ljung. System identification: theory for the user. Prentice Hall, Englewood Cliffs, 2nd
edition edition, 1999.
[86] LMS International NV. LMS CADA-X manuals Modal, 2002.
[87] E. Lukacs. Stochastic convergence. Academic Press, New York, USA, 1975.
[88] N. M. M. Maia and J. M. M. Silva, editors. Theoretical and experimental modal analysis.
Research Studies Press, LTD., Somerset, England, 1997.
[89] S. L. Marple. Digital spectral analysis. Prentice-Hall, New York, USA, 1987.
[90] T. McKelvey. An efficient frequency domain state space identification algorithm. In Proceedings of the 33rd IEEE Conference on Decision and Control, pages 33593364, Lake
Buena Vista, FL, USA, 1994.
[91] T. McKelvey. Identification of state space models from time and frequency data. PhD
thesis, dept. of Electrical Engineering, Link
oping University, Link
oping, Sweden, 1994.
[92] E. A. Medina, R. D. Irwin, J. R. Mitchell, and A. P. Bukley. MIMO system identification
using frequency response data. Journal of the Astronautical Sciences, 42:113129, 1994.
[93] L. Meirovitch. Analytical methods in vibrations. The MacMillan Company, London, UK,
1967.
226
Bibliography
[94] M. Mergeay. Least squares complex exponential method and global system parameter
estimation used by modal analysis. In Proceedings of the 8th International Seminar on
Modal Analysis, Leuven, Belgium, 1983.
[95] L. D. Mitchell and J. C. Deel. An unbiased frequency response function estimator. In
Proceedings of the 5th International Modal Analysis Conference. SEM, 1987.
[96] R.S. Pappa, G.H. James III, and D.C. Zimmerman. Autonomous modal identification of the
space shuttle tail rudder. In Proceedings of the Design Engineering Technical Conferences
(DETC), pages DETC97/VIB4250, Sacramento, CA, USA, 1997. ASME.
[97] R.S. Pappa, S.E. Woodard, and J Juang. A benchmark problem for development of autonomous structural modal identification. In Proceedings of the 15th International Modal
Analysis Conference, pages 10711077, Orlando, FL, USA, 1997.
[98] E. Parloo, P. Guillaume, and M. Van Overmeire. Damage assessment using mode shape
sensitivities. Mechanical systems and Signal Processing (in press), 2002.
[99] E. Parloo, P. Verboven, P. Guillaume, and M. Van Overmeire. Autonomous structural
health monitoring part II: Vibration-based operation damage assessment. Mechanical
systems and Signal Processing (in press), 2002.
[100] B. Peeters. System dentification and damage detection in civil engineering. PhD thesis,
dept. of Civil Engineering, KULeuven, Leuven, Belgium, december 2000.
[101] J. W. Pendered. Theoretical investigation into the effects of close natural frequencies in
resonance testing. Journal of Mechanical Engineering Science, 7(4):372379, 1965.
[102] A. W. Phillips, R. J. Allemang, and W. A. Fladung. The complex mode indicator function
(CMIF) as a parameter estimation method. In Proceedings of the 16th International Modal
Analysis Conference, pages 705710, Santa Barbara, CA, USA, 1998. SEM.
[103] A. W. Phillips, R. J. Allemang, and A. T. Zucker. A new excitation method: combining
burst random excitation with cyclic averaging. In Proceedings of the 16th International
Modal Analysis Conference, Santa Barbara, CA, USA, 1998. SEM.
[104] A. W. Phillips, A. T. Zucker, and R. J. Allemang. Frequency resolution effects on FRF
estimation: Cyclic averaging vs. large block size. In Proceedings of the 17th International
Modal Analysis Conference, Kissimmee, FL, USA, 1999. SEM.
[105] R. Pintelon, P. Guillaume, Y. Rolain, J. Schoukens, and H. Van Hamme. Parametric
identification of transfer functions in the frequency domain - a survey. IEEE Transactions
an Automatic Control, 39(11):22452260, November 1994.
[106] R. Pintelon, P. Guillaume, and J. Schoukens. Measurement of noise (cross-) power spectra
for frequency-domain system identification purposes: large-sample results. IEEE Transactions on Instrumentation and Measurement, 45(1):1220, February 1996.
[107] R. Pintelon, P. Guillaume, G. Vandersteen, and R. Rolain. Analysis, development and
applications of TLS algorithms in frequency domain system identification. Recent advances
in total least squares techniques and errors-in-variables modeling, pages 341358, 1997.
[108] R. Pintelon, P. Guillaume, G. Vandersteen, and Y. Rolain. Analysis, development and
application of TLS algorithms in frequency domain system identification. Siam Journal
on Matrix Analyses and Applications, 19(4):9831004, 1998.
[109] R. Pintelon and J. Schoukens. System identification: A frequency domain approach. IEEE
Press, New Jersey, USA, 2001.
[110] R. Pintelon, J. Schoukens, and P. Guillaume. Parametric frequency domain modeling in
modal analysis. Mechanical Systems and Signal Processing, 3(4):389403, 1989.
[111] R. Pintelon, J. Schoukens, and G. Vandersteen. Frequency-domain system identification
using arbitrary signals. IEEE Transactions on Automatic Control, 42(12):17171720, 1997.
[112] R. Prony. Essai experimental et analitique sur les lois de la dilatabilite des fluides elastiques
et sur celled de la force expansive de la vapeur de leau et de la vapeur de lalkool, a
differentes temperatures. Journal de Lecole Polytechnique, Vol 1(Cahier 2, Floreal et
Prairial, An III):2476, February 1795.
Bibliography
227
[113] M. H. Richardson and D. L. Formenti. Parameter estimation from frequency response

measurements using rational fraction polynomials. In Proceedings of the 1st International
Modal Analysis Conference, pages 167181, Orlando, USA, 1982.
[114] M. H. Richardson and D. L. Formenti. Global curve fitting of frequency response measurements using the rational fraction polynomial method. In Proceedings of the 3rd International Modal Analysis Conference, pages 390397, Orlando, USA, 1985.
[115] J. Rissanen. Modeling by shortest data description. Automatica, 14:465471, 1978.
[116] P.A. Rizzi. Microwave Engineering: Passive Circuits. Prentice Hall, 1988.
[117] Y. Rolain, R. Pintelon, K.Q. Xu, and H. Vold. Best conditioned parametric identification
of transfer function models in the frequency domain. IEEE Transactions an Automatic
Control, 40(11):19541960, November 1995.
[118] Y. Rolain, J. Schoukens, and R. Pintelon. Order estimation for linear time-invariant systems using frequency domain identification methods. IEEE Transactions an Automatic
Control, 42(10):14081417, October 1997.
[119] C. K. Sanathanan and J. Koerner. Transfer function synthesis as the ratio of two complex
polynomials. IEEE Transactions Automatic Control, 9(1):5658, 1963.
[120] P. Sas. Relation between acoustic intensity and modal deformaton patterns of vibrating
structures. Academiae Analecta, 46(4), 1984. Royal Academy of Sciences of Belgium.
[121] J. Schoukens, T. Dobrowiecki, and R. Pintelon. Parametric and non-parametric identification of linear systems in the presence of nonlinear distortions. IEEE Transactions an
Automatic Control, 43(2):176191, 1998.
[122] J. Schoukens and R. Pintelon. Identification of linear systems. A practical guide to accurate
modeling. Pergamon Press, London, UK, 1991.
[123] J. Schoukens, R. Pintelon, and H. Van Hamme. Identification of linear dynamic systems
using piecewise constant excitations: use, misuse and alternatives. Automatica, 30(7):1153
1169, 1994.
[124] J. Schoukens and J. Renneboog. Modeling the noise influence after a discrete Fourier
transform. IEEE Transactions on Instrumentation and Measurement, 35(3):278286, 1986.
[125] J. Schoukens, Y. Rolain, F. Gustaffson, and R. Pintelon. Fast calculation of least-squares
estimates for system identification. In Proceedings of the 32nd Decision and Control Conference, pages 34083410, Tampa, FL, USA, 1998.
[126] J. Schoukens, Y. Rolain, and R. Pintelon. Improved frequency response function measurements for random noise excitations. IEEE Transactions on Instrumentation and Measurement, 47(1):322326, February 1998.
[127] J. Schoukens, Y. Rolain, G. Simon, and R. Pintelon. Fully automated spectral analysis of
periodic signals. In Proceedings of the IEEE Instrumentation and Measurement Technology
Conference, Anchorage, AK, USA, 2002.
[128] M. D. Sidman, F. E. DeAngelis, and G. C. Verghese. Parametric system identification on logarithmic frequency response data. IEEE Transactions on Automatic Control,
36(9):10651070, 1991.
[129] T. S
oderstr
om. On model structure testing in system identification. Automatica, 11:537
541, 1975.
[130] T. S
oderstr
om. Tests of pole-zero cancellation in estimated models. Automatica, 11:537
541, 1975.
[131] T. S
oderstr
om and P. Stoica. System identification. Prentice Hall, Hemel Hempstead,
Hertfordshire, UK, 1989.
[132] A.S.J. Swamidas and Y. Chen. Monitoring crack growth through the change of modal
parameters. Journal of Sound and Vibration, 186(2):325343, 1995.
228
Bibliography
[133] H. Van der Auweraer. Development and evaluation of advanced measurement methods
for experimental modal analysis. PhD thesis, dept. PMA, KULeuven, Leuven (Belgium),
1987.
[134] H. Van der Auweraer and P. Guillaume. A maximum likelihood parameter estimation
technique to analyse multiple input - multiple output flutter test data. In Structures
and Materials Panel Specialists Meeting on Advanced Aeroservoelastic Testing and Data
Analysis, page paper no. 12, Rotterdam, The Netherlands, 1995. AGARD.
[135] H. Van der Auweraer and J. Leuridan. Multiple input orthogonal polynomial parameter
estimation. Mechanical Systems and Signal Processing, 1(3):259272, 1987.
[136] H. Van der Auweraer, W. Leurs, P. Mas, and L. Hermans. modal parameter estimation from
inconsistent data sets. In Proceedings of the 19th International Modal Analysis Conference,
pages 763771, San Antonio, Texas, USA, 2000.
[137] H. Van Hamme and R. Pintelon. Application of the bootstrapped total least squares
(BTLS) estimator in linear system identification. Signal Processing VI: Theories and
Applications, pages 731734, 1992.
[138] S. Van Huffel, editor. Recent advances in total least squares techniques and errors-invariables modeling. SIAM, Philadelphia,PA, USA, 1997.
[139] S. Van Huffel and J. Vandewalle. The total least squares problem: computational aspects
and analysis. SIAM, Philadelphia, PA, USA, 1991.
[140] P. Van Overschee and B. De Moor. Continuous-time frequency domain subspace system
identification. In Proceedings of 13th Triennal World Congress, San Francisco, CA, USA,
1996.
[141] P. Van Overschee and B. De Moor. Subspace identification for linear systems. Kluwer
Academic Publishers, Dordrecht, The Netherlands, 1996.
[142] P. Vanhonacker. The use of modal parameters of mechanical structures in sensitivity
analysis, system synthesis and system identification methods. PhD thesis, dept. PMA,
KULeuven, Leuven (Belgium), 1982.
[143] S. Vanlanduit. High Spatial Resolution Experimental Modal Analysis. PhD thesis, vakgroep
ELEC, VUB, Brussel, (Belgium), may 2001.
[144] S. Vanlanduit, P. Guillaume, J. Schoukens, and P. Verboven. An automatic frequencydomain modal parameter estimation algorithm. In Proceedings of the International Conference on Structural System Identification, Kassel, Germany, 2001.
[145] P. Verboven, P. Guillaume, and M. Van Overmeire. Modal parameter identification : estimation of starting values for MLE-like algorithms. In Proceedings of the 23rd International
Seminar on Modal Analysis, pages 409417, Leuven, Belgium, 1998.
[146] P. Verboven, E. Parloo, P. Guillaume, and M. Van Overmeire. Autonomous modal parameter estimation based on a statistical frequency domain maximum likelihood approach.
In Proceedings of the 19th International Modal Analysis Conference, pages 15111517,
Kissimmee, FL, USA, 2001.
[147] 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 (in press), 2002.
[148] H. Vold, L. Kundrat, G. T. Rocklin, and R. Russel. A multi-input modal estimation
algorithm for mini-computers. SAE Technical Paper series, 820194, 1982.
[149] 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 Transactions
on Audio and Electroacoustics, 15:7073, June 1967.
[150] S. Woodard and R. Pappa. Development of structural identification accuracy indicators using fuzzy logic. In Proceedings of the Design Engineering Technical Conferences (DETC),
pages DETC97/VIB4258, Sacramento, CA, USA, 1997. ASME.
[151] L. A Zadeh. Fuzzy logic toolbox for use with MATLAB. The MathWorks Inc., 1995.
[152] O. C. Zienkiewicz and R. L. Taylor. The finite element method. MacGrawHill, 1986.

Frequency-Domain System Identification For Modal Analysis

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

Frequency-Domain System Identification For Modal Analysis

Uploaded by

Copyright:

Available Formats

BR

VRIJE UNIVERSITEIT BRUSSEL

FREQUENCY-DOMAIN SYSTEM IDENTIFICATION

Proefschrift ingediend tot

VRIJE UNIVERSITEIT BRUSSEL

FREQUENCY-DOMAIN SYSTEM IDENTIFICATION

Proefschrift ingediend tot

c Vrije Universiteit Brussel Faculteit Toegepaste Wetenschappen

Pleinlaan 2, B-1050 Brussel (Belgium)

eigenvalue of a square matrix A

var(x) = E{|x E{x}|2 }

generalized transform variable

AutoRegressive Moving Average

Prediction Error Method

Focus and Outline of the Thesis . . . . . . . . . . . . . . . . . . . .

2 FRF Identification in the Presence of Measurement Noise

Errors-in-Variables Stochastic Framework . . . . . . . . . . . . . .

Hgtls a Maximum Likelihood Approach . . . . . . . . . . . . . . .

Practical Issues of EMA . . . . . . . . . . . . . . . . . . . . . . . .

Effects of Environmental Noise . . . . . . . . . . . . . . . .

FRF identification for EMA . . . . . . . . . . . . . . . . . . . . . .

Hev estimator Hiv Using Periodic Excitation . . . . . . .

Effects of Non-linear Distortions on FRF Identification . . . . . . .

Appendix A The GTLS Estimator per Output . . . . . . . . . .

3 FRF Identification in the Presence of Leakage

Simulated Vibration Experiment . . . . . . . . . . . . . . . . . . .

Power Spectra Estimators . . . . . . . . . . . . . . . . . . . . . . .

Exponential Windowing of FRF Estimates . . . . . . . . . . . . . .

Hexp FRF Estimator . . . . . . . . . . . . . . . . . . . . . .

Nonparametric Least Squares IRF Estimator . . . . . . . .

Iterative Nonparametric FRF Estimator . . . . . . . . . . . . . . .

Cyclic Averaging Approach . . . . . . . . . . . . . . . . . . . . . .

Experimental Case Study . . . . . . . . . . . . . . . . . . . . . . .

4 Introduction to Modal Parameter Estimation

Overview of MPE Algorithms . . . . . . . . . . . . . . . . . . . . .

5 Frequency-domain MPE from FRF Data

Parametric model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

(Weighted) Linear Least Squares . . . . . . . . . . . . . . . . . . . 102

LS Formulation based on Jacobian Matrix . . . . . . . . . . 102

LS Formulation based on Normal Matrix . . . . . . . . . . 107

Optimal LS Implementation for MPE . . . . . . . . . . . . 109

Transformation to Modal Model . . . . . . . . . . . . . . . 124

Constructing Stabilization Charts

LSCF Estimator . . . . . . . . . . . . . . . . . . . . . . . . 129

LS Residue Estimator . . . . . . . . . . . . . . . . . . . . . 130

Comparison of LSCF with LSCE . . . . . . . . . . . . . . . 131

Weighted Generalized Total Least Squares . . . . . . . . . . . . . . 133

Cost Function Formulation . . . . . . . . . . . . . . . . . . 133

Generalized Eigenvalue Problem Formulation . . . . . . . . 135

Maximum Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . 137

Parameter Uncertainty Bounds . . . . . . . . . . . . . . . . 140

ML Stabilization Chart . . . . . . . . . . . . . . . . . . . . 141

Choice of Frequency Weighting . . . . . . . . . . . . . . . . . . . . 142

Validation of Stochastic Estimators for Experimental Data . . . . . 148

Asymptotical Properties of Stochastic Estimators . . . . . . . . . . 148

5.10 Appendix A Jacobian and Normal Matrix based LS Algorithms . 152

Errors-in-Variables Noise Model . . . . . . . . . . . . . . . . . . . . 158

(Weighted) Linear Least Squares . . . . . . . . . . . . . . . . . . . 158

Weighted Generalized Total Least Squares . . . . . . . . . . . . . . 161

Cost Function Formulation . . . . . . . . . . . . . . . . . . 161

Generalized Eigenvalue Problem Formulation . . . . . . . . 161

Validation of Linear Approximation . . . . . . . . . . . . . 164

Validation of (WG)TLS for Experimental Data . . . . . . . . . . . 167