Dep Mohanty 20050110

Operational Modal Analysis in the
Presence of Harmonic Excitations
Proefschrift
ter verkrijging van de graad van doctor

aan de Technische Universiteit Delft,
op gezag van de Rector Magnificus prof. dr. ir. J.T. Fokkema,
voorzitter van het College voor Promoties,
in het openbaar te verdedigen op maandag 10 januari 2005 om
10:30 uur
door Prasenjit MOHANTY,
Bachelor in Technology, Indian Institute of Technology,
Kharagpur, India,
geboren te Balasore, India
Dit proefschrift is goedgekeurd door de promotor:
Prof. dr. ir. D.J. Rixen, M.Sc.
Samenstelling promotiecommissie:
Prof. dr. ir. D.J. Rixen, Technische Universiteit Delft, promotor
Prof. Dr.-Ing. M. Link, Universität Kassel, Duitsland
Prof. dr. ir. A. de Boer, Universiteit Twente
Prof. dr. ir. M. Steinbuch, Technische Universiteit Eindhoven
Prof. dr. ir. P.M.J. Van den Hof, Technische Universiteit Delft
Prof. dr. ir. J.A. Mulder, Technische Universiteit Delft
Prof. dr. ir. R. de Borst, Technische Universiteit Delft
iii
ACKNOWLEDGEMENTS
Since November 2000, I have enjoyed being a member of Engineering Mechanics
group of the Mechanical Engineering and Marine Technology department, Delft
University of Technology. I am very grateful to all my colleagues for creating
such a wonderful atmosphere for research. I particularly appreciated the timely
help of Jos van Driel and Jan Sterk in setting up for the experiments. I want to
thank Nadine Conza, Ester Vlaanderen, Corinne du Burck, Cornelis van’t Hof,
Dr. Paul van Woerkom, Prof. Leo Ernst and Dr. Arend L. Schwab for their
continuous support during my stay in Delft.
I want to express my sincere gratitude to Prof. Daniel J. Rixen for his
continuous innovating and challenging supervision. It was the luckiest moment
in my career to have him as my supervisor as he is such a wonderful person as
well as a wonderful professor.
I also want to thank Dr. Thomas G. Carne of Sandia National Laboratories,
Albuquerque for his valuable suggestion during my research.
Delft, January 2005

Prasenjit Mohanty
vii
Samenvatting
De stabiliteit, betrouwbaarheid en levensduur van een constructie wordt voor
een belangrijk deel bepaald door het dynamisch gedrag van de constructie.
Daarom is het in veel geavanceerde technische toepassingen essentieel om: de
dynamica van nieuwe ontwerpen door simulaties te kunnen voorspellen, de dy-
namische kenmerken van prototypen te kunnen identificeren en bestaande syste-
men te kunnen controleren. Operationele Modale Analyse (OMA) is een prak-
tische methode om de dynamische eigenschappen van bestaande constructies
tijdens bedrijf te onderzoeken. Het staat namelijk toe eigenparameters van con-
structies te identificeren onder willekeurige excitatie, dit terwijl de krachtinput
niet kan worden gemeten zoals vaak het geval is tijdens bedrijf.
Nochtans hebben de OMA methode beperkingen wanneer deze wordt toegepast
op praktische gevallen. Één van die beperking van OMA is dat de opgelegde ex-
citatie aan het systeem een stochastische witte ruis moet zijn. Dit betekent dat,
als er naast de witte ruis ook harmonische belastingen aanwezig zijn, de stan-
daard OMA procedure niet kan worden toegepast. Bij de identificatie kunnen de
extra harmonische responsies soms als virtuele modes worden beschouwd, maar
wanneer de harmonische excitatie frequenties dicht bij eigenfrequenties van de
constructies liggen is de OMA methode onbruikbaar geworden.
In dit proefschrift worden drie nieuwe procedures gepresenteerd om de toepas-
baarheid van de standaard OMA methode uit te breiden. Indien de onbekende
harmonische krachten kunnen worden samengevoegd tot een combinatie van sto-
chastische witte ruis en harmonische excitatie kunnen, gebaseerd op het principe
van de ”Natural Excitation Technique” (NExT), de drie procedures worden
toegepast om de modale parameters van structuren te bepalen. Een dergelijke
combinatie stochastische en harmonische excitatie is een vrij algemeen verschi-
jnsel: bij voorbeeld de excitatie van de wind op de wieken van een windmolen
en het akoestische lawaai in een auto kunnen worden beschouwd als de som van
een harmonische en een stochastische excitatie.
In dit proefschrift worden drie identificatietechnieken besproken die een aan-
passing zijn van de standaard methode zodanig dat bij de identificatie rekening
kan worden gehouden met harmonische responsiecomponenten die een bekende
frequentie hebben. In het bijzonder is de efficiency bekeken van de nieuwe pro-
cedures welke de eigenfrequenties, de modale demping en de trillingsvormen
bepalen.
De belangrijkste en originele bijdragen van dit proefschrift zijn:
• In hoofdstuk 3 wordt een gewijzigd Least Square Complexe Exponentiële

(LSCE) methode gepresenteerd, welke de toepassing van OMA bij har-
monische excitatie behandelt met als ingangssignaal de auto- en kruiscor-
relatie van de responsies.
• In overeenstemming met de gewijzigde LSCE wordt in hoofdstuk 4 een
gewijzigde ”Ibrahim Time Domain” (ITD) methode gepresenteerd. Deze
viii
methode is gebaseerd op de ”Single Station Time Domain” (SSTD) meth-

ode en is in het bijzonder geschikt voor de OMA methode in de aan-
wezigheid van harmonische excitaties.
• In hoofdstuk 5 wordt een gewijzigde ”Eigensystem Realization Algorithm”
(ERA) methode voorgesteld om efficiënt de OMA methode te kunnen
toepassen bij de aanwezigheid van harmonische excitaties. In tegenstelling
tot de eerste twee methodes worden bij deze derde methode de modale pa-
rameters niet direct geı̈dentificeerd maar eerst wordt de lineaire systeem-
matrix bepaald waaruit de eigenparameters dan kunnen worden berekend.
• In hoofdstuk 6 wordt een methode voorgesteld om de trillingsvormen te
bepalen via de OMA methode in de aanwezigheid van harmonische exci-
taties. De modale parameters zoals bepaald met de methode van Hoofd-
stuk 3, 4 of 5 zijn een startpunt voor deze berekeningen.
De voorbeelden die in dit proefschrift worden gepresenteerd (een scharnierend

ondersteunde balk en een vrije plaat) hebben aangetoond dat de voorgestelde
methodes de nauwkeurigheid van de identificatie aanzienlijk verbeteren in vergelijk-
ing met standaardtechnieken. Het toepassen van dezelfde technieken op meer
complexe constructies zal waarschijnlijk dezelfde verbetering laten zien. De
toepassing van de gepresenteerde methode en technieken op bestaande indus-
triële constructies viel echter buiten de inhoud van dit proefschrift.
ix
Summary
The dynamic behavior of structures is a primordial factor determining its struc-
tural stability, reliability and lifespan. Therefore, in many advanced engineering
applications, it has become essential to predict the dynamics of new designs
through simulations, to identify the dynamic characteristics of prototypes and
finally to monitor systems in operation. Operational Modal Analysis (OMA)
provides an interesting manner to investigate the dynamic properties the struc-
ture while in operation. Indeed it allows identifying eigenparameters of struc-
tures excited randomly and for which the force inputs cannot be measured, as
is often the case in operation.
However OMA methods have limitations when applied to practical cases.
One limiting constraint of OMA is that the non-measured excitation to the sys-
tem in operation must a stochastic white noise. This implies that if harmonic
excitations are present in addition to random forces, the standard OMA pro-
cedures cannot be applied in a straightforward way. The harmonic response
components can sometimes be considered as virtual modes in the identification,
but when the harmonic excitation frequencies are close to eigenfrequencies the
standard OMA approaches may break down.
In this thesis, three new procedures are proposed to extend the applicability
of the standard OMA methods. Based on the principle of Natural Excitation
Technique (NExT), all the procedures can be used to extract modal parameters
of structures in operation condition when the unknown input forces can be
assimilated to a combination of stochastic white noise and harmonic excitation.
Such combination of random and harmonic excitations are quite common in
real life: wind excitations and acoustic noise for instance can be considered
as stochastic and harmonic excitation forces can arise due for instance to the
presence of rotating components.
This thesis discusses three identification algorithms adapted from standard
methods in order to explicitly take into account harmonic response components
of known frequency. In particular we have investigated the efficiency of the
new procedures to properly identify eigenfrequencies, modal damping and mode
shapes.
The major and original contributions of this thesis are:
• In chapter 3, a modified Least Square Complex Exponential (LSCE) method

is proposed, which handles as input the auto- and cross-correlation of the
response signals to perform OMA in the presence of harmonic excitations.
• Similar to the modified LSCE, in chapter 4, a modified Ibrahim Time

Domain (ITD) method, based on the Single Station Time Domain (SSTD)
method is proposed which also performs OMA in the presence of harmonic
excitations.
• In chapter 5, a modified Eigensystem Realization Algorithm (ERA) method

is presented to perform efficiently OMA in the presence of harmonic exci-
tations. Unlike the first two methods, this third approach does not directly
x
identify modal parameters but it first identifies the linear system matrix
from which eigenparameters are then found.
• In chapter 6, a method has been proposed to identify mode shapes by

OMA in the presence of harmonic excitations. In this chapter modal
parameters computed by the methods presented in the previous chapters
are used as starting point to derive mode shapes.
The examples used in this thesis work (a simply supported beam and a free
plate) have shown that the proposed methods significantly improve the accu-
racy of the identification compared to standard OMA techniques and therefore
applying them to complex structures is expected to yield similar improvements.
Nevertheless application to industrial case studies was beyond the scope of the
present thesis.
Contents
Summary in Dutch vii
Summary in English ix
1 Introduction 1
1.1 Research context . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Scientific challenge and thesis topic . . . . . . . . . . . . . . . . . 7
1.3 Outline and structure of the thesis . . . . . . . . . . . . . . . . . 8
2 Operational Modal Analysis 11

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 OMA: an overview . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 The Natural Excitation Technique (NExT) . . . . . . . . . . . . 14
2.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.2 Modal analysis of linear structures . . . . . . . . . . . . . 14
2.3.3 Derivation of correlation functions as impulse response
functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 OMA in the presence of harmonic excitation . . . . . . . . . . . . 20
2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3 Modified LSCE method for harmonics 23

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 Least Square Complex Exponential method . . . . . . . . . . . . 24
3.3 Modification of LSCE method to Include Harmonic Components 26
3.4 Experimental example . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4.1 Pure stationary white noise excitation . . . . . . . . . . . 28
3.4.2 Single sine harmonic at 23.225Hz . . . . . . . . . . . . . . 29
3.4.3 Two square shaped harmonics . . . . . . . . . . . . . . . . 30
3.4.4 Three additional harmonics . . . . . . . . . . . . . . . . . 32
3.4.5 Robustness of the modified LSCE . . . . . . . . . . . . . 33
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.6 Possible further developments . . . . . . . . . . . . . . . . . . . . 37
xi
xii CONTENTS
4 Modified ITD method for harmonics 41

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2 A variant of the ITD . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 A SIMO-SSTD method . . . . . . . . . . . . . . . . . . . . . . . 44
4.3.1 Modified SIMO-SSTD for harmonics . . . . . . . . . . . . 45

4.4 Experimental examples . . . . . . . . . . . . . . . . . . . . . . . . 47
4.4.2 Two additional harmonics . . . . . . . . . . . . . . . . . . 48
4.4.4 Robustness of the modified SIMO-SSTD . . . . . . . . . . 51
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5 Modified ERA method for harmonics 57

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.2 OMA and modified ERA . . . . . . . . . . . . . . . . . . . . . . . 58
5.2.1 Basics of the ERA algorithm . . . . . . . . . . . . . . . . 58
5.2.2 Modified ERA for harmonics . . . . . . . . . . . . . . . . 61
5.3 Experiment and Signal Processing . . . . . . . . . . . . . . . . . 64
5.3.2 Two additional harmonics . . . . . . . . . . . . . . . . . . 66
5.3.4 Robustness of the modified ERA method . . . . . . . . . 69
5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6 Mode Shapes in the Presence of Harmonics 75

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.2 Operational modal analysis in the presence of harmonics . . . . . 76
6.2.1 Extraction of mode shapes . . . . . . . . . . . . . . . . . 76
6.3 Steel plate test case . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.3.1 Pure noise excitation . . . . . . . . . . . . . . . . . . . . . 79
6.3.2 Three additional harmonics with the standard LSCE . . . 81
6.3.3 Three additional harmonics with modified LSCE . . . . . 82
6.3.4 Discussion of results . . . . . . . . . . . . . . . . . . . . . 83
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
7 Conclusions 89
7.1 Summary and main contribution . . . . . . . . . . . . . . . . . . 89
List of Figures
1.1 Examples of structures in operating condition. bridge (a), aero-

plane (b), offshore oil platform (c), rocket (d) . . . . . . . . . . . 3
1.2 Organization of the text. . . . . . . . . . . . . . . . . . . . . . . . 9
2.1 Peak picking method for modal analysis . . . . . . . . . . . . . . 13

2.2 Sketch of the NExT method for doing OMA . . . . . . . . . . . . 15
3.1 (a)Experimental setup for the beam structure. (b) slider arrange-
ment at one end . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2 PSD of acceleration at position 1 for pure random excitation and
stability diagram for the polyreference LSCE method. . . . . . . 29
3.3 PSD of acceleration at position 1 for a combined single harmonic
and random excitation, and stabilization diagram of the polyref-
erence LSCE method . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.4 PSD of acceleration at position 1 for a combined single harmonic
and random excitation, and stabilization diagram of the modified
single reference LSCE . . . . . . . . . . . . . . . . . . . . . . . . 32
3.5 PSD of the acceleration at position 1 for random loads and two
additional periodic square excitations, and stabilization diagram
from the polyreference LSCE method . . . . . . . . . . . . . . . . 34
additional periodic square excitations, and stabilization diagram
from the modified single reference LSCE method . . . . . . . . . 35
3.7 PSD of the acceleration at position 1 for random loads and three
additional harmonic excitations, and stabilization diagram from
the polyreference LSCE method . . . . . . . . . . . . . . . . . . . 37
additional harmonic excitations, and stabilization diagram from
the modified single reference LSCE method . . . . . . . . . . . . 38
3.9 Identified eigenfrequency (a) and damping ratio (b) for the second
ω −ω bending
bending mode as a function of harmonic
ω bending . . . . . . . . . 40
3.10 Identified modal frequency (a) and damping ratio (b) as a func-
tion of the harmonic frequency specified in the modified LSCE. . 40
xiii
xiv LIST OF FIGURES
4.1 Experimental Setup of a Plate . . . . . . . . . . . . . . . . . . . . 48

4.2 Stability diagram for white noise excitation only computed by
the SIMO-SSTD method . . . . . . . . . . . . . . . . . . . . . . . 49
LSCE method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
additional harmonic excitations, and stabilization diagram for the
SIMO-SSTD (4.10) and for the SIMO-SSTD modified to explic-
itly include harmonic components (4.18) . . . . . . . . . . . . . . 52
additional harmonic excitations, and stabilization diagram ob-
tained from the SIMO-SSTD and the modified SIMO-SSTD. . . 54
4.6 Eigenfrequency vs. assumed harmonic frequency in the modified
SIMO-SSTD (harmonic excitations at 274Hz and 276Hz) . . . . . 55
4.7 Damping vs. assumed harmonic frequencies in the modified SIMO-
SSTD (harmonic excitations at 274Hz and 276Hz). . . . . . . . . 55

the ERA method . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2 Stability diagram for white noise loading and additional harmonic
excitations at 23.30 Hz and 23.80 Hz (standard and modified ERA
methods). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3 Stability diagram for white noise loading and additional harmonic
excitations at 23.30 Hz, 23.80 Hz and 24.30Hz (standard and
modified ERA methods). . . . . . . . . . . . . . . . . . . . . . . . 69
5.4 Identified eigenfrequency for the second bending mode as a func-
ω −ω bending
tion of the Frequency Resolution(FR), F R = harmonic ω bending . 71
5.5 Eigenfrequency and damping ratio identified by the modified ERA
method when the assumed harmonic frequency is different from
the true harmonic frequency (23.25 Hz) . . . . . . . . . . . . . . 72
6.1 Experimental setup for a steel-plate placed in free-free condition,

excited by a shaker . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.2 Stability diagram for the steel-plate while it is excited by the
white-noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.3 Mode shapes identified in the presence of pure stationary noise
and taken as reference . . . . . . . . . . . . . . . . . . . . . . . . 86
additional periodic excitations, stabilization diagram from the
standard LSCE method and the modified LSCE method . . . . . 87
6.5 Identified harmonic responses corresponding to harmonic frequency
144 Hz and 148 Hz. The harmonic response at 146 Hz is very
similar. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.6 First mode shape computed by the modified LSCE method for
the example, where three harmonics are present . . . . . . . . . . 88
List of Tables
3.1 Frequencies and associated damping (white noise only) . . . . . . 30

3.2 Frequencies and associated damping identified by the polyref-
erence LSCE and the modified single reference LSCE methods
(random and single harmonic excitation). . . . . . . . . . . . . . 33
3.3 Frequencies and associated damping for both polyreference LSCE
and single reference LSCE method (two additional periodic exci-
tations with square waveform) . . . . . . . . . . . . . . . . . . . . 36
3.4 Frequencies and associated damping from the polyreference LSCE
and the modified single reference LSCE methods (three additional
harmonic excitations). . . . . . . . . . . . . . . . . . . . . . . . . 39
4.1 Frequencies and associated dampings (white noise only) . . . . . 51

4.2 Frequencies and associated damping identified by the SIMO-SSTD
and the modified SIMO-SSTD. . . . . . . . . . . . . . . . . . . . 53
4.3 Frequencies and associated damping identified by the SIMO-SSTD
and the modified SIMO-SSTD. . . . . . . . . . . . . . . . . . . . 54
5.1 Frequencies and associated dampings identified by ERA (white

noise excitation only) . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.2 Frequencies and associated damping identified by the standard
ERA and the modified ERA. System excited by stationary white
noise and harmonic loads at 23.30Hz and 23.80Hz. . . . . . . . . 68
ERA and the modified ERA. System excited by stationary white
noise and harmonic loads at 23.30Hz, 23.80Hz and 24.30Hz. . . . 70
6.1 Identified frequencies and damping by white-noise excitation . . 80

LSCE method in the presence of harmonic excitations at 144Hz,
146Hz and 148Hz . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.3 Poles used for identification of the eigenmodes and MAC values
between the reference modes and the modes identified by the
standard LSCE procedure applied to the signals with harmonics. 82
xv
xvi LIST OF TABLES
6.4 Frequencies and associated damping identified by the modified

LSCE method in the presence of harmonic excitations at 144Hz,
146Hz and 148Hz . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.5 Poles used for identification of the eigenmodes and MAC values
between the reference modes and the modes identified by the
modified LSCE procedure applied to the signals with harmonics. 84
Chapter 1
Introduction
In this general introduction we will outline the major issues investigated within
the present research. The research context is described in Sec. 1.1. The objective
of the thesis, as well as the main contributions, are discussed in Sec. 1.2. The
structure of the manuscript and the content of its chapters are discussed in Sec.
1.3
1
2 CHAPTER 1. INTRODUCTION
1.1 Research context

Dynamic design of a system aims at obtaining the desired dynamic character-
istics of the system by specifying the right shape, size, configuration, material
and manufacturing steps of its various parts. Desired dynamic characteristics
include low vibration and noise levels meaning for instance tuning its natural
frequencies to be well separated from the excitation spectrum. Other dynamic
properties that are often required for systems include high dynamic stability and
prescribed mode shapes or vibration patterns. Ultimately, the designer will have
as an objective to develop quieter and more comfortable products and increase
reliability and efficiency (ex: noise and vibrations in cars, reliability of long span
bridges, efficiency of loudspeakers, accuracy of wafer-steppers, avoid flutter of
aircrafts). While fulfilling the above mentioned requirements, the designer has
to look for solutions within specific economical constraints and in a highly evo-
lutive global market. Thus there is a need for lower design and development
time calling for advanced analysis approaches such as optimum dynamic design.
Sound dynamic design of real-life complex systems, like those mentioned
above, cannot be ensured only on the basis of numerical and analytical investi-
gations. Experimental Modal Analysis (EMA) of prototypes provides essential
information on dynamical properties of the system during the process of de-
signing, which is impossible to verify only by computer simulations. Sometimes
EMA is carried out only on the part of the system, which has been redesigned
and thus for which dynamic properties have to be identified [61]. This process is
quite standard in industry and, combined with advanced numerical simulation
techniques, helps to develop new products in a short time for the design cycle.
Modal analysis methods are based on the assumption that dynamic response
of a system can be totally represented by its eigenparameters. Response at any
point on the system is considered as the superposition of responses from the
individual modes of the system at which the system vibrates naturally.
During the last few years, EMA technology has evolved in a very efficient
and reliable method in the quest for accurate determination of dynamic prop-
erties of the structures. It has become a standard procedure in many engineer-
ing fields such as mechatronical design, transport technology, aeronautics and
aerospace, civil engineering, biomechanics and biomedical engineering, vibro-
acoustics, wind mills and microsystems. For all these applications, predicting
and analyzing the dynamical behavior is essential.
In this modern era, structural design in fields like aircrafts, satellites, cars
and buildings is characterized by the trend to become lighter, smaller, cheaper
and mulitifunctional while satisfying stringent safety and reliability criteria.
Availability of efficient optimization tools on one hand and ever increasing con-
sumer expectation on the other have lead to tremendous changes in the way
structures look today. The severe requirement on mechanical components have
resulted in unwanted vibrations. The multifunctionality of products has called
for the need to include multiphysical effects in the analysis.
The demand for better products and higher reliability of structures has in-
creased the necessity to properly account for the dynamics during design, man-
1.1. RESEARCH CONTEXT 3
Figure 1.1: Examples of structures in operating condition. bridge (a), aeroplane

(b), offshore oil platform (c), rocket (d)
ufacturing and operation of new concepts. Most commonly used finite element
(FE) analysis provides robust modelling capability to analyze a structure for its
dynamic and static properties thanks to the advent of modern high-performance
computer. However it is often seen that dynamic properties computed by the
FE analysis differ from the actual dynamic properties of the structure. The
difference between simulations and reality can be traced back to the following
reasons:
• FE analysis is based on a discretization of the reality, meaning that the

displacement fields are approximated by predefined shape functions within
each element. Further assumptions such as mass lumping are also used in
practice.
• Eventhough bulk properties of materials such as mass density and elastic-

ity coefficients are often well known, damping characteristics of material
is difficult to properly account for in simulation. Evenmore structural
damping which is related not only to material properties but also to dissi-
pation in joints or in interaction with the surroundings can not be properly
included in the models.
• The geometry of a specific project can significantly differ from what is

indicated on drawings and used for the FE model.
• Nonlinearity of the structures also remains as a source of uncertainties

while doing numerical modal analysis which by essence is based on lin-

earized models.
To fill the gap between reality and modelling, experimental techniques are
necessary in order to confront measurement with simulated results. Updating
techniques can then be applied to tune the models such that they represent real-
ity more accurately. Note that complete dynamic analysis has become common
practice thanks to
• advances in measurement techniques, sensors and signal processing,
• increase of computer power and decrease of computer costs which allows

computing detailed dynamical behavior based on complex and large mod-
els.
• the increase of computational power has also significantly broadened the

scope of experimental analysis by allowing to measure and process many
signals in real time. Obviously this is essential for health monitoring and
control of systems.
Determining the linear dynamic behavior of systems from the test is called
Experimental Modal Analysis (EMA). EMA procedures are based on the as-
sumption that the vibration response of a linear time-invariant dynamical sys-
tem can be expressed as the linear combination of a set of simple harmonic
motions called natural modes of vibration. It identifies dynamic properties of
the structure in terms of natural frequencies, damping ratios and mode shapes.
Natural modes of vibration are inherent to the dynamic system and are de-
termined completely by its physical properties (mass, stiffness and damping).
Response at any point on the structure can be considered as the sum of all modal
responses of each of its individual modes. Based on eigenparameters identified
from the modal analysis, a mathematical model of the dynamic behavior of the
structure can be constructed. Such procedures are known as model synthesis
and results in a modal model of the structure. A modal model of the structures
plays an important role in dynamical behavior prediction for the structure.
In EMA, structures are usually excited with one or more excitation sources
and responses of the structure are measured at one or more locations. Modal
parameters can be identified from the frequency response function (FRF) of the
response signal to the input signal or from the impulse response of the system.
In such a frequency domain approach, eigenmodes and eigenfrequencies are iden-
tified by fitting the measurement to a theoretical expression of the response as a
superposition of modes. Another way consists in analyzing time signals such as
impulse response and identify modal parameters by fitting a mode superposition
expression on the time data. Such a procedure is known as time-domain modal
identification.
In practice, within their operational environment, structures are excited nat-
urally due to external or internal excitations. Think of buildings excited by the
wind, bridges under the action of traffic or engine vibration on aircrafts or cars
1.1. RESEARCH CONTEXT 5
(see Figure 1.1). Such excitations can be considered as white-noise in nature.

Even though the input force to the system can generally not be measured in
such cases, it is still possible to apply modal analysis techniques to evaluate
some of the modal parameters of the structures. This modal analysis procedure
is called Operational Modal Analysis (OMA). OMA is also sometimes denoted
as ambient vibration analysis or output-only analysis.
Different methods have been developed to do the modal parameters estima-
tion using output-only data. The concepts underlying these methods are often
similar, but they differ in the aspects of data reduction, mathematical modelling
of equations or signal processing. Few methods work directly on the raw mea-
sured data while others require pre-processing of the measured response. OMA
methods can mostly be subdivided in the following categories:
• Methods performing auto- and cross-correlation of the time-signals and

handling the results as impulse response functions (the Natural Excitation
Technique or NExT [28–30]. Time domain identification methods are
then used to extract modal parameters from impulse responses of the
system. To mention a few: Least Square Complex Exponential (LSCE)
[9], Polyreference Complex Exponential method (PRCE) [62, 63], Ibrahim
Time Domain Method (ITD) [25–27], Single Station Time Domain method
[64] and Eigensystem Realization Algorithm (ERA) [31, 32].
• Methods based on discrete-time data models. Examples of such methods

are the Auto Regressive Moving Average Vector (ARMAV) [16–21] and
the Auto Regressive Vector Time Series Modelling [48].
• Stochastic subspace methods such as the Canonical Variate and Balance

Realization Techniques [4, 49].
• The Maximum Likelihood frequency domain method, which works on

auto-and cross-spectrum functions of the response signals [23, 50, 55].
In OMA procedures, like in classical EMA, natural frequencies, modes and

damping ratios are evaluated for a structure. However due to the absence of
input force to the system, modal participation factors can not be computed. Re-
cently however [53] a procedure was propose for deriving scaling factors based
on sensitivity analysis for the perturbed system. Typically the modal partici-
pation factors are found by performing OMA on the system with known added
masses.
In practice modal analysis is a succession of three steps:
1. Planning and executing the experiments: in the initial phase, experiments

must be planned based on information known at forehand such as fre-
quency range, flexibility distribution, accessibility of measurement points,
vibration nodes [1, 3] ... An important issue is to choose the boundary
condition to apply on the test object. Two types of boundary condition
can be used: free and fixed boundaries. Usually free surface boundary
condition can be achieved by hanging the structure from soft springs or

putting the structure on a very soft surface like foam. Fixed boundary
can be achieved by attaching one end of the structure to a heavy, highly
stiff external object. Note that for bulky structures or structures that
have to be tested in situ the boundary conditions are mostly imposed.
Proper locations and orientations of the actuators and sensor must then
be defined based on preliminary information one may have from previous
tests or from models. Data acquisition parameters like measurement time,
sampling frequencies must also be chosen properly depending on the ex-
perimental goals. In classical EMA, excitation is provided either by an
impact hammer or by a shaker. A hammer produces input in a broad
frequencies range whereas shakers are more often used to apply specific
excitation inputs in terms of prescribed amplitude and frequencies. In
OMA, excitation is inherently provided by uncontrolled sources like wind,
waves or acoustic noise.
2. Processing of the data and identification of modal parameters: During the
measurement, signals are measured in time domain. Output data are then
formatted according to the specific identification utilized. For frequency
domain identification methods FRF curves are derived using for instance
Fast Fourier Transforms (FFT) and windowing [3, 36]. For time domain
identification approaches, the time signal might be filtered or used as such
in the identification algorithms. Often the time-signal is transformed into
an impulse response before performing the identification. This can be
done by FFT and inverse FFT of the raw signal.
3. Validation of the modal models: In this process, modal parameters ex-
tracted from the measurements are visualized and analyzed. So-called
“numerical” modes have to be discarded.
Once the modal model has been found, it can be used in several different
ways. For an engineer it is a matter of interest to investigate different proper-
ties of a structure using a good modal model of the structure by experimental
analysis. Some of the applications where the modal model is used are:
• Troubleshooting: this is a very general application of modal analysis.
Structure’s vibrational properties like natural frequencies, damping ratios
and mode shapes measured from the experiments are used to understand
the dynamics behavior of the structure to find out the root cause of prob-
lems such as resonance and noise often encountered in real life [6, 61].
• Model updating: a mathematical model of a structure (e.g. using FE)
may not represent the correct dynamical behavior of the structure due to
the fact that material properties, geometry, joint properties and bound-
ary conditions for the model may not represent the real system given its
complexity and uncertainties arising during the production. Experimen-
tal modal analysis provides a tool to evaluate the modal parameters in
practice and helps in updating the FE model to make it more realistic [2].
1.2. SCIENTIFIC CHALLENGE AND THESIS TOPIC 7
• Structural modification and sensitivity analysis: structural modification

consists in predicting the dynamical behavior of the structure assuming
that structural modifications are brought on a nominal configuration. It
allows to evaluate the effect of changes on the dynamics without having
to actually modify the structure. This is particularly useful at the early
stage of a new design to optimize its dynamic characteristics starting from
a pre-existing system. The sensitivity analysis provides the rate of change
in dynamical behavior for specified small system perturbations [14, 60].
It thus provides the information necessary to guide the designer in his
decisions during the redesign phase.
• Structural health monitoring and damage detection: structural integrity
can be assessed by comparing modal parameters from the current state
of a structure with the modal parameters of the structures in nominal
conditions [4, 11, 51, 52]. Modal parameters can give indications on the
severity of the damage as well as on its location (using model updating
techniques).
• Performance evaluation: modal parameters and mode shapes can be valu-
able information to evaluate the dynamic performance of systems. For
instance the quality and performance of a tennis racquet (described in
terms of so-called sweet spots) are closely related to its eigenmodes (see
e.g. [39])
• Force identification: this is an inverse problem compared to the response
prediction from the known force input [10, 22]. Force can be identified by
measuring only the response of the structure, using a modal or FE model.
• Controller synthesis: a last common utilization of modal models is to
synthesize controllers. For instance in house appliance or in CD players,
the control loop is typically designed based on the characteristic of the
system such as measured in EMA.
As described here above, there are many important applications, where

modal parameters provide essential information to solve complex engineering
problems. So it is essential to perform the modal parameter estimation of a
structure as accurately as possible.
1.2 Scientific challenge and thesis topic

This thesis addresses an essential issue in Operational Modal Analysis, namely
accounting for non-random excitations that often appears in addition to white
noise. Indeed one major disadvantage of output-only identification is that the
input force to the system must be a stationary white noise while in practice
periodic excitations inherent to the system are also present.
Some examples of structures were standard OMA techniques are well-suited
are bridges, buildings, aeroplanes, spacecrafts, oil platforms or ships. In these
examples, the excitation forces during operation are caused by wind, waves and
other broadband excitations such as road roughness or jet propulsion noise. In
those cases, the input forces during operation can not be properly measured. On
the other hand exciting those structures with external excitations (e.g. shakers)
like in classical EMA is hardly feasible due to the fact that the structures are
very large and heavy or because they can not be accessed easily. Especially
when the dynamic properties are to be measured in operation (think about the
vibration modes of a wing during flight) applying external excitation is often
difficult, expensive and can even cause safety hazards. In this case, operational
modal analysis is a suitable method to estimate modal properties of these kind
of structures as input force to these systems are very difficult or often impossible
to measure.
In many other structures such as wind mills, milling machines, turbines or
cars, periodic exciting forces are present in addition to the white-noise equiva-
lent excitations. Harmonically exciting forces can arise from moving components
like the blades of an helicopter, the rotation frequency of turbines, teeth con-
tact frequencies in gears or milling processes or car engine orders. In theory,
for such problems, operational modal analysis techniques can still be used to
extract modal properties. Indeed considering that harmonic excitation gener-
ates harmonic frequencies in the response signals, the harmonic responses will
be identified during the OMA identification as non-damped modes with very
low damping (theoretically null), which distinguish them from true eigenmodes.
However such a procedure fails to identify eigenmodes when the harmonic fre-
quencies are close to the eigenfrequencies of the structures or when the strength
of the signal arising from harmonic excitations is very high compared to the
response associated with the eigenfrequencies of the structures excited through
the noise input.
Addressing the above drawback of OMA is primordial if OMA techniques
are to be used in a wide range of practical applications. To tackle this important
engineering and scientific challenge, three major novel methods are proposed in
this thesis work to enable OMA in the presence of harmonically exciting forces
in addition to stochastic force in a structure:
• a new method based on the Least Square Complex Exponential method

is proposed (Chapter 3 ).
• a new method related to the Ibrahim Time Domain method is described

(Chapter 4 ).
• a novel extension to the Eigensystem Realization Algorithm is outlined

in Chapter 5.
1.3 Outline and structure of the thesis

In this section a brief overview of the content of the different chapters is given
in a Chapter-by-Chapter outline. (see also Figure 1.2).
1.3. OUTLINE AND STRUCTURE OF THE THESIS 9
Figure 1.2: Organization of the text.
After the present chapter describing the general context and the scientific
challenges of our research, chapter 2 gives a short overview of OMA approaches
and in particular explains the procedure known as the Natural Excitation Tech-
nique (NExT). The NExT approach to OMA will be considered for the algo-
rithmic developments exposed in the subsequent chapters.
Chapter 3 discusses OMA identification procedures based on the Least Square
Complex Exponential method (LSCE). First we recall the theory and procedure
of the standard LSCE algorithm. Then we propose a modified version of the
LSCE that explicitly accounts for harmonic excitations during the identifica-
tion. The effectiveness and robustness of the method is illustrated on a beam
structure.
In chapter 4 a second method is developed relying on the principle of the
Ibrahim Time Domain (ITD) method and on one of its variants, the Single
Station Time Domain technique (SSTD). First, the SSTD method is generalized
for multiple correlation signals considering several response channels. Then a
novel technique is described that modifies the SSTD-like algorithm to account
for harmonic excitations in OMA. A plate like structure is used as a testing
structure to show the performance of the modified ITD method in the presence
of harmonic excitations.
In chapter 5 a third approach to OMA identification is described, based on
the popular Eigensystem Realization Algorithm (ERA). The extension proposed
for the ERA allows to properly account for harmonic excitations of known fre-
quencies. Once more the beam structure as used in chapter 3 is used to show
the performance of the method.
In chapter 6 a procedure to evaluate mode-shapes is discussed. Assuming
that the modal parameters have been obtained by one of the methods described
in the previous chapters, we can use the standard modeshape identification
procedure in the time-domain based on the expression of responses in terms
of complex modal contributions. The mode identification post-processing is
illustrated on the plate structure in the presence of strong harmonic signals.
Modal Assurance Criteria (MAC) values are considered in order to compare the
modes identified with or without harmonic inputs.
Finally in chapter 7 we conclude the thesis by a highlighting the main results
of our research and by indicating relevant and challenging research direction for
the future.
Chapter 2
Operational Modal Analysis
In this chapter, the conceptual background of operational modal analysis is intro-

duced. In the section 2.2 different procedures for doing operational modal analy-
sis are shortly reviewed. In the section 2.3, the Natural Excitation Technique
(NExT) method is described in detail. The NExT method allows to combine
OMA techniques with standard algorithms for modal parameter identification in
the time domain for impulse responses. That is the analysis procedure that will
be investigated further in the following chapters.
11
12 CHAPTER 2. OPERATIONAL MODAL ANALYSIS
2.1 Introduction
Dynamic analysis of structures is an essential engineering tool since it allows to
construct a mathematical model of the structure based on measured modal para-
meters. The experimental results can also be compared to numerical models and
used for model updating. Reliable mathematical models obtained experimen-
tally or by updating numerical models are primordial for designers and engineers
since they can represent the dynamical behavior of the structure under various
dynamic loadings (earthquake, winds, explosions) and therefore allow analyzing
issues such as reliability, integrity or lifetime prediction of systems.
Experimental analysis can be performed in a traditional way, namely in
a laboratory environment. In standard Experimental Modal Analysis (EMA)
externally applied input forces are measured together with dynamic responses
of the system at different locations. The forces applied by the excitations should
be large enough to produce measurable response at the measuring points, but
yet small enough for the structure to be considered as a linear system. EMA
techniques for non-linear systems exist but are not yet mature and thus rarely
applied in practice.
Exciting large engineering structures is cumbersome due to the excitation
force levels required and do to accessibility issues. Simply exciting at one point
of the structure may not be enough. For these reasons, when measuring large
structures, multiple shakers at different locations need to be used, implying high
costs and long experimentation time.
Some structures experience vibration due to natural excitations. Sometimes
it becomes essential to test the dynamic properties of a structural system while
it is functioning in its operational environment in terms of its internal stress
state, its coupling with the environment or its thermal condition. In opera-
tional condition, it is often impossible to apply standard EMA since on one
hand applying excitations in a controlled way might be intricate and dangerous
and, on the other hand, the presence of unmeasurable internal and ambient
forces jeopardizes the application of identification techniques. When so-called
natural excitations are present during operation of a structural system, ambient
vibration testing or operational modal analysis (OMA) needs to be considered.
The main advantages of OMA are:
• Testing is cheap and fast since no excitation equipment is needed.
• Measurements are done in the actual operation conditions for the struc-
ture. Identified modal parameters represents the dynamic behavior of the
structure such as used in reality.
• Testing does not interfere with the operation of the structure meaning for
instance that a power generator need not to be laid still to measure its
turbine, or a bridge needs not to be closed to traffic when analyzed.
In this chapter we will briefly outline common OMA techniques and discuss
in particular the Natural Excitation Technique (NExT) that will be considered
throughout this thesis.
2.2. OMA: AN OVERVIEW 13
2.2 Operational modal analysis techniques:

an overview
Peak-Picking techniques
A simple way of doing modal analysis from the response-only data is based
on a Peak-Picking technique applied to the auto and cross-spectrum of the re-
sponse signal [5]. This is a Single-Degree-Of-Freedom (SDOF) method, which
involves selecting peaks in the spectra corresponding to resonance frequencies.
Peak-Picking can be a reliable approach when the modes of the system are well
separate and when the damping associated with each mode is low so that the in-
fluence each mode on the dynamical behavior can be considered independently.
An advanced version of Peak-Picking method is known as the Complex Mode
Indicator Function (CIMF) [57]. This method is based on singular value de-
composition (SVD) of the matrix of power spectra. The steps involved in the
Figure 2.1: Peak picking method for modal analysis
Peak-Picking method are:

• Individual peak regions are visually chosen and marked from the cross-
spectra between responses and a reference signal. In the vicinity of each
peak, the maximum is found and associated to the eigenfrequency of the
mode.
one below the rth eigenfrequency
• Two points are identified on the cross-spectra curves, √
ω r and one above, both having an amplitude Cωr / 2 where Cωr is the
amplitude at the natural frequency ω r . These two points are called half-
power points.
• The damping ratio ζ r of the associated mode is then computed by
ω 2a − ω 2b ∼ ∆ω
ζ= = (2.1)
4ω 2r 2ω r
Other methods for OMA

Several other methods based on discrete-time data models have been proposed.
The Maximum Likelihood Estimation(MLE) technique is a frequency domain
method [23,54,56]. Other methods include the Auto Regressive Moving Average
(ARMAV) [16–21], the Auto Regressive Vector Time Series Modelling [48] and
Stochastic Subspace methods such as the Canonical Variate and the Balance
Realization Techniques [4,49]. Describing those methods is beyond the scope of
the present manuscript. The reader is referred to the references for a detailed
description of those approaches.
2.3 The Natural Excitation Technique (NExT)
2.3.1 Introduction
The Natural Excitation Technique (NExT) [28–30] is a method using the mea-
sured time response to white noise excitation to construct an impulse-like re-
sponse signal of the structure. The NExT approach is based on the property
that auto- and cross-correlation functions between response signals can be used
as the impulse response functions of the system when the system is excited
by a white noise force. This property will be outlined in this section. Figure
2.2 illustrates this concept underlying the NExT procedure. This fundamen-
tal property pertains to linear systems that can be represented by the modes
of the associated conservative system, namely systems that are lightly damped
and which eigenfrequencies are well-separate [35]. Note that for linear systems
where modal coupling through damping can not be neglected and where com-
plex eigenmodes should be considered, the NExT method can, in theory, not be
used since no proof exists to relate signal correlations to impulse-like responses.
Using the fundamental property on which the NExT method is based, auto-
and cross-correlation functions of the responses to white noise can be handled as
being impulse response functions and the modal parameters of the system can be
identified using standard time domain identification methods like Least Square
Complex Exponential (LSCE) [9], the Ibrahim Time Domain(ITD) method [25–
27] and the Eigensystem Realization Algorithm (ERA) [32]. The time-domain
identification techniques will be discussed in later chapters.
2.3.2 Modal analysis of linear structures

Considering a Multiple-Degree-Of-Freedom (MDOF) system, the linear(ized)
equation of motion assuming viscous damping can be can be written as:
[M ] {ẍ (t)} + [C] {ẋ (t)} + [K] {x (t)} = {f } (2.2)

where
2.3. THE NATURAL EXCITATION TECHNIQUE (NEXT) 15
Figure 2.2: Sketch of the NExT method for doing OMA
[M ] mass matrix
[C] damping matrix
[K] stiffness matrix
{f } force vector
{x} displacement vector
{ẋ} velocity vector
{ẍ} acceleration vector
The solution of the dynamic equation can be written as a combination of

the modal basis [Φ] of the conservative system,
N
X
{x(t)} = {φr } qr = [Φ] {q (t)} (2.3)
r=1
where {q} is the vector of modal coordinates and where [Φ] is a matrix collecting
N eigenmodes of the system and satisfies the eigenvalue problem
([K] − ω 2r [M ]) {φr } = 0 (2.4)
The dynamic equation (2.2) then writes
[M ] [Φ] {q̈ (t)} + [C] [Φ] {q̇ (t)} + [K] [Φ] {q (t)} = {f } (2.5)
T
Pre-multiplying this equation by the mode shape matrix [Φ] to project the
dynamic equilibrium on the modal basis, we obtain
T T T T
[Φ] [M ] [Φ] {q̈ (t)} + [Φ] [C] [Φ] {q̇ (t)} + [Φ] [K] [Φ] {q (t)} = [Φ] {f } (2.6)
The eigenmodes [Φ] of the conservative system satisfy orthogonality with respect
to the mass and stiffness matrices, namely
T £r ¤ T £r ¤
[Φ] [M ] [Φ] = mr and [Φ] [K] [Φ] = kr (2.7)
where [r mr ] and [r kr ] are the modal mass and stiffness diagonal matrices
respectively. Note that, according to (2.4), the modal mass and stiffness are
related by kr /mr = ω 2r .
Assuming that the eigenfrequencies of the system are well separated and that
they are associated with low damping ratios, the real modes of the associated
conservative system can be assumed to be identical to the complex modes of
the damped system and we can assume that the eigenmodes [Φ] can be used to
diagonalize the damping matrix [C] namely [35, 36]
T £r ¤
[Φ] [C] [Φ] = cr (2.8)
where [r cr ] is the modal damping matrices. Finally the dynamic equations in

the modal basis are obtained from (2.6):
£r ¤ £ ¤ £ ¤ T
mr {q̈ (t)} + r cr {q̇ (t)} + r kr {q (t)} = [Φ] {f } (2.9)
(N ×N ) (N ×1) (N ×N ) (N ×1) (N ×N ) (N ×1) (N ×1)
Eq. (2.9) are known as the normal equations and can be written for a par-
ticular mode r as
T
mr q̈ r (t) + cr q̇r (t) + kr qr (t) = {φr } {f (t)} (2.10)
th
where {φr } is r mode shape vector. Dividing the equation (2.10) by the
rth modal mass mr , we get:
1 T
q̈r (t) + 2ζ r ω r qr (t) + ω 2r qr (t) =
{φ } {f (t)} (2.11)
mr r
where ζ r and ω r are the damping ratio and natural frequencies associated with
rth mode.
The solution of Eq. (2.11), assuming a general force {f } is obtained from
the convolution or Duhamel integral [58]:
Z t
T
qr (t) = {φr } {f (τ )} gr (t − τ ) dτ (2.12)
−∞
where gr (t) is the impulse response of a SDOF system
1 ¡ ¢
gr (t) = d
e−ζ r ωr t sin ω dr t (2.13)
mr ω r
¡ ¢1/2
and ω dr = ω r 1 − ζ 2r is the damped natural frequency.
Now the response vector {x (t)} can be formulated as the summation of
modal responses by using Eq. (2.12) and Eq. (2.3) as:
N
X Z t
T
{x (t)} = {φr } {φr } {f (τ )} gr (t − τ ) dτ (2.14)
r=1 −∞
Eq. (2.14) can now be written for a single response, xik (t) at a point i, due
to single input force, fk (t), at point k.
N
X Z t
xik (t) = φri φrk fk (τ ) gr (t − τ ) dτ (2.15)
r=1 −∞
The impulse response function between input k and output i is obtained by

assuming f (τ ) in Eq. (2.15) to be a unit Dirac delta function at τ = 0. So
the response at the point i due to an input impulse at k, can be expressed as a
summation of modal impulse responses:
XN
φri φrk −ζ r ωr t ¡ ¢
xik (t) = d
e sin ω dr t (2.16)
r=1
mr ω r
2.3.3 Derivation of correlation functions as impulse re-

sponse functions
In this section we derive the fundamental property used in the NExT method
such as outlined in the milestone publications [28–30].
Let us call Rijk (T ) the cross-correlation function between two responses xik
and xjk at i and j due to a white-noise input at a location k. By definition
Rijk (T ) is the expected value of the product between xik and xjk evaluated
with a time delay T , namely [59]
Rijk (T ) = E [xik (t + T ) xjk (t)] (2.17)

where E is the expectation operator. For measured signals, the correlations can
be computed as
Z
1 T /2
Rijk (T ) = lim xik (t)xjk (t − T )dt
T →∞ T −T /2
n
1X
' xik (tα + T ) xjk (tα ) (2.18)
n α=1
Substituting (2.15) into (2.17),

ÃN N Z t Z t+T !
XX
Rijk(T )=E φri φrk φsj φsk gr (t+T− σ) gs(t − τ ) fk(σ) fk(τ ) dσdτ
r=1 s=1 −∞ −∞
(2.19)
Let us note that the expectation of the product between a random function a(t)
and a non-random function b(t) satisfies [15]
E [a(t)b(t)] = b(t)E [a (t)] (2.20)
Using this property in (2.19) where fk (t) is a random function and x (t) is a
non-random time function, one obtains
N X
X N Z t Z t+T
Rijk (T ) = φri φrk φsj φsk gr (t+T− σ) gs(t − τ ) E [fk(σ) fk(τ )] dσdτ
r=1 s=1 −∞ −∞
(2.21)
Using the definition of the autocorrelation function (2.17) and assuming f (t)
to be white noise we have [59]
Rf f k (τ − σ) = E [fk (τ ) fk (σ)] = αk δ (τ − σ) (2.22)

where αk is a constant and δ (t) Dirac delta function.
Substituting (2.22) in (2.21) we get:
N X
X N Z t
Rijk (T ) = αk φri φrk φsj φsk gr (t + T − σ) gs (t − σ) dσ (2.23)
r=1 s=1 −∞
Equation (2.23) can be further simplified by making a change of the variable

for the integration variable: let λ = t − σ, then dλ = −dσ and the limits of
integration for λ are +∞ and 0. Equation (2.23) then becomes:
N X
X N Z ∞
Rijk (T ) = αk φri φrk φsj φsk gr (λ + T ) gs (λ) dλ (2.24)
r=1 s=1 0
Using the definition of gr from Eq. (2.13) we can separate the terms depend-
ing on λ and the ones depending on T as follows
1 ¡ ¢
gr (λ + T ) = d
e−ζ r ωr (λ+T ) sin ω dr (λ + T ) (2.25)
mr ω r
1 ¡ ¢
= d
e−ζ r ωr (λ+T ) sin ω dr λ cos ω dr T + sin ω dr T cos ω dr λ
mr ω r
¡ ¢
£ −ζ ωr T ¡ d ¢¤ e−ζ r ωr λ sin ω dr λ
= e r cos ω r T
mr ω dr
¡ ¢
£ −ζ ωr T ¡ d ¢¤ e−ζ r ωr λ cos ω dr λ
+ e r sin ω r T
mr ω dr
Substituting (2.25) in (2.24) and using the definition (2.13) for gs (λ), on finds
N
X £ r −ζ ωr T ¡ ¢ ¡ ¢¤
Rijk (T ) = Gijk e r cos ω dr T + Hijk
r
e−ζ r ωr T sin ω dr T (2.26)
r=1
where Grijk and Hijk

r
are independent of T and are given by
½ r ¾ X N Z ½ ¡ ¢¾
Gijk αk φri φrk φsj φsk ∞ ¡ d ¢ sin ω dr λ
r = e(−ζ r ω r −ζ s ω s )λ
. sin ω s λ ¡ ¢ dλ (2.27)
Hijk mr ω dr ms ω ds 0 cos ω dr λ
s=1
Equation (2.26) is the key result of this derivation. Examining Eq. (2.26),
we can see that the cross-correlation function is indeed a linear combination of
decaying sinusoids similar to the impulse response function of the original sys-
tem (see Equation (2.16)); thus, the cross-correlation functions can be handled
as impulse response functions and processed with time-domain identification
techniques to estimate the modal parameters of the system. Note that in this
approach the amplitude of the random inputs are not measured so that the con-
stant αk in (2.27) is unknown. As a consequence the modal masses mr cannot
be estimated in (2.27) and one will say that the modes so-obtained are unscaled.
It should not come as a surprise since the actual transfer function can not be
evaluated since the inputs are not measured. Some perturbation techniques
have however been proposed to obtain the scaling of the modes based on several
analysis of the system with added mass [53]. Applying sensitivity analysis, the
modal masses can be evaluated.
In (2.27) Grijk and Hijk
r
can be further simplified by evaluating the definite
integral, leading to
N
X · ¸
αk φri φrk φsj φsk Irs
Grijk = (2.28)
s=1
mr ms ω dr 2 + I2
Jrs rs
N
X · ¸
r
αk φri φrk φsj φsk Jrs
Hijk = (2.29)
s=1
mr ms ω dr 2 + I2
Jrs rs
³ 2 ´
2 2
where Irs = 2ω dr (ζ r ω r + ζ s ω s ) and Jrs = ω ds − ω dr + (ζ r ω r + ζ s ω s ) . Let
us then define γ rs such that:
tan (γ rs ) = Irs /Jrs (2.30)

Equation (2.28) and (2.29) then write
 N


 r φri X rs ¡ 2 ¢
2 −1/2
 Gijk = mr ω d
 β jk Jrs + Irs sin (γ rs )
r s=1
N (2.31)

 r
 φri X rs ¡ 2 ¢
2 −1/2
 H
 ijk = d
β jk Jrs + Irs cos (γ rs )
m ω r r s=1
where we define
αk φrk φsj φsk
β rs
jk = (2.32)
ms
Substituting (2.31) into Eq. (2.26), and summing over k, namely over all M
input locations to find the cross-correlation function due to all inputs, we find:
XN N M
φri X X rs ¡ 2 ¢
2 −1/2 −ζ r ω r T
¡ ¢
Rij (T ) = d
β jk Jrs + Irs e sin ω dr T + γ rs
r=1
mr ω r s=1
k=1
(2.33)
The inner summations on s and k are merely a summation of constants times
the sine function, with variable phase but fixed frequency. Equation (2.33) can
therefore be rewritten as a single sine function with a new phase angle (Θr ) and
new constant multipliers (Arj ):
XN
φri Arj −ζ r ωr T ¡ ¢
Rij (T ) = d
e sin ω dr T + Θr (2.34)
r=1
mr ω r
This last result shows that the cross-correlation function Rij (T ) between
the response signal at j and a reference signal at i at an interval of time T
can be expressed as a sum of N decaying sinusoids. Each decaying sinusoid is
characterized by a natural frequency ω r and damping ratio ζ r of the structure.
2.4 OMA in the presence of harmonic excitation

Operational Modal Analysis is based on the assumption that the input to the
structure is stationary white noise. In practice however, structural vibration ob-
served in operation cannot always be considered as pure white noise excitation.
In many practical cases, vibrations are induced by a combination of white noise
and harmonic excitations. Harmonic excitations in addition to random inputs
can occur due to rotating components or fluctuating forces in electric actuators
for instance. In these particular cases, OMA procedures are, strictly speaking,
not applicable. However, one can view the harmonic response as the vibration of
the structure according to a virtual eigenmode having zero damping and consider
that only white-noise excitation is present. In this way, harmonic excitation re-
sponses are taken into account during OMA as non-damped modes in addition
to the genuine natural eigenmodes of the structure. Unfortunately in that case,
the identification methods fail to identify modal parameters properly especially
for modes with eigenfrequencies close to the harmonic frequencies. Special nu-
merical filters can also be applied to filter-out harmonic components from the
measured response. Unfortunately in practice, filters are not perfect and if the
harmonic frequency is close to eigenfrequencies, the filtering will pollute the
measured response so that the identified modal parameters are perturbed.
In [7], a procedure based on statistical properties of the outputs [34] was
proposed in order to distinguish between harmonic response and narrow-band
stochastic response for output-only modal testing. The authors in [7] define an
indicator to identify harmonics from the signal when the Frequency Domain De-
composition is applied, but the example presented assumes harmonic frequencies
well separate from the eigenfrequencies.
2.5. CONCLUSION 21
2.5 Conclusion
This chapter briefly outlines the concepts of response-only modal testing. In
particular we have described the theoretical basis for the Natural Excitation
Technique (Sec. 2.3) working directly in the time domain and considering auto-
and cross-correlation of the response signals. One of the major limitation of
OMA has been described namely the fact that in the presence of harmonic
excitation the procedures can, in many cases, no longer be applied. In the
following chapters this issue will be further investigated.
Chapter 3
Modified LSCE method to

include known harmonic
excitations
In this chapter a modified version of Least Square Complex Exponential (LSCE)

method is proposed to perform operational modal analysis in the presence of
harmonic excitations. In the section (Sec. 3.2), a brief theoretical background is
presented for the LSCE. The auto- and cross-correlation functions of the system
as used in the Natural Excitation Technique (NExT, see section 2.3) are used
as impulse response functions of the system to estimate the modal parameters.
In the section (Sec. 3.3), the standard LSCE algorithm is extended to incorpo-
rate harmonic excitations. Section (Sec. 3.4), contains a brief description of
the test object. In section (Sec. 3.4.1), parameter values used as reference for
the experimental tests are obtained from standard OMA procedures. Then in
sections (Sec. 3.4.2, 3.4.3, 3.4.4) system parameters obtained in the presence of
harmonic excitations are presented and discussed. In the section (Sec. 3.4.5),
the robustness of the modified method is tested with respect to the difference be-
tween excitation and eigen frequencies and with respect to the accuracy of the
harmonic frequencies specified during the analysis.
This work has been published in [40, 41, 44, 47]
23
24 CHAPTER 3. MODIFIED LSCE METHOD FOR HARMONICS
3.1 Introduction
Least Square Complex Exponential (LSCE) method is a time domain method
which is applied to the impulse response function (IRF) of a system to compute
modal parameters. In this method impulse response functions of the systems
are modelled in a non-parametric way on the basis of Prony’s equation.
3.2 Least Square Complex Exponential method

Assuming that a system is excited by a stationary white noise, it has been
shown in the Natural Excitation Technique (NExT)(Sec. 2.3) that the correla-
tion function Rij (T ) between the response signals i and j at a time interval of T
is similar to the response of the structure at i due to an impulse on j. Assuming
damping to be small, this is expressed by the relation (2.34)
XN
φri Arj −ζ r ωr T ¡ ¢
Rij (T ) = d
e sin ω dr T + Θr (3.1)
r=1
mr ω r
where φri is the ith component of the eigenmode number r of the conservative
system, Arj is a constant associated to the j th response signal taken as reference,
mr is the rth modal mass, ζ r and ω r are respectively
q the rth modal damping ratio
and non-damped eigenfrequency, ω dr = ω r 1 − ζ 2r and Θr is the phase angle
associated with the rth modal response. Hence the correlation between signals is
a superposition of decaying oscillations having damping and frequencies equal to
the damping and frequencies of the structural mode. As a consequence, modal
parameter identification techniques like the Least Square Complex Exponential
method (LSCE) [9] can be used to extract the modal parameters from the
correlation functions between measured responses to the noise input. In terms
of the complex modes of the structure, the correlation function (3.1) can be
written as (see for instance [13])
N
X N
X ∗
Rij (k∆t) = esr k∆t Crj + esr k∆t Crj
∗
(3.2)
r=1 r=1
q¡ ¢
where sr = −ω r ζ r + iω r 1 − ζ 2r and where Crj is a constant associated with
the rth mode for the j th response signal, which is the reference signal. ∆t is the
sampling time step and the superscript ∗ denotes the complex conjugate. Note
that in conventional modal analysis, these constant multipliers are modal par-
ticipation factors. Numbering all complex modes and eigenvalues in sequence,
3.2. LEAST SQUARE COMPLEX EXPONENTIAL METHOD 25
equation (3.2) can be written as

2N
X
0
Rij (k∆t) = Crij esr k∆t (3.3)
r=1
As sr appears in complex conjugate forms in this expression, there exists a

polynomial of order 2N (known as Prony’s equation) of which expsr ∆t are roots:
β 0 + β 1 Vr1 + β 2 Vr2 + ... + β 2N −1 Vr2N −1 + Vr2N = 0 (3.4)
where Vr = esr ∆t and where β 2N = 1. {β} is the coefficient matrix of the

polynomial. To determine the values of β i , let us multiply the impulse response
(3.3) for sample k by the coefficient β k and sum up these values for k = 0, . . . 2N
[12, 24, 36]:
2N 2N
Ã 2N ! 2N
Ã 2N
!
X X X X X
0
β k Rij (k∆t) = βk Crij Vrk = 0
Crij β k Vrk = 0 (3.5)
k=0 k=0 r=1 r=1 k=0
Hence the coefficients {β} satisfy a linear equation which coefficients are the
impulse response (or correlation functions) at 2N + 1 successive time samples.
In order to determine those coefficients, relation (3.5) is written at least for 2N
times, starting at successive time samples. In other words, many Prony’s equa-
tions can be written to build up a linear system that determines the coefficients
{β}:
n n+1 n+2N −1 n+2N
β 0 Rij + β 1 Rij + ... + β 2N −1 Rij = −Rij (3.6)
k
where we note Rij = Rij (k∆t) and where n = 1...L. One build therefore a linear
system of L equations based on the Hankel matrix to determine the coefficients
{β}:
[R]ij {β} = −{R0 }ij (3.7)
The procedure explained above uses a single correlation function. Since {β} are
global quantities related to modal parameters, the same parameters can be de-
rived from different correlation functions between any two response signals and
from the auto-correlation functions. Hence, let us consider p response stations
and adopt q of them as reference locations. Grouping the associated equations
(3.7) for all correlation functions in one system yields
   
[R]11 
 {R0 }11 

 [R]12  
 
  {R0 }12 
 ..  {β} = − .. (3.8)
 .  
 . 


 

[R]qp {R0 }qp
A solution in a least square sense can be found for {β} from (3.8) using pseudo-
inverse techniques. Observing that [R]ij = [R]ji , the least square problem will
determine the parameters {β} provided L(qp − q(q−1) 2 ) ≥ 2N .
The LSCE method as explained above is different from the standard LSCE
since we process the correlation functions between any two response locations
whereas the standard LSCE considers only one reference. Note however that,
although correlation functions with respect to several references are considered,
the procedure described here is different from the so-called Polyreference LSCE
[12,36] where the response at one location is assumed to result from simultaneous
excitations at multiple points. In the present case, every time a new location is
considered as a reference, one reconstruct experimental data where, seemingly,
a single impulse is given at that reference point.
Nevertheless, even though the procedure defined by (3.8) is not a polyref-
erence method stricto sensu, we will still call it the polyreference LSCE. This
specific LSCE version allows us to explicitly introduce harmonic components in
the identification procedure as explained in the next section.
3.3 Modification of LSCE method to Include Har-

monic Components
If the signal contains a forced harmonic part due to a harmonic excitation on top
of the response to the noise excitation, the correlation between measured data
will, in addition to the impulse response described earlier, contain a non-damped
harmonic part. This extra component can be looked at as a non-damped virtual
eigenmode of the system. Hence, the operational modal analysis procedure
described above could be used even in the presence of harmonic excitation.
However, explicitly including the harmonic in the identification procedure leads
to a more efficient and robust approach as explained next. When a harmonic
excitation of known (or a priori identified) frequency ω is present, it can easily
be shown that the correlation functions have the form (3.3) with two extra
term corresponding to eigenvalues sr = ±iω r . There are thus two extra roots
of Prony’s polynomial, namely Vr = esr ∆t = e±iωr ∆t . Let us now explicitly
express that Vr = cos(ω r ∆t) − i sin(ω r ∆t) and Vr = cos(ω r ∆t) + i sin(ω r ∆t) are
roots of (3.6). Re-arranging the Prony’s equations for these two eigenvalues, we
find
 

 β0 
· 
¸  ½ ¾
0 sin(ω r ∆t) ... sin(ω r (2N − 1)∆t) β1  sin(2N ω r ∆t)
. = −
1 cos(ω r ∆t) ... cos(ω r (2N − 1)∆t)   .. 
 

cos(2N ω r ∆t)
 
β 2N −1
(3.9)
These two independent linear relations for the coefficients {β} must be satisfied
by Prony’s coefficients to represent exactly the known harmonic components. In
general, we will assume there are m harmonic frequencies in the signal within
the range of frequencies considered. Adding the linear relations (3.9) to the
linear system (3.8), we get
3.4. EXPERIMENTAL EXAMPLE 27
  
R0 ... R2m−1 R2m ... R2N −1  β0
 

 .. .. .. .. 
 β1 

 . A . . C . 
 

 
 .. 

 RLp−1 ... RLp+2m−2 RLp+2m−1 ... RLp+2N −2 
 . b1 

 





 sin(ω 1 (2N − 1)∆t) 
 0 ... sin(ω 1 (2m − 1)∆t) sin(ω 1 2m∆t) ...  β 2m−2
 cos(ω 1 (2N − 1)∆t) 
 1 ... cos(ω 1 (2m − 1)∆t) cos(ω 1 2m∆t) ...  β 2m−1
 

 .. .. .. .. 
 β 

 
 2m 

 . B . . D . 

 .. 


 0 ... sin(ω m (2m − 1)∆t) sin(ω m 2m∆t) ... 
sin(ω m (2N − 1)∆t)  . b2 


 

1 ... cos(ω m (2m − 1)∆t) cos(ω m 2m∆t) ... sin(ω m (2N − 1)∆t) β 2N −1
 

 R2N 


 .. 


 


 . E 


 


 RLp+2N −1 

 
sin(ω 1 2N ∆t)
=−

 cos(ω 1 2N ∆t) 


 .. 


 


 . F 


 


 sin(ω m 2N ∆t) 

 
cos(ω m 2N ∆t)
(3.10)
In symbolic form we can write
[A] {b1 } + [C] {b2 } = {E} (3.11)

(Lp×2m)(2m×1) (Lp×2N −2m)(2N −2m×1) (Lp×1)
and
[B] {b1 } + [D] {b2 } = {F } (3.12)
(2m×2m)(2m×1) (2m×2N −2m)(2N −2m×1) (2m×1)
From equation (3.12)
{b1 } = [B]−1 [{F } − [D]{b2 }] (3.13)
and substituting in (3.11), we obtain

£ ¤
[C] − [A][B]−1 [D] {b2 } = {E} − [A][B]−1 {F } (3.14)
From the over-determined system (3.14), {b2 } can be found as a least square
solution. {b1 } is then retrieved from (3.13). Together {b1 } and {b2 } provide
the coefficients of Prony’s polynomial. The roots of the polynomial will include
the harmonic frequencies since the procedure presented here enforces it exactly.
3.4 Experimental example

In order to investigate the efficiency of the proposed procedure, experiments
have been done for a pinned-pinned beam depicted in Figure 3.1. The beam was
simply supported on two lubricated pin-joints. At one end, the pin was placed on
a lubricated slider so that the longitudinal motion was not constrained. Hence
we can assume that, for small vibrations as encountered in our experiments, the
system behaves linearly. A shaker was used to provide stationary noise excita-
tion as well as harmonic loads. A total of 8192 discrete time samples have been
obtained for each experiment. The data was sampled at 128Hz. A preliminary
analysis was carried out to know the mode shapes and the frequencies. Accord-
ingly we placed four accelerometers to ensure good level of measured response
signal for the second bending mode. Harmonics were introduced close to the
second bending mode frequency. In the examples, modal parameters will first be
Figure 3.1: (a)Experimental setup for the beam structure. (b) slider arrange-
ment at one end
computed only with stationary white noise input to get the modal parameters
as would be done in a normal OMA procedure. Then, we will consider the cases
when harmonic excitations are present along with the random loads. Modal
parameters are then identified both with the polyreference LSCE method and
by the single reference LSCE modified to explicitly account for harmonic com-
ponents (see previous section). When the polyreference LSCE method is used,
correlation functions are computed with reference to the first two accelerometer
signals, leading to eight sets of correlation functions in total. When the single
reference LSCE method is employed, we will consider correlation functions with
reference to the first accelerometer only, leading to four sets of correlation func-
tions. For both identification methods, the number of rows L in the data matrix
is taken as L = 200 so that the total number of discrete correlation values used
in the identification is 200 + 2N for each correlation function, N being the order
of the identified model.
3.4.1 Pure stationary white noise excitation

First the experiment was done with stationary white noise introduced between
18Hz to 28Hz, to excite the second bending mode properly. Eigenfrequencies
and damping values were identified with the polyreference LSCE method. In
Figure 3.2: PSD of acceleration at position 1 for pure random excitation and
stability diagram for the polyreference LSCE method.
Figure 3.2 and Table 3.1 we report the convergence of the poles to the identi-
fied first, second and third eigenfrequencies when the order N of the system is
increased. It can be observed that the first three eigenfrequencies are identified.
Note however that, since the stationary white noise input was set for the experi-
ment between 18Hz and 28Hz, only the second eigenfrequency can be assumed
to be identified properly by the OMA procedure. The second eigenfrequency
was found to be 23.15Hz and its associated damping is taken as 0.26%.
3.4.2 Single sine harmonic at 23.225Hz

Let us now consider the case when a single harmonic frequency is added to
the response signal close to the second bending frequency. This is achieved by
adding a sinusoidal harmonic frequency to the input of the shaker along with
the stationary white noise. As discussed in the theory, we have introduced two
more equations in the Hankel matrix of the single reference LSCE method to
take into account the harmonic frequency. In Figure 3.3 the harmonic frequency
at 23.225Hz is clearly visible in the PSD plot because of its high peak. There
is also a smaller peak at a frequency slightly below the harmonic frequency.
From the PSD it is not clear if that secondary peak corresponds to the second
bending mode frequency. Applying the polyreference LSCE, the stabilization
diagram identifies the dominating harmonic peak with an associated damping
of the order of 0.05%. But the polyreference LSCE fails to identify the actual
eigenfrequency properly even for an identification order of 40. Figure 3.3 and
Table 3.1: Frequencies and associated damping (white noise only)
Polyreference LSCE
Modes Freq.1 Damp.1 Freq.2 Damp.2 Freq.3 Damp.3
(Hz) (%) (Hz) (%) (Hz) (%)
30 6.59 0.23 23.16 0.27 50.57 0.09
28 6.59 0.23 23.16 0.29 50.57 0.09
26 6.59 0.23 23.16 0.30 50.57 0.09
24 6.59 0.23 23.17 0.28 50.57 0.09
22 6.59 0.22 23.15 0.26 50.57 0.09
20 6.59 0.22 23.15 0.26 50.57 0.09
18 6.59 0.22 23.15 0.26 50.58 0.09
16 6.60 0.23 23.15 0.26 50.58 0.08
14 6.60 0.23 23.15 0.26 50.57 0.07
12 6.60 0.23 23.15 0.26 50.57 0.07
10 6.61 0.24 23.15 0.24 50.57 0.07
8 6.61 0.25 23.16 0.21 50.54 0.09
6 6.61 0.40 23.18 0.21 50.52 0.17
4 6.61 0.60 23.23 0.32 50.37 0.61
Table 3.2 indicate a clear stable frequency line at 23.21Hz computed by the
polyreference LSCE method. Reminding that the eigenfrequency is at 23.15Hz
for the second bending mode and that the harmonic frequency is at 23.225Hz,
one concludes that the polyreference LSCE identification procedure is strongly
influenced by the presence of the harmonic component.
Figure 3.4 and Table 3.2 report the results obtained when applying the
modified single reference LSCE approach proposed in this work. Obviously, since
the harmonic component is explicitly introduced in the identification procedure,
the harmonic has an associated zero damping. The identified eigenfrequency and
its damping is very close to the second bending mode frequency and damping.
Both frequency and damping values are nearly invariant at different modal order,
which can be seen from Table 3.1 and Figure 3.4. Hence, even for low modal
orders, the modified single reference LSCE can identify the eigen-parameters
accurately.
3.4.3 Two square shaped harmonics at 23.30Hz and 23.80Hz

In order to investigate the efficiency of the method when non-sinusoidal har-
monics are present, we have introduced two periodic square wave excitations on
top of the random loads, at frequencies 23.30Hz and 23.80Hz. Note that the
additional periodic loads have fundamental frequencies very close to the second
bending mode frequency, namely 23.15Hz. Also, because the periodic excita-
tions have a square waveform, they have a very strong fundamental component
so that, in the response, mainly the harmonic response corresponding to the
Figure 3.3: PSD of acceleration at position 1 for a combined single harmonic

and random excitation, and stabilization diagram of the polyreference LSCE
method
fundamental frequencies will appear. In Figure 3.4.3 two peaks at 23.30Hz and
23.80Hz are clearly observed. Again, a secondary peak is present slightly below
23.30Hz, which might be due to the second bending mode frequency. From
Table 3.3 and Figure 3.4.3 it can be concluded that, when using the polyref-
erence LSCE, the identified frequencies are associated to the two fundamental
harmonic frequencies. Note that the identified damping corresponding to the
harmonic frequencies are not small, while there should be null in theory. Hence,
in practice, it would be difficult at this point to assimilate the identified modes
to harmonic responses and the analyst would probably conclude that the iden-
tified parameters correspond to true eigenmodes. When applying the modified
single reference LSCE, we have introduced in the Hankel matrix two harmonic
components of frequencies corresponding to the periodic excitations, namely
23.30Hz and 23.80Hz. The stabilization diagram is shown in Figure 3.6. The
last two columns in Table 3.3 give the identified frequencies and damping val-
ues. The identified harmonic parameters are not reported in the table since,
by construction, they have the exact harmonic frequency and zero damping. It
is seen that the identified eigenfrequency and the associated damping are very
close to the expected values for the second bending mode. Therefore, whereas
we failed to get any correct modal parameters from the polyreference LSCE
method, the modified single reference approach allows to identify the eigenpa-
rameters in a robust manner. This example indicates that the modified single
Figure 3.4: PSD of acceleration at position 1 for a combined single harmonic and
random excitation, and stabilization diagram of the modified single reference
LSCE
reference LSCE method can be expected to be efficient for many general periodic
loads that appear additionally to the random excitation. Indeed, if the periodic
excitation has a waveform such that significant super-harmonic components are
present, one can easily introduce in the modified single reference LSCE several
super-harmonics for the identification.
3.4.4 Three additional harmonics at 22.40Hz, 22.65Hz and

22.90Hz
Now three sine harmonics at 22.40Hz, 22.65Hz and 22.9Hz are given as input
to the shaker along with the stationary white noise in order to investigate the
effectiveness of the new method in the presence of strong multi-harmonics. As
mentioned in the theory, we have introduced all three harmonics in the Hankel
matrix of the single reference LSCE method to take into account the harmonic
frequencies. In Figure 3.7 three harmonic frequencies can be clearly identified at
22.40Hz, 22.65Hz and 22.90Hz. We also observe just after the third harmonic
frequency a peak which corresponds to the second bending mode frequency. As
shown in Table 3.4 and Figure 3.7, the polyreference LSCE method exhibits two
stabilization lines. Those lines are at 22.47Hz and 22.90Hz, which corresponds
to two of the harmonic frequencies introduced in the signal. There is no sign of
the second harmonic frequency in the graph. Although in theory those frequen-
Table 3.2: Frequencies and associated damping identified by the polyreference

LSCE and the modified single reference LSCE methods (random and single
harmonic excitation).
Polyreference LSCE Modified Ref. LSCE

(Hz) (%) (Hz) (%) (Hz) (%)
40 23.18 0.08 – – 23.15 0.31
38 23.18 0.05 – – 23.15 0.29
36 23.20 0.08 – – 23.15 0.29
34 23.21 0.05 – – 23.16 0.30
32 23.21 0.05 – – 23.15 0.29
30 23.21 0.06 – – 23.15 0.29
28 23.21 0.06 – – 23.14 0.32
26 23.21 0.07 – – 23.16 0.26
24 23.21 0.09 – – 23.16 0.23
22 23.22 0.10 – – 23.16 0.24
20 23.24 0.22 – – 23.15 0.25
18 23.23 0.04 23.32 0.28 23.15 0.22
16 23.23 0.03 23.35 0.65 23.15 0.21
14 23.22 0.03 23.35 0.81 23.16 0.21
12 23.16 0.22 – – 23.15 0.22
10 23.21 0.03 – – 23.15 0.23
8 23.21 0.12 – – 23.17 0.25
6 23.22 0.03 23.26 0.01 23.18 0.28
4 23.22 0.03 – – – –
cies should be associated to zero-damping, the results in Table 3.4 indicate that
the damping of the identified harmonic parts becomes small only for high modal
orders. The polyreference LSCE does not identify the actual eigenfrequency.
Figure 3.8 and Table 3.4 represent the stability diagram for the modified
LSCE method proposed. The first three frequency lines belong to the three
harmonics, and their frequencies have been set so as to exactly match the har-
monic frequencies, the associated damping being zero (those frequencies and
damping values are not given in the table). The other identified frequency and
damping is listed in Table 3.4 and can be seen to be very close to the frequency
and damping of the second mode identified earlier. Note however that the damp-
ing ratio is slightly changing depending on the identification order N , but it is
close to the reference damping of 0.26%.
3.4.5 Robustness of the modified LSCE

From the experiments described above, it can be concluded that the polyrefer-
ence LSCE method cannot identify the eigenfrequencies accurately when har-
Figure 3.5: PSD of the acceleration at position 1 for random loads and two ad-
ditional periodic square excitations, and stabilization diagram from the polyref-
erence LSCE method
monic frequencies close to the natural frequencies are present in the signal.
However the modified single reference LSCE approach proposed here allows to
identify the modal parameters with high confidence in that case. In order to
further illustrate the robustness of the new approach and to investigate its sensi-
tivity to the interval between eigenfrequency and harmonic frequency, we show
in Figure 3.9 the computed modal parameters as a function of the harmonic
frequency. We recall that the eigenfrequency corresponding to the second bend-
ing mode is 23.15Hz. In this example, the level of the harmonic input relative
to the level of random excitation was kept constant. Figure 3.9 shows that the
modified single reference LSCE can identify the eigenfrequency with an accuracy
of about 0.01Hz for a harmonic component 0.2% lower or 0.4% higher than the
eigenfrequency. It also identifies the damping ratio within 0.05% for a harmonic
frequency 0.5% lower or 0.3% higher than the eigenfrequency. The polyreference
method fails to identify the eigenfrequency and the corresponding damping ac-
curately when the harmonic frequency is so close to the eigenfrequency. Hence,
clearly, the modified single reference LSCE significantly improves the robustness
of the identification when the harmonic frequency is close to an eigenfrequency.
Obviously, in the modified LSCE method proposed here, it is important to in-
troduce the exact harmonic frequency in the identification a priori. In Figure
3.10 we consider the case where the harmonic frequency specified in the identi-
fication with the modified LSCE algorithm is different from the exact harmonic
3.5. CONCLUSION 35
Figure 3.6: PSD of the acceleration at position 1 for random loads and two addi-
tional periodic square excitations, and stabilization diagram from the modified
single reference LSCE method
frequency, namely 23.225Hz in this case. One observes that if a harmonic fre-
quency different from the actual 23.225Hz of the input excitation is specified,
the identified frequencies varies linearly with respect to the harmonic frequency
input in the vicinity of the exact harmonic frequency. When the harmonic fre-
quency used in the identification is significantly higher than the exact harmonic
frequency, the identified eigenfrequency converges to the one identified by the
standard polyreference LSCE method whereas the identified damping ratio de-
creases significantly. Hence, it is essential to introduce in the modified LSCE
algorithm a harmonic frequency very close to the actual one. Otherwise the
harmonic component explicitly included in the identification is not capable of
representing the harmonic part of the signal properly and the modification of
the LSCE gives no benefit. In the worse case, the modification in the LSCE has
no effect and the identified parameters are the ones obtained by standard LSCE
algorithms.
3.5 Conclusion
We have proposed a modification of the single reference Least Square Complex
Exponential identification method used in the context of Operational Modal
Analysis. This modification allows to explicitly account for harmonic contents in
the measured signal. When the harmonic excitation part has a frequency close
Table 3.3: Frequencies and associated damping for both polyreference LSCE
and single reference LSCE method (two additional periodic excitations with
square waveform)
Polyreference LSCE Modified LSCE

(Hz) (%) (Hz) (%) (Hz) (%)
40 23.26 0.11 23.90 0.07 23.17 0.23
39 23.26 0.11 23.90 0.03 23.17 0.24
37 23.26 0.11 23.90 0.04 23.17 0.25
35 23.26 0.12 23.91 0.04 23.17 0.25
33 23.26 0.12 23.93 0.14 23.17 0.25
31 23.26 0.11 23.92 0.10 23.17 0.26
29 23.26 0.11 23.92 0.09 23.17 0.28
27 23.25 0.08 23.89 0.12 23.18 0.30
25 23.27 0.04 23.89 0.06 23.17 0.27
24 23.26 0.09 23.92 0.11 23.17 0.25
23 23.26 0.10 23.92 0.11 23.18 0.27
22 23.25 0.02 23.87 0.19 23.18 0.27
20 23.29 0.01 23.84 0.09 23.19 0.29
18 23.27 0.24 23.47 0.70 23.19 0.34
16 23.27 0.16 – – 23.19 0.31
14 23.27 0.16 – – 23.18 0.37
12 23.26 0.16 – – 23.16 0.43
10 23.26 0.15 24.05 0.07 23.14 0.26
9 23.26 0.15 24.03 0.29 23.13 0.28
8 23.26 0.14 23.96 0.28 23.34 0.79
7 23.26 0.14 24.01 1.26 – –
to eigenfrequencies that one tries to identify, standard LSCE methods fail to

identify the modal parameters accurately. In that very challenging situation the
modified LSCE proposed here yields accurate identification even if the harmonic
excitation frequency is very close to an eigenfrequency. Even in the case where
the standard polyreference LSCE can identify the modal parameters accurately
(that is when the harmonic frequency is not too close to an eigenfrequency), the
proposed modified LSCE algorithm yields accurate parameters for a much lower
assumed system order. The experimental example described here illustrates the
efficiency and robustness of the modified LSCE. It is also shown that, when
the harmonic frequency input in the modified LSCE is different from the actual
harmonic frequency in the signal, the modified LSCE simply degenerates to a
standard algorithm.
3.6. POSSIBLE FURTHER DEVELOPMENTS 37
Figure 3.7: PSD of the acceleration at position 1 for random loads and three ad-
ditional harmonic excitations, and stabilization diagram from the polyreference
LSCE method
3.6 Possible further developments

The method proposed in this work was found to be efficient in identifying modal
parameters in the presence of harmonic excitation with a frequency close to an
eigenfrequency. It is therefore an interesting method for applications where har-
monic loads are inherently present together with operational noise excitation. It
can easily be extended to include any general periodic component by consider-
ing its Fourier expansion. The modified LSCE procedure can also be extended
to explicitly include damped oscillations known a priori. This could be helpful
in order to allow the identification of modes only weakly present in a signal
and which are masked by eigenmodes strongly dominant. The procedure would
then consists in first identifying the dominant components and then, in a second
step, include the dominant modes in the LSCE explicitly to identify the modes
weakly present in the signal. Other research directions currently considered also
include the application of the proposed modified LSCE to the identification of
super- and sub-harmonics in non-linear systems.
Figure 3.8: PSD of the acceleration at position 1 for random loads and three
additional harmonic excitations, and stabilization diagram from the modified
single reference LSCE method
3.6. POSSIBLE FURTHER DEVELOPMENTS 39
Table 3.4: Frequencies and associated damping from the polyreference LSCE
and the modified single reference LSCE methods (three additional harmonic
excitations).
Polyreference LSCE Modified LSCE

(Hz) (%) (Hz) (%) (Hz) (%)
40 22.44 0.00 22.90 0.04 23.13 0.18
39 22.46 0.10 22.90 0.06 23.13 0.17
38 22.47 0.19 22.90 0.06 23.13 0.17
37 22.46 0.23 22.90 0.07 23.13 0.20
36 22.47 0.23 22.90 0.06 23.14 0.20
35 22.47 0.25 22.90 0.06 23.13 0.25
34 22.47 0.29 22.90 0.07 23.14 0.27
33 22.47 0.31 22.90 0.07 23.14 0.31
32 22.47 0.27 22.90 0.07 23.13 0.31
31 22.48 0.26 22.90 0.06 23.12 0.36
30 22.48 0.27 22.90 0.06 23.14 0.50
29 22.49 0.26 22.90 0.06 23.13 0.54
28 22.49 0.27 22.90 0.06 23.13 0.54
27 22.49 0.28 22.90 0.06 23.15 0.54
26 22.50 0.26 22.90 0.06 23.15 0.52
24 22.52 0.35 22.90 0.06 23.18 0.50
22 22.51 0.40 22.90 0.06 23.17 0.70
20 22.54 0.43 22.90 0.06 23.18 1.50
18 22.54 0.49 22.90 0.06 – –
16 22.57 0.72 22.90 0.06 – –
14 22.43 0.40 22.90 0.05 – –
12 22.45 0.39 22.91 0.05 – –
23.3
Frequency, Hz
(a)
Single Ref. LSCE 23.25
Polyreference LSCE
23.2
23.15
actual frequency for
23.1
the second mode
23.05
23
-0.01 -0.008 -0.006 -0.004 -0.002 0 0.002 0.004 0.006 0.008 0.01
relative frequency difference
(b)
Damping, %
1
Single Ref. LSCE 0.9
Polyreference LSCE 0.8
0.7
0.6 actual damping
associated with the
0.5
second mode
0.4
0.3
0.2
0.1
0
-0.01 -0.008 -0.006 -0.004 -0.002 0 0.002 0.004 0.006 0.008 0.01
relative frequency difference
Figure 3.9: Identified eigenfrequency (a) and damping ratio (b) for the second
ω −ω bending
bending mode as a function of harmonic
ω bending
(a) 23.24
frequency line, computed by (b) 0.35
polyreference LSCE method
23.22 0.3
damping line for the
23.2 0.25 second mode
Damping, %
Frequency, Hz
23.18 0.2
true frequency line for
23.16 the second mode 0.15
23.14 0.1
actual harmonic
23.12 actual harmonic 0.05 frequency
frequency
23.1 0
23.21 23.22 23.23 23.24 23.25 23.26 23.27 23.28 23.21 23.22 23.23 23.24 23.25 23.26 23.27 23.28
Harm onic Frequency, Hz Harmonic Frequency, Hz
Figure 3.10: Identified modal frequency (a) and damping ratio (b) as a function
of the harmonic frequency specified in the modified LSCE.
Chapter 4
Modified ITD method for

harmonic excitations
In this chapter, a modified version of Ibrahim Time Domain (ITD) method

is proposed to perform operational modal analysis in the presence of harmonic
excitations. Like in the modified LSCE method (Ch. 3), the auto- and cross-
correlation functions of the system are used as impulse response functions of the
system to estimate the modal parameters (Sec. 2.3). In section (Sec. 4.2) the
Single Station Time Domain (SSTD), variant version of ITD method is briefly
described. Then in the section (Sec. 4.3), the SSTD method is extended to the
Single Input Multi Output (SIMO) version, which can be used to process many
correlation functions to compute modal parameters. In section (Sec. 4.3.1) the
SIMO version of the SSTD method (SIMO-SSTD) is modified to incorporate
harmonic responses in addition to white noise output. The experimental setup
is described in section (Sec. 4.4). In the sections (Sec. 4.4.1, 4.4.2, 4.4.3) the
modal parameters of the test system is computed when the standard SIMO-SSTD
and the modified SIMO-SSTD method are applied to the correlation functions.
A sensitivity analysis of the modified SIMO-SSTD method is presented for this
particular system in order to evaluate the accuracy of the identification procedure
with respect to the harmonic excitation frequencies specified to the algorithm
(Sec. 4.4.4). Conclusions of the chapter are found in the section (Sec. 4.5).
This work has been published in [42, 45, 46]
41
42 CHAPTER 4. MODIFIED ITD METHOD FOR HARMONICS
4.1 Introduction
The Ibrahim Time Domain (ITD) method is a time domain method like the
LSCE method (Ch. 3), which computes modal parameters of the system. Ini-
tially the ITD method was developed for displacement and velocity responses
[25, 26]. The same author developed a method based on free acceleration re-
sponse [27]. The ITD method can compute mode shape vectors together with
the modal frequencies and associated damping ratios. However these methods
are restricted by the fact that we must have at least 2N number of response
channels, when the system is assumed to have a modal order of N . This means
that many response channels are required to complete an experiment.
The Single Station Time Domain (SSTD) [64] method is a modified version
of the ITD method. In the SSTD method, only one response signal is used to
compute modal parameters. In this chapter, first a generalized version of the
SSTD method is developed (Sec. 4.3), which can take many response signals
to process modal parameters. The generalized version of the SSTD method is
a Single Input and Multi Output (SIMO) method. In the section (Sec. 4.3.1),
the generalized version of SSTD method is modified to perform OMA in the
presence of harmonics.
4.2 A variant of the Ibrahim Time Domain Method
When a system is excited by pure stationary white noise, the correlation be-
tween dynamic responses can be exploited to identify modal parameters. One
procedure to analyze structures under random excitation is based on the fact
that the correlation function Rij (T ) between the response signals i and j at
a time interval of T is similar to the response of the structure at i due to an
impulse on j. The theoretical basis for this results has been outlined in the
Natural Excitation Technique (NExT) [30] (see Ch. 2).
The Ibrahim Time Domain algorithm (ITD) [27] uses the time response of
several outputs in order to find modal parameters. In the standard ITD method,
at least 2N response location need to be measured to identify a model of order
N . The Single-Station Time Domain method proposed in [64] uses an approach
similar to the ITD, but the identification is based on the response at a unique
location, considered for different time intervals. The method presented next is
in fact a blend of the ITD and SSTD methods to get an SIMO algorithm like
the standard ITD, but where several time intervals per response are used as in
the SSTD.
Re-arranging equation (3.3) expressing the correlation functions as a modal
4.2. A VARIANT OF THE ITD 43
superposition in a Hankel matrix form, we can write

 
R1 R2 ... ... RL
 R2 R3 .. ... RL+1 
 
 .. .. .. =
 . . ... ... . 
(4.1)
R2N R2N+1 ... ... RL+2N-1
(2N x L)
[X]
 0
 
C1 C20 ... 0
C2N es1 t1 es1 t2 ... ... es1 tL
 C1 e
0 s 1 4t 0
C2 e s 2 4t
... C2N e 2N 4t
0 s  es2 t1 es2 t2 .. ... es2 tL 
  
 .. .. ...... ..  .. .. .. 
 . . ... .  . . ... ... . 
C10 es1 (2N-1)4t C20 es2 (2N-1)4t ... C2N
0 s2N (2N-1)4t
e es2N t1 es2N t2 ... ... es2N tL
(2N x 2N) (2N x L)
[A] [Λ]
where tk = k∆t, N is the total number of modes considered for the identifi-
cation and L is the number of correlation values per row. Note that we have
dropped the index for the reference and response locations in order to simplify
the notations, namely Rk = Rij (k∆t). The identification order N is usually not
known a priori but can be increased until the identified parameters converge.
In symbolic form, we write (4.1) as
[X] = [A] [Λ] (4.2)

(2N x L) (2N x 2N)(2N x L)
We can write a similar equation by shifting all the discrete response values by
∆t as follows:
 
R2 R3 ... ... RL+1
 R3 R4 .. ... RL+2 
 
 .. .. .. = (4.3)
 . . ... ... . 
R2N+1 R2N+2 ... ... RL+2N
(2N x L)
[X̂ ]
  
C10 es1 4t C20 es2 4t ... C2N 0 s2N 4t
e es1 t1 es1 t2 ... ... es1 tL
 C10 es1 24t C20 es2 24t ... C2N 0 s2N 24t
e  s2 t1 s2 t2 
  e e .. ... es2 tL 
 ...... ..  . . .. 
 ... ... ... .  . .. 
. ... ... .
C10 es1 2N 4t C20 es2 2N 4t ... C2N0 s2N 2N 4t
e es2N t1 es2N t2 ... ... es2N tL
(2N
h xi2N) (2N x L)
Â [Λ]
which can be noted as h i h i

X̂ = Â [Λ] (4.4)
(2N x L)
(2N x L) (2N x 2N)
Now let us define a system matrix [S] so that

h i
[S] [A] = Â (4.5)
(2N x 2N)(2N x 2N)
(2N x 2N)
Pre-multiplying equation (4.2) by [S], we get
[S] [X] = [S] [A] [Λ] (4.6)

(2N x 2N)(2N x L) (2N x 2N)(2N x 2N)(2N x L)
and inserting the definition (4.5),

h i
[S] [X] = Â [Λ] (4.7)
(2N x 2N)(2N x L) (2NxL)
(2N x 2N)
Finally, taking account of (4.4) one obtains

h i
[S] [X] = X̂ (4.8)
(2N x 2N)(2N x L)
(2N x L)
Equation (4.8) uses a single-output signal due to a single input (correlation

between i and j) and corresponds to the SSTD method, itself similar to the
ITD [64].
4.3 A SIMO-SSTD method
In order to find a SIMO variant of the SSTD, let us observe that the system
matrix [S] is independent of the location of measurements and thus that equa-
tion (4.8) remains valid for any SISO combination. Hence the values of [S] can
also be obtained by considering several responses to an input and by satisfying
(4.8) in a least square sense. So let us write (4.8) for n responses to a single
impulse (SIMO) as:
" #  
£ 1¤ £ 2¤ h ih i h i
[S] X X ..... [X n ] =  X̂ 1 X̂ 2 .... X̂ n  (4.9)
(2N x 2N) (2N x L)(2N x L) (2N x L)
(2N x L)(2N x L) (2N x L)
(2N x Ln) (2N x Ln)
where [X j ] is the matrix of correlation functions defined in ( 4.1) expressed

between the reference i and an output j. In more compact form:
h i
[S] [Y ] = Ŷ (4.10)
(2N x 2N)(2N x Ln)
(2N x Ln)
A least-square solution [S] of (4.10) can be computed (for instance using a

pseudo-inverse technique). Finally the eigenvalues are deduced from [S] be
observing that Eq. (4.5) can be decomposed as
4.3. A SIMO-SSTD METHOD 45
   

 ar esr 04t 
 
 ar esr 04t 

[S] .. = .. esr 4t r = 1, 2....2N
 .   . 
(2N 
x 2N)   
ar esr (2N −1)4t ar esr (2N −1)4t
(2N x 1) (2N x 1)
(4.11)
and therefore
 

 ar esr 04t 

£ ¤ ..
[S] − esr 4t [I] . = {0} (4.12)

 

sr (2N −1)4t
ar e
Equation (4.12) is a standard eigenvalue problem, from which sr can be found.

Modal frequencies and damping can be computed from the values of sr .
4.3.1 Modification of SIMO-SSTD in the presence of har-

monic excitation
When in addition to the stationary white noise the structure is excited by
harmonic excitations, the dynamic response will include the associated forced
harmonic components which can be seen as virtual non-damped modes in the
correlation functions. These virtual modes can in theory be identified as be-
fore by an SIMO-SSTD method, but, as illustrated in [47] , identification of
those non-damped virtual modes can be very difficult in practice. Let us as-
sume that the harmonic excitation has a frequency ω. The modal superpo-
sition (3.3) describing the impulse response now includes a forced harmonic
part equivalent to a virtual modal response with frequency s = ±iω so that
esr ∆t = e±iω∆t = cos(ω r ∆t) ± i sin(ω r ∆t). This solution being known a priori
we will force the system matrix [S] to have e±iω∆t as an eigensolution: writing
(4.11 ) for s = ±iω and rearranging, [S] must satisfy
   
0 1 sin (ω 1 4t) cos (ω 1 4t)
 sin (ω 1 4t) cos (ω 1 4t)   cos (2ω 1 4t) 
   sin (2ω 1 4t) 
[S]   : : =
  : : 

(2N x 2N) sin ((2N-2) ω 1 4t) cos ((2N-2) ω 1 4t)   sin ((2N-1) ω 1 4t) sin ((2N-1) ω 1 4t) 
sin ((2N-1) ω 1 4t) cos ((2N-1) ω 1 4t) sin (2Nω 1 4t) cos (2Nω 1 4t)
(2N x 2) (2N x 2)
(4.13)
which in symbolic form writes
£ 1¤ h i
[S] H = Ĥ 1 (4.14)
(2N x 2N )(2N x 2)
(2N x 2)
If m harmonic excitation frequencies exist, (4.14) can be written generalized to

" #  
£ 1¤ £ 2¤ h i h i h i
[S] H H ... [H m ] =  Ĥ 1 Ĥ 2 ... Ĥ m  (4.15)
(2N x 2N ) (2N x 2)(2N x 2) (2N x 2)
(2N x 2)(2N x 2) (2N x 2)
(2N x 2m) (2N x 2m)
symbolically written as
h i
[S] [H] = Ĥ (4.16)
(2N x 2N )(2N x 2m)
(2N x 2m)
The system matrix [S] must satisfy the dynamic equation (4.10) and the har-
monic relation (4.16). So [S] is solution of
" #  
h i h i
[S] [H] [Y ] =  Ĥ Ŷ  (4.17)
(2N x 2N ) (2N x 2m)(2N x Ln)
(2N x 2m)(2N x Ln)
(2N x (2m+Ln))
(2N x (2m+Ln))
Our objective is to solve the equation (4.17) exactly for the harmonic part and
therefore we partition the matrices in equation (4.17) as
  
[S1 ] [S2 ] [H1 ] [Y1 ]
 (2m x 2m) (2m x 2N-2m)   (2m x Ln) 
  (2m x 2m) 
 [S3 ] [S4 ]  [H2 ] [Y2 ] 
(2N-2m x 2m) (2N-2m x 2N-2m) (2N-2m x 2m) (2n-2m x Ln)
(2N × 2N ) (2N × 2m+Ln)
 h i h i 
Ĥ1 Ŷ1
 
 (2m 
= i )
h x 2m (2m
h x iLn)  (4.18)
 
 Ĥ2 Ŷ2 
(2N-2m x 2m) (2n-2m x Ln)
(2N × 2m+Ln)
Equation (4.18) is first solved for [S1 ] and [S3 ] so that the harmonic part
(4.16) is satisfied exactly, namely
³ ´
S1 = Ĥ1 −S 2 H2 H1−1 (4.19)
³ ´
S3 = Ĥ2 −S 4 H2 H1−1 (4.20)
The remainder of [S] is then computed from the response equations in (4.9)
(corresponding to (4.10)):
¡ ¢ ³ ´
S2 H2 H1−1 Y1 −Y 2 = Ĥ1 H1−1 Y1 −Ŷ 1 (4.21)
4.4. EXPERIMENTAL EXAMPLES 47
¡ ¢ ³ ´
S4 H2 H1−1 Y1 −Y 2 = Ĥ2 H1−1 Y1 −Ŷ 2 (4.22)
These last relations are solved in a least square sense for [S2 ] and [S4 ]. Once [S]
is known, the system eigenvalues are obtained by solving the eigenvalue problem
(4.12). Note that s = ±iω will obviously be one of the solution by construction.
4.4 Experimental examples

The technique discussed above was applied to the analysis of a steel, rectangular
plate in the laboratory. As shown in the figure 4.1, the plate was hanged using
soft stings. A shaker was applied on the structure to provide stationary noise
excitation as well as harmonic loads. A total of 32768 discrete time samples per
response channel have been obtained for each experiment. Sampling was done
at 2560Hz.We identified 6-locations as shown in the diagram, where response
signals are taken. Due to the limited number of acquisition channels (four),
two sets of measurements have been done for each test described hereafter.
Accelerometer one has been kept at the same location in each measurement set
and is used as reference signal (for discussion of practicalities, see [24]). The level
of vibration has been maintained constant throughout each test. Correlation
functions in the tests presented here have been computed with reference to
the response of the first accelerometer. Since the first accelerometer is at the
corner of the plate, it is assumed that this location does not correspond to
a vibration node for the modes of interest. For the same reason, the shaker
excitation is applied at location No. 3, another corner of the plate (see Figure
4.1). It will be assumed that the shaker excites all fundamental modes of the
plate. In the modal identification technique, we will be using correlation signal
as impulse response function. Since a single reference signal has been used,
all identification method will work on the principle of a Single Input Multi
Output (SIMO) identification. In the examples, modal parameters will first be
computed only with stationary white noise input to get the modal parameters
as would be done in a normal OMA setting. Then, we will consider the cases
when harmonic excitations are present along with the random loads. For all
identification methods, the number of correlation data per row and per signal (L
in equation (4.1)) is taken as L = 200 . The total number of discrete correlation
values used in the identification is L + 2N for each correlation function, N being
the order of the identified model.

Here, the experiment was done with stationary white noise introduced between
0Hz to 1000Hz, to excite all the modes properly. Eigenfrequencies and damping
values were identified with the SIMO-SSTD version of the ITD method described
by (4.9) and the LSCE method. In Figure 4.2 and in Table 4.1 we report natural
frequencies and associated damping. Note that the parameters identified by the
SIMO-SSTD are nearly equal to the ones computed by the LSCE approach.
Figure 4.1: Experimental Setup of a Plate
It can be seen from the figure that the modes corresponding to peaks in the
PSD plot are found in a stable manner by both identification methods. It can
be observed that there is an additional stability line at 59Hz, which is not
listed in the Table 4.1. It was found later that this excitation frequency was
due to a the amplifier of the shaker. Now we will concentrate one the mode at
271.047Hz. In the subsequent examples, in addition to the random excitations,
we will introduce harmonic excitations near to that 271.047Hz mode and we
will try to identify modal parameters of this particular mode. In particular, we
are interested in investigating the accuracy of the identification when harmonic
excitations with frequencies close to the 271.047Hz eigenfrequency are present.
4.4.2 Two harmonics at 274Hz and 276Hz

In this experiment, two harmonic frequencies additional to the white noise are
present at 274Hz and 276Hz, which are close to the 271.047Hz modal fre-
quency of the system. In Figure 4.4 two peaks at 274Hz and 276Hz are clearly
observed from the power spectral density (PSD) plot. One secondary peak is
present slightly below 274Hz which might be representing the 271.047Hz nat-
ural frequencies of the system. As described in the theory, we introduced four
additional rows of sin and cos terms to explicitly account for the two harmonic
frequencies in the response matrices (see Eqs. (4.14) and (4.18)). Figure 4.4
shows the stability diagram superimposed on the auto-spectrum diagram. In
Table 4.2 frequencies and dampings for modes in the neighborhood of 271.047Hz
Number N of modes used in the identification damping

90 40.00
0.00% 0.60% 1.20% 1.80% 2.40% 3.00%
80 30.00
Power Spectral Density, dB

70 20.00
60 10.00
50 0.00
40 -10.00
30 -20.00
20 -30.00
10 -40.00
0 -50.00
100 200 300 400 500 600 700 800 900 1000
Frequency,
Hz
Figure 4.2: Stability diagram for white noise excitation only computed by the
SIMO-SSTD method
are listed. These modal parameters were found with the SIMO-SSTD and the
SIMO-SSTD variant including harmonics as presented in this paper. From
Table 4.2 and Figure 4.4 it can be concluded that, when using the normal
SIMO-SSTD, the identified frequencies are associated to the two fundamental
harmonic frequencies. Note that the identified damping corresponding to the
harmonic frequencies are not small, while there should be null in theory. Hence,
in practice, it would be difficult at this point to assimilate the identified modes to
harmonic responses and the analyst would probably conclude that the identified
parameters correspond to true eigenmodes. When applying the new method, we
have introduced in the Hankel matrix two harmonic components of frequencies
corresponding to the periodic excitations, namely 274Hz and 276Hz. The sta-
bilization diagram is shown in Figure 4.4. In the Table 4.2 the identified modal
frequency corresponding to the mode in the neighborhood of 271.047Hz and its
associated damping are reported. Note that harmonic frequencies are identified
exactly by construction in the modified SIMO-SSTD method and are therefore
not listed in the table. It is seen that the identified eigenfrequency and the as-
sociated damping are very close to the expected values. Modal parameters were
also computed using the previously published modified LSCE method [41, 47]
which also explicitly takes account of the harmonic components, but which is
based on the LSCE algorithm. The modified LSCE yields 270.911Hz and the
associated damping is 0.169%. Therefore, whereas we failed to get any cor-
rect modal parameters from the standard SIMO-SSTD method, the modified
SIMO-SSTD and LSCE methods yield accurate eigenfrequencies even for low
damping
90 40.00
0.00% 0.80% 1.60% 2.40% 3.20% 4.00%
80 30.00
Number N of modes used in the identification

70 20.00
60 10.00
Power Spectral Density, dB

50 0.00
40 -10.00
30 -20.00
20 -30.00
10 -40.00
-50.00
100 200 300 400 500 600 700 800 900 100
Frequency, Hz 0
Figure 4.3: Stability diagram for white noise excitation only computed by LSCE
method
identification orders. Note that the damping identified by the modified SIMO-
SSTD is slightly different from the one identified in the absence of harmonic
excitation (namely 0.3%, see Table 4.1). However it seems that the modified
SIMO-SSTD yields more accurate damping ratios than the modified LSCE.
4.4.3 Three additional harmonics at 277Hz, 282Hz and

287Hz
Now three sine harmonic frequencies at 277Hz, 282Hz and 287Hz are given
as input to the shaker along with the stationary white noise in order to inves-
tigate the effectiveness of the proposed method in the presence of harmonics
components of high amplitude. As mentioned in the theory, we have introduced
all three harmonics in the Hankel matrix of the proposed method to take into
account the harmonic frequencies. In Figure 4.5 three harmonic frequencies can
be clearly identified at 277Hz, 282Hz and 287Hz. We also observe just before
the first harmonic frequency a peak which corresponds to the natural frequency.
As shown in Table 4.3 and Figure 4.5, the basic SIMO-SSTD method exhibits
two stabilization lines. Those lines correspond to 277Hz and 287Hz, which is
related to two of the harmonic frequencies introduced in the signal. There is
no indication of the presence of the second harmonic frequency in the graph.
Although in theory those frequencies should be associated to zero damping, the
results in Table 4.3 indicate that the damping of the identified harmonic parts
are quite significant. One can thus conclude that, when using standard identifi-
SIMO-SSTD LSCE
Freq. Damp. Freq. Damp.
(Hz) (%) (Hz) (%)
28.754 1.726 28.665 1.794
141.731 0.326 141.754 0.327
271.047 0.300 271.093 0.297
328.105 0.226 328.150 0.225
408.485 0.132 408.546 0.132
560.472 0.486 560.563 0.483
624.050 0.470 624.130 0.481
678.931 0.164 679.059 0.180
731.271 0.162 731.642 0.153
844.701 0.282 844.826 0.267
892.166 0.280 892.337 0.275
Table 4.1: Frequencies and associated dampings (white noise only)
cation procedure such as the SIMO-SSTD version of the ITD, the identification
procedure breaks down in the presence of harmonic excitations with frequencies
close to eigenfrequencies. Figure 4.5 and Table 4.3 also show the stability di-
agram for the modified SIMO-SSTD method proposed in this paper. The last
three frequency lines belong to the three harmonics, and their frequencies have
been set to exactly match the harmonic frequencies, the associated damping
being zero (those frequencies and damping values are not given in Table 4.3).
The other identified frequency and damping is listed in Table 4.3 and can be
seen to be very close to the frequency and damping of the searched mode. We
have also computed the modal parameters by the modified LSCE method. The
frequency found is 271.175Hz and the associated damping is 0.22%.
4.4.4 Robustness of the modified SIMO-SSTD with re-

spect to assumed harmonic frequencies
From all the examples, it is observed that classical identification algorithms ap-
plied to the correlation functions obtained from OMA do not work well in the
presence of harmonic excitations with frequencies close to the natural frequen-
cies of the system. The new method proposed here is able to provide reliable
modal parameters in those cases. Obviously, in the new method proposed here,
it is important to introduce the exact harmonic frequencies in the identification
a priori (see (4.18)). Let us once more consider the experimental example where
two harmonic excitations are acting at 274Hz and 276Hz (as in section 4.4.2).
In order to investigate the sensitivity of the identification to the accuracy of the
harmonic frequency specified in the modified SIMO-SSTD, we will now spec-
ify assumed harmonic frequencies different from the actual ones. In figures 4.6
and 4.7, we plot the eigenfrequency identified by the modified SIMO-SSTD as
Figure 4.4: PSD of the acceleration at position 1 for random loads and two
additional harmonic excitations, and stabilization diagram for the SIMO-SSTD
(4.10) and for the SIMO-SSTD modified to explicitly include harmonic compo-
nents (4.18)
a function of the assumed harmonic frequencies, for an identification order of

46. The actual harmonic frequencies are 274Hz and 276Hz while the assumed
ones introduced in the identification are varied simultaneously. One observes
that if the assumed harmonic frequencies are different from the actual ones, the
identified eigenfrequency varies linearly with respect to the harmonic frequency
input in the vicinity of the exact harmonic frequency. Also the identified damp-
ing ratio is very sensitive to the accuracy of the harmonic frequencies specified
in the identification procedure. Hence, it is essential to introduce in the new
algorithm assumed harmonic frequencies accurately. Otherwise the harmonic
components explicitly included in the identification are not capable of repre-
senting the harmonic part of the signal properly and the new method gives no
better results than standard procedures.
4.5 Conclusion
In this chapter, we have considered a SIMO-SSTD variant of the Ibrahim Time
Domain algorithm to identify modal parameters from measurements in Opera-
tional Modal Analysis. In particular, a modification is introduced in the iden-
tification procedure to explicitly account for harmonic components that might
be present in addition to the response to stationary white noise. The proposed
modified SIMO-SSTD methods allows to identify accurately eigenfrequencies of
4.5. CONCLUSION 53
Table 4.2: Frequencies and associated damping identified by the SIMO-SSTD

and the modified SIMO-SSTD.
SIMO-SSTD Modified SIMO-SSTD
Modes Freq.1 Damp.1 Freq.2 Damp.2 Freq.3 Damp.3 Freq. Damp.
(Hz) (%) (Hz) (%) (Hz) (%) (Hz) (%)
80 – – 273.849 0.069 276.031 0.136 270.031 0.341
76 – – 273.813 0.080 275.965 0.168 271.009 0.247
72 – – 273.856 0.099 276.042 0.202 270.945 0.246
68 – – 273.860 0.075 276.065 0.140 270.975 0.256
64 268.936 0.078 273.948 0.121 276.295 0.248 270.970 0.249
60 268.749 0.543 274.039 0.135 276.611 0.100 271.037 0.200
56 269.325 1.358 274.110 0.160 276.916 0.468 271.095 0.257
52 267.396 1.223 274.068 0.177 276.459 0.546 270.964 0.190
48 – – 274.123 0.223 275.75 0.814 271.072 0.316
44 – – 274.184 0.279 274.800 0.729 271.357 0.425
40 – – 274.176 0.295 274.764 0.664 271.301 0.545
36 – – 273.947 0.378 274.783 0.419 272.125 1.638
32 – – 273.773 0.356 274.938 0.357 – –
28 – – 273.875 0.330 275.020 0.411 – –
24 – – 273.999 0.236 275.532 0.570 – –
the system even when harmonic components with frequencies close to an eigen-
frequency are present. Damping parameters are however difficult to identify
accurately if harmonic excitations hide the actual eigenresponse of the system.
Standard identification methods fail to identify the system parameters prop-
erly in that case. Results of the method proposed here are slightly better than
identification results obtained with a previously proposed modified LSCE ap-
proach [47] (typically dampings are better identified). In our tests, we found
that in order to identify eigenparameters accurately, the harmonic frequencies
introduce a priori in the identification should be known precisely. In case the
assumed harmonic frequencies are different from the actual ones, the identifica-
tion with the modified algorithm yield results similar to the ones obtained from
standard methods.
additional harmonic excitations, and stabilization diagram obtained from the
SIMO-SSTD and the modified SIMO-SSTD.
Table 4.3: Frequencies and associated damping identified by the SIMO-SSTD

and the modified SIMO-SSTD.
standard SIMO-SSTD Modified SIMO-SSTD
Modes Freq.1 Damp.1 Freq.2 Damp.2 Freq.3 Damp.3 Freq. Damp.
(Hz) (%) (Hz) (%) (Hz) (%) (Hz) (%)
80 271.807 0.354 280.257 1.028 287.659 0.007 271.002 0.236
76 271.769 0.180 276.723 0.024 287.972 0.192 270.992 0.233
72 271.835 0.079 276.725 0.048 287.794 0.258 270.979 0.229
68 – – 276.805 0.060 287.763 0.396 270.992 0.232
64 – – 276.773 0.080 287.472 0.314 271.058 0.232
60 – – 276.831 0.077 287.344 0.373 270.990 0.252
56 – – 276.866 0.080 287.172 0.386 270.948 0.284
52 – – 276.882 0.075 287.099 0.411 270.891 0.278
48 – – 276.878 0.075 287.105 0.418 270.945 0.503
44 – – 276.891 0.072 287.021 0.440 270.877 1.010
40 270.921 0.066 276.921 0.072 286.854 0.442 272.311 1.727
36 270.762 0.248 276.949 0.058 286.703 0.554 – –
32 271.987 0.286 276.863 0.030 287.677 0.807 – –
28 275.099 0.998 – – 287.942 1.625 – –
24 – – 277.041 0.230 277.123 0.265 – –
4.5. CONCLUSION 55
Figure 4.6: Eigenfrequency vs. assumed harmonic frequency in the modified

SIMO-SSTD (harmonic excitations at 274Hz and 276Hz)
(b) 0.5
0.45
0.4
0.35
actual damping
Damping, %
0.3
0.25
0.2
0.15
exact harmonic
0.1
frequencies
0.05
0
X 273.6 273.8 274 274.2 274.4 274.6 274.8
Y 275.6 275.8 276 276.2 276.4 276.6 276.8
Harmonic Frequency, Hz
Figure 4.7: Damping vs. assumed harmonic frequencies in the modified SIMO-
SSTD (harmonic excitations at 274Hz and 276Hz).
Chapter 5
Modified ERA method for

harmonic excitations
In this chapter a modified version of the Eigensystem Realization Algorithm(ERA)

method is proposed to perform operational modal analysis in the presence of har-
monic excitations. Like for the LSCE (Ch. 3) and modified SIMO-SSTD (Ch.
4) methods, auto- and cross-correlation functions of the system are used as
impulse response functions following the Natural Excitation Technique (NExT)
(Sec. 2.3). Section (Sec. 5.1) contains the introduction. In section (Sec. 5.2),
subsection (Sec. 5.2.1), the theory behind ERA methods for modal parameter
estimation is outlined. In the subsection (Sec. 5.2.2) a modified version of the
standard ERA method is proposed in order to incorporate harmonic excitations.
Section (Sec. 5.3) contains a brief description of the experimental setup and of
signal processing issues.
In sections (Sec. 5.3.1, 5.3.2, 5.3.3), modal parameters of the experimental
system have been computed and comparisons are made between the parameters
identified with the standard ERA and with the modified ERA method for har-
monics. Section (Sec. 5.3.4) describes the sensitivity analysis of the proposed
modified ERA method with respect to the accuracy of the harmonic value given
to the algorithm. Conclusions of the chapter are given in the last section (Sec.
5.4).
The work of this chapter has been published in [38, 43].
57
58 CHAPTER 5. MODIFIED ERA METHOD FOR HARMONICS
5.1 Introduction
Like the LSCE and ITD methods, described in the chapters (Ch. 3) and (Ch. 4)
respectively, the Eigensystem Realization Algorithm (ERA) method works on
the principle of damped complex exponential modelling of structural responses
in time domain. The ERA method is a Multi Input and Multi Output (MIMO)
method, where forces are measured at multiple points and at the same time,
response signals are measured at multiple points on the structure. The ERA
method is based upon realization theory (state-space), which uses the concept
of controllability and observability of state-space modal model [31–33].
In this chapter at (Sec. 5.2) ERA method is outlined and applied to the
auto and cross-correlation signals taken as Impulse response functions (IRF),
as described in (Sec. 2.3). Than a modified version of the ERA method is
proposed (Sec. 5.2.2) to take into consideration multiple harmonic excitation
signals present in the response signals in addition to the response from the
white-noise. The modified ERA method is a Single Input Multi Output (SIMO)
version of the ERA method.
The modified ERA procedure can be applied to the correlations of response
signals based on the principle of NExT [28]. In section (Sec. 5.3), experimental
results obtained with the modified ERA algorithm will be compared to results
obtained with the standard ERA method in the presence of harmonics.
5.2 Operational Modal Analysis and the modi-

fied Eigensystem Realization Algorithm
In chapter (Ch. 2) the correlation between signals has been shown to be equiv-
alent to a superposition of decaying sinusoids having damping and frequencies
equal to the damping and frequencies of the structural mode. According to Eq.
(3.3) the correlation function can be written as
2N
X
R(k∆t) = Rk = ar esr k∆t (5.1)
r=1
where Rk = Rij (k∆t) is the discrete correlation function at k th discrete time, ar

is a constant associated with the rth mode and sr is the rth complex frequency.
5.2.1 Basics of the ERA algorithm

In order to explain how the ERA method is modified in our work, we will
first recall the main concepts underlying the ERA algorithm (a more thorough
explanation of the method can be found for instance in [32]. Then we will
indicate how the standard ERA is modified.
A linear system is characterized by a system matrix [A] relating its state
space variables {u} at two successive instances. The measured outputs {x} of
5.2. OMA AND MODIFIED ERA 59
the system (a restricted number of the state space variables) at a time instant
tk+1 then satisfies the relation
½
{u}k+1 = [A] {u}k + {b}k
(5.2)
{x}k+1 = [C] {u}k+1
where {u}k defines the state space of the system at time tk = k∆t, {b}k is
the input to the system (assumed null after the initial time for an impulse
response) and [C] the output operator (typically a Boolean restriction operator).
It can be shown that the eigenvalues of the system matrix [A] correspond to
e(−ζ r ωr ±iωr )∆t , ∆t being the sampling time interval assumed to be constant.
d
The true system matrix [A] has a dimension equal to twice the number of
physical degrees of freedom in the system, namely 2N . The number of measured
outputs is assumed to be m.
In the identification procedures one tries to construct the realized matrix [Ar ]
approximating [A] from measurements {x}k , k = 1, . . . . Once the system matrix
[Ar ] has been “realized” based on the measurements, computing its eigenvalues
will then yield the eigenparameters of the system.
Since the ERA algorithm will be applied in the context of the NExT tech-
nique, we assume that an impulse is given before time 0 and that the system
vibrates freely. In order to identify the system matrix, one builds the Hankel
matrices [H (0)] and [H (1)] of the response functions (correlation functions in
the case of NExT) starting for t = 0 and t = ∆t respectively, namely
 
{x}k {x}k+1 · · · {x}k+q−1
 {x}k+1 {x}k+2 · · · {x}k+q 
 
[H (k)] =  . .. .. ..  (5.3)
mp×q  .. . . . 
{x}k+p−1 {x}k+p · · · {x}k+q+p−2
where p and q are integers defining the number of time instances considered
in the Hankel matrix. Here we assume that, in agreement with the NExT
approach, the m output {x} are impulse responses associated to an unknown
single input. Hence, as indicated by (5.1), during OMA analysis {x}k in (5.3)
will be given by the correlations {R}k .
Since the outputs at two successive instances are implicitly related through
the system matrix, the main idea of the ERA algorithm consists in building a
system matrix [Ar ] (the realized matrix) based on two successive Hankel ma-
trices [H(0)] and [H(1)] for instance. First one observes that, according to the
system equations (5.2), the Hankel matrix [H(0)] can be written as
[H(0)] = [Q][W ] (5.4)
where
 
[C]
 [C] [A] 
 
[Q] =  [C] [A]
2 
 
..
.
£ ¤
[W ] = {u}0 [A] {u}0 [A]20 {u} · · · (5.5)
To find [A], let us observe that the next Hankel matrix [H(1)] can be expressed
as
[H(1)] = [Q][A][W ] (5.6)
Note that at this point [Q] and [W ] are unknown but satisfy relation (5.4). If a
pseudo-inverse of [Q] and [W ] would be known, the system matrix [A] could be
computed.
A pseudo inverse for [Q] and [W ] is naturally obtained by considering the
Singular Value Decomposition (SVD) of [H(0)] (see e.g. [36]). Considering m
responses contributes to the construction of Hankel matrix [H(0)], it can be
decomposed as:
[H(0)] = [U ] [Σ] [V ]T (5.7)
mp×q mp×mp mp×q q×q
where [Σ] is the diagonal matrix of the singular values and [U ] and [V ] are
orthonormal matrices such that
[U ]T [U ] = [I] [V ]T [V ] = [I] (5.8)

mp×mp q×q
Hence defining
[Q] = [U ][Σ]1/2 [W ] = [Σ]1/2 [V ]T (5.9)
the relation (5.4) is satisfied and one can readily define the pseudo-inverses
[Q]+ = [Σ]+1/2 [U ]T [W ]+ = [V ][Σ]+1/2 (5.10)
such that [Q]+ [Q] = [I] and [W ][W ]+ = [I]. Note that [Σ]+1/2 is the pseudo-
inverse obtained simply by inverting and taking the square root of the diagonal
part of [Σ] and, by definition, it satisfies
[Σ]1/2 [Σ]+1/2 [Σ]1/2 = [Σ]1/2
Pre-multiplying by [Q]+ and post-multiplying by [W ]+ the expression (5.6) for

[H(1)], one finds
[Q]+ [H(1)][W ]+ = [A] (5.11)
The ERA method thus consists in building the system matrix A from (5.11)
where [W ]+ and [Q]+ are defined by (5.7, 5.10).
One should note that, if the system matrix A has an order 2N , there are
only 2N non-zero singular values in [Σ]. Hence in the definitions (5.10) of [Q]+
and [W ]+ , we should consider only the 2N non-zero singular values and the
corresponding 2N singular vectors:
−1/2 T −1/2
[Q2N ]+ = [Σ2N ] [U2N ] [W2N ]+ = [V2N ] [Σ2N ] (5.12)
2N ×2N 2N ×mp q×2N 2N ×2N
Unfortunately, in practice systems have an infinite order and output mea-

surements are corrupted by noise and errors. So that choosing the number of
2N singular values in [Σ] that are significant for realizing the system matrix
[A] is not straightforward since, contrary to the theory, no singular value will
actually be null. Nevertheless, if the measurements are performed properly, one
often observe a jump in the singular values on the diagonal of [Σ] so that one
can choose an appropriate order N for the system at hand. Having then chosen
2N singular values, the realized system matrix [Ar ] is actually computed as
[Q2N ]+ [H(1)][W2N ]+ = [Ar ] (5.13)
As explained before, computing the eigenvalues of [Ar ] leads then to the system
eigenparamaters.
5.2.2 A modified ERA scheme to explicitly include har-

monic responses
Let us now assume that, in the OMA, harmonic excitations are present. When
computing the auto and cross-correlation between signals, one obtains on top of
the impulse-like responses, harmonic excitation components of frequency ω j,ex ,
j = 1...mh , the mh harmonic excitation frequencies exciting the system in ad-
dition to the random loads. In theory the ERA algorithm can still be applied
to such a response, in which case one would obtain in addition to the eigen-
paramaters, mh virtual non-damped modes with poles sj = ±iω j,ex , so that
esj ∆t = e±iωj,ex ∆t = cos(ω j,ex ∆t) ± i sin(ω j,ex ∆t). However, in practice, the
identified damping of those modes is not exactly null so that distinguishing be-
tween genuine modes and harmonic responses might be difficult or impossible.
This is especially true when a harmonic frequency is close to an eigenfrequency
of the system.
In this chapter we propose to modify the ERA algorithm such that the
identified system matrix [A] will be forced to have 2mh number of exact roots
for mh number of harmonics, the harmonic poles being assumed to be known
a priori.1 Hence we will force the algorithm to identify the known harmonic
frequencies properly. To that purpose we add to the Hankel matrices [H(0)]
and [H(1)] the response of virtual measurements representing single oscillators
1 Note that if the harmonic frequencies are not known a priori, they can be evaluated
accurately by analyzing the cross-correlation functions for times when the impulse response
is damped out.
vibrating with the known harmonic frequencies:

 
0 sin (ω 1 4t) · · · sin ((q − 1) ω 1 4t)
 1 cos (ω 1 4t) · · · cos ((q − 1) ω 1 4t)  
 
 .. .. .. ..
  
 . . . . 
  [Hh (0)]
 
[H̃(0)] = 0 sin (ω mh 4t) · · · cos ((q − 1) ω mh 4t) =  (5.14)
 1 cos (ω mh 4t) · · · cos ((q − 1) ω mh 4t) 
(mp+2mh )×q  [H (0)]
 2mh ×q 
 
 
[H (0)]
mp×q
and
 
sin (ω 1 4t) sin (2ω 1 4t) ··· sin (qω 1 4t)
 cos (ω 1 4t) cos (2ω 1 4t) ··· cos (qω 1 4t)
 
 .. .. .. .. 
  

 . . . .  [Hh (1)]

 sin (ω m 4t) sin (2ω mh 4t)
[H̃(1)] = · · · cos (qω mh 4t) = 

 cos (ω m 4t) cos (2ω mh 4t) · · · cos (qω mh 4t) 
(mp+2mh )×q   [H (1)]
 2mh ×q 
 
 
[H (1)]
mp×q
(5.15)
where [Hh (0)] and [Hh (1)] are the matrices for the harmonic components. Note
that extending the Hankel matrix in this way does not modify the original modal
content of the signal. Let us then apply the SVD to this extended Hankel matrix
and replace (5.7) by
[H̃(0)] = [Ũ ] [Σ̃] [Ṽ ]T = [Q̃][W̃ ] (5.16)

(mp+2mh )×q (mp+2mh )×(mp+2mh ) (mp+2mh )×q q×q
Truncating the singular values to the 2N relevant ones, we can write
[Q̃2N ] = [Ũ2N ][Σ̃2N ]1/2 [W̃2N ] = [Σ̃2N ]1/2 [Ṽ2N ]T (5.17)

(mp+2mh )×2N 2N ×q
Note that now N is the number of degrees of freedom relevant for the system
plus the mh , the number of harmonic response components in the signal.
Since the SVD has to be truncated, [Q̃2N ][W̃2N ] will not be exactly equal to
[H̃(0)]:
· ¸
[Hh (0)]
[Q̃2N ][W̃2N ] = + Error (5.18)
[H(0)]
In particular, partitioning [Q̃2N ] in

· ¸
[Q̃h ]
[Q̃2N ] =
[Q̃]
(5.18) shows that

[Q̃h ][W̃2N ] = [Hh (0)] + Errorh
meaning that the harmonic components will no longer be correctly reproduced
˜ ]
by [Q̃2N ][W̃2N ]. Hence we propose to replace [Q̃2N ] by a corrected [Q̃2N
h i · ¸
˜ [Q̃h ] + [Qcor ]
Q̃ 2N = (5.19)
[Q̃]
so that
([Q̃h ] + [Qcor ])[W̃2N ] = [Hh (0)] (5.20)
Re-arranging this last equation, the correction [Qcor ] is solution of
[Qcor ] [W̃2N ] = [Hh (0)] − [Q̃h ][W̃2N ] (5.21)

2mh ×2N 2N ×q
Since in practice q > 2N , an exact solution might not exist, but a least square
approximate for [Qcor ] is found if we use the pseudo-inverse of [W̃2N ], namely
³ ´
[Qcor ] = [Hh (0)] − [Q̃h ][W̃2N ] [V2N ][Σ2N ]−1/2
= [Hh (0)][V2N ][Σ2N ]−1/2 − [Q̃h ] (5.22)
Now the system matrix can be realized following the ERA algorithm by
solving
˜ ][Ar ][W̃ ]
[H̃(1)] = [Q̃ (5.23)
2N 2N
˜ ] is computed as in (5.19 - 5.22), and where [W̃ ] is obtained as in

where [Q̃2N 2N
(5.17). Defining as before the pseudo-inverse of [W̃2N ] as
[W̃2N ]+ = [V2N ][Σ2N ]−1/2 (5.24)
the equation (5.23) for the realized matrix writes
˜ ][Ar ] = [H̃(1)][W̃ ]+ = [S]

[Q̃ (5.25)
2N 2N
In the standard ERA procedure [Ar ] is easily found by considering the pseudo
˜ ] has no readily available pseudo-inverse, we solve
inverse of Q̃ , but since [Q̃
2N 2N
(5.25) in a least square sense, making sure that at least the harmonic part of
the relation is satisfied exactly: let us partition (5.25) as:
  
˜ ]
[Q̃ ˜ ]
[Q̃ [Ar ]11 [Ar ]12
2N 11 2N 12
 2mh ×2mh 2mh ×(2N −2mh ) 
  2mh ×2mh 2mh ×(2N −2mh ) 
 ˜ ˜  [Ar ]21 [Ar ]22 
[Q̃2N ]21 [Q̃2N ]22
mp×2mh mp×(2N −2mh ) (2N −2mh )×2mh (2N −2mh )×(2N −2mh )
 
[S]11 [S]12
 2m ×2mh 2mh ×(2N −2mh ) 
= h  (5.26)
[S]21 [S]22
mp×2mh mp×(2N −2mh )
Now Equation (5.26) can be decomposed as:
˜ ] [Ar ] + [Q̃
[Q̃ ˜ ] [Ar ] = [S] (5.27)
2N 11 11 2N 12 21 11
˜ ] [Ar ] + [Q̃
[Q̃ ˜ ] [Ar ] = [S] (5.28)
2N 11 12 2N 12 22 12
˜ ] [Ar ] + [Q̃
[Q̃ ˜ ] [Ar ] = [S] (5.29)
2N 21 11 2N 22 21 21
˜ ] [Ar ] + [Q̃
[Q̃ ˜ ] [Ar ] = [S] (5.30)
2N 21 12 2N 22 22 22
Let us observe that the first 2mh equations in (5.23) (and thus equations (5.27)
and (5.28)) relate to the harmonic response components. Since we want to
build a system matrix [Ar ] exactly reproducing the roots associated to the har-
monic response, equations (5.27) and (5.28) will be solved exactly. After simple
manipulations we thus obtain:
³ ´
˜ ]−1 [S] − [Q̃
[Ar ]11 = [Q̃ ˜ ] [Ar ] (5.31)
2N 11 11 2N 12 21
³ ´
˜ ]−1 [S] − [Q̃
[Ar ]12 = [Q̃ ˜ ] [Ar ] (5.32)
2N 11 12 2N 12 22
Replacing in (5.29) and (5.30), one obtains

³ ´ ³ ´
˜ ] [Q̃
[Q̃ ˜ ]−1 [Q̃
˜ ] − [Q̃˜ ] r ˜ ˜ −1
2N 21 2N 11 2N 12 2N 22 [A ]21 = [Q̃2N ]21 [Q̃2N ]11 [S]11 − [S]21
³ ´ ³ (5.33)
´
˜ ˜ −1 ˜
[Q̃ ] [Q̃ ] [Q̃ ] − [Q̃ ] ˜ r ˜ ˜ −1
[A ] = [Q̃ ] [Q̃ ] [S] − [S]
2N 21 2N 11 2N 12 2N 22 22 2N 21 2N 11 12 22
(5.34)
The last two relations are solved in a least-square sense for [Ar ]21 and [Ar ]22 .
Finally, substituting in (5.31) and (5.32), the realized matrix is obtained. Since
equations (5.27) and equation (5.28) have been solved exactly, the roots of the
system matrix will include the harmonic components. Additional roots will
represent the modal parameters of the system.
5.3 Experiment and Signal Processing

In order to evaluate the effectiveness of the proposed modified ERA procedure,
the beam structure described in section (Sec. 3.4, Figure 3.1) will be analyzed.
5.3. EXPERIMENT AND SIGNAL PROCESSING 65
First a preliminary experiment was done to obtain estimates for eigenmodes

and eigenfrequencies. Harmonic excitations will be applied on the beam with a
frequency close to the second eigenfrequency later in the experiment and there-
fore 4 accelerometers were placed along the length to properly measure the
transverse displacements of the second mode. The time signals were sampled at
128Hz, the second eigenfrequency being close to 23Hz. A total of 8192 time-
samples were taken for each response accelerometer in each experiment. Cor-
relation functions of the systems were computed taking the first accelerometer
signal as reference. As required by the theory presented here, only a single ref-
erence analysis will be used both for standard ERA algorithm and the modified
ERA algorithm to show the difference of results computed by both algorithms.
For all analyses discussed hereafter, a total of p + q = 200 discrete correlation
data have been used per correlation function.

First stationary white noise with a spectrum between 18Hz to 28Hz was applied
to the beam to excite the second bending mode properly. Eigenfrequencies and
damping values were identified with the standard ERA method. In Figure
5.1 and Table 5.1 convergence of the poles to the identified first, second and
third eigenfrequencies are reported for increasing orders N considered in the
identification. Three modes can clearly be identified from the stability plot.
Figure 5.1: Stability diagram for white noise excitation only computed by the
ERA method
Table 5.1: Frequencies and associated dampings identified by ERA (white noise
excitation only)
Standard ERA
System order Freq.1 Damp.1 Freq.2 Damp.2 Freq.3 Damp.3
N (Hz) (%) (Hz) (%) (Hz) (%)
30 6.59 0.22 23.15 0.30 50.57 0.10
28 6.59 0.22 23.15 0.31 50.57 0.11
26 6.59 0.22 23.15 0.31 50.57 0.10
24 6.59 0.22 23.15 0.31 50.57 0.09
22 6.59 0.22 23.15 0.31 50.58 0.09
20 6.59 0.22 23.14 0.30 50.58 0.09
18 6.59 0.22 23.14 0.29 50.58 0.10
16 6.59 0.22 23.15 0.32 50.58 0.08
14 6.59 0.22 23.15 0.32 50.58 0.10
12 6.59 0.22 23.15 0.30 50.58 0.11
10 6.59 0.22 23.15 0.30 50.56 0.11
8 6.59 0.22 23.15 0.30 50.56 0.11
6 6.59 0.22 23.15 0.30 50.57 0.10
4 6.59 0.53 23.16 0.26 50.58 0.10
As the beam has been excited with white noise excitation between 18Hz and
28Hz, we assume that the modal parameter associated with the second mode are
identified properly. From Table 5.1 it is observed that the second eigenfrequency
is evaluated as 23.15Hz and the associated damping is 0.30%. Note that the
frequency of the second mode identified by the ERA procedure is nearly the
same as obtained from the LSCE analysis in section (Sec.3.4.1). The associated
damping ratio found by the ERA is however slightly higher since the LSCE
procedure resulted in a damping ratio of 0.26%.
5.3.2 Two harmonics at 23.30Hz and 23.80Hz

We now add to the white noise load two harmonic excitations at frequencies
23.30Hz and 23.80Hz, which are close to the second mode frequency of the
beam (23.15Hz). We first perform an identification using the standard time-
domain technique ERA. Then the modified ERA procedure as presented in the
previous section is applied. In the stability plot Figure 5.2, modal parameters
computed by both the standard and the modified ERA procedures are shown.
The numerical values are listed in Table 5.2. In Figure 5.2, two peaks at 23.30Hz
and 23.80Hz are clearly observed on the PSD plot of the signal from the 1st
accelerometer. A secondary peak is present slightly below the 23.30Hz har-
monic, which corresponds to the the second bending mode frequency. From
Table 5.2 and Figure 5.2, it is observed that the modal parameters computed
by the standard ERA method in the neighborhood of the second eigenfrequency
Figure 5.2: Stability diagram for white noise loading and additional harmonic
excitations at 23.30 Hz and 23.80 Hz (standard and modified ERA methods).
only indicate one stability line which is associated to the first harmonic excita-
tion frequency, 23.30Hz. In Table 5.2 we read that the frequency identified by
the standard ERA is 23.27Hz with an associated damping of 0.11%. Note that,
while the damping should be zero in theory, it is not small in the results and
therefore an analyst could be misled and interpret this pole as a true eigenpole
of the system. Note also that no other stability lines representing the second
harmonic excitation frequency and the second eigenmode of the beam appear
when the standard ERA is used. So it can be concluded that the standard
ERA procedure fails to identify the actual modal parameter in the present case.
Consider now in Figure 5.2 and Table 5.2 the results obtained with he modified
ERA. As described in the theory of the previous section, when applying the
modified ERA approach 4 rows of sine and cosine terms have been added in
the Hankel matrix for the two harmonic excitations (see equation (5.14-5.15)).
In Figure 5.2, two stability lines (represented by + symbols) exactly match the
two harmonic frequencies and the associated dampings are exactly zero as ex-
pected. These values are not listed in the Table 5.2 for simplicity. The third
stability line observed in Figure 5.2 when applying the modified ERA corre-
sponds to the second eigenfrequency of the beam. The corresponding frequency
and damping (see Table 5.2) are very close to the eigenparameters of the second
mode as identified earlier in the experiment where pure white noise was used
(Table 5.1). Therefore, whereas we failed to get any correct modal parameters
from the standard ERA method, the modified ERA approach allows to identify
Table 5.2: Frequencies and associated damping identified by the standard ERA
and the modified ERA. System excited by stationary white noise and harmonic
loads at 23.30Hz and 23.80Hz.
Standard ERA Modified ERA
System order Freq. Damp. Freq. Damp.
N (Hz) (%) (Hz) (%)
30 23.27 0.11 23.18 0.29
28 23.27 0.11 23.15 0.33
26 23.27 0.11 23.15 0.32
24 23.27 0.11 23.15 0.27
22 23.27 0.11 23.15 0.28
20 23.27 0.12 23.16 0.28
18 23.27 0.11 23.15 0.27
16 23.27 0.12 23.16 0.26
14 23.27 0.12 23.16 0.27
12 23.27 0.12 23.16 0.28
10 23.27 0.12 23.16 0.19
8 23.27 0.10 23.16 0.20
6 23.27 0.15 23.17 0.25
4 23.27 0.15 – –
the eigenparameters accurately.
5.3.3 Harmonics at 23.30Hz, 23.80Hz and 24.30Hz

Similarly to the previous test, we now introduce 3 additional harmonic exci-
tations at 23.30Hz, 23.80Hz and 24.30Hz. The harmonic components can be
clearly seen on the PSD plot in Figure 5.3. Again the standard and the mod-
ified ERA methods will be applied to extract the modal parameters. From
Figure 5.3 and Table 5.3, it is observed that the standard ERA method identi-
fies poles associated with the first two harmonics and that those poles have an
associated damping that is not small (while it should be zero in theory). No
stability line associated with the second bending mode of the beam is found.
When the modified ERA procedure is applied, the stability lines corresponding
to the harmonic components are indicated by + symbols in the Figure 5.3. The
harmonic frequencies are exact by construction of the identification algorithm
and have zero damping. In Figure 5.3 a stability corresponding to the second
bending mode is also visible when the modified ERA method is used. In Table
5.3, the corresponding frequencies and computed damping are given. Clearly
this identified pole is close to the real second mode frequency and damping as
computed in the reference test (pure noise, Table 5.1). Note however that the
damping ratio is slightly under-estimated.
Figure 5.3: Stability diagram for white noise loading and additional harmonic
excitations at 23.30 Hz, 23.80 Hz and 24.30Hz (standard and modified ERA
methods).
5.3.4 Robustness of the modified ERA method

From the two examples above it can be concluded that standard time-domain
methods like ERA applied to correlation functions in the presence of harmonic
excitations do not compute modal parameters properly if harmonic excitation
frequencies close to eigenfrequencies are present. The modified ERA method
however is able to compute modal parameters of the structures even when har-
monic excitation frequencies are close to natural frequencies. In order to investi-
gate the robustness of the identification procedure with respect to the difference
between harmonic excitation frequency and eigenfrequency, more tests on the
beam structure were performed where a single harmonic excitation was added to
the stationary white noise. The experiment was repeated for different harmonic
excitation frequencies in the neighborhood of the second bending frequency. The
level of the harmonic input relative to the level of random excitation was kept
constant. The eigenfrequency is identified using the standard and the modi-
fied ERA approach. In Figure 5.4, we report the identified eigenfrequencies and
damping as a function of the frequency resolution, namely the relative difference
between harmonic and eigenfrequencies (ω 2 − ω)/ω 2 . We recall that the refer-
ence second eigenfrequency identified in the presence of white noise excitation
only was ω 2 = 23.15Hz. Figure 5.4 shows that the modified ERA can iden-
tify the eigenfrequency with great accuracy even when the harmonic excitation
frequency is close to the second modal frequency of the beam. The standard
Table 5.3: Frequencies and associated damping identified by the standard ERA
and the modified ERA. System excited by stationary white noise and harmonic
loads at 23.30Hz, 23.80Hz and 24.30Hz.
Standard ERA Modified ERA

System order Freq.1 Damp.1 Freq.2 Damp.2 Freq. Damp.
N (Hz) (%) (Hz) (%) (Hz) (%)
30 23.30 0.02 23.86 0.05 23.15 0.24
28 23.30 0.02 23.86 0.00 23.16 0.22
26 23.30 0.02 23.86 0.01 23.16 0.19
24 23.30 0.02 23.85 0.04 23.16 0.19
22 23.30 0.02 23.85 0.04 23.16 0.20
20 23.30 0.01 23.85 0.08 23.16 0.17
18 23.30 0.01 23.84 0.06 23.16 0.17
16 23.30 0.01 23.84 0.07 23.16 0.17
14 23.30 0.01 23.82 0.09 23.16 0.16
12 23.30 0.04 23.70 0.29 23.16 0.16
10 23.30 0.02 23.93 0.58 23.16 0.14
8 23.30 0.02 – – 23.19 0.31
6 23.30 0.02 24.25 0.00 – –
4 23.29 0.05 – – – –
ERA method however fails to identify the modal parameter accurately when
the harmonic excitation frequency is close to the natural frequency.
An important aspect of the modified ERA method presented here is that the
harmonic excitation frequency is assumed to be known accurately and a priori.
This is often the case in practice. Nevertheless it is interesting to investigate
how the modified algorithm behaves when the harmonic excitation specified
in the identification procedure is not fully exact. Therefore, using the same
experimental setup as before, and applying in addition to the random excitation
a harmonic excitation at 23.25Hz, the identification is now performed with the
modified ERA where a harmonic frequency slightly different from 23.25Hz is
assumed. In Figure 5.5 the identified eigenfrequency and damping are plotted
for different values of assumed harmonic frequencies assumed in the modified
ERA identification. One observes that if a harmonic frequency different from the
true value (i.e. 23.25Hz) is specified in the algorithm, the identified frequencies
varies linearly with respect to the harmonic frequency input in the vicinity of
the exact harmonic frequency. In Figure 5.5 it is also seen that the damping
value gets very inaccurate when the assumed harmonic frequency is set between
the true harmonic frequency and the neighboring eigenfrequency. It can also
be noted that when the harmonic frequency assumed in the modified ERA is
very different from the actual harmonic frequency, the poles identified by the
modified ERA are similar to those obtained by the standard ERA method.
Figure 5.4: Identified eigenfrequency for the second bending mode as a function
ω −ω bending
of the Frequency Resolution(FR), F R = harmonic ω bending
Hence, it is essential to introduce in the modified ERA algorithm a harmonic

frequency very close to the actual one. Otherwise the harmonic component ex-
plicitly included in the identification is not capable of representing the harmonic
part of the signal properly and the modification of the ERA method gives no
benefit. In the worst case, the modification in the ERA method has no effect and
the identified parameters are the ones obtained by standard ERA algorithms.
Comparing the results in Figure 3.9 obtained with the LSCE approach to the
results of the ERA procedure (Figure 5.4), it can be conclude that both the mod-
ified LSCE algorithm and the modified ERA methods can compute the modal
parameters accurately in the presence of harmonic excitation. The robustness
of both approaches are comparable both in terms of frequency resolution and
in terms of sensitivity to errors in assumed harmonic frequency.
Figure 5.5: Eigenfrequency and damping ratio identified by the modified ERA
method when the assumed harmonic frequency is different from the true har-
monic frequency (23.25 Hz)
Nevertheless it is our experience that the results obtained by the ERA

method are less dependent on the model order N chosen in the identification
whereas the LSCE approach (and similarly the ITD method) might yield sig-
nificantly different results for different model orders.
5.4 Conclusion
In this chapter we propose a modified ERA method where the harmonic exci-
tation components are explicitly taken into account during the identification,
assuming the harmonic frequencies to be known. In the beam example presented
here, the modified ERA method was shown to compute modal parameters accu-
rately even if the harmonic frequencies are close to eigenfrequencies. Modifying
the ERA is more challenging since, unlike the LSCE or ITD methods, the ERA
procedure does not directly compute eigenparameters of the system but rather
system information in terms of singular values and vectors of the Hankel matrix
5.4. CONCLUSION 73
that do not explicitly relate to the system modal response.

The efficiency of the ERA method in the context of harmonic excitation and
OMA is similar to the one of the LSCE method discussed in Ch. 3.
Chapter 6
Mode Shapes in the

Presence of Harmonics
In this chapter, a method is proposed to perform mode shapes computations in

the presence of harmonic excitations. In section (Sec. 6.1) we shortly discuss
the issue of mode shape identification for Operational Modal Analysis (OMA) in
the presence of harmonic excitations. Especially earlier published work on doing
OMA. In section (Sec. 6.2) a procedure is proposed to find mode shapes. The
procedure is applicable wether or not harmonic excitations are present during
the OMA analysis. Section (Sec. 6.3) presents a test case for which extracted
mode shapes obtained based on the eigenvalues computed by the standard and the
modified identification methods are compared. Section (Sec. 6.3.4) discusses the
results of this chapter, followed by section (Sec. 6.4) which draws the conclusion
of the chapter.
This work has been published in [37].
6.1 Introduction
Earlier three methods to tackle the identification of eigenfrequencies and damp-
ing in OMA in the presence of harmonic excitation were proposed (Ch. 3, 4 and
5) [38,45,47]. The underlying principle of those procedures consists in explicitly
including harmonic frequencies known a priori into the identification algorithms.
It was shown that those methods allow to properly identify eigenfrequencies and
dampings if harmonic loads with unknown amplitude but known frequencies are
present, even if the harmonic frequencies are very close to eigenfrequencies of
the system.
In this chapter a procedure to deduce modal parameters from operational re-
sponse measurements is outlined. Structural eigenfrequencies and modal damp-
ing are computed using the modified Least Square Complex Exponential method
75
76 CHAPTER 6. MODE SHAPES IN THE PRESENCE OF HARMONICS
(Ch. 3).1 Note that the procedure to retrieve the mode shapes as discussed here
is independent of the algorithm used to compute the frequencies and modal
damping. In fact once the complex poles of the system have been found by
any of the methods described in the previous chapters, the procedure describe
here can be applied to evaluate the eigenmodes as a postprocessing step. The
robustness and accuracy of the approach is illustrated by performing tests on a
plate structure.
In order to indicate how well the modes can be identified in the presence of
harmonic excitation, experimental results obtained for a small steel plate are
reported in section (Sec. 6.3). A shaker was used to provide both random noise
excitation as well as harmonic excitations. Along with the computed natural
frequencies and associated damping ratio, mode-shapes are computed. Modal
parameters and mode shapes are obtained from pure stationary white noise
excitation and from a combined random and harmonic loading. Results from
the combined excitation tests will be compared to the pure noise excitation test
in order to evaluate the efficiency of the modified approach.
6.2 Operational modal analysis in the presence

of harmonics
In classical experimental dynamics, dynamic transfer functions or measured
time-responses are used to find eigenfrequencies, damping and eigenmodes of a
system. In operational modal analysis (OMA), the dynamic response of systems
are measured during operation, meaning that, in general, the actual excitation
inducing the operational response is unknown (e.g. road profile inducing vi-
brations in a car, cutting process exciting a tooling machine, wind loads on
constructions). Hence, transfer functions can not be measured. Nevertheless,
assuming the excitation to be random, one can use the responses of the system
to estimate the modal properties of the system. Note that the mode shapes
obtained in OMA are said to be “unscaled”, meaning that the associated modal
mass can not be obtained since the excitation is not measured. This prevents
to use the results of the OMA for synthesizing transfer functions. Nevertheless,
specific techniques were proposed lately to estimate the modal scalings during
OMA (see e.g. [8, 53]).
Here we will indicate how the eigenmodes can be evaluated once the eigen-
parameters are known.
6.2.1 Extraction of mode shapes

Once the poles sr of the system have been identified (Sec. 3.3, 4.3, 5.2.2), the
0 0
residues Crij in (3.3) can be evaluated. The residues Crij are in fact the eigen-
1 One could think of devising modeshape identification techniques based on the ERA
method. However the extension proposed in Chapter (Ch. 5) does currently not allow us
to retrieve eigenmodes. Note also that the SSTD method discussed and adapted in Chapter
(Ch. 4) can not be used to compute eigenmodes, unlike the basic ITD method.
6.2. OPERATIONAL MODAL ANALYSIS IN THE PRESENCE OF HARMONICS77
modes of the system as explained in section (Sec. 3.2).

If one considers p response stations and adopt q of them as reference loca-
tions, a set of qp − q(q−1)
2 auto- and cross-correlation functions in one system
yields. Than correlation function Rij (k∆t) for a unique reference signal j, then
(3.3) yields
Rij (k∆t) = Rik

2M
X
= φri Arj esr k∆t
r=1
2M
X
0 sr k∆t
= Cri e (6.1)
r=1
½
i = 1, . . . q
k = 0, . . . L − 1
where φri is the complex mode shape vectors. Arj is the modal participation
factor. L is the number of time instances of the correlation functions considered
0
for the mode extraction. We have dropped the subscript j in Cri to indicate
0
that the residues correspond all to the same reference and thus Cri represents
the modal amplitude of mode r for degree of freedom i. To compute the residues
0
Cri let us define Vr = esr ∆T so that (3.3) can be put in the matrix form
 
1 1 ... 1 1  0 0

  C11 ... C1q
 V1 V2 ... V2M −1 V2M  0 0 
  C21 ... C2q 
 V12 V22 ... V2M2
V2M2
  =
−1 .. ..
 .. .. .. ..  . ... . 
 . . ... . .  0 0
C2M ... C2M
V1L−1 V2L−1 ... L−1
V2M −1
L−1
V2M 1 q
 
R10 ... Rq0
 R11 ... Rq1 
 
 .. .  (6.2)
 . ... .. 
R1L−1 ... RqL−1
In this expression, Rik are obtained from the measurements, Vr are known
from the pole identification step and M is the number of modes considered.
Hence if there are twice as many time instances of correlation functions Rik as
0
identified poles (i.e. L = 2M ), (6.2) can be solved to yield the residues Cri and
thus the mode shapes corresponding to the poles. Let us note that:
• Equation (6.2) is a complex system of equation where the coefficients Vr

0
and Cri appear as complex conjugate pairs. Hence it would be sufficient
to consider L = M time instances of the correlations.
• In practice, in order to average out numerical errors and measuring noise,

one chooses to consider many more time instances Rik (L >> 2M ) and
the residues are computed as least square solution of (6.2). Using a large
number of signal correlations instances L also allows to average out modes
that are not accounted for in the identification process.
• In the presence of harmonic excitations, V1 ... Vm correspond to the

harmonic frequencies and thus the first m residues yield the corresponding
harmonic response.
6.3 Quality of modal extraction: a steel plate

test case
In this work we investigated the quality of the mode shapes identified in the case
when harmonic excitations are present in operational modal analysis. Modal
analysis was performed on a steel plate in our laboratory (Figure 6.1). The
plate is hanged with soft strings. A shaker was used to provide both white
Figure 6.1: Experimental setup for a steel-plate placed in free-free condition,

excited by a shaker
noise and harmonic excitations. Dimensions of the plate were 43cm x 25.5cm
x 0.6cm. Dynamic responses perpendicular to the plate were measured at 16
locations using accelerometers (see Figure 6.1). The shaker was attached close
6.3. STEEL PLATE TEST CASE 79
to the location number 3 of the structure. A data acquisition system with

4-channels was used. The accelerometer at point 16 on the plate was kept
as reference signal while the other 3 accelerometers were used to measure the
remaining 15 points. Thus 5 runs were needed to measure all points to complete
an experiment. Each accelerometer has a mass of about 15-grams. In order to
avoid varying mass loading effects while roaming the accelerometers on the plate,
additional masses of 15 -grams each were placed at all measuring positions where
no accelerometer was used. The signals were sampled at 1260Hz, assuming
that frequencies up to 500Hz will be analyzed. A total of 8192-samples of
data were taken for each measuring point. During a given experiment, every
run was performed with the same random noise. Correlation functions were
computed with reference to the first channel signal (point 16). Although the
five runs per experiment were performed sequentially and thus the signals were
not measured all simultaneously, cross-correlation signals can be processed for
all response locations as if all impulse responses result from a single impulse of
the system [24]. In order to evaluate the quality of the extracted mode shapes,
we will first identify the mode shapes of a plate structure excited by pure random
noise. The poles will be identified by the standard Least Square Exponential
method (LSCE). The modes φr so obtained will be considered as reference. Then
additional harmonic excitation will be added in the experimental measurements
and the mode shapes will again be identified following the operational modal
analysis procedure described here. The poles will be identified first using a
standard LSCE method (yielding modes φ̃r ). Then the LSCE method modified
to explicitly include the harmonic frequencies [47] will be applied and the modes
resulting from that identification will yield φ̃rm . In order to compare the mode
shapes φ̃r and φ̃rm computed in the presence of harmonic excitation to the
reference ones φr obtained with noise only, we will compute the Modal Assurance
Criterion number (MAC) defined by
¯ ∗ ¯2
¯ ¯
¯φ̃r φr ¯
M AC = ³ ∗ ´
φ̃r φ̃r . (φ∗r φr )
¯ ∗ ¯2
¯ ¯
¯φ̃rm φr ¯
M AC m = ³ ∗ ´
φ̃rm φ̃rm . (φ∗r φr )
∗
the superscript denoting the complex conjugate.
6.3.1 Modal parameters obtained with pure noise excita-

tion
First the plate was excited with white noise generated in the range of 0Hz
to 500Hz by the shaker. Since the shaker is exciting the structure close to a
corner of the plate, it is assumed that all lower eigenmodes up to 500Hz will be
properly excited. Figure 6.2 reports the values of the natural frequencies and
associated damping ratio identified by the LSCE for increasing modal orders
N. In Table 6.1 all converged natural frequencies and associated damping
Figure 6.2: Stability diagram for the steel-plate while it is excited by the white-
noise
Modes Frequency Damping

(Hz) (%)
1 152.166 0.811
2 176.546 1.466
3 299.796 0.799
4 367.104 1.151
5 392.832 0.829
6 462.416 0.654
7 489.831 0.316
Table 6.1: Identified frequencies and damping by white-noise excitation
have been listed up to 500Hz. Considering those 7 poles for mode extraction
(M = 7), the system (6.2) was solved for 500 time instances (L = 500). The
auto-correlation was not used in this work so that q = 15, meaning that the
residues were identified only for the points 1 to 15 (see Figure 6.1). The first 5
identified mode shapes are shown in Figure 6.3.
6.3.2 Additional harmonics at 144Hz, 146Hz and 148hz:

identification with the standard LSCE
Now harmonic excitations will be added to the random load. The harmonic
excitation frequencies are chosen close to the first mode (152.2Hz). In this par-
ticular example three harmonic excitation frequencies are used: 144Hz, 146Hz
and 148Hz. Figure 6.4 shows the PSD of the measured signal and zooms in the
region of the harmonic frequencies since outside that region the stability plot
is similar to the case when only noise was applied (Figure 6.2). The poles are
identified using first the standard LSCE method, then using the modified LSCE
where harmonic frequencies are explicitly accounted for. Figure 6.4 shows the
stability diagram resulting from both methods. It can be seen that the standard
LSCE procedure fails to identify the first pole of the plate properly. Clearly the
identified poles are dominated by the harmonic frequencies. The actual eigen-
frequency at 152.2 Hz is masked and can not be found as a stable pole even for
an identification of order N = 80. The stable poles found by the standard LSCE
method are listed in Table 6.2. For the higher modes, the eigenfrequencies and
damping identified by the LSCE are close to those found in the presence of
pure noise (Table 6.1). Using the identified poles listed in Table 6.2, the modes

(Hz) (%)
1 177.664 0.740
2 299.565 0.841
3 367.535 1.123
4 391.988 0.976
5 463.044 0.555
6 486.794 0.570
7 488.529 0.326
Table 6.2: Frequencies and associated damping identified by the standard LSCE
method in the presence of harmonic excitations at 144Hz, 146Hz and 148Hz
can be identified from (6.2) with q = 15 and L = 500. The modes identified
can then be compared to the reference mode shapes obtained with pure noise
excitation (Figure 6.3). The eigenparameters and the MAC number between
the modes identified in the presence of harmonics and the reference ones are
listed in Table 6.3. Note that for identification of the modes, we have included
the exact harmonic frequencies as poles in (6.2) instead of the poles identified
by the standard LSCE around 152 Hz (see also Table 6.3). In this way the
higher modes are better identified. The identified virtual modes, namely the
harmonic responses corresponding to the harmonic frequencies, are very similar
to the first eigenmode of the system (see figure 6.5 and compare to mode 1 in
Figure 6.3). Indeed, the MAC value between the first reference mode and the
harmonic responses at 144, 146 and 148 Hz is about 0.9. This is explained by
the fact that the harmonic response is dominated by the first eigenmode since
Modes Frequency Frequency MAC

(standard LSCE) (standard LSCE)
only noise with harmonics
(Hz) (Hz)
- — 144.000 —
- — 146.000 —
- — 148.000 —
1 152.166 — —
2 176.546 177.664 0.145
3 299.796 299.565 0.837
4 367.104 367.535 0.971
5 392.832 391.988 0.951
6 462.416 463.044 0.937
7 489.831 488.529 0.923
Table 6.3: Poles used for identification of the eigenmodes and MAC values
between the reference modes and the modes identified by the standard LSCE
procedure applied to the signals with harmonics.
the harmonic frequencies are very close to the first eigenfrequency. The second
mode identified from the poles obtained with the modified LSCE has a surpris-
ingly low MAC value (see Table 6.3). The most probable cause for failing to
identify the second mode properly has been traced back to the fact that the
measuring point 16 used as reference in all cross-correlation functions is close to
a vibration node of the second mode as indicated by the mode shape 2 in Figure
6.3. Hence, for this experiment several reference location should be considered
in order to improve the identification of the modes.
6.3.3 Additional harmonics at 144Hz, 146Hz and 148hz:

identification with the modified LSCE
In the Figure 6.4, the stability diagram of the pole identification when using
the modified LSCE method is also shown. The harmonic frequencies explic-
itly accounted for in the identification and assumed to be known a priori are
restituted exactly by the procedure. They are indicated by + signs in Figure
6.4. It is also observed in the same figure that a stable pole line at around
152Hz appears in addition to the three harmonic lines. This corresponds to the
true first eigenmode of the system. Table 6.4 lists the modal parameters of the
plate computed by the modified LSCE procedure in the presence of harmonics.
Comparing the parameters identified here to the reference ones identified with
noise only, one observes that
• The identified modal frequencies are very close to the reference ones. Even
the frequency close to the harmonic frequencies is found accurately.
• The modal damping of the first two modes (Table 6.4) significantly differ
from the reference values. For the other modes, damping ratios agree
rather well.

(Hz) (%)
- 144.000 0.000
- 146.000 0.000
- 148.000 0.000
1 151.865 1.065
2 177.580 0.828
3 299.565 0.825
4 367.549 1.115
5 392.195 0.985
6 463.092 0.557
7 488.628 0.314
Table 6.4: Frequencies and associated damping identified by the modified LSCE
method in the presence of harmonic excitations at 144Hz, 146Hz and 148Hz
Using the values listed in Table 6.4, the mode shapes have been computed by
solving (6.2) with q = 15 and L = 500. Note that the harmonic response modes
also result from this analysis. The MAC values between the reference modes and
the modes identified using the poles from the modified LSCE are listed in Table
6.5. For the first identified mode, a MAC value of 0.93 is found, indicating that
the mode is very similar to the reference first mode (see figure 6.6). The second
mode identified here has a low MAC value when compared to the reference
second mode. As in the previous section where modes where identified from
poles obtained by the standard LSCE, the most probable reason for failing to
identify the second mode accurately is that the reference measurement point is
close to a vibration node. Higher modes identified with the poles of the modified
LSCE have a very good MAC values.
6.3.4 Discussion of results

From the results obtained in this experimental test, the following points are
noteworthy:
• In the presence of harmonic excitations with frequencies close to the first

eigenfrequency, the standard LSCE method fails to identify the first pole
properly whereas the modified LSCE identifies the eigenpole even for low
orders N . In that case the modified LSCE can identify the eigenfrequencies
accurately but the associated damping are significantly different from the
damping found in the absence of harmonic excitation.
Modes Frequency Frequency MAC

(standard LSCE) (modified LSCE)
only noise with harmonics
(Hz) (Hz)
- — 144.000 0.000
- — 146.000 0.000
- — 148.000 0.000
1 152.166 151.865 0.930
2 176.546 177.580 0.215
3 299.796 299.565 0.859
4 367.104 367.549 0.961
5 392.832 392.195 0.948
6 462.416 463.092 0.956
7 486.545 486.829 0.916
Table 6.5: Poles used for identification of the eigenmodes and MAC values
between the reference modes and the modes identified by the modified LSCE
procedure applied to the signals with harmonics.
• Since only one reference is taken for computing the cross-correlation func-
tions and because the reference point in our test might be close to a
vibration mode of the second mode, this second mode can not be properly
identified. Hence, in such a case, cross-correlation with respect to several
references should be used in the identification (in a way similar to MIMO
identification in standard modal analysis). However, in our test procedure
only location 16 was kept constant in all measurement sets so that, with
our measurements, no other cross-correlation functions can be evaluated.
• Except for the case when reference locations are close to vibration nodes
(see above), eigenmodes of the system can be identified accurately in oper-
ational modal analysis even in the presence of strong harmonic excitations.
This is achieved by considering for the identification the poles obtained
from the modified LSCE [47].
6.4 Conclusion
Operational Modal Analysis (OMA) techniques provide very valuable tools to
identify modal parameters of structures in practical situations. When, in addi-
tion to ambient or operational stationary white noise excitation, harmonic loads
are present specific techniques must be used in order to identify eigenparame-
ters accurately. In this work we indicate how modal frequencies, damping and
eigenmodes can be identified in OMA even when harmonic excitations having
frequencies close to eigenfrequencies are present. It is shown on a test example
of a plate that eigenfrequencies and eigenmodes can be identified accurately as
6.4. CONCLUSION 85
long as the reference location with respect to which the signal cross-correlations
are computed is not close to a vibration node. With the current analysis meth-
ods, modal dampings can not always be identified accurately when harmonic
excitations are present.
Mode 1 Mode 2
152.2 Hz, 0.81% damping 176.5 Hz, 1.47% damping
Mode 3 Mode 4
299.8 Hz, 0.80% damping 367.1 Hz, 1.15% damping
Mode 5
392.8 Hz, 0.83% damping
Figure 6.3: Mode shapes identified in the presence of pure stationary noise and
taken as reference
6.4. CONCLUSION 87
additional periodic excitations, stabilization diagram from the standard LSCE
method and the modified LSCE method
Harmonic response (144.0 Hz) Harmonic response (148.0 Hz)
Figure 6.5: Identified harmonic responses corresponding to harmonic frequency

144 Hz and 148 Hz. The harmonic response at 146 Hz is very similar.
Figure 6.6: First mode shape computed by the modified LSCE method for the
example, where three harmonics are present
Chapter 7
Conclusions
7.1 Summary and main contribution

This thesis investigated the possibility to apply Operational Modal Analysis
(OMA) in the case often encountered in practice where the input force to the
system is a combination of stationary white noise and harmonically exciting
force of known frequency.
The major and important contributions of this thesis can be summarized as
follows:
• A modified Least Square Complex Exponential (LSCE) method has been

proposed, which takes as input the auto- and cross-correlation of the re-
sponse signals to perform OMA in the presence of harmonic excitations.
The standard LSCE method involves solving an overdetermined system
constructed on the Hankel matrix of correlation function. Solving the
overdetermined problem in a least square sense reduces the influence of
measurement noise. In the proposed modified LSCE method, the Hankel
matrix of the correlation functions is complemented with harmonic compo-
nents and the solution of the identifying equations is carried out so that the
harmonic contributions are solved for exactly whereas the eigenparameters
are still obtained from a least square problem. As a result, the exact un-
damped pseudo-modal parameters pertaining to the harmonic excitations
are exactly restituted. Since the harmonic part of the response is properly
accounted for during the identification procedures, the genuine eigenpa-
rameters are obtained with good accuracy and in a numerically stable
manner. Several harmonic excitations can be considered in the proposed
approach. In the example of chapter (Ch. 3), the modified LSCE is able
to identify accurately eigenparameters even when the harmonic excitation
frequency is up to 0.2% lower or 0.4% higher than an eigenfrequency of the
structure. Also it was observed that in the presence of non-sine periodic
excitations, one needs to include in the identification procedure only those
harmonic frequencies that are close to any eigenfrequency of the system
89
90 CHAPTER 7. CONCLUSIONS
one wants to identify and/or those for which the associated power in the
signal is high. It is essential that the harmonic frequencies assumed in the
modified LSCE algorithm accurately represent the true harmonic excita-
tions present in the signal. It was observed in the test example that the
error on the identified eigenparameters gradually increases when the error
on the estimated harmonic frequencies increases. In particular when the
error on the harmonic frequency assumed in the algorithm is large, the
modified LSCE method simply degenerates to the standard LSCE and the
identified modal parameters converge to the modal parameters identified
by standard LSCE method.
• In a way similar to the modified LSCE method, a modified Ibrahim Time
Domain (ITD) method, based on the Single Station Time Domain (SSTD)
method has been proposed which processes the auto- and cross-correlation
of the response signals as inputs, conforming to the NExT approach (Sec.
2.3.2). First the SSTD method has been generalized to include multiple
responses. Then a method has been developed base on the multi-response
version of the SSTD method to identify modal parameters for OMA in
the presence of harmonic excitations. Like for the LSCE method, several
harmonic excitation frequencies can be considered in the algorithm. In the
example of (Ch. 4) it was shown that the modified ITD method is efficient
and allows to identify modal parameters accurately even in the presence
of harmonic excitations. The robustness of the modified ITD method was
found to be similar to the robustness of the modified LSCE approach.
• A third method based on the Eigensystem Realization Algorithm(ERA)
has been developed for OMA in the presence of harmonic excitations.
This is also a time-domain method that can be used in the context of
the NExT approach to OMA. Since the ERA method reduces the data
contained in the Hankel matrix by considering a few number of singular
values from the singular value decomposition of the matrix. Since the ERA
approach is not identifying directly the eigenparameters but rather the
singular value decomposition allowing to reconstruct the system matrices,
constructing the Modified ERA method in order to explicitly account for
harmonic excitations is significantly challenging and abstract compared
to the modification of the LSCE and ITD. However the concepts used to
integrate explicitly the harmonic contributions are similar for all methods
in the sense that the modifications brought to the standard methods ensure
that the exact harmonic frequencies known a priori are accounted for
during the identification, therefore allowing a accurate identification of the
genuine eigenparameters. The efficiency and robustness of the modified
ERA approach are very similar to what was observed for the modified
LSCE and ITD approach. Nevertheless in the course of our many tests
and identification runs, we have found that the ERA approach is superior
for what concerns consistency of the results with varying model orders.
• In order to evaluate the eigenmodes in OMA, we have used standard mode
7.1. SUMMARY AND MAIN CONTRIBUTION 91
identification techniques in the time-domain based on the complex expo-

nential representation of the response. In our work we have used this mode
identification procedure as a post-processing step following the identifica-
tion of the eigenparameters with the modified identification techniques
proposed here. In particular we have illustrated the efficiency of combin-
ing the modified LSCE and the mode identification for analyzing a plate
structure. In that experimental example, the MAC values were consid-
ered to compare the mode-shapes and eigenparameters obtained on one
hand by the standard NExT/LSCE procedure and on the other hand by
the modified NExT/LSCE method. It was found that the accuracy as
measured by the MAC values is significantly superior for the modified
identification approach advocated in our work.
Numerical stability remains an issue for all modified methods de-
veloped here. The origin of the bad conditioning can be traced back to the cases
where several harmonic contributions are explicitly introduced in the identifica-
tion algorithms. Indeed in that case, the partial problem consisting in solving
exactly for the harmonic sub-problem implies matrices constructed on harmonic
responses at successive time instants. Such systems are known to be badly con-
ditioned especially if the harmonic frequencies are close to one another. It is
thus advisable to consider only a few harmonic excitations in any single analy-
sis. Considering the above limitations, the following guidelines should be kept
in mind:
• only those particular harmonics that are in the vicinity of eigenvalues to
be identified should be included in the analysis.
• if possible, one should avoid including in the analysis several harmonics
having frequencies close to one another.
In our work, mode-shapes computation is performed as a post-
processing step, separate from the eigenparameters identification. The proce-
dure for mode extraction is based on the modal superposition of the impulse-like
response. It is also common to extract the modes of the system from the stan-
dard ERA method. However, in the modified ERA method as proposed in this
work, modes can not be obtained directly together with the eigenparameters.
The post-processing procedure has thus to be used. Investigating the possibil-
ity to extend the ERA approach in order to properly extract modes within the
ERA computations is certainly an interesting and challenging issue for future
research.
In theory, the concepts used in our work to modify standard identi-
fication methods can also be extended to explicitly include damped oscillations
related to poles known a priori. This could be helpful in order to allow the
identification of modes only weakly present in the signal and which is masked
by strongly dominant eigenmodes. The procedure would then consists in first
identifying the dominant components and then, in a second step, include the
dominant modes in the LSCE explicitly to identify the modes weakly present in
the signal.
92 CHAPTER 7. CONCLUSIONS
Other research directions currently considered also include the ap-

plication of the proposed modified identification procedures to the identification
of super- and sub-harmonics in non-linear systems.
Bibliography
[1] Handbook on Modal Testing. Dynamic Testing Agency, 1993.
[2] Finite element model updating in structural dynamics. Kluwer Publisher,

1995. ISBN 0-7923-3431-0.
[3] Modal Analysis Theory and Testing. Katholieke Universiteit Leuven, Leu-
ven, Belgium, 1997. ISBN 90-73802-61-X.
[4] M. Basseville, A. Benveniste, B. Gach-Devauchelle, M. Goursat, D. Bon-

necase, P. Dorey, M. Prevosto, and M. Olagnon. Damage monitoring in
vibration mechanics: Issues in diagonstics and predictive maintenence. Me-
chanical Systems and Signal Processing, 7(5):401–403, Sept. 1993.
[5] J. Bendat and A. Piersol. Engineering Applications of Correlation and

Spectral Analysis. John Wiley, 1980.
[6] S. Berg, P. Mohanty, and D. J. Rixen. Investigating the causes of non-

uniform cookie flow in vibratory conveyors. Experimental Techniques, Ac-
cepted for Publication.
[7] R. Brincker, P. Andersen, and N. Møller. An indicator for separation of

structural and harmonic modes in output-only modal testing. In Interna-
tional Modal Analysis Conference (IMAC XVIII), pages 1649–1654, San
Antonio,Texas, February 7-10 2000. Society for Engineering Mechanics.
[8] Rune Brincker and Palle Andersen. A way of getting scaled mode shapes
in output only modal testing. In International Modal Analysis Conference
(IMAC XXI), Kissimmee, Florida, USA, February 3-6 2003. Society for
Engineering Mechanics.
[9] D. L. Brown, R. J. Allemang, R. Zimmerman, and M. Mergeay. Parameter

estimation techniques for modal analysis. SAE Technical Paper Series,
(790221), 1979.
[10] B. Dobson and E. Reider. A review of the indirect calculation of excita-

tion forces from measured structural response data. Journal of Mechanical
Engineering Science, 204(2):69–75, 1990.
93
94 BIBLIOGRAPHY
[11] S. Doebling and C. Farrar. Computation of structural exibility for bridge

health monitoring using ambient modal data. In Proceedings of the 11th
ASCE Engineering Mechanics Conference, pages 1114–1117. American So-
ciety of Civil Engineers, 1996.
[12] D. J. Ewins. Modal Testing: theory, practice and application. Research

Studies Press LTD, 2000. ISBN 0 86380 218 4.
[13] Charbel Farhat and Daniel Rixen. Encyclopedia of Vibration, chapter Lin-
ear Algebra, pages 710–720. Academic Press, 2002. ISBN 0-12-227085-1.
[14] R. L. Fox and M. P. Kapoor. Rate of change of eigenvalues and eigenvectors.

AIAA Journal, 6:2426–2429, 1968.
[15] George R. Cooper and Clare D. McGillem. Probabilistic Methods of Signal

and System Analysis. Oxford University Press, 3rd edition, 1999. ISBN
0-19-512354-9.
[16] W. Gersch. Estimation of the autoregressive parameters of a mixed au-

toregressive moving-average time series. IEEE Transactions on Automatic
Control, AC-15:583–588, Oct. 1970.
[17] W. Gersch and T. Brotherton. Estimation of stationary structural system

parameters from non-stationary random vibration data: A locally station-
ary model method. Journal of Sound and Vibration, 81(2):215–227, 1982.
[18] W. Gersch and R. S. Z. Liu. Time series methods for the synthesis of
random vibration systems. Transactions of teh ASME. Journal of Applied
Mechanics, 43(1):159–165, Mar. 1976.
[19] W. Gersch and S. Luo. Discrete time series synthesis of randomly excited
structural system response. The Journal of Acoustical Society of America,
51(1 (Part 2)):402–408, 1972.
[20] W. Gersch and D. R. Sharpe. Maximum likelihood estimation of structural

parameters from random vibration data. Journal of Sound and Vibration,
31(3):295–308, 1973.
[21] W. Gersch and D. R. Sharpe. Parameter identification: Stochastic process

techniques. The Shock and Vibration Digest, 7(11):71–86, Nov. 1975.
[22] S. Granger and L. Perotin. An inverse method for the identification of

a distributed random excitation acting on a vibrating structure part 1:
theory. Mechanical System and Signal Processing, 13(1):53–65, Jan. 1999.
[23] P. Guillaume. Identification of multi-input multi-output systems using

frequency-domain models, PhD thesis. Dept. ELEC, Vrije Universiteit Brus-
sel, Belgium., 1992.
BIBLIOGRAPHY 95
[24] L. Hermans and H. Van der Auweraer. Modal testing and analysis of struc-
tures under operational conditions: Industrial applications. Mechanical
System and Signal Processing, 13(2):193–216, October 1999.
[25] S. R. Ibrahim and E. C. Mikulcik. A time domain modal vibration test

technique. The Shock and Vibration Bulletin, 43(4):21–37, 1973.
[26] S. R. Ibrahim and E. C. Mikulcik. The experimental determination of vi-

bration parameters from time responses. The Shock and Vibration Bulletin,
46(5):186–196, 1976.
[27] S. R. Ibrahim and E. C. Mikulcik. A method for the direct identification

of vibration parameters. The Shock and Vibration Bulletin, 47(4):183–198,
1977.
[28] G. H. James, T. G. Carne, and J. P. Lauffer. The natural excitation tech-

nique (next) for modal parameter extraction from operating structures.
Journal of Analytical and Experimental Modal Analysis, 10(4):260–277, Oc-
tober 1995.
[29] G. H. James, T. G. Crane, and R. S. Edmunds. Star missile analysis of first-

flight data using the natural excitation technique, next. In International
Modal Analysis Conference, Ilikai in Honolulu, HI, Jan 31 Feb 3 1994.
Society for Engineering Mechanics.
[30] George H. James, Thomas G. Carne, James P. Lauffer, and Arlo R. Nord.
Modal testing using natural excitation. In International Modal Analysis
Conference (IMAC XX), San Diego, CA, February 3-7 1992. Society for
[31] J. N. Juang. Mathematical correlation of modal-parameters-identification

methods via system realisation theory. International Journal of Analytical
and Experimental Modal Analysis, 2(1), Jan. 1987.
[32] J. N. Juang and R. S. Pappa. An eigensystem realization algorithm for

modal parameters identification and model reduction. Journal of Guidance,
Control and Dynamics, 8(5):620–627, Sept.-Oct. 1985.
[33] J. N. Juang and R. S. Pappa. Effects of noise on modal parameters identified

by the eigensystem realisation algorithm. Journal of Guidance, Control and
Dynamics, 9(3):294–303, May-June 1986.
[34] Lagö, T. The difference between harmonics and stochastic narrow band
responses. oral presentation at the SVIBS symposium, Stockholm, 1997.
[35] M. Géradin and D. Rixen. Mechanical Vibrations. Theory and Application

to Structural Dynamics. Wiley & Sons, Chichester, 2d edition, 1997. ISBN
0-471-93927-7.
96 BIBLIOGRAPHY
[36] N. M. M. Maia and J. M. M. Silva, editors. Theoretical and Experimental

Modal Analsysis. Research Studies Press Ltd., Somerser, England, 1997.
ISBN 0 86380 208 7.
[37] P. Mohanty and D. J. Rixen. Identifying mode shapes and frequencies by

operational modal analysis in the presence of harmonic excitation. Exper-
imental Mechanics, In Press.
[38] P. Mohanty and D. J. Rixen. Modified era method for operational modal
analysis in the presence of harmonic excitations. Mechanical Systems and
Signal Processing, In Press.
[39] P. Mohanty and D. J. Rixen. Measuring sweet spots of tennis racket. In

20th International Modal Analysis Conference (IMAC XX), Los Angeles,
USA., February 4-7 2002. Society for Engineering Mechanics.
[40] P. Mohanty and D. J. Rixen. Operational modal analysis in the presence

of harmonic excitation. In 47th Congress of Indian Society of Theoretical
and Applied Mechanics, IIT Guwahati, India, December 23-26 2002. Indian
Institute of Technology, Kharagpur.
[41] P. Mohanty and D. J. Rixen. Accounting for harmonic excitations in op-

erational modal analysis. In 21st International Modal Analysis Conference
(IMAC XXI), Kissimmee, Florida, USA., February 3-6 2003. Society for
[42] P. Mohanty and D. J. Rixen. Including harmonic responses in the ibrahim

time domain algorithm for operational modal analysis. In Tenth Interna-
tional Congress on Sound and Vibration, Stockholm, Sweden, July 7-10
2003.
[43] P. Mohanty and D. J. Rixen. Modifying the era identification for opera-
tional modal analysis in the presence of harmonic perturbations. In 16th
ASCE Engineering Mechanics Conference, University of Washington, Seat-
tle, USA., July 16-18 2003. American Society of Civil Engineers.
[44] P. Mohanty and D. J. Rixen. Operational dynamic testing in the presence

of harmonic excitation. In Second M.I.T. Conference on Computational
Fluid and Solid Mechanics, MIT, Cambridge, U.S.A., June 17-20 2003.
Massachusetts Institute of Technology.
[45] P. Mohanty and D. J. Rixen. A modified ibrahim time domain algorithm

for operational modal analysis including harmonic excitation. Journal of
Sound and Vibration, 275(Issues 1-2):375–390, Aug. 2004.
[46] P. Mohanty and D. J. Rixen. Modified sstd method to account for harmonic
excitations during operational modal analysis. Mechanism and Machine
Theory, 39(12):1247–1255, December 2004.
BIBLIOGRAPHY 97
[47] P. Mohanty and D. J. Rixen. Operational modal analysis in the presence of

harmonic excitations. Journal of Sound and Vibration, 270(Issues 1-2):93–
109, Feb. 2004.
[48] D. Otte, Q. Yang, and P. Sas. Analysis of multivariate operating response

data by means of arv time series modelling techniques. In 19th International
Seminar on Modal Analysis (ISMA 19), Leuven (B), Belgium, Sep. 1994.
KU Leuven.
[49] P. Van Overschee and B. de Moor. A unifying theorem for three subspace
system identification algorithms. In American Control Conference, Balti-
more, USA, June 1994.
[50] E. Parloo, P. Guillaume, and B. Cauberghe. Maximum likelihood identifi-

cation of non-stationary operational data. Journal of Sound and Vibration,
268(5):971–991, 2003.
[51] E. Parloo, P. Guillaume, and M. Van Overmeire. Autonomous structural

health monitoring - part ii: Vibration-based in-operation damage asses-
ment. Mechanical Systems and Signal Processing, 16(4):659–675, 2002.
[52] E. Parloo, P. Guillaume, and M. Van Overmeire. Damage assessment us-

ing mode shape sensitivities. Mechanical Systems and Signal Processing,
17(3):499–518, 2003.
[53] E. Parloo, P. Verboven, P. Guillaume, and M. Van Overmeire. Sensitivity-

based operational mode shape normalization. Mechanical System and Sig-
nal Processing, 16(5):757–767, 2002.
[54] R. Pintelon, P. Guillaume, Y. Rolain, J. Schoukens, and H. Van Hamme.

Parametric identification of transfer function in the frequency domain - a
survey. IEEE Transcations on Automatic Control, 39(11):2245–2260, 1994.
[55] R. Pintelon and J. Schoukens. System Identification: a Frequency Domain

Approach. IEEE Press, 2001.
[56] J. Schoukens and R. Pintelon. Identification of Linear Systems: a Practical

Guideline to Accurate Modelling. Pergamon Press, 1991.
[57] C. Shih, Y. Tsuei, R. Allemang, and D. Brown. Complex mode indici-

tion function and its application to spatial domain parameter estimation.
Mechanical Systems and Signal Processing, 2(4):367–377, 1988.
[58] Singiresu S. Rao. Mechanical Vibrations. Addison-Wisley Publishing Com-

pany, 3rd edition, 1995. ISBN 0-201-52686-7.
[59] Sudhakar M. Pandit. Modal and Spectral Analysis: Data Dependent Sys-
tems in State Space. Wiley & Sons, 1991. ISBN 0-471-63705-X.
98 BIBLIOGRAPHY
[60] P. Vanhonacker. The Use of Modal Parameters of Mechanical Structures

in Sensitivity Analysis, System Synthesis and System Identification Meth-
ods, PhD thesis, page 250. Dept. of Mech. Eng., KULueven, Belgium.,
Septmber, 1988.
[61] C. van’t Hof, P. Mohanty, and D. J. Rixen. Testing a dynamic mechanical
analyzer-influence of the measuring column dynamics. In 21st International
Modal Analysis Conference (IMAC XXI), Kissimmee, Florida, USA., Feb-
ruary 3-6 2003. Society for Engineering Mechanics.
[62] H. Vold, J. Kundrat, G. T. Rocklin, and R. Russel. A multi-input modal
estimation algorithm for mini-computers. SAE Technical Paper Series,
(820194), 1982.
[63] H. Vold and G. T. Rocklin. The numerical implication of a multi-input
modal estimation method for mini-computers. In Proceedings of the 1st In-
ternational Modal Analysis Conference (IMAC I), Orlando, Fl,USA, 1985.
Society for Engineering Mechanics.
[64] S. A. Zaghlool. Single-station time-domain (sstd) vibration testing tech-
nique: Theory and application. Journal of Sound and Vibration, 72(2):205–
234, 1980.
Curriculum Vitae
Mr. Prasenjit Mohanty was born on 10th February 1975 in Balasore, India.
From August 1995 to May 1999, he pursued his undergraduate study in the field
of Ocean Engineering and Naval Architecture at Indian Institute of Technology,
Kharagpur, India. He than joined the Indian Register of Shipping, Mumbai as
an assistant ship surveyor and continued working there until January, 2000. In
March 2000 he came to the Netherlands to pursue his PhD at Delft University
of Technology. During the same year, in November 2000, he started his PhD
under the supervisor of Prof. Daniel J. Rixen. At this moment Prasenjit is a
Research Assistant at the University of Sheffield, United Kingdom. He is now
working on an EPSRC research project: ”Dynamic Behavior of Stadia Under
Human Occupation and Excitation”
In 1999 Prasenjit received the Shri Sam Dotiwala Cash Award from the
Institution of Naval Architects at the Indian Institute of Technology, Delhi,
India. In 2003 he was honored by the Dominick J. DeMichele Scholarship
Award from the Society for Experimental Mechanics, Inc. (SEM), USA.

Dep Mohanty 20050110

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

Dep Mohanty 20050110

Uploaded by

Copyright:

Available Formats

Operational Modal Analysis in the

Presence of Harmonic Excitations

ter verkrijging van de graad van doctor

Delft, January 2005

De belangrijkste en originele bijdragen van dit proefschrift zijn:

• In hoofdstuk 3 wordt een gewijzigd Least Square Complexe Exponentiële

methode is gebaseerd op de ”Single Station Time Domain” (SSTD) meth-

De voorbeelden die in dit proefschrift worden gepresenteerd (een scharnierend

• In chapter 3, a modified Least Square Complex Exponential (LSCE) method

• Similar to the modified LSCE, in chapter 4, a modified Ibrahim Time

• In chapter 5, a modified Eigensystem Realization Algorithm (ERA) method

• In chapter 6, a method has been proposed to identify mode shapes by

Summary in Dutch vii

2 Operational Modal Analysis 11

3 Modified LSCE method for harmonics 23

4 Modified ITD method for harmonics 41

4.3.1 Modified SIMO-SSTD for harmonics . . . . . . . . . . . . 45

5 Modified ERA method for harmonics 57

6 Mode Shapes in the Presence of Harmonics 75

1.1 Examples of structures in operating condition. bridge (a), aero-

2.1 Peak picking method for modal analysis . . . . . . . . . . . . . . 13

4.1 Experimental Setup of a Plate . . . . . . . . . . . . . . . . . . . . 48

5.1 Stability diagram for white noise excitation only computed by

6.1 Experimental setup for a steel-plate placed in free-free condition,

3.1 Frequencies and associated damping (white noise only) . . . . . . 30

4.1 Frequencies and associated dampings (white noise only) . . . . . 51

5.1 Frequencies and associated dampings identified by ERA (white

6.1 Identified frequencies and damping by white-noise excitation . . 80

6.4 Frequencies and associated damping identified by the modified

1.1 Research context

Figure 1.1: Examples of structures in operating condition. bridge (a), aeroplane

• FE analysis is based on a discretization of the reality, meaning that the

• Eventhough bulk properties of materials such as mass density and elastic-

• The geometry of a specific project can significantly differ from what is

• Nonlinearity of the structures also remains as a source of uncertainties

while doing numerical modal analysis which by essence is based on lin-

• advances in measurement techniques, sensors and signal processing,

• increase of computer power and decrease of computer costs which allows

• the increase of computational power has also significantly broadened the

(see Figure 1.1). Such excitations can be considered as white-noise in nature.

• Methods performing auto- and cross-correlation of the time-signals and

• Methods based on discrete-time data models. Examples of such methods

• Stochastic subspace methods such as the Canonical Variate and Balance

• The Maximum Likelihood frequency domain method, which works on

In OMA procedures, like in classical EMA, natural frequencies, modes and

1. Planning and executing the experiments: in the initial phase, experiments

condition can be achieved by hanging the structure from soft springs or

• Structural modification and sensitivity analysis: structural modification

As described here above, there are many important applications, where

1.2 Scientific challenge and thesis topic

• a new method based on the Least Square Complex Exponential method

• a new method related to the Ibrahim Time Domain method is described

• a novel extension to the Eigensystem Realization Algorithm is outlined

1.3 Outline and structure of the thesis

Figure 1.2: Organization of the text.

Operational Modal Analysis

In this chapter, the conceptual background of operational modal analysis is intro-

2.2 Operational modal analysis techniques:

Figure 2.1: Peak picking method for modal analysis

Peak-Picking method are:

Other methods for OMA

2.3 The Natural Excitation Technique (NExT)

2.3.2 Modal analysis of linear structures